abstracts from the 10th american conference on ... combi… · this limit should be considered for...
TRANSCRIPT
Abstracts from the 10th American Conference on Pharmacometrics
ACoP10
ISSN 2688-3953
ACoP10, Orlando, FL 20-23 October
Abstracts Presented on Wednesday 23rd October
Citation: Author names, Abstract title, ACoP10, Orlando FL, ISSN:2688-3953,
2019, Vol 1
W-001
How predictive is short-term body-weight loss in the longer term for decision making during clinical drug
development?
Franziska Kluwe1, Monika Maas2, Anita Hennige3, Benjamin Weber4, Charlotte Kloft1
1Department of Clinical Pharmacy and Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, Berlin,
Germany; 2Translational Medicine and Clinical Pharmacology, Boehringer Ingelheim Pharma GmbH & Co. KG,
Biberach an der Riss, Germany; 3Therapeutic Area Metabolism Medicine, Boehringer Ingelheim International
GmbH; Biberach an der Riss, Germany; 4Boehringer Ingelheim Pharmaceuticals Inc., Ridgefield, CT, USA.
Objectives: Obesity is a complex disease that leads to adverse health consequences and premature death. Currently
there are only a few approved anti-obesity treatments, but new compounds are being evaluated. To facilitate
selection of promising candidates and streamline development, innovative and efficient clinical trial designs are
needed. The key aspect in the differentiation to existing therapies is the reduction in mean body weight relative to
baseline (ΔWT) after 12, 24 or 52 weeks; hence, large and lengthy trials are warranted. The objectives of this work
were (i) to investigate the relation between short-term and long-term ΔWT for different incretin-based therapies in
patients with and without type 2 diabetes mellitus (T2D) and based on this relation, (ii) to evaluate the feasibility
and potential of shorter clinical trials or interim data analyses after only 4 weeks of treatment for decision-making in
the selection of most promising drug candidates during clinical development.
Methods: A regression-based meta-analysis was performed amalgamating relevant summary-level data from
published randomized controlled Phase I-IV clinical trials in adult patients with and without T2D, receiving incretin-
based therapies, published from 2010 to 2018. For model development, data from all trials was analyzed using R
(v3.4.3). The relation between short-term and long-term ΔWT was jointly evaluated for all incretin-based therapies,
as well as placebo. To evaluate model performance, internal and external model validation were performed using
trials excluded prior to model development.
Results: In the present analysis, data from 232 trials with 523 trial arms, representing over 123000 patients and 16
treatments, were used. Median patient age, median baseline bodyweight, primary indication and primary drug class
across all trials were 56.3 years, 90.2 kg, T2D and GLP-1 agonists, respectively. A strong correlation between early
4-week (4w) ΔWT4w and 3-month (3m) ΔWT3m (r=0.87, with 95% confidence interval 0.84-0.89), as well as 6-
month (6m) ΔWT6m (r=0.82, with 95% confidence interval 0.77-0.87), was observed (Figure 1). Based on this
identified relationship, a ΔWT4w of -2.0% would translate into an average reduction of -4.0% ΔWT3m and -5.5%
ΔWT6m (Figure 1).
Conclusions: The performed analysis revealed a strong correlation between early and longer-term ΔWT for all
investigated treatments independent of mechanism of action or dosing regimen. Further exploration of these
correlations including potential predictors, such as drug class, patient population, baseline bodyweight or
background therapy, may contribute to further elucidate and quantify this relation and its variability. Overall, these
results can be easily visualized, interpreted, communicated and made available (e.g. using R-Shiny) to contribute to
the optimization of future clinical trial design and interim data analyses to ultimately facilitate decision-making
during clinical development of new compounds, so that the most efficacious may be identified as early as after 4
weeks of treatment.
Figure 4: Short-term (4 weeks) body-weight change from baseline (∆WT, %) versus long-term (10-14 weeks)
body-weight change from baseline (panel A) and versus long-term (23-26 weeks) body-weight change from
baseline (panel B) for all trial arms. Data point size corresponds to trial arm size. Black line: Linear regression
line; shaded grey area: 95% confidence interval; dashed red line: 95% prediction interval.
W-002
Target-Mediated Exposure Enhancement – A New Limit of TMDD?
Authors: Patrick M. Glassman, Vladimir R. Muzykantov
Affiliation: Department of Systems Pharmacology & Translational Therapeutics, Center for Targeted Therapeutics
and Translational Nanomedicine, University of Pennsylvania, Philadelphia, PA
Objectives: Molecules with high affinity for a pharmacologic target are often subject to target-mediated drug
disposition (TMDD), where the general expectation is greater than dose-proportional increases in exposure and
dose-dependent decreases in volume of distribution (1). Molecules that are non-specifically eliminated more rapidly
than their target may have unique pharmacokinetic features within the TMDD system. In this work, we utilize
modeling and simulation to characterize this limit of TMDD.
Methods: Data describing blood concentration vs. time profiles were digitized for two structurally unrelated –
caplacizumab (bivalent anti-von Willebrand Factor) (2) and linagliptin (small molecule anti-DPP-4) (3). Non-
compartmental analysis (NCA) and the general model of TMDD (gTMDD) (1) were used to obtain estimates of
relevant parameters and to characterize the pharmacokinetic profiles. Sensitivity analyses were performed to assess
the range of parameters that would allow for this pharmacokinetic phenomenon to occur.
Results: Non-compartmental analysis showed dose-dependent increases in clearance for both molecules –
caplacizumab (1.29 – 66.7 mL/h/kg, 0.02 – 8 mg/kg) and linagliptin (2.51 – 14.3 mL/min, 0.5 – 10 mg. The
proposed model was able to well-characterize observed data for both caplacizumab and linagliptin. In both cases,
non-specific clearance (CLns) was estimated to be significantly greater than target-mediated clearance (CLtmd) – 94.3
(caplacizumab) and 345 (linagliptin). Percent prediction errors of AUCinf were 1.32% (0.02 mg/kg), 11.0% (0.4
mg/kg), and 6.19% (8 mg/kg) for caplacizumab and 7.58% (0.5 mg), 20.1% (2.5 mg), 32.6% (5 mg), and 19.6% (10
mg) for linagliptin. Sensitivity analyses suggest that this phenomenon is likely to occur when CLns is greater than
CLtmd (Figure 1).
Conclusions: Experimental results and pharmacokinetic modeling suggest that molecules with very high non-
specific clearance relative to that of drug-receptor complexes can display pharmacokinetics inconsistent with the
general assumptions of TMDD. At this limit, less than dose-proportional increases in exposure and dose-dependent
increases in clearance would be anticipated. This limit should be considered for small molecule, peptide, and small
protein therapeutics with high affinity for targets with a relatively slow constitutive turnover rate.
References: 1. Mager DE and Jusko WJ. J Pharmacokin Pharmacodyn. 28(6), 507-532 (2001). 2. Ulrichts H et al.
Blood. 118, 757-765 (2011). 3. Retlich S et al. Clin Pharmacokin. 49(12), 829-840 (2010).
Figure 1:
W-003
Dual Target-Mediated Drug Disposition (TMDD) Model in Support of the Optimization of MEDI5752, a
Monovalent Bispecific Antibody targeting PD-1 and CTLA-4
Xuyang Song1, Jing Li2, Yariv Mazor1, Simon Dovedi3, Chunning Yang1, Ikbel Achour1, Raffaella Faggioni1, Lorin
Roskos1, Rajesh Narwal1, Anis Khan1
1AstraZeneca, Gaithersburg, MD, USA; 2AstraZeneca, South San Francisco, CA, USA. 3AstraZeneca, Cambridge,
U.K.
Objectives: MEDI5752 is a monovalent bispecific antibody (DuetMab) with an Fc-domain engineered to reduce
effector function, targeting two clinically validated negative T cell regulators, Programmed death 1 (PD-1) and
cytotoxic T-lymphocyte-associated protein-4 (CTLA-4). The bispecific antibody has been designed to suppress the
PD-1 pathway and provide modulated CTLA-4 inhibition to uncouple CTLA-4 dependent peripheral toxicity from
tumor efficacy. The objective of this modeling was to develop a dual TMDD model to investigate the impact of
monovalent vs. concurrent bivalent binding of PD-1 and CTLA-4 on single-positive and double-positive cells in
comparison to anti-PD-1 and anti-CTLA-4 monotherapies and combination.
Methods: The dual TMDD model included receptor occupancy in CTLA-4 single-expressing cells (to simulate
periphery) and double-positive cells (to simulate Tumor-Infiltrating Lymphocytes, TILs). The in vitro cell binding
data including antibody affinity and antigen density was incorporated in the model to predict potency of MEDI5752
for PD-1 and CTLA-4 + receptor occupancy in the periphery and TILs, respectively. Using the model, additional
parameters such as optimal cell density ratio and antibody affinity for the target binding were explored in a
sensitivity analysis.
Results: The dual TMDD model adequately predicted the observed in vitro cell based binding data. Model-
predicted IC50 values were consistent with the observed IC50 values for CTLA-4 and PD-1 in cell binding inhibition
studies for MEDI5752 (Figure 1). The receptor occupancy predictions confirmed that the optimized bispecific
design increased the binding potency approximately 11-250 fold at PD-1/CTLA-4 cell density ratios of 10:1 to 40:1,
respectively. The increased potency for CTLA-4 due to the increased avidity in the TILs targeting CTLA-4/PD-1
double positive cells was further predicted using the dual TMDD model by incorporating the receptors’ turn over
rates and bispecific PK, which allowed higher occupancy and inhibition of CTLA-4 while decreasing binding to the
monovalent CTLA-4 in the periphery. The model simulated data demonstrated that bivalent targeting of CTLA-4
with MEDI5752 enables saturation of CTLA-4 at lower concentrations of MEDI5752 as compared to the saturation
of CTLA-4 when anti-PD-1 and anti-CTLA-4 monoclonal antibodies (mAbs) are given as combo therapy.
Subsequent sensitivity analyses revealed that increased cell density ratio between PD-1 and CTLA-4 can result in
enhanced potency of MEDI5752 for CTLA-4.
Conclusions: A dual TMDD model was developed and qualified. Simulations in mono- and bivalent binding
settings for MEDI5752 demonstrated the potential for MEDI5752 to inhibit CTLA-4 on TILs whilst reducing
peripheral T cell targeting and proving favorable safety profile. The model also has facilitated molecule selection
and can be used generally for design, optimization and human dose prediction of similar class of bispecific
antibodies.
Figure 1 Left Panel: Observed (left) and Model-Predicted (right) Binding and receptor occupancy of MEDI5752 to
CTLA-4 and PD-1 on CHO PD-1+CTLA-4+ Concentration; Right Panel: Observed(Left) and Model-Predicted
(right) Binding and receptor occupancy of parental anti-PD-1 and anti-CTLA-4 mAbs to CTLA-4 and PD-1 on CHO
PD-1+CTLA-4+ (10:1) cells
W-004
Individual Model Averaging for estimating treatment effects
Estelle Chasseloup1, Adrien Tessier2, Mats O. Karlsson1
1Dept. Of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden; 2Dept. Of Clinical Pharmacokinetics
and Pharmacometrics, Institut de Recherches Internationales Servier, Suresnes, France
Objectives: Model misspecifications in longitudinal data analysis may affect both the type 1 error and the estimates
of treatment effects. We used different real data sets to compare these aspects of the standard pharmacometric
approach (STD) and a new approach, Individual Model Averaging (IMA), based on mixture models.
Methods: Three real placebo-only data sets were used: ADAS-cog scores from patients with Alzheimer’s disease
[1], pain scores from patients with neuropathic pain[2], and seizure counts from patients with epilepsy[3]. To create
data sets mimicking trials where the drug would have no effect, repeated (n=1000) 1:1 randomizations of the
subjects to “treatment” vs “control” were implemented. The data set sizes were varied (N=50-800) and the type I
error rate at p<5% obtained for each scenario based on the likelihood ratio test. Different combinations of models
describing the placebo effect (smaller or bigger than published models) and the drug effect (offset or disease
modifying), with different models for inter-individual variability were used. STD contrasted nested models without
(base) or with (full) the drug effect. IMA assumed two sub-models, the placebo effect with or without the drug
effect. Individual objective function value (OFVi) was an average of OFVi for each sub-models weighted by their
respective probabilities. IMA compared nested models (i) with fixed equal probability for all patients for the two
sub-models (base), and (ii) with a function describing the probability of each sub-model estimated using the
allocation arm as covariate (full).
Results: Each ADAS-cog, pain and seizure dataset were analysed with 66, 32 and 32 base-full model pairs. IMA
showed better performances in type I error control than STD on all three data types at all study sizes. Median type I
error at the largest study size across all scenarios for ADAS-cog, pain and seizures were 26.4%, 96.9% and 45.5%
(STD), 3.5%, 5.0% and 5.0% (IMA).
For all three data sets, STD showed considerable bias in the drug effect estimates in the majority of the scenarios but
no bias was evident for IMA in any scenario. When both STD and IMA estimates were unbiased, IMA in most cases
provided more precise estimates.
Conclusions: IMA was more robust towards model misspecifications and over-parameterization with better control
of the type I error and more accurate effect size estimates. IMA seems promising to evaluate the treatment effect in
longitudinal data analysis.
References: [1] Ito K et al. (2011). Alzheimers Dement. 7(2):151–60. [2] Schindler E, Karlsson MO (2017). AAPS
J. 19(5):1424–35. [3] Trocóniz IF et al. (2009). J Pharmacokinet Pharmacodyn. 36(5):461-77.
W-005
A Modeling and Simulation Study of Less Frequent Dosing of Pembrolizumab
Authors: Cody J. Peer1, Daniel A. Goldstein2, William D. Figg1, Mark J. Ratain3
Affiliations: 1Clinical Pharmacology Program, National Cancer Institute, Bethesda, MD, USA
2Davidoff Cancer Center, Rabin Medical Center, Israel 3Center for Personalized Therapeutics, University of
Chicago Comprehensive Cancer Center, Chicago, IL, USA
Objectives: Pembrolizumab is currently approved at 200 mg q3w, yet there is sufficient evidence that less frequent
dosing could maintain effective serum concentrations. Given the high cost of pembrolizumab and the lack of a dose-
response relationship, we hypothesized that less frequent dosing of 200 mg would maintain therapeutic serum
concentrations. The objective of this study was to use modeling and simulation to develop alternative dosing
strategies.
Methods: A simulation model was built from a published population pharmacokinetic model, incorporating time-
dependent clearance1. Various alternative dosing schedules were simulated, beginning with the third dose (doses 1
and 2 were 200 mg at wk 0 and 3). We chose 10 g/mL as the target concentration (TC), around the mean simulated
Ctrough at 1 mg/kg q3w based on observed efficacy2. The simulated dose schedules were q6w, q8w, and q10w,
beginning with the third dose. Simulations were performed on 50 simulated patients.
Results: The simulated Ctrough following doses 2-4 are presented in the table below. Dosing q6w should maintain TC
in >70% of patients.
Conclusions: Modeling and simulation provide evidence that pembrolizumab can be effectively dosed q6wk (after
the first 2 doses), resulting in a potential 60-70% cost savings. Randomized trials of this interventional
pharmacoeconomic strategy are indicated, where either a one size fits all approach, or a personalized strategy based
on trough levels could be employed. Similar opportunities may exist for other checkpoint inhibitors.
Table 1. Simulated Trough Concentrations for Pembrolizumab
Dose q6w q8w q10w
2 10.7 (9.0 – 12.7) 6.9 (5.6 – 8.5) 4.5 (3.6 – 5.8)
3 11.6 (9.8 – 13.9) 7.3 (5.9 – 8.9) 4.7 (3.8 – 5.9)
4 12.8 (10.8 – 15.2) 7.9 (6.5 – 9.9) 5.0 (3.9 – 6.3)
*values represent geo mean (95% CI) of simulated serum trough levels (ug/mL)
References: 1 Li H et al. J Pharmacokin Pharmacodyn. 44:403-414 (2017). 2 Patnaik A et al. Clin Cancer Research.
21(9):4286-93 (2015).
W-006
PKPD modeling of platelet time course in acute myeloid leukemia patients under MDM2 inhibitor HDM201
to support dosing regimen selection
Authors: Shu Yang1, Nelson Guerreiro2, Astrid Jullion2, Claire Fabre2 , Christophe Meille2
Institutions: 1Novartis Pharmaceuticals, East Hanover, NJ, USA 2Novartis Pharmaceuticals, Basel, Switzerland
Objectives: In acute myeloid leukemia (AML) patients, platelet levels are linked to both efficacy and safety. The
objective is to develop a PK/PD model describing the longitudinal time course of drug-induced platelet changes in
AML patients treated with the HDM2-p53 protein interaction inhibitor HDM201, in order to identify an optimized
dosing regimen maximizing the total dose while mitigating the safety risk of severe thrombocytopenia.
Methods: The PK and PK-platelet models were established in a two-step approach using non-linear mixed-effects
modeling implemented in Monolix 2018. First, the PK model was established. Secondly, individual PK parameters
were considered as fixed to estimate the PK/PD model that describes the longitudinal time course of drug-induced
platelet changes. This model, based on Friberg and Al. [2], was modified to take into account the perturbed
characteristics of hematopoiesis in AML by adding a surreptitious disease component, represented by leukemic
myeloblasts, to mimic a system with normal platelet production. As an extension of the methodology presented in
[1], several dosing regimens were simulated, and metrics were derived to support the optimized dosing regimen
selection (i.e. percentage of patients with platelet improvement or worsening from baseline).
Results: The established population PK model is one-compartment with a delayed parallel zero- and first-order
absorption process, and linear clearance (Cl/F). The PK/PD model [1] for platelet includes a drug toxic effect on
platelet production and a disease kinetics component, dependent of drug concentrations, influencing the feedback
mechanism. In addition, platelet transfusion events were taken into account in the models as 0.5h infusions with
estimation of individual platelet amount and half-life.
Conclusion: With this extension to AML, the platelet PK/PD model not only describes drug effect on hematopoiesis
and the drug-induced toxicity but also the drug effect on the disease and the resulting improvement on hematopoietic
function. In this complex situation, the population PK/PD model was utilized as an integrated quantitative tool for
suggesting an optimal dosing regimen strategy,
References: [1] Meille et al, 2016. “Revisiting Dosing Regimen Using Pharmacokinetic/ Pharmacodynamic
Mathematical Modeling: Densification and Intensification of Combination Cancer Therapy, Clin
Pharmacokinet. 2016 Aug;55(8):1015-25. [2] Friberg et al, 2002. “Model of chemotherapy-induced
myelosuppression with parameter consistency across drugs.” J. Clin. Oncol. 20:4713–4721
W-007
Population Pharmacokinetic and Pharmacodynamic Analysis to Evaluate a Switch to Doravirine
(DOR)/Lamivudine (3TC)/Tenofovir disoproxil fumarate (TDF) in Participants with HIV-1
Pavan Vaddady1, Bhargava Kandala1
Merck & Co. Inc., Kenilworth, NJ, USA
Objectives: To develop a population PK model and evaluate the exposure-response relationship for the proportion
of participants maintaining HIV-1 RNA < 50 copies/mL in virologically suppressed participants switching from a
stable antiretroviral regimen to DOR/3TC/TDF.
Methods: This analysis dataset includes 20 Phase 1 studies, 1 Phase 2b trial, and 2 Phase 3 trials, comprising 341
healthy volunteers and 959 participants with HIV-1 receiving oral doses of DOR as a single entity or as
DOR/3TC/TDF with the addition of 443 HIV-1-infected switch participants. The DOR population PK model
developed for treatment naïve population [1] was updated using DOR concentration data from the switch
population. The final model was used to estimate individual empirical Bayes PK parameter values for switch and
treatment naïve populations. The exposure-response relationship of DOR was explored graphically to assess
response at different quantiles of PK exposure in switch participants. Logistic regression was performed to evaluate
the exposure-response relationship. The proportion of participants maintaining HIV-1 RNA <50 copies/mL by the
Abbott Real Time HIV-1 Assay at study week 48 in the immediate switch group (ISG) using FDA snapshot
approach was used as the primary endpoint for this evaluation.
Results: Consistent with the prior results in treatment naïve participants, the PK of DOR in switch participants was
characterized with a one-compartment model described by first-order absorption (Ka), apparent volume of
distribution (V/F), and apparent linear clearance (CL/F) from the central compartment. The final population PK
model included statistically significant effects of body weight and healthy versus HIV-1-infected participant status
on the apparent volume of distribution and age on apparent clearance. The final pharmacokinetic parameters of the
model were well estimated with small relative standard errors (<15%). These estimates were very similar to the
previously developed population PK model [1]. The PK exposures (Cmax, C24 and AUC0-24) from the ISG
population were comparable to the treatment naïve population. Graphical analysis and logistic regression analysis of
the primary pharmacodynamics endpoint i.e., proportion of participants maintaining <50 copies/mL versus quantiles
of steady-state DOR C24 from the ISG (Figure 1) showed a flat exposure-response relationship. These results
indicate that the responses observed at the 100 mg dose lack exposure-dependency supporting appropriateness of the
selected dose in the switch population.
Conclusions: DOR PK exposures were consistent between treatment naïve and switch populations. A flat exposure-
response relationship was obtained between the proportion of participants maintaining <50 copies/mL and steady-
state DOR C24 at a 100 mg once daily dose in the switch population.
References: [1] Yee KL et. al. Population Pharmacokinetics of Doravirine and Exposure-Response in Individuals
with HIV-1. Antimicrob Agents Chemother. 2019 doi: 10.1128/AAC.02502-18.
Figure 1 Predicted and observed proportion of participants maintaining HIV-1 RNA <50 copies/mL using FDA
snapshot approach as a function of DOR steady-state C24 quartiles following administration of DOR 100 mg QD
(ISG, N=443)
W-008
Model-based Elaboration of a Limited Sampling Strategy in the Bioequivalence Assessment of Dabigatran
Cassandre Legault1, Fahima Nekka2,3, Jun Li2,3
1Faculté de Pharmacie, Université de Montréal, Montréal, Canada; 2Chidren’s Hospital of Eastern Ontario,
University of Ottawa, Ottawa, Canada; 3Centre de Recherches Mathématiques, Université de Montréal, Montréal,
Canada.
Objectives: The bioequivalence (BE) assessment of generic (Test) and brand name (Reference) formulations of
dabigatran represent an expensive challenge for pharmaceutical companies because of the drug’s high variability
and narrow therapeutic index. Using the population pharmacokinetic (PK) approach, the current study investigates
the potential to improve this procedure with a reduced number of samples of the Test formulations.
Methods: The dataset of a BE study of dabigatran including 16 subjects, corresponding to 640 plasma
concentrations, was used for the retrospective population PK analysis. The population PK models were accordingly
developed for each formulation. Sampling scenarios with increasingly reduced sampling time points (19, 15, 10, 8,
6, 5 and 4 samples included) were selected based on known PK properties and clinical knowledge of the drug.
Population PK models were refitted for each sampling scenarios and PK profiles were simulated and so forth. Then,
BE tests were performed based on these PK profiles to identify the most likely reduced sampling scenario that
preserved the BE conclusions obtained from the original dataset.
Results: The two-compartment model with first order elimination and a lagged absorption best described the plasma
concentration data for dabigatran, while sex was identified as a significant covariate for bioavailability. For the
seven reduced sampling scenarios, all simulated PK profiles were similar to the PK profile generated from the
complete sampling, except for the Cmax values. The BE test results proved that the BE verdict could be maintained
until the reduced sampling scenario of five samples using the current regulatory BE standards and criteria.
Conclusions: The population PK modeling approach can help reduce the number of samplings used for the BE
assessment of dabigatran, thus can potentially lower the costs of future BE trials.
REFERENCES: 1. Mekaj YH, M. A. (2015). New oral anticoagulants: their advantages and disadvantages compared with vitamin K antagonists in the prevention and treatment of patients with thromboembolic events. Therapeutics and Clinical Risk Management, 967–977. 2. CGPA), C. G. (2019). Market Trends. Retrieved from Canadian Generic Pharmaceutical Association (CGPA): http://canadiangenerics.ca/. 3. Upton, D. M. (2013). Basic Concepts in Population Modeling, Simulation, and Model-Based Drug Development—Part 2: Introduction to Pharmacokinetic Modeling Methods. CPT: Pharmacometrics & Systems Pharmacology. 4. Center for Drug Evaluation and Research (CDER) (2001). Statistical Approaches to Establishing Bioequivalence. U.S. Department of Health and Human Services, Food and Drug Administration. 5. Pradaxa® [Label] (Revised: 2013/12). NDA 022512, Drugs@FDA, FDA Web site. http://www.accessdata.fda.gov/scripts/cder/drugsatfda.
W-009
Population Pharmacokinetic Analysis of Esomeprazole in Japanese Subjects with CYP2C19 Phenotype
Mario Nagase1, Hitoshi Shimada2, Masahiro Nii2, Shinya Ueda2, Mitsuo Higashimori2, Katsuomi Ichikawa2, Li
Zhang1, Li Zhou1, Diansong Zhou1, Jim Dunyak1 and Nidal Al-Huniti1
1 CPSS, Biopharmaceuticals, AstraZeneca, Boston, United States; 2AstraZeneca KK, Osaka, Japan.
Objectives: Esomeprazole, the S‐isomer of omeprazole, is a potent proton pump inhibitor that has been approved
for treating multiple acid-peptic disorders. Esomeprazole is metabolized predominantly by hepatic CYP2C19, a
polymorphically expressed enzyme that show marked interindividual and interethnic variation. Individuals who are
deficient in CYP2C19 are poor metabolizers (PM), and which occurs in approximately 3% of Caucasians and 15-
20% of Asians. The objective of this population pharmacokinetic (PK) analysis was to describe esomeprazole PK
and to assess the influence of CYP2C19 phenotype on PK parameters in healthy Japanese population.
Methods: A total of 82 healthy Japanese male subjects plasma concentrations data pooled from 2 Phase I clinical
studies (D961HC00004 and D961HC00009), were used for model development. Esomeprazole (10, 20 or 40 mg)
was orally administered once daily for five days and plasma samples were collected on Day 5 for PK analysis.
NONMEM 7.4 was used for the population PK analysis. After development of base structural model, forward
selection and backward elimination were utilized to evaluate following covariates on esomeprazole PK: age, body
weight, and CYP2C19 phenotypes (i.e. homozygote extensive metabolisers (homo-EM), heterozygote extensive
metabolisers (hetero-EM), and PM). Model diagnostic plots were also used to identify the model with significant
covariate effects.
Results: The PK of esomeprazole in Japanese subjects were adequately described by one compartment model with
first-order absorption, and first order elimination. The estimated population means for esomeprazole apparent
clearance, apparent volume of distribution, first order absorption rate constant, were 11.6 (L/hr), 13.3 (L), 0.9 (1/hr),
respectively. Among the tested covariates, CYP2C19 phenotypes was identified as the statistically significant
covariate on the apparent clearance of esomeprazole. The inter-individual variability of esomeprazole apparent
clearance decreased from 45.8% to 34.6% after adding covariate of CYP2C19 phenotype. The results showed that
the apparent clearance in subjects with CYP2C19 hetero-EM and PM were 64.4% and 48.2% of that in individuals
with CYP2C19 homo-PM, respectively.
Conclusions: Based on developed Population PK model, CYP2C19 phenotype had significant impact on the
apparent clearance of esomeprazole in healthy Japanese male population. Further studies are warranted to evaluate
the effects of other covariates on esomeprazole exposure and to guide clinical study design in future.
W-010
Simulation Investigation of the Integrated Pharmacokinetic (PK) Model for Antibody-Drug-Conjugates
(ADC)
Leonid Gibiansky, Ekaterina Gibiansky
QuantPharm LLC, North Potomac, MD
Background: ADCs consist of dynamic mixtures of antibodies with different number of attached toxin molecules
(DAR, drug to antibody ratio). Measurements of components with different DARs are rarely available. Instead, total
antibody (tAB) and unconjugated toxin (T) are always measured. In addition, either number of all toxins attached to
antibodies with different DARs (antibody-conjugated toxin, acT) or number of antibodies with at least one toxin
attached (ADC=∑iADCi) are measured. Theoretical investigation [1] indicated that tAB and acT can be described by
two two-compartment models. These models were successfully applied to clinical data [2]. No theoretical
justification exists for PK ADC (=∑iADCi) equations, although linear two or three compartment models described
ADC in [3].
Objectives: To investigate, using simulations, relationships between acT, ADC, tAB, and T concentrations and to
suggest empirical simplifications for use in population PK modeling of ADC.
Methods: Typical concentration-time courses of all components (ADCi, i=0,..,8) of ADC mixture were simulated
using a linear 19-compartment model developed in [1]. Models with deconjugation in the central compartment (Case
1) and in both, central and peripheral compartments (Case 2) were investigated. Model parameters consistent with
values estimated for brentuximab vedotin [3] were used for simulations. Consistent with load-independent
assumption [1], deconjugation rate kdec and PK parameters of ADCDAR were assumed to be independent of DAR.
Simulated tAB, acT, ADC, DAR=acT/ADC, and T concentrations were used as observations (in different
combinations) to find simplest models that provide adequate fit.
Results: Figure 1 illustrates the results. As expected, tAB and acT were described by two-compartment models that
shared all parameters (except additional acT elimination due to deconjugation). ADC was well described by a two-
compartment model, but parameters differed from those used for simulations. In Case 2, DAR=acT/ADC was
described by a mono-exponential function as in [3], while in Case 1, a bi-exponential function provided a better fit.
However, results could be dependent on the choice of model parameters. As expected, toxin concentrations were
well described by a combined acT-T model (with acT elimination directed to T compartment). Surprisingly, ADC-T
model (with ADC elimination directed to T compartment without accounting for DAR) was as good as acT-T
model, although it was not able to correctly estimate toxin PK parameters. tAB-T model was only marginally worse
than ADC-T model.
Conclusions: Models that use ADC rather than acT measurements can be applied to describe ADC PK. However,
parameters estimated by these models may differ from true parameters. ADC DAR ratio decreases as a mono- or bi-
exponential function of time after dose.
Reference: [1] Gibiansky L, Gibiansky E, J Pharmacokinet Pharmacodyn. 2014 Feb;41(1):35-47. [2] Lu D et al,
CPT Pharmacometrics Syst Pharmacol. 2016, 5(12):665-673. [3] Li H et al, J Clin Pharmacol. 2017, 57(9):1148-
1158.
W-011
Creatinine Based Renal Function Formula: Can They be Used to Predict Drug Clearance?
Yifei Zhang1, Alena Zhang1, Neha Mehta1, Catherine Sherwin3, Xiangyu Liu2, Mona Khurana1, Yaning Wang2, Jian
Wang1
1Office of Drug Evaluation IV, OND, U.S Food and Drug Administration; 2Office of Clinical Pharmacology, OTS,
U.S Food and Drug Administration; 3Department of Pediatrics, Dayton Children's Hospital
Objectives: Clinical assessment of renal function mainly relies on calculation of estimated glomerular filtration rate
(eGFR) from serum creatinine (SCR) concentration. Our goal is to compare the accuracy of currently available
eGFR equations, and to evaluate their ability to predict the clearance of predominantly renal-eliminated drugs in
pediatric population.
Methods: Individual drug clearance was obtained from population PK models using NONMEM. Eight different
SCR-based equations, such as Schwartz equation, were used to calculate eGFR for each individual. The obtained
eGFR values were compared with the normal range of GFR for each age group. The eGFR were also compared with
observed clearance of drugs that are >90% renal eliminated, including Gadobutrol, Gadoterate, Amikacin and
Vancomycin. The performance of eGFR equations in predicting drug CL was evaluated by linear regression and
concordance analysis.
Results: The observed clearance of all four renal-eliminated drugs in children was significantly correlated but
consistently lower than the eGFR value calculated by the Schwartz equations. The Leger and Cockcroft-Gault
equations over-predicted drug clearance by more than two-fold for most individuals. The Schwartz equations also
over-predicted drug clearance, but to a less extent. Further, individual eGFR calculated by the Modified Schwartz
equation (k=0.413) had multiple outliers dramatically exceeds upper limit of the normal range of GFR (mean +
2SD), however, their drug clearance did not show a significant difference from other individuals. Those subjects
typically have SCR < 0.3 mg/dL and relatively low in body weight and body surface area.
Conclusions: The Schwartz equation results in a significant proportion of subjects that are above the normal GFR
range in pediatrics. Using the bedside creatinine-based formula over-predicts clearance for the tested renally-
eliminated drugs. Caution should be taken if applying the calculated eGFR to derive the dose of renally eliminated
drugs in children with low serum creatinine levels.
W-012
Can exposure-response analysis provide greater power to detect changes in vitals in a sparse sampling
paradigm as compared to standard statistical analysis?
Sarita Koride1, Sridhar Duvvuri1, and Arthur Bergman1 1Early Clinical Development- Clinical Pharmacology, Pfizer, Inc., Cambridge, MA, USA
Objectives: Exposure-response (E-R) analysis is a standard tool to detect and quantify signals of interest in clinical
trials. In phase 2a studies, with sparse sampling, use of standard statistical analysis to detect changes in ECG/Vitals
might be limited. The objective was to understand the ability of E-R modeling to detect changes in vitals - blood
pressure and heart rate - and to optimize for PK and vital collection times for Phase 2 studies.
Methods: A typical 16-week Phase 2a type study with three treatment groups and number of subjects, n=30/44 was
considered for clinical trial simulations. Assuming a half-life of around 11 hours and a peak to trough ratio of 5, a
standard two compartmental PK model with combined proportional and additive residual error was used to simulate
concentration-time profiles. Vitals (blood pressure and heart rate) were simulated using estimates for inter-
individual variability and residual error from placebo data of a previous study and assuming different slopes for
effect of concentration. Three practical sampling schemes were considered. All schemes had three peaks and three
trough collection times for PK and vitals. In addition, schemes 2 and 3 had one and two 4-hour post dose samples
respectively. Additional positive control sampling scenarios were also considered (cross-over designs and extensive
PK sampling). A linear model as suggested in the scientific white paper for concentration-QTc modeling was used
to fit simulated PK-vitals data1. Success was defined as lower limit of 90% CI of estimated vitals measure at
maximum concentration observed in the study being above zero.
Results: Smallest detectable changes in BP and HR for the three sampling schemes with the samples sizes
considered did not vary significantly and were similar to the values obtained from standard statistical analysis with
80% power and 5% significance level, 1-sided. As expected, probability to detect similar magnitude signal increased
with positive control scenarios.
Table 1: Smallest detectable mean changes relative to placebo with 80% power for Systolic blood pressure (SBP)
Parameter E-R simulation analysis
Statistical Analysis Scheme 1[n=30,44] Scheme 2[n=30,44] Scheme 3[n=30,44]
SBP (CFB, mmHg) 6.71, 5.67 6.5, 5.46 6.5, 5.38 6.7, 5.5
Conclusions: In standard Phase 2a studies, with sparse sampling and limited number of doses, exposure response
analysis may not always confer greater power in detecting BP and HR signals than standard statistical analysis.
Future work includes conducting this exercise in studies with sparse sampling but with a wider dose range.
References: 1. Garnett, C. et al. Scientific white paper on concentration-QTc modeling. J. Pharmacokinet. Pharmacodyn. 45, 383–397 (2018).
W-013
Characterizing Impact of Interoccasion Variability in Dasatinib Exposure on Efficacy in Chronic Phase
Chronic Myeloid Leukemia (CP-CML) Patients
Sai Praneeth Bathena1, Xiaoning Wang1,2, Amit Roy1
1Clinical Pharmacology and Pharmacometrics, Bristol-Myers Squibb, Lawrenceville, NJ, USA 2Current affiliation: Metrum Research Group, Tariffville, CT USA
Objectives: Interoccasion variability (IOV) in dasatinib exposure is higher under fasted conditions compared to fed
conditions, and the higher IOV has been attributed to higher variability in bioavailability (F) under fasted
conditions. Our objective was to investigate the potential association between IOV in F and efficacy in subjects with
CP-CML.
Methods: Dasatinib relative F across occasions was estimated for each subject using a previously developed PPK
model that incorporated both interindividual variability (IIV) and IOV in F [1]. The extent of IOV in F for each
subject was then quantified by the ratio of the maximum and the minimum value of F across all occasions (IOV
ratio). The potential association between the IOV ratio and efficacy endpoints was investigated graphically, as well
as by model-based analyses separately for first line (1L) and second- line (2L) CP-CML patients. The following 3
efficacy endpoints were described by logistic regression models of exposure-response (E-R): major cytogenetic
response at 6 months (MCyR-6mo), complete cytogenetic response at 12 months (CCyR-12mo), and major
molecular response at 12 months (MMR-12mo). IOV ratio was assessed as both a categorical and continuous valued
predictor, and the analyses included the following previously identified predictors of efficacy [2]: time-averaged
concentration at steady state weighted by dose modification (wCavgss), % dose interruption duration, age, sex, race,
and imatinib status (resistant or intolerant).
Results: The graphical exploratory analyses did not suggest an association between IOV ratio and efficacy. Figure 1
shows a representative exploratory plot of IOV ratio vs MCyR-6mo in 1L and 2L CP-CML. Additionally, E-R
analyses showed that IOV ratio (continuous and categorical forms) did not have a significant effect on MCyR-6mo,
CCyR-12mo, or MMR-12mo (odds ratio 95% CI includes 1.0). Consistent with previous findings that in 2L CP-
CML the probability of achieving response was lower for imatinib-resistant subjects compared to imatinib-intolerant
patients and was higher in subjects who had less dose interruption in both 1L and 2L CP-CML for all efficacy
endpoints. There was a trend towards a higher probability of response with increasing wCavgss in both 1L and 2L
CP-CML for all endpoints.
Conclusions: The graphical and model-based E-R analyses utilizing IOV ratio as a predictor did not show a
significant effect of IOV in F (and therefore exposure) on response for all endpoints. Higher IOV in F associated
with fasted conditions, does not have a significant effect on efficacy.
References: [1] Dai G et al. J Clin Pharmacol. 2008 Nov;48(11):1254-69 [2] Wang et al. Clin Pharmacol. 2013 Jun
10; 5:85-97
Figure 1: IOV Ratios by MCyR-6mo Response Status in 1L and 2L CP-CML Patients
W-014
Dose Correction for the Michaelis-Menten Approximation of the Target-Mediated Drug Disposition Model
for Multiple Dosing Regimen
Xiaoyu Yan1, Juan Jose Perez Ruixo2, Wojciech Krzyzanski3
1 School of Pharmacy, The Chinese University of Hong Kong, Hong Kong; 2 Janssen Research & Development,
Beerse, Belgium; 3 Department of Pharmaceutical Sciences, School of Pharmacy and Pharmaceutical Sciences,
University at Buffalo, Buffalo, New York, USA
Objective: The Michaelis-Menten (M-M) approximation of the target-mediated drug disposition (TMDD) model
with intravenous (IV) bolus administrations requires a correction of the input dose to account for the free drug lose
before reaching the rapid binding equilibrium. The objective of this research provides a method for implementation
of dose correction in the M-M approximation of the TMDD model upon multiple IV bolus administrations.
Methods: We derived the formula of a correction factor (Fcorr) for each dose in a multiple dosing regimen with IV
administration for M-M model. Fcorr depends on the residual free drug amount prior dosing event and administered
IV bolus dose. Previously published clinical PK data of recombinant human erythropoietin (rHuEPO) from four
clinical trials in healthy subjects receiving multiple IV bolus doses were analyzed by both M-M model with and
without dose correction (MMC and MMNC). Simulations were further conducted to evaluate the influence of dose,
concentration and target-binding parameters on Fcorr for the IV dose of rHuEPO. The analysis was performed using
NONMEM 7.4 with the FOCEI estimation method.
Results: Fcorr was calculated on the fly for each IV dose in a multiple dosing regimen and dynamically accounted
for the drug amount which is already present in the system at each dosing time. Our analysis showed that MMC
outperformed MMNC as demonstrated by basic model diagnostics and visual predictive check. There was also a
clear trend in individual PK parameters with dose for MMNC, which suggested that MMNC was not able to
adequately characterize the nonlinearity in the PK data of rHuEPO. The estimation of random effect by MMNC
appeared to be inflated due to the model deficiency. For instance, the random effect associated with the central
volume of distribution estimated by MMNC was about 6-fold of that estimated by MMC. The fixed effect
parameters estimated by MMNC also deviated substantially from those by MMC. The simulation demonstrated that
for HuEPO at thricely weekly 10 IU/kg dosing regimen yielded Fcorr = 0.5. This result suggested that the lower
than expected exposure for rHuEPO at low dose is due to target binding.
Conclusion: For M-M approximation of TMDD model, dose correction should be implemented to avoid model
misfit and potential bias in estimates of PK parameters. Inappropriate characterization of drug disposition may
further lead to erroneous estimation of absolute bioavailability.
W-015
A Physiological Model of Anticancer Drug-Induced Neutropenia: Clinical Translation of in vitro Bone
Marrow Toxicity Induced by Cytotoxic Agents
Wenbo Chen1, Britton Boras2, Tae Sung2, Jenny Zheng3, Yanke Yu2, Diane Wang2, Wenyue Hu2, Mary E. Spilker2,
David Z. D’Argenio1
1Department of Biomedical Engineering, University of Southern California, Los Angeles, CA, USA; 2Pfizer, Inc.,
San Diego, CA, USA; 3Pfizer, Inc., Collegeville, PA, USA.
Objectives: Neutropenia is the most common side effect and dose-limiting toxicity of chemotherapy. In vitro bone
marrow toxicity (BMT) assays have been used to assess the potential of anticancer agents to induce neutropenia in
clinical trial patients. To provide a mechanistic basis for such in vitro to clinical translation, we have incorporated
BMT assay results from several cytotoxic drugs into a physiological model of granulopoiesis [1] based on the
distinct mechanism of action (MOA) of each drug and evaluated the resulting model’s ability to predict neutropenia.
Methods: An in vitro bone marrow proliferation and differentiation assay [2] was used to determine the time course
of viable bone marrow cells during exposure to paclitaxel, docetaxel, and gemcitabine over wide dose ranges. An in
vitro version of a previously reported physiological model of granulopoiesis and its regulation [1] was constructed
that incorporated the MOA of each of the cytotoxic drugs based on the BMT assay results. The in vitro model
included the hematopoietic stem cell (HSC), proliferating progenitor cell (PC) with G1, S, and G2/M phase cells,
along with the differentiation process, which are each part of the complete physiological model. The drug specific in
vitro model was then incorporated into the whole-body physiological model and used to predict, via a population
simulation, absolute neutrophil count (ANC) in response to each of the cytotoxic agents.
Results: Paclitaxel and docetaxel were modeled as inducing apoptosis of M-phase PC, while gemcitabine was
modeled as inducing apoptosis of S-phase PC. The estimated IC50s and other drug specific parameters obtained
from fitting the models to the BMT assay dose-response curves for each drug were used in the physiological model
to predict the ANC time profiles (nadir, time to reach nadir, and recovery) following single dose administration.
These results were consistent with the ANC time course profiles predicted by previously reported models for
paclitaxel and docetaxel [3]. Moreover, the incidences of neutropenia grades 1-4 predicted from the population
simulation of the physiological model, after stratifying for baseline ANC, were in overall agreement with the
corresponding neutropenia grades we obtained based on simulations of the models reported in [3] for the cytotoxic
agents studied.
Conclusions: The proposed model-based approach presented may serve as a mechanistic framework for predicting
neutropenia based on in vitro BMT assay results, including the effects of treatment with multiple agents.
References: [1] Chen et al., J Pharmacokinet Pharmacodyn 45:S127,2018 [2] Hu et al., Clin Cancer Res
22:2000:2008,2016 [3] Friberg et al., J Clin Oncol 20:4713-4721,2002
W-016
Benefit/Risk Assessment of Sarilumab Using Exposure/Response Modeling of Key Efficacy and Safety
Endpoints in Patients with Rheumatoid Arthritis
Bernard Sebastien1, Sunny Sun2, Clemence Rigaux-Lampe1, Hui Quan3, Anne Paccaly4, Yong Lin3, and Christine
Xu3
1Sanofi, Chilly-Mazarin, France 2Sanofi Research and Development, Beijing, China
3Sanofi Genzyme, Bridgewater, NJ, USA 4Regeneron Pharmaceuticals, Inc, Tarrytown, NY, USA
Objectives: Sarilumab, a human mAb blocking cytokine IL-6Rα, is approved for the treatment of rheumatoid
arthritis (RA) and has shown efficacy in improving the signs and symptoms of RA, physical function and
radiographic deterioration. Exposure/response (E/R) analyses of selected efficacy and safety endpoints were
conducted to evaluate E/R relationships of subcutaneous (SC) sarilumab and support the benefit-risk assessment of
sarilumab SC dose regimens (150 and 200 mg every 2 weeks [q2w]).
Methods: Efficacy data (Disease Activity Score 28–C-Reactive Protein [DAS28–CRP], Clinical Disease Activity
Index [CDAI], Health Assessment Questionnaire Disability Index [HAQ-DI], and van der Heijde modified total
Sharp [mTSS]) were included from 1915 patients with RA treated with sarilumab SC 150 or 200 mg q2w in two
Phase 3 studies (NCT01061736/NCT01709578). Safety data (percent change in low density lipoprotein [LDL]
levels, alanine aminotransferase (ALT; x upper limit of normal [ULN]), and time to first ALT >3x ULN) were
collected from an additional 306 patients treated with sarilumab SC 100 and 150 mg every week or 100, 150, and
200 mg q2w from a Phase 2 study (NCT01061736). For each endpoint, linear, log-linear and Emax E/R models, with
appropriate covariates, were evaluated to select the model that best fitted the data.
Results: Log-linear or Emax models best described the E/R relationships between the efficacy endpoints, DAS28–
CRP, CDAI, HAQ-DI and mTSS, and sarilumab trough serum concentration (Ctrough) at Week 24. With the
exception of HAQ-DI (which indicated only a small improvement with increase in exposure from 150 to 200 mg
q2w), higher exposure resulted in increased efficacy for all endpoints. Percent change in LDL, ALT x ULN and time
to first ALT >3x ULN at Week 12 and Week 24 were best explained using an Emax relationship: effects plateaued in
the lower range of Ctrough, meaning marginal increase in effect or risk from sarilumab 150 mg to 200 mg q2w SC.
Conclusions: E/R relationships for efficacy and safety endpoints support a sarilumab SC starting dose of 200 mg
q2w. This is consistent with the approved label that recommends a starting dose of 200 mg q2w with a decrease to
150 mg q2w in the event of laboratory abnormalities1.
References: 1. Meng et al, Poster II-69. Population Approach Group in Europe Meeting, Budapest, Hungary; June
6-9, 2017.
Acknowledgements: Sanofi and Regeneron provided study funding and medical writing support (Sarah Feeny,
Adelphi Communications).
Disclosures:
Bernard Sebastien, Sunny Sun, Clemence Rigaux-Lampe, Hui Quan, Yong Lin, and Christine Xu are employees of
Sanofi and may hold stock and/or stock options in the company. Anne Paccaly is an employee of Regeneron
Pharmaceuticals, Inc, and may hold stock and/or stock options in the company.
W-017
Exposure-Response Analysis of Efficacy and Safety for Pexidartinib in Patients with Tenosynovial Giant Cell
Tumor (TGCT)
Ophelia Yin1, Jonathan French2, Daniel Polhamus2, Hamim Zahir1, Michiel van de Sande3, William D. Tap4, Hans
Gelderblom3, Andrew J. Wagner5, Jon Greenberg6,Dale Shuster6, Silvia Stacchiotti7
1Quantitative Clinical Pharmacology and Translational Sciences, Daiichi Sankyo, Inc., Basking Ridge, NJ; 2Metrum
Research Group, Tariffville, CT; 3Leiden University Medical Center, Leiden, Netherlands; 4Memorial Sloan
Kettering Cancer Center, New York, NY; 5Dana-Farber Cancer Institute, Boston, MA; 6Global Oncology R&D,
Daiichi Sankyo, Inc., Basking Ridge, NJ; 7Fondazione IRCCS Instituto Nazionale dei Tumori, Milan, Italy
Objectives: Pexidartinib (PLX3397) is a novel tyrosine kinase inhibitor that targets colony-stimulating factor 1
receptor and shows significant antitumor activity in patients with TGCT. Exposure-response (E-R) analyses for
efficacy and safety were conducted to support the pexidartinib dose selection for patients with TGCT.
Methods: Exposure-efficacy analysis was based on data from 113 TGCT patients in the phase 3 ENLIVEN
(PLX108-10) study, which included the primary efficacy endpoint of response evaluation criteria in solid tumors
(RECIST) at Week 25, and the secondary efficacy endpoint of tumor volume score (TVS) response at Week 25.
Pharmacokinetic-pharmacodynamic (PK-PD) modeling was performed to assess the relationship between
pexidartinib exposure and longitudinally-measured tumor size (i.e., tumor size measurements collected at different
time points over treatment course), using data of 141 TGCT patients from the phase 1 PLX108-01 study and the
ENLIVEN study. Exposure-safety relationships were assessed in 214 patients with TGCT or other solid tumors from
studies PLX108-01 and ENLIVEN, including the endpoints related to significant elevations in alanine
aminotransferase (ALT), aspartate aminotransferase (AST), and total bilirubin (TBIL). For each endpoint, logistic
regression or time-to-event modeling was performed, and various pexidartinib exposure metrics were tested.
Results: The E-R relationships for both the RECIST- and TVS-based overall response rate (ORR) were adequately
described by proportional odds logistic regression models, using pexidartinib average daily AUC as the descriptor.
PK-PD analysis of longitudinal tumor size showed similar findings as logistic regression analysis of ORR, where
higher average daily AUC was associated with greater reduction in the tumor size. Further simulations suggested
predicted ORR at Week 25 increased from 0.25 (90% CI: 0.15, 0.36) at 400 mg/day to 0.32 (90% CI: 0.23, 0.42) at
800 mg/day, with no discernable difference between two dose regimens from the ENLIVEN study (800 mg/day
versus 1000 mg/day for 14 days followed by 800 mg/day). Overall, clinical observation from the ENLIVEN study
suggested a lower rate of hepatic adverse reactions (30.0% versus 41.0%) in the crossover subjects who received
pexidartinib 800 mg/day, as compared to subjects who received 1000 mg/day for 14 days followed by 800 mg/day.
The associations between the time to hepatic enzyme elevations and time-varying average pexidartinib concentration
were well-characterized by piecewise-exponential survival models. Higher pexidartinib average concentration (Cav)
was statistically significantly associated with a higher incidence of elevated ALT (>3 x ULN or >5 x ULN) and AST
(>3 x ULN or >5 x ULN), and therefore the predicted incidence of ALT and AST elevations increased from 400
mg/day to 1000 mg/day regimens. No statistically significant E-R relationship was identified for TBIL (>2 x ULN
or >2 x baseline).
Conclusion: E-R analyses together with clinical data supported the recommendation of pexidartinib 800 mg/day
without a loading dose for patients with TGCT.
W-018
Population Pharmacokinetics of Sarilumab in Japanese and Non-Japanese Patients with Rheumatoid
Arthritis
Christine Xu1, Yoshihisa Shitara2, Anne Paccaly3, Vanaja Kanamaluru1
1Sanofi Genzyme, Bridgewater, NJ, USA 2Sanofi, Tokyo, Japan 3Regeneron Pharmaceuticals, Inc, Tarrytown, NY, USA
Objectives: To develop and validate a population pharmacokinetic (PopPK) model of sarilumab, a human
monoclonal antibody blocking the interleukin (IL)-6α receptor, in Japanese and Non-Japanese patients with
rheumatoid arthritis (RA), to determine sources of pharmacokinetic variability, and to identify covariates that are
potential sources of variability in exposure.
Methods: A PopPK model was developed from sarilumab serum concentration data pooled across eight Phase 1,
one Phase 2, and seven Phase 3 studies (including NCT01328522, NCT01850680, NCT02097524, NCT02017639,
NCT02404558, NCT01061736, NCT01768572, NCT02057250, NCT02121210, NCT01709578, NCT02293902,
NCT02373202), resulting in a final dataset of 12088 observations from 2,453 patients (76% Caucasian, 16% Asian
including 285 Japanese, 3% Black and 5% other race) with RA. Leveraging the prior PopPK analyses and
knowledge gained from cross-study comparisons, evaluation of covariates focused on demographic characteristics
(including race, and Japanese vs non-Japanese), renal function, anti-drug antibody (ADA) and baseline disease
activity. Potential covariates were identified according to a forward-addition, backward-deletion strategy. Validation
of the final PopPK model was performed using bootstrapping and visual predictive checks.
Results: The pharmacokinetics of sarilumab were adequately described by a 2-compartment, target-mediated drug
disposition model with parallel linear and non-linear (ie Michaelis-Menten) elimination and first order absorption.
Body weight, ADA-status, albumin, sex, creatinine clearance and baseline C-reactive protein were statistically
significant covariates influencing sarilumab pharmacokinetics. Sarilumab exposure increased with a lower body
weight (analysis range 31.5 kg–183 kg). Compared with a typical 70 kg patient, with sarilumab at 150 mg q2w
subcutaneously (SC) and 200 mg q2w SC, respectively, AUC0-14 increased by 32% and 25% for a 59 kg patient, and
decreased by 23% and 20% for an 82 kg patient. Other statistically significant covariates were not considered to be
clinically relevant. Race and Japanese vs non-Japanese had no statistically significant effect on the pharmacokinetics
of sarilumab in the presence of the other covariates. A post hoc graphical inspection after repeated dosing suggested
little or no impact of baseline Disease Activity Score 28-CRP or prior use of biologics on sarilumab exposure.
Conclusions: PopPK analysis indicated the absence of differences in the pharmacokinetics of sarilumab between
Japanese and non-Japanese patients, aside from differences in body weight, which is the main intrinsic covariate in
pharmacokinetics. No adjustment in sarilumab dose is required for body weight, or any other demographic
characteristics assessed.
References: 1. Xu et al, Poster II-63. Population Approach Group in Europe Meeting, Budapest, Hungary; June 6-9,
2017.
Acknowledgements: Sanofi and Regeneron provided study funding and medical writing support (Sarah Feeny,
Adelphi Communications).
Disclosures: Christine Xu, Yoshihisa Shitara and Vanaja Kanamaluru are employees of Sanofi and Anne Paccaly is
an employee of Regeneron Pharmaceuticals, Inc. All authors may hold stock and/or stock options in their company.
W-019
Population Pharmacokinetic and Exposure Safety Analyses of Deutetrabenazine in Patients with Moderate to
Severe Tardive Dyskinesia
Micha Levi1, Frank Schneider2, Nathalie H. Gosselin3, Donna Cox1, Juha-Matti Savola1 1Teva Pharmaceuticals, 2Ratiopharm GmbH, 3Certara
Objective: A population PK (popPK) model was developed to assess the individual systemic exposure to the active
deuterated metabolites of deutetrabenazine, α-HTBZ and β-HTBZ in patients with moderate to severe tardive
dyskinesia (TD) and, to evaluate the relationship between total (α+β)-HTBZ exposure and adverse events (AEs).
Background: Austedo (deutetrabenazine) is a highly selective VMAT2 inhibitor approved by the FDA in 2017 as a
treatment for chorea associated with Huntington's disease and TD.
Design/Methods: Rich PK data from phase 1 studies of healthy volunteers and patients with Tourette syndrome,
and sparse PK data from phase 3 studies of moderate to severe TD patients who received a total daily dose of 12-48
mg deutetrabenazine were pooled to construct the popPK model. The dataset for the exposure-safety analysis
included AEs and individual exposure parameters of total (α+β)-HTBZ for TD patients in the phase 3 safety
populations. Logistic regressions were performed to evaluate the associations of total (α+β)-HTBZ exposure (AUC,
Cmin and Cave) for patients with the selected AEs (akathisia, depression, diarrhea, insomnia and somnolence) and
their probability.
Results: The popPK model described the exposures in TD patients reasonably well. Of the 410 patients in the safety
population, 13% experienced one of the selected AEs. No patients on placebo reported akathisia, 1 reported
depression, 5 diarrhea, 1 insomnia and 9 somnolence. Four patients on deutetrabenazine reported akathisia, 5
depression, 12 diarrhea, 10 insomnia and 12 somnolence. As shown in Table 1, there was no relationship between
the occurrence of these AEs and exposure to total (α+β)-HTBZ, perhaps due to the small number of AEs. There was
no relevant trend on the probability of the selected AEs over time.
Conclusions: Overall, this study showed no exposure-response relationship between total (α+β)-HTBZ exposure
following daily dosing of deutetrabenazine and selected adverse events such as akathisia, depression, diarrhea,
insomnia and somnolence for patients with TD.
Table 1. Logistic regressions for probability of adverse events based on Cave of total (α+β)-HTBZ
Adverse Events PK exposure of
(α+β)-HTBZ
Chi-square p value Odds Ratio 95% CI of
Odds Ratio
Akathisia Cave (ng/mL) 2.9520 0.0858 1.0164 [0.9971,
1.0315]
Depression Cave (ng/mL) 0.0575 0.8106 0.9969 [0.9622,
1.0171]
Diarrhea Cave (ng/mL) 2.3244 0.1274 1.0090 [0.9971,
1.0191]
Insomnia Cave (ng/mL) 1.0953 0.2953 1.0078 [0.9921,
1.0198]
Somnolence Cave (ng/mL) 0.4946 0.4819 1.0042 [0.9916,
1.0144]
W-020
Novel population pharmacokinetic approach to explain the differences between cystic fibrosis patients and
healthy volunteers via protein binding
Nirav R Shah1, Jürgen B Bulitta1, Martina Kinzig2, Cornelia B. Landersdorfer3, Yuanyuan Jiao1, Dhruvitkumar S
Sutaria1, Xun Tao1, Rainer Höhl4, Ulrike Holzgrabe5, Frieder Kees6, Ulrich Stephan2,7, & Fritz Sörgel2,7
1Department of Pharmacotherapy and Translational Research, College of Pharmacy, University of Florida, Orlando,
FL, USA; 2IBMP – Institute for Biomedical and Pharmaceutical Research, Nürnberg-Heroldsberg, Germany; 3Drug
Delivery, Disposition and Dynamics, Monash Institute of Pharmaceutical Sciences, Monash University, Parkville,
VIC, Australia; 4Institute of Clinical Hygiene, Medical Microbiology and Infectiology, Klinikum Nürnberg,
Paracelsus Medical University, Nürnberg, Germany; 5Institute for Pharmacy and Food Chemistry, University of
Würzburg, Würzburg, Germany; 6Department of Pharmacology, University of Regensburg, Germany; 7Department
of Pharmacology, University of Duisburg, Essen, Germany
Objectives: The pharmacokinetics in patients with cystic fibrosis (CF) has long been thought to differ considerably
from that in healthy volunteers. While profound differences were found in comparatively morbid patients with CF
for highly protein bound β-lactams, these differences could be explained by body weight and body composition for
β-lactams with low protein binding. This study aimed to develop a novel population modeling approach to describe
the pharmacokinetic differences between both subject groups by estimating protein binding.
Methods: Eight patients with CF (lean body mass [LBM]: 39.8±5.4kg) and six healthy volunteers (LBM:
53.1±9.5kg) received 1,027.5 mg cefotiam intravenously. We determined cefotiam in plasma and urine by HPLC-
UV and used the S-ADAPT software package for population PK modeling. We accounted for the effect of body size
and body composition by linear and allometric scaling based on total body weight (WT) and lean body mass (LBM).
Sex and disease state were studied as covariate on fraction of unbound cefotiam.
Results: A three-compartment model was defined for unbound cefotiam. Unscaled total clearance and volume of
distribution were 3% smaller in patients with CF compared to those in healthy volunteers. After accounting for body
size and composition via allometric scaling by LBM, renal clearance was 21% larger, non-renal clearance 11%
larger, and volume of distribution 38% larger in patients with CF compared to healthy volunteers, when calculating
these PK parameters based on total drug concentrations. We further explained the remaining pharmacokinetic
differences by estimating the unbound fraction of cefotiam in plasma. The latter was fixed to 50% in male and
estimated as 54.5% in female healthy volunteers as well as 56.3% in male and 74.4% in female patients with CF.
Conclusions: This study was the first to compare the population PK of cefotiam in patients with CF to that in
healthy volunteers. Overall, the proposed novel population modeling approach holds promise to describe potential
PK differences in special patient populations with altered protein binding.
Figure 1. Visual predictive check for cefotiam concentrations in patients with CF (left) and healthy volunteers
(right). The plots show the observations (markers), the 50th percentile (i.e. median) of the model predictions (black
line) along with the 80% prediction interval [10th to 90th percentile] and the interquartile range [25th to 75th
percentile]. Ideally, the median of the observations and of the predictions should superimpose and 10% of the
observations should fall outside the 80% prediction interval on either side.
W-021
Modeling the In Vivo Target Suppression of a Bispecific Antibody Against Human Tumor Necrosis Factor
(TNF) Alpha and Interleukin (IL)-17A
Songmao Zhenga, Fang Shenb, Damien Finka, Ivo Nnanec, Brian Geista, Qun Jiaoa, Lanyi Xiea, Nando Bansalb,
Tatiana Ortb, Weirong Wanga
aBiologics Development Sciences, Janssen Biotherapeutics, bImmunology Discovery, cClinical Pharmacology and
Pharmacomatrics, Janssen R&D. 1400 McKean Road, Spring House, PA 19477
Objectives: Tumor necrosis factor alpha (TNFα) and interleukin (IL)-17A are pleiotropic cytokines implicated in
the pathogenesis of several autoimmune diseases. JNJ-61178104 (JNJ-8104) is a novel human anti-TNFα and anti-
IL-17A monovalent, bispecific antibody. The objective of the study is to characterize the pharmacokinetics (PK) and
target engagement (TE) of JNJ-8104 and comparing to its parental antibodies in cynomolgus monkeys (Cyno) and in
human.
Methods: In the main PK/TE study in Cyno, the animals received a single intravenous (IV) bolus dose of either
JNJ-8104 at 0.3, 1 or 10 mg/kg, CNTO 148 (golimumab, anti-TNFα parental antibody) at 0.15 or 0.5 mg/kg, or
CNTO 6785 (anti-IL-17A parental antibody) at 0.15 mg/kg, on study Day 0. Blood samples were collected at
selected time points for PK and TNFα/IL-17A TE measurements. Similarly, PK and TE samples were collected and
analyzed from the Cyno toxicology study where the animals received weekly IV dose of either CNTO 148 at 50
mg/kg, CNTO 6785 at 50 mg/kg or JNJ-8104 at 100 mg/kg. JNJ-8104 was also investigated in a first-in-human
(FIH) study. Fifty-four healthy subjects were enrolled in one of 5 single IV dose levels (0.1, 0.3, 1, 3, and 10 mg/kg)
or a single SC dose (1 mg/kg). Blood samples were collected to measure JNJ-8104 PK, total TNFα and total IL-17A.
A quasi-equilibrium target-mediated drug disposition (TMDD) model was developed in NONMEM 7.3 to describe
the interaction between and the antibodies and their ligands in the systemic circulation of Cyno. The scaled PK/TE
parameters were then used to simulate the PK/TE profiles in human.
Results: JNJ-8104 was shown to engage its targets, TNFα and IL-17A, in systemic circulation in normal
cynomolgus monkeys. Surprisingly, the accumulation of total target following JNJ-8104 dosing was substantially
lower than that following the molar-equivalent doses of its bivalent parental antibody. Quantitative PK/TE
assessment suggested substantially lower apparent in vivo target-binding affinities for JNJ-8104 when comparing to
its bivalent parental antibodies, despite their similar in vitro monovalent target-binding affinities and potencies. The
target engagement profiles of JNJ-8104 in humans were in general agreement with the predicted profiles based on
cynomolgus monkeys, suggesting it may bind to human TNFα and IL-17A with lower affinities than its bivalent
parental antibodies as well.
Conclusions: Quantitative PK/TE assessment suggested interesting monovalent target binding differences between
the monovalent bispecific antibody JNJ-8104 and its bivalent parental antibodies. This difference may be related to
the differential binding characteristics of monovalent and bivalent antibodies when they bind to low-abundance
dimeric (e.g. IL-17A) or trimeric (e.g. TNFα) target in vivo. These findings provide valuable insights into the design
of biologics against dimeric or trimeric soluble targets.
W-022
Development and External Validation of a Nivolumab Population Pharmacokinetic (PPK) Model in
Previously Treated Advanced Non-Small Cell Lung Cancer (NSCLC) Patients With Disease Control
Xiaochen Zhao1, Elizabeth Ludwig2, Jill Fiedler-Kelly2, Luann Phillips2, Erin Dombrowsky1, Amit Roy1
1Bristol-Myers Squibb, Princeton, NJ, USA; 2Cognigen Corporation, a Simulations Plus Company, Buffalo, NY,
USA
Objectives: Nivolumab pharmacokinetics (PK) were previously reported to be time-varying and the decrease in
nivolumab clearance (CL) was associated with the extent of disease improvement in response to treatment (Liu et
al., 2017). A previously developed PPK model for nivolumab-naive patients (Bajaj et al., 2017) was revised to
describe nivolumab PK in patients who achieved disease control on nivolumab (complete response [CR], partial
response [PR], or stable disease [SD]). The revised PPK model was externally validated with data from CheckMate
384 (NCT02713867), a study evaluating nivolumab 480 mg once every 4 weeks (Q4W) dosing vs. 240 mg once
every 2 weeks (Q2W) dosing in patients with previously treated advanced NSCLC with disease control on prior
nivolumab (Edward et al., 2019).
Methods: The PPK analysis data set included nivolumab concentration data from 956 patients in 6 clinical studies
that enrolled patients with NSCLC. The previously developed PPK model was revised with data from 5 of the
studies (N = 654), in which patients received nivolumab weight-based dosing (ranging from 0.3 mg/kg to 10 mg/kg
Q2W), to incorporate the effect of tumor response on the temporal change in nivolumab CL. The revised model was
externally validated with data from the remaining study (CheckMate 384) using prediction-corrected visual
predictive checks.
Results: The PPK of nivolumab was described by a 2-compartment model with time-varying CL, and included the
effects of baseline body weight (BBWT), baseline serum albumin, estimated glomerular filtration rate, performance
status, sex, and race on CL; BBWT and sex on volume of the central compartment; and tumor response category on
maximal change in CL. The magnitude of reduction in nivolumab CL over time was higher in patients with CR/PR
and lower in those with progressive disease, relative to SD patients. External validation demonstrated adequate
predictive performance of the PPK model with respect to the median and 95th percentiles of concentration, and a
modest underprediction for the 5th percentile values. Notably, the performance of the model was similar for both
nivolumab 240 mg Q2W and 480 mg Q4W.
Conclusions: The PPK model of nivolumab developed from weight based Q2W dosing provided an adequate
description of the data with flat dosing of 240 mg Q2W and 480 mg Q4W, with deviations that were similar and not
clinically relevant. This evaluation further supports the application of the revised nivolumab PPK model to predict
the exposures in patients with disease control.
References: Bajaj, G., Wang, X., Agrawal, S., et al. CPT Pharmacometrics Syst Pharmacol. 2017;6(1)58–66.
Edward, B.G., Niels, R., Lionel, F., et al. J Clin Oncol. 2019;37(Abstract 100).
Liu, C., Yu, J., Li, H., et al. Clin Pharmacol Ther. 2017;101(5)657–666.
W-023
Longitudinal Pharmacokinetic/Pharmacodynamic Analysis for Urinary Aminolevulinic Acid (ALA) from
Phase I, II, and III Clinical Studies of Givosiran in Acute Hepatic Porphyria Patients
Jongtae Lee, Varun Goel, Sagar Agarwal, Megan Melch, Gabriel J. Robbie
Alnylam Pharmaceuticals, Cambridge, MA, USA
Objectives: Givosiran is an investigational RNA interference (RNAi) therapeutic agent that specifically inhibits
hepatic aminolevulinate synthase 1 (ALAS1) mRNA and is currently under development for the treatment of acute
hepatic porphyria (AHP). AHP is a family of rare, serious genetic disorders that involve a defect in the heme
biosynthetic pathway in liver and is characterized by severely debilitating neurovisceral attacks due to accumulation
of the toxic heme intermediates of aminolevulinic acid (ALA) and porphobilinogen (PBG). A population
pharmacokinetic/pharmacodynamic (PK/PD) analysis was performed to relate predicted liver concentrations of
givosiran and its active metabolite, AS(N-1)3' givosiran, with urine ALA levels and evaluate covariates effects.
Methods: Longitudinal urinary ALA data were pooled from chronic high excreter (CHE) subjects and AHP patients
in clinical studies of givosiran treated with placebo or givosiran (0.035–5 mg/kg, single dose, monthly (QM), or
quarterly (QQ) regimens). Since it was not practical to measure liver PK in human, givosiran PK in human liver was
predicted from preclinical species using allometric scaling. Population PK/PD model structure was based on
mechanistic understanding of drug effect on PD lowering from pre-clinical studies of givosiran.
Results: The final PK/PD model consisted of a semi-mechanistic model linking allometrically scaled concentrations
of active siRNA in liver to synthesis rate of ALA through an intermediary RNA-induced silencing complex (RISC)
effect compartment. The effect of total RISC loaded siRNA concentrations on the synthesis rate of ALA was
modeled with an inhibitory effect relationship. The effects of hemin and givosiran were included in the model as a
multiplicative inhibitory effect on the synthesis rate of liver ALA. Model estimated IC50 of RISC loaded siRNA was
0.476 ng/g (IIV: 144%). Following 2.5 mg/kg QM regimen, model predicts median ALA levels reach steady-state
by 3 months and are within normal range in AHP patients. Synthesis rate of ALA was estimated to be 11.47
mmol/mol·hr and was 2–fold higher in AHP patients relative to CHE subjects. IC50 was estimated to be lower in
CHE patients relative to AHP patients. Higher baseline ALA was associate to have slighter higher Imax and mild
hepatic impairment status was associated with slightly higher baseline ALA.
Conclusions: A semi-mechanistic PK/PD model related predicted liver and RISC concentrations to ALA levels, and
adequately described observed time-course of ALA reduction after givosiran treatment. The 2.5 mg/kg QM dose
reduces urinary ALA towards the normal range in AHP patients. Similar ALA lowering was estimated across all
investigated covariates.
W-024
Population Pharmacokinetic Analysis of Vutrisiran in Healthy Volunteers
Jongtae Lee, Varun Goel, Bahru Habtemariam, Megan Melch, Gabriel J. Robbie
Alnylam Pharmaceuticals, Cambridge, MA, USA
Objectives: Vutrisiran is an investigational RNA interference (RNAi) therapeutic agent that specifically inhibits
hepatic transthyretin (TTR) synthesis and is currently under development for the treatment of TTR-mediated
amyloidosis. It is composed of a double stranded small interfering RNA (siRNA) conjugated to a triantennary N-
acetylgalactosamine (GalNAc) ligand that leads to efficient hepatic uptake via asialoglycoprotein receptors
(ASGPR). In liver, vutrisiran catalytically cleaves wild type and mutant TTR mRNA leading to lowering of serum
TTR. A population pharmacokinetic (PK) model of vutrisiran was developed to describe plasma and urine PK
profiles and to evaluate significant covariates after single subcutaneous (SC) administration of vutrisiran in healthy
adult subjects.
Methods: A Phase 1 study of vutrisiran was carried out in healthy adults of Japanese and non-Japanese descent. A
total of 60 subjects were enrolled at in 6 ascending dose cohorts (5, 25, 50, 100, 200, and 300 mg). Blood PK
samples collected at 0, 10, 30 minutes, 1, 2, 4, 6, 8, 12, 24, and 48 hours post-dose. In addition, 24-hour urine PK
samples were collected. Plasma PK and urinary excretion data were simultaneously fitted together using a mixed
effect model (NONMEM, ver. 7.4.1).
Results: A two-compartment linear PK model with renal and hepatic uptake pathways best described the PK of
vutrisiran. The apparent clearance (CL/F) and associated inter-patient variability (IIV) for renal and hepatic routes
were estimated to be 4.03 (18.9%) and 17.3 L/h (25.1%), respectively. The apparent volume of distribution (Vd/F)
for central and peripheral compartment were estimated to be 9.99 (70.6%) and 52 L, respectively. The absorption
rate constant (ka=0.152 h-1, absorption t1/2=4.56 h) is substantially lower than the terminal elimination rate constant
(beta elimination t1/2=0.32 h) suggesting a flip-flop kinetics after SC dosing. Allometric scaling based on body
weight was applied to all clearance and volume terms for vutrisiran. None of following covariates affected PK
exposures: baseline age, sex, race (Japanese/non-Japanese), body mass index, liver and renal biomarker.
Conclusions: Population analyses adequately described plasma and urine PK profiles of vutrisiran across a wide
range of dose levels. Overall PK of vutrisiran was linear, dose-proportional, time-invariant. Hepatic uptake via
ASGPR was rapid accounted for approximately 80% of the total plasma clearance; renal clearance was
approximately equal to baseline estimated glomerular filtration rate. Modeling suggests that approximately 80% of
the drug was taken up into liver and approximately 20% was excreted unchanged in urine.
W-025
Pharmacokinetic/Pharmacodynamic Analysis of Givosiran in Rats
Jongtae Lee, Varun Goel, Sagar Agarwal, Megan Melch, Amy Guan, Jing Li, Sean Dennin, Michael Arciprete,
Gabriel J. Robbie
Alnylam Pharmaceuticals, Cambridge, MA, USA
Objectives: Givosiran is an investigational RNA interference (RNAi) therapeutic agent that specifically inhibits
hepatic aminolevulinate synthase 1 (ALAS1) mRNA and is currently under development for the treatment of acute
hepatic porphyria (AHP). Pharmacokinetic/Pharmacodynamic (PK/PD) analysis in rats was performed to predict
liver concentrations of givosiran and AS(N-1)3' givosiran (active siRNA) in human using allometric scaling.
Methods: Givosiran PK and ALAS1 mRNA levels were evaluated in Sprague-Dawley rats after administration of a
single subcutaneous dose of givosiran at 1, 5 or 10 mg/kg and multiple weekly doses at 1 and 2.5 mg/kg (8 times).
Liver samples were collected from 2 animals/sex/group/time point for quantitation of givosiran at approximately 4
and 24 hours on Day 1 and 4, 7, 14, and 28 days post-dose. Liver and RNA-induced silencing complex (RISC)
concentrations of givosiran and AS(N-1)3' givosiran and liver ALAS1 mRNA levels were determined. Non-linear
mixed effects modeling (NONMEM, version 7.4.1) was used for PK/PD analysis and a sequential approach was
applied to develop PK/PD model.
Results: The final PK/PD model consisted of a semi-mechanistic model linking concentrations of active siRNA in
liver to degradation rate of ALAS1 mRNA through an intermediary RISC effect compartment. Givosiran
concentrations in liver exhibited a biphasic curve with the alpha phase elimination half-life of 3 days followed by a
slow terminal beta phase with a half-life of 30 days. The RISC concentration for a 50% decrease in ALAS1 mRNA
was estimated to be 0.478 ng/g.
Conclusions: A semi-mechanistic PK/PD model related liver and RISC concentrations to ALAS1 mRNA levels in
rats, and adequately described observed time-course of ALAS1 reduction after givosiran treatment. The PK/PD
model in rats was used to predict liver concentrations of givosiran for PK/PD model in human.
W-026
Relationship Between Urinary Aminolevulinic Acid (ALA) Levels and Porphyria Attacks in Acute Hepatic
Porphyria Patients in Clinical Trials with Givosiran
Authors: Varun Goel1, Michael Dodds2, Sagar Agarwal1, Amy Simon1, and Gabriel Robbie1
Affiliations: [1] Alnylam Pharmaceuticals, Cambridge MA; [2] Certara, Seattle, WA
Objectives: Acute hepatic porphyrias (AHPs) are a family of ultra-rare, genetic diseases characterized by potentially
life-threatening attacks resulting from accumulation of neurotoxic heme intermediates including ALA. Givosiran is
an investigational, subcutaneously administered RNA interference therapeutic that specifically cleaves
aminolevulinic acid synthase 1 mRNA leading to decreased hepatic production of ALA. Our aim was to quantify
the relationship between urinary ALA levels and annualized attack rates (AAR) in AHP patients.
Methods: Longitudinal urinary ALA and porphyria attack data were pooled from AHP patients across phase 1, 2
and 3 clinical studies of givosiran. A mixed effects two-state Markov model was developed to describe the
relationship between levels of ALA and the patient attack status i.e. in attack versus no-attack. Several relationships
including linear, log linear, Emax and sigmoidal Emax were evaluated for best fit. Disease specific covariates
including historical attack rate, baseline ALA levels, prior use of hemin prophylaxis, age, and gender were evaluated
in the model. Simulations were conducted from the final model to understand the impact of ALA levels on
annualized attack rate, mean days in attacks, mean time to next attack in patients treated with givosiran or placebo.
Results: The relationship between ALA levels and the probability of transitioning into attacks was statistically
significant. At the population level, higher ALA levels were associated with a greater probability of having an
attack. Covariate effect of baseline ALA, age, and gender were not statistically significant. In absence of treatment
with givosiran, patients with higher historical attack rate and prior use of heme prophylaxis were predicted to have
higher AAR. Model predicts that givosiran administration significantly reduced AAR and increased mean days
between attacks and mean time to next attack across all disease specific covariates.
Conclusion: A mixed effects two-state Markov model adequately describes the relationship between incidence of
porphyria attacks and urinary ALA levels with a clear relationship between magnitude of urinary ALA lowering and
reduction in porphyria attacks. Givosiran significantly reduced AAR across all investigated covariates in AHP
population.
W-027
Population Pharmacokinetics of Brincidofovir and Cidofovir in Adults and Pediatric Subjects With Viral
Infection
Virna Schuck1, Adekemi Taylor1, Tim Bergsma1, Yuan Xiong1, Marion Morrison2, Maggie Anderson2, Tim Tippin2,
Odin Naderer2, Mark Lovern1
1Certara USA, Inc, 2Chimerix Inc.
Objectives: Brincidofovir (BCV) is an orally available lipid conjugate of cidofovir (CDV) with in vitro antiviral
activity against double-stranded DNA (dsDNA) viruses, including adenoviruses, herpesviruses, orthopoxviruses,
and polyomaviruses. BCV has been administered to adult and pediatric patients at risk of or infected with dsDNA
virus in various clinical trials. After oral administration, BCV is metabolized to CDV intracellularly, which is
converted to the active antiviral CDV-diphosphate. Population pharmacokinetic (POPPK) models were developed
to characterize BCV and CDV PK in adults and pediatric patients and evaluate the impact of intrinsic and extrinsic
factors on PK to inform optimal efficacious and safe exposures.
Methods: The analysis dataset included data from six patient studies and one healthy volunteer (HV) study.
Subjects with at least one BCV dose and one postdose BCV or CDV concentration time point were included in the
analysis. In total, 941 subjects receiving BCV doses between 8 and 300 mg once or twice a week were included in
the dataset. Analysis was performed using NONMEM v7.3. Separate POPPK models were developed for BCV and
CDV. Body weight effects were included on BCV and CDV clearance and distribution parameters to allow scaling
from pediatrics to adults. Candidate covariates (age, sex, race, disease, cyclosporine use, creatinine clearance,
diarrhea, formulation, and prandial status) were selected based on scientific interest, mechanistic plausibility, or
clinical relevance. Covariates were tested in an iterative process, using a forward addition process followed by
backward deletion.
Results: Plasma BCV PK in adults and pediatric patients was well-described by a two-compartment model with
zero-order appearance in a depot absorption compartment, followed by first-order absorption into the central
compartment with first-order elimination. Diarrhea and food decreased BCV bioavailability by <21%. BCV
clearance estimate was 30% lower in patients receiving cyclosporine. HV had twice the BCV clearance of patients.
Because sequential parent-metabolite models proved to be unstable and very sensitive to initial estimates, BCV
dosing was linked directly to CDV concentration profiles. Creatinine clearance scaling component was also added to
CDV clearance. Diarrhea and food effect on bioavailability identified in the BCV POPPK model were fixed in the
CDV POPPK. Age effect was identified in subjects <18 years of age, with CDV clearance in 2-year olds being
38.8% of a typical adult clearance. HV had 97.1% higher CDV clearance than patients.
Conclusions: The analysis identified POPPK models that can describe the PK of BCV and CDV in adult and
pediatric patients. Important covariates were identified, such as age and cyclosporine co-dosing, a common co-
medication in the patient population.
W-028
Developing a Semi-mechanistic Pharmacokinetic/Pharmacodynamic (PK/PD) Model of
Ceftolozane/Tazobactam against E. coli Strains of Differing Drug Susceptibility
Lyrialle W. Han1, Abhay Joshi2, Xiaohui (Tracey) Wei2
1Department of Pharmaceutics, University of Washington, Seattle, WA, USA 2Division of Clinical Pharmacology IV, Office of Clinical Pharmacology, Office of Translational Sciences, Center
for Drug Evaluation and Research, US Food and Drug Administration, Silver Spring, MD, USA
Objectives: Ceftolozane (TOL)/tazobactam (TAZ) is a β-lactam (BL)/β-lactamase inhibitor (BLI) combination used
to treat gram-negative infections, including complicated intra-abdominal infection and complicated urinary tract
infection. TAZ as a BLI restores the susceptibility of its BL partner toward β-lactamase-producing bacteria. A
quantitative approach to describe the time-course of BL/BLI bactericidal activity may provide key information for
BL/BLI dose optimization. The objective of this study was to develop a semi-mechanistic PK/PD model to describe
the in vitro time-killing effect of TOL/TAZ against two E. coli strains of differing drug susceptibility.
Methods: A dataset from in vitro static time-killing experiments using two E. coli strains with different drug
susceptibility was created by digitizing results from previously published work1. The two investigational strains
include MERCK 2805 (no β-lactamase) and MERCK 2890 (AmpC β-lactamase), with MICs to TOL of 0.25 and 4
mg/L with no TAZ, and MICs of 0.25 and 1 mg/L with 4 mg/L TAZ, respectively. Both strains were exposed to a 6
by 5 array of TOL (0, 1, 4, 16, 64, and 256 mg/L) and TAZ (0, 1, 4, 16, and 64 mg/L) over 48 h using starting
inocula of 106 CFU/mL. A stepwise approach was used to build a semi-mechanistic PK/PD model to describe the
time-killing effect of TOL/TAZ on strains MERCK 2805 and MERCK 2890 separately, using NONMEM 7.
Results: A two-state model with logistic growth was used to describe the natural growth and death of E. coli. Effect
of TOL alone on bacterial growth/kill was characterized by a sigmoidal Emax model. The interaction of TAZ on the
bactericidal activity of TOL was characterized by TAZ concentration-dependent shifting on IC50 of TOL using a
non-linear equation, and a concentration-dependent change in delayed regrowth of bacteria. Compared to the non-
resistant strain MERCK 2805, the model estimated a higher IC50 and a lower Emax for strain MERCK 2890, which
is consistent with its increased resistance toward TOL. The final model achieved good individual fits to both strains.
The population PD parameters were estimated with acceptable precision (standard error <30%) for both strains.
TAZ alone at high concentration (64 mg/L) showed minor suppression on bacterial growth which was captured in
the final PK/PD model using a sigmoidal Emax equation.
Conclusion: A two-state logistic growth semi-mechanistic PK/PD model was developed and successfully described
the in vitro time-killing effect of TOL/TAZ on one TOL susceptible E. coli strain and one TOL resistant E. coli
strain. This model will be further validated with bacterial strains of greater β-lactam resistance.
References: 1. RL Soon et al. Antimicrobial Agents and Chemotherapy. 2016, 60(4): 1967-1973 2. SKB Sy et al.
CPT Pharmacometrics Syst. Pharmacol. 2017, 6:197–207
W-029
At the interface of physiologically-based and nonlinear mixed-effects modelling: Enhancing the quantitative
understanding of the complex pharmacokinetics and -genetics of tamoxifen and its major metabolites
Lena Klopp-Schulze1, Anna Mueller-Schoell1, Markus Joerger2, Patrick Neven3, Stijn L.W. Koolen4, Ron H.J.
Mathijssen4, Stephan Schmidt5 & Charlotte Kloft1
1Dept. of Clinical Pharmacy and Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, Berlin, Germany; 2Dept. of Medical Oncology and Haematology, Cantonal Hospital, St. Gallen, Switzerland; 3Vesalius Research
Center – VIB, University Hospitals Leuven, KU Leuven-University of Leuven, Leuven, Belgium; 4Dept. of Medical
Oncology, Erasmus MC Cancer Institute, Rotterdam, The Netherlands; 5Center for Pharmacometrics and Systems
Pharmacology, College of Pharmacy, University of Florida, Orlando, FL, USA
Objectives: To enhance the pharmacokinetic (PK) and pharmacogenetic knowledge of the bioactivation processes
of tamoxifen to its therapeutically relevant metabolite endoxifen via two parallel metabolic pathways, a joint
minimal physiologically based parent-metabolites model was developed. This approach ultimately aimed to
synthesize clinically relevant information, facilitating the identification of patient groups at risk of treatment failure
and highlighting the need for dosing adaptations to improve tamoxifen treatment.
Methods: To develop a minimal physiologically-based model of tamoxifen and its three major metabolites, data
from multiple sources were integrated: Namely, (i) scaled in vitro data from enzyme kinetic experiments [1-2]
applying an in vitro-in vivo correlation approach, (ii) PK parameters from previous clinical studies [3] and (iii) in
vivo data from three different investigator-initiated clinical trials [4-6]. By utilizing an approach at the interface
between physiologically-based PK (PBPK) and nonlinear mixed-effects (NLME) PK modelling, the model focused
on capturing relevant physiologically-mechanistic processes, while allowing to further characterize interindividual
variability considering relevant patient-specific factors.
Results: CYP2D6 activity scores (AS) and age from 406 patients were implemented using mechanistic covariate
relationships. The model successfully captured the typical behavior, as well as the large variability in the PK of
tamoxifen and its three metabolites. Applying the model, low hepatic extraction ratios (overall <0.3) in the primary
and secondary tamoxifen metabolism were determined, classifying tamoxifen as a low extraction drug, i.e.
elimination is limited by protein binding and metabolic capacity. While in a typical CYP2D6 normal metabolizer
(AS=2) 10% of tamoxifen was bioactivated to endoxifen, it was only 2.2% in a typical CYP2D6 poor metabolizer
(AS=0). Using simulations, subpopulations at highest risk of sub-therapeutic endoxifen concentrations were
identified (applying the proposed therapeutic threshold of 5.97 ng/mL endoxifen [7]): Patients with impaired
CYP2D6 enzyme activities (AS≤ 0.5) showed highest risk values. Irrespective of the CYP2D6 activity, younger
patients were at higher risk (25-35 years and AS=0.5: 62% vs. 85-95 years and AS=0.5: 35%).
Conclusions: By leveraging the synergy between physiologically based and clinical data-driven (NLME) PK
approaches, a minimal NLME-PBPK model framework was developed, capturing important pharmacogenetic
factors, physiological changes of the target population and ‘real-world’ interindividual variability on relevant PK
parameters of tamoxifen and its sequential bioactivation process. This framework provides a sound basis for further
simulation-based explorations as well as model extensions to guide future individual dose recommendations for
breast cancer patients who rely on tamoxifen treatment success.
References: 1. Desta et al. J. Pharmacol Exp. (2004) 2. Coller et al. J. Clin. Pharmacol. (2002) 3. Ahmad et al.
Clin. Pharmacol. Ther. (2010) 4. P. Neven et al. Clin. Cancer. Res. (2018) 5. P. Neven. CYPTAM-3 trial
NCT00966043 (2011) 6. Binkhorst et al. Breast Cancer Res. Treat. (2015) 7. Madlensky et al. Clin. Pharmacol.
Ther. (2011)
W-030
A Quantitative Relationship Between CAR-Affinity, Target Abundance, Target-Cell Depletion and Cytokine
Release: Implications Towards Discovery and Development of CAR-T cell-therapy
Aman P. Singh1, Xirong Zheng1, Donald Heald1 and Weirong Wang1
1Discovery and Translational Research (DTR), Biologics Development Sciences, Janssen BioTherapeutics (JBIO),
Janssen R&D, PA.
Background: Adoptive cell transfer of T-cells genetically modified with tumor-associated antigen (TAA) reactive
chimeric antigen receptors (CARs) is a rapidly emerging field in clinical oncology. Although the preliminary
clinical trials have demonstrated remarkable therapeutic benefits, the key drug-specific and system-specific
determinants leading to CAR-T-mediated target cell-depletion, in vivo expansion/persistence as well as occurrence
of cytokine release syndrome (CRS) are still not well understood. Amidst of all these uncertainties, development of
platform-level translational PK-PD models can be a valuable tool for successful discovery and development of
CAR-T cell therapies.
Objectives: To determine the relationship between CAR-binding affinity, target antigen density, CAR-density on T-
cells and different effector: target (E: T) ratios to quantitatively describe target-cell depletion, cytokine release and
T-cell expansion in vitro.
Methods: An in-vitro PK-PD model was formalized (Figure 1), which accounts for a dynamic population of CAR-
T cells and tumor cells growing with their individual growth rates, respectively. Upon interaction, there is formation
of CAR-Target complex, which mediates the killing of target (tumor) cells, further expansion of CAR-T cells and
induction of cytokine release syndrome. The mathematical model was used to characterize the datasets for two
affinity variant (Kd of 20 nM and 0.2 nM respectively) anti-EGFR CAR-T cells, having a ~100-fold difference in
Kon (association rate) to TAA. The datasets comprised of target cell depletion, induction of CAR-T cell expansion
and release of interferon-γ (cytokine) at E:T ratios ranging from 0.1:1 to 20:1 and antigen expression ranging from
30K to 600K [1]. The developed model was later employed to describe the effects of variable CAR affinities (0.58
nM – 1.1 µM), target antigen densities (30,899 - 628,265 receptors/cell) and E:T ratios (0.5:1—16:1) using a
comprehensive dataset on four different affinity variant anti-HER2 CAR-T cells [2].
Results and Conclusion: Model was able to simultaneously capture the dynamic interaction of target-cell killing
and CAR-T cell expansion for both affinity variant anti-EGFR and anti-HER2 CAR-T cells. Additionally, the
formation of CAR-Target complexes/target cells was able to characterize the rate and extent of cytokine release.
Model predicted a very fast CAR-mediated killing rate (Kmax) and highly potent IC50 values values across two
different CAR constructs (EGFR and HER2). The model estimated IC50 values were 1.4 and 0.098 CAR-target
complexes/tumor cell for anti-EGFR and anti-HER2 CAR-Ts respectively. Additionally, the cytokine induction
(Emax and EC50) parameters associated with the two CAR-constructs were estimated to be very similar. Developed
model allowed integrated analysis of the key determinants of CAR-T activity and provided a basis to extrapolate in
vitro and in vivo CAR-T activities. It can be implemented to guide the early candidate selection for CAR-T
programs.
References: 1) Caruso et. al, Cancer Res, 2015 75(17):3505-18; 2) Liu et al., Cancer Res, 2015 75(17): 3596-3607
Figure 1: Schematics of the mechanism-based model to describe CAR-T cell in vitro activity.
W-031
A Mechanistic Mixed-effects Beta Regression Model for the Time-course of Leukemic Cell Burden in Acute
Myeloid Leukemia (AML) Patients Treated with CPX-351
Sarah F. Cook1, Qi Wang2, Scott A. Van Wart1, Donald E. Mager1
1Enhanced Pharmacodynamics, LLC, Buffalo, NY, USA; 2Jazz Pharmaceuticals, Inc., Palo Alto, CA, USA.
Objectives: CPX-351 (Vyxeos®), a dual-drug liposomal encapsulation of daunorubicin and cytarabine at a
synergistic ratio, is approved by the US FDA and EMA for treatment of adults with newly diagnosed therapy-related
AML and AML with myelodysplasia-related changes. In this study, a mechanistic mixed-effects pharmacokinetic-
pharmacodynamic (PK-PD) model was developed for the number of leukemic cells in bone marrow following CPX-
351 administration.
Methods: PK-PD data from two phase 2-3 studies in adults with AML (N=140) were modeled in NONMEM 7.3.
Individual patient-predicted PK profiles were obtained by applying previously developed PK models for CPX-351–
derived cytarabine and daunorubicin. A hybrid physiologically-based PK model was used to drive drug
concentrations in bone marrow. Sparse data were available from bone marrow aspirates/biopsies; patients had pre-
treatment and 1-4 post-treatment samples collected. The number of leukemic cells in bone marrow samples was
calculated as the product of cellularity fraction, myeloblast fraction, and maximal disease burden, which was
estimated at 1012 cells for AML. The total number of leukemic cells was modeled as the sum of cells in G1, S, and
G2M phases [1]. A logistic function was imposed on progression from G1 to S phase, so proliferation slows as the
system approaches the maximal disease burden. Mechanistic inhibition of proliferation was achieved by applying
phase-specific drug effects via Imax functions (Figure). Beta regression was employed to constrain model predictions
to the 0-1012 range. Prior to model fitting, parameters were fixed at values derived from literature sources [1-3], and
a global sensitivity analysis was conducted (R 3.5.1) to guide selection of parameters to be estimated.
Results: Sensitivity analysis indicated that model output was most sensitive to IC50 of daunorubicin, S-phase
duration (TS), and G1-phase duration (TG1); therefore, these parameters were prioritized for estimation during mixed-
effects modeling. The best model fit was achieved by estimating fixed effects for IC50 of daunorubicin and initial
number of leukemic cells (NLEUK0), estimating inter-individual variability (IIV) for TS and NLEUK0, and fixing
IIV for IC50 of daunorubicin to a literature value. Population estimates (IIV) were 214 µg/L (118% CV, fixed) for
IC50 of daunorubicin, 13 h (fixed) (168% CV) for TS, and 0.265×1012 cells (85% CV) for NLEUK0. All other cell
cycle and PD parameters were fixed to literature values.
Conclusions: A mechanistic mixed-effects PK-PD model characterizing the effects of CPX-351 on leukemic cell
burden in AML patients was developed and adequately described the dynamics of bone marrow leukemic cells, as
well as IIV in leukemic cell dynamics, following chemotherapy treatment. To our knowledge, this represents the
first population model to characterize leukemia dynamics following phase-specific chemotherapeutic action.
Figure Mechanistic cell cycle-PD model for bone marrow leukemic cells following CPX-351 administration
References: 1. Pefani, et al. Comput Chem Eng. 2013;57:187-195. 2. Quartino, et al. J Clin Pharmacol.
2007;47(8):1014-1021. 3. Raza, et al. Blood. 1990;76(11):2191-2197.
W-032
Expansion of the Target-Biologic-Effector (TBE) Complex-Based Cell-Killing Model for Multivalent T-Cell
Redirecting Antibodies
Xirong Zheng1, Donald Heald1, Weirong Wang1 and Aman P. Singh1
1Discovery and Translational Research, Biologics Development Sciences, Janssen BioTherapeutics (JBIO), Janssen
R&D, PA.
Background: T-cell redirecting bispecific antibodies have shown promises in cancer treatment. By simultaneously
binding to CD3 on T cells and antigens on cancer cells, such antibodies redirect T cells to kill tumor cells. However,
the tumor antigens targeted by these antibodies often have lower levels of normal tissue expression, which may
result in on-target/off-tumor toxicities. Recent advancement in antibody engineering have made it possible to
develop multivalent tumor antigen binding T-cell redirecting antibodies, which can enable avidity-driven, more
selective binding to tumor tissues in comparison to normal tissues.
Objectives: 1) To expand the previously developed target-biologic-effector (TBE) complex-based cell killing
model to incorporate multivalent tumor antigen binding T-cell redirectors, to characterize avidity-driven selective
killing of high expressing c.f. low expressing target cells in-vitro, and 2) To integrate the expanded in-vitro model
with a physiologically-based PK (PBPK) model for mAbs and T-cells to characterize the tumor growth inhibition
(TGI) in vivo.
Methods: An in-vitro binding kinetic model was developed (Figure 1.) to characterize the binding and target cell
killing of a T cell redirecting bispecific antibody with bivalent tumor antigen binding. The antibody has one binding
site for CD3 on T cells, while having two binding sites for the tumor antigen expressed on target cell surface. The
binding of first tumor antigen molecule is assumed to be only dependent on the concentrations of free antibody and
free tumor antigen, which is assumed to be evenly distributed on the tumor cell surface. Following binding to the
first antigen molecule, the antibody is tethered to the cell surface, and hence is only accessible to a fraction of the
total antigen, i.e. the antigen molecules within the cross-arm distance between the two bivalent tumor target-binding
sites.
Results and Conclusions: Developed model was utilized to characterize the reported binding and cytotoxicity of a
monovalent and a series of bivalent anti-HER2/CD3 antibodies [1], varying in binding affinities to HER2 (Kd: 0.3
nm-49 nM) with a fixed CD3 affinity of 70 nM. This model was able to capture the selective binding and cell
killing of lower affinity, bivalent anti-HER2/CD3 antibody against tumor cells with varying antigen densities
(40,000-1,000,000 HER2/cell) and predict the in vivo efficacy of these constructs in xenograft models. This model
can be used to facilitate design of multivalent T cell redirecting antibodies, lead candidate selection and predict in
vitro and in vivo activities of such multi-specific antibodies.
Reference: 1) Slaga, D. et al. (2018). 10 (463) Sci. Transl Med.
Figure 1. Schematic of a target-biologic-effector model for bivalent bispecific antibodies:
W-033
Development of a Translational PK-PD Model for CAR-T Cell Therapy to Facilitate In Vitro-In Vivo
Correlation
Xirong Zheng1, Donald Heald1, Weirong Wang1, and Aman P. Singh1
1Discovery and Translational Research, Biologics Development Sciences, Janssen R&D, PA.
Background: Adoptive immunotherapy using chimeric antigen receptor (CAR)-modified T cells have been rapidly
emerged for cancer treatment. CARs specifically bind to tumor antigens, leading to the depletion of tumor cells as
well as activation and proliferation of CAR-T cells. However, despite initial clinical success of CAR-T in the
treatment for hematopoietic malignancies, the basic pharmacokinetic and pharmacodynamic relationships of CAR-T
cells are not well understood.
Objectives: To develop a physiologically based pharmacokinetic (PBPK)-pharmacodynamic (PD) model capable of
1) predicting the distribution of CAR-T cells within different tissues of interest and 2) developing a dose-response
relationship to characterize tumor growth inhibition (TGI) induced by CAR-T cells.
Methods: A PBPK-PD model (Figure 1.) has been developed to characterize the quantitative relationship between
the biodistribution of CAR-T cells and CAR-T cell-mediated tumor growth inhibition in tumor-bearing mice. Nine
tissues of interest, including a tumor compartment, are divided into vascular and extravascular sub-compartments,
and are connected via blood and lymph flows. In normal tissue compartments, CAR-T cells transmigrate to the
extravascular sub-compartment from the vascular sub-compartment. Transmigrated CAR-T cells recirculate back to
lymph node via lymphatic flow. In contrast, CAR-T cells are accumulated in tumor extravascular compartment by
binding with tumor antigen, which further leads to tumor growth inhibition. Only unbound CAR-T cells are
recirculated in the lymphatic system. CAR-T cells are assumed to be eliminated only in the liver extravascular
compartment.
Results and Conclusions: Developed PBPK-PD model was utilized to characterize the biodistribution of
untransduced T-cells and CAR + T cells targeting CD19 [1] and EGFR [2] receptors. Validated PBPK model for
CAR-T cells was later utilized to characterize tumor growth inhibition (TGI) for CAR-T cells (varying in affinities)
in xenograft and orthotopic mouse models with different target antigens (i.e. EGFR [3], CD19 [4, 5] and HER2 [6]),
varying antigen densities and different route of administration (intravenous versus intratumoral). The developed PK-
PD relationship can serve as a cornerstone in developing an in vitro-in vivo correlation, which can facilitate lead
candidate selection and preclinical-to-clinical translation of CAR-T cells.
References: 1) Khawli, L. et al. (2018). Poster Presentation 2) Chapelin, F. et al. (2017). Scientific reports 3)
Caruso, H. G. et al. (2015). Cancer Res. 4) Berahovich, R. et al. (2017). Front. Biosci. 5) Li, S., Siriwon, et al.
(2017). Clin Cancer Res. 6) Liu, X. et al. (2015). Cancer Res. 7) Khot, A. et al. (2019). JPET.
Figure 1. Schematic of a PBPK-PD model for CAR-T cells:
W-034
Assessment of Variance Reporting in the Literature with Implications for Model Application: Four Cases
Authors: James A. Clary, PharmD (Trainee)1 Andrew Castleman, PharmD, M.S.1 Jill Fiedler-Kelly, M.S.2
Joel S. Owen, Ph.D.2
Affiliations: 1. Union University College of Pharmacy
1050 Union University Dr
Jackson, TN 38305
2. Cognigen Corporation, a Simulations Plus Company
1780 Wehrle Dr #110
Buffalo, NY 14221
Objectives: Two methods have been reported for the calculation of percent coefficient of variation (%CV): an
exact method (%CV = 100*sqrt(exp(ω2)-1) and an approximate method (%CV = 100*sqrt(ω2)). Many published
models have not reported the methodology used leading to possible misuse of the model.
Methods: Four models were selected for simulation exercises: 1) Linear model with large %CV (~70%)1, 2) Non-
linear model with large %CV2, 3) Linear model with etas on 4 parameters with medium %CV (~30%)3, and 4) PK-
PD model with medium %CV on PK parameters and large %CV on PD parameters4. If the paper reported the
method of %CV calculation or the ω2 estimate, those estimates were used, otherwise the ω2 estimate from the exact
method was used. Simulations using estimates obtained from calculating the %CV using either the exact or
approximate method and then using the incorrect back calculation were compared to the original value. Simulated
concentrations and responses were plotted to show the error that may occur as a result of incorrect ω2 estimates.
Results: With large %CV values4, a larger deviation from the original value was observed at the 5th and 95th
confidence intervals than was observed with smaller %CV values. CL, a parameter with 29.8% CV only expressed a
4.494% under and over-prediction in the PK-PD model. CD34, a parameter with 75.6% CV, expressed a 20.979%
under-prediction and a 34.965% over-prediction.
Conclusion: When ω2 estimates are less than or equal to 30% CV, an incorrect assumption regarding the
methodology used does not significantly impact the ability to simulate correctly from the model, but as the %CV
increases, the ability to use the model accurately is significantly impacted.
References: 1. Feng Y, Pollock BG, Ferrell RE, Kimak MA, Reynolds CF, Bies RR. (2006) Paroxetine: population
pharmacokinetic analysis in late-life depression using sparse concentration sampling. British Journal of Clinical
Pharmacology 61(5):558-569. doi:10.1111/j.1365-2125.2006.02629.x 2. Moore, J. , Healy, J. , Thoma, B. , Peahota,
M. , Ahamadi, M. , Schmidt, L. , Cavarocchi, N. and Kraft, W. (2016), A Population Pharmacokinetic Model for
Vancomycin in Adult Patients Receiving Extracorporeal Membrane Oxygenation Therapy. CPT Pharmacometrics
Syst. Pharmacol., 5: 495-502. doi:10.1002/psp4.12112 3. Wang, B. , Yan, L. , Yao, Z. and Roskos, L. (2017),
Population Pharmacokinetics and Pharmacodynamics of Benralizumab in Healthy Volunteers and Patients With
Asthma. CPT Pharmacometrics Syst. Pharmacol., 6: 249-257. doi:10.1002/psp4.12160 4. Bihorel, S. , Raddad, E. ,
Fiedler‐Kelly, J. , Stille, J. , Hing, J. and Ludwig, E. (2017), Population Pharmacokinetic and Pharmacodynamic
Modeling of LY2510924 in Patients With Advanced Cancer. CPT Pharmacometrics Syst. Pharmacol., 6: 614-624.
doi:10.1002/psp4.12221
Table 1: Magnitude of Difference in ω2 When Wrong Methodology is Utilized: LY2510924
Parameter4 True ω2
%CV using
Approx.
Method
(Reported)
Incorrectly
calculated ω2
using Exact
Method
%CV using
Exact Method
Incorrectly
calculated ω2
using Approx.
Method
Difference in
true ω2 and
incorrect
Exact Method
(%)
Difference in
true ω2 and
incorrect
Approx.
Method (%)
CL 0.089 29.8 0.085 30.509 0.093 4.494 4.494
V2 0.071 26.6 0.068 27.126 0.074 4.225 4.225
RV1 0.05 22.3 0.049 22.643 0.051 2 2
RV2 0.168 41 0.155 42.771 0.183 7.738 8.929
CD34 0.572 75.6 0.452 87.853 0.772 20.979 34.965
KOUT 4.84 220 1.765 1120.131 125.469 63.533 2492.335
SMAX 0.212 46 0.192 48.595 0.236 9.434 11.321
SC50 1.796 134 1.028 224.176 5.025 42.762 179.788
RV3 0.218 46.7 0.197 49.355 0.244 9.633 11.927
W-035
A Composite Quantitative Modeling Approach Including PK, Safety and Activity Biomarkers to Inform the
Recommended Phase II Schedule Selection of BET Inhibitor CC-90010 in Solid Tumors
Authors: Bishoy Hanna1, Chee Ng2, Zariana Nikolova3, Jorge DiMartino4, Ida Aronchik4, Ellen Filvaroff4, Rafael
Sarmiento3, Juan De Alvaro3, Simon Zhou1, Maria Palmisano1, Manisha Lamba1
1Celgene Corporation, Summit, NJ, USA 2NewGround Consulting LLC, San Francisco, CA, US 3Celgene Institute
for Translational Research Europe, Seville, Spain 4Celgene Corporation, San Francisco, CA, USA
Objectives: CC-90010 is an oral, potent and reversible bromodomain and extra-terminal (BET) inhibitor in
development for treatment of a variety of advanced solid tumors. The objectives of these analyses were to develop a
semi-mechanistic model linking exposure to safety and a pharmacodynamic (PD) biomarker to guide the
dose/schedule selection from a first in human study. For safety optimization, platelet (PLT) counts in peripheral
blood were used, while C-C Motif Chemokine Receptor 1 (CCR1) in blood was used as a PD target engagement
marker.
Methods: Data from a Phase I dose escalation study in patients with advanced solid tumors and relapsed/refractory
non-Hodgkin’s lymphomas were analyzed. Three different schedules were tested using weekly, bi-weekly, and
monthly intervals with total dose intensities ranging from 90-280 mg. The PK model utilized was a two-
compartment model with first-order oral absorption and first-order elimination. A semi-mechanistic model1 that
incorporated a proliferation, three transit compartments, and a compartment representing PLT counts in peripheral
blood was fit to the observed PLT counts with individual-predicted concentrations as the input function in an
inhibitory Emax model. System-specific parameters (baseline PLT levels, mean transit time (MTT), and feedback
governing proliferation of PLTs (𝜸)) and drug specific potency parameters (Emax and EC50) were estimated. An
indirect effect model2 was fit to the observed PD activity biomarker (CCR1) data and consisted of an input rate (kin),
loss rate (kout), and feedback governing formation of CCR1 (𝜸), as well as an inhibitory Emax model driven by
plasma concentrations of CC-90010. Simulations were performed to select recommended dose and schedule.
Results: The PK model adequately described the observed plasma concentration data and estimated a volume of
distribution of 115 L and a total clearance value of 1.46 L/hr, resulting in a terminal half-life of 73 hrs. The PLT
model was deemed adequate based on the goodness of fit plots, VPC, and individual patient profiles. Estimated
baseline PLT levels were 264 (109/L), MTT was 104 hours, and 𝜸 was 0.089, in line with literature reported values.
The CCR1 model was adequate in describing the individual observed data, suggesting the decline rate of the
biomarkers was governed by the cumulative drug exposure.
Conclusions: Three models were developed and adequately described the observed PK, PLT, and CCR1 data from a
Phase I study of CC-90010. These models were used to simulate the impact of different dose and regimens of CC-
90010; results supported the selection of 4 consecutive days in a 28-day cycle to optimize potential for platelet
recovery and maximum decline of CCR1.
References: 1. Friberg LE (2002). Model of chemotherapy-induced myelosuppression with parameter consistency
across drugs. Clin Oncol, 4713-21. 2. Natalie L Dayneka (1993). Comparison of Four Basic Models of Indirect
Pharmacodynamic Responses. Pharmacokinetics and Biopharmaceutics 21, 4: 457–478.
Figure:
Figure 1. Model schemes of PK, PLT, and CCR1 model utilized.
W-036
Fixed Dose Selection Based On Population Pharmacokinetic/Pharmacodynamic Modeling
Aleksandra Chertkova 1, Elena Marnopolskaya1, Daria Zhuravleva1, Pavel Yakovlev1
1Computational Biology Department, JSC BIOCAD, Russia
Objectives: Most monoclonal antibodies (mAbs) have been dosed relied on body weight due to its contribution in
pharmacokinetic variability. A huge variety of mAb clinical trials aims to investigate properties of dosage regimens
that depend on body size. However, fixed dosing could measurably simplify treatment regimens and result in drug
sustainability and thus is more desirable in certain circumstances. In this study we provide an analysis supporting
suitability of a number of fixed dose regimens using population modeling approach.
Methods: Population pharmacokinetic and pharmacodynamic models were built in order to define a fixed dose
interval that results in exposure and response similar to one of the clinically tested doses. We modify population
pharmacokinetic model to analyze whether the BCD-100 distribution in blood significantly depends on weight. To
determine whether fixed dosing would maintain exposures within the range of clinical experience, the individual PK
properties of fixed dose regimens are compared with the range of exposures from the BCD-100 clinically tested
doses.
Results: With the help of the population PK and PD models the individual treatment response was predicted in
order to study a number of fixed dosing regimens that were not clinically tested. We considered a range of fixed
dosing levels that correspond to 1, 2, 2.5 and 3 mg/kg doses multiplied by median body weight and administered
both q2w and q3w.
Conclusions: The PK data analysis showed no measurable dependency on body weight, as far as the drug exposure
properties from the chosen regimens are comparable with the exposure properties from the clinically experienced
dosing levels. The Marascuillo procedure applied to the predicted and observed data detects no statistical difference
between the proportion of patients with a particular response type (complete and partial response, stable and
progressive disease) for the examined dose levels. Thus, we show that all the explored dosing regimens could be
considered as acceptable in the upcoming trials and the subsequent drug registration.
W-037
Modeling the Effect of Antiangiogenic/Antiproliferative Scheduling Optimization in Lung Cancer
Author: Benjamin K Schneider1, Arnaud Boyer2,3, Joseph Ciccolini2, Fabrice Barlesi3, Kenneth Wang4, Sebastien
Benzekry*1,5 and Jonathan P Mochel*1
Institutions: (1) SMART Pharmacology, Iowa State University College of Veterinary Medicine, Ames, IA, U.S.A,
(2) SMARTc Unit, Centre de Recherche en Cancerologie de Marseille UMR Inserm U1068, Aix Marseille
University, Marseille, France (3) Multidisciplinary Oncology and Therapeutic Innovations Department, Assistance
Publique Hôpitaux de Marseille, Marseille, France, (4) Mayo Clinic, Rochester, MS, U.S.A, (5) Team MONC, Inria
Bordeaux Sud-Ouest, Institut de Mathématiques de Bordeaux, France. (*) co-last authors.
Objectives: Bevacizumab-pemetrexed/cisplatin (BEV-PEM/CIS) is a combination therapy for non-small cell lung
cancer (NSCLC). Previous studies have reported that sequential scheduling of BEV-PEM/CIS, i.e. administering
BEV before PEM/CIS, improves the efficacy of BEV-PEM/CIS (1). In this study, we used a large dataset generated
from xenograft NSCLC tumor-bearing mice in Imbs et al. 2017 to validate those findings and fit a semi-mechanistic
PKPD model of tumor growth vs. BEV-PEM/CIS exposure (2). We then used relevant literature values to scale the
model fit to make predictions of tumor growth vs. BEV-PEM/CIS administration in humans.
Methods: PK models and parameter estimates for BEV, PEM and CIS were adapted from literature values.
Competing PKPD models were written as NLME models and parameter estimates were obtained using the SAEM
algorithm. Standard goodness-of-fit metrics and figures were used for diagnosing fit. Scaling of PK to humans was
done by substituting mouse PK models and parameter estimates with values from the literature. The PD portion of
the model was then scaled by using literature estimates of human NSCLC tumor growth parameters. After adapting
the model to make predictions in humans, 1000 MC simulations were performed using mlxR in R 3.4.4 to estimate
the optimal scheduling of sequential BEV-PEM/CIS in humans (3).
Results: We predicted the optimal scheduling gap in mice to be 2.0 days, which is consistent with findings in
previous preclinical studies. The optimal scheduling gap in humans was estimated at 1.2 day. In our human
simulations, administrating BEV-PEM/CIS at a 1.2-day gap rather than concomitantly improved therapeutic efficacy
(defined as relative tumor volume reduction) by 52% over 85 days of treatment.
Conclusion: Using mathematical modeling to explore a range of practical scheduling regimens allowed us to
estimate the optimal scheduling gap in sequential BEV-PEM/CIS in humans without the considerable time and
resource investment required to conduct a suite of in vivo experiments. These estimates can be used to guide future
studies in optimal BEV-PEM/CIS administration, while the model parameter estimates can be used to aid in future
systems pharmacology modeling of tumor growth and response vs. antiangiogenic-antiproliferative combination
therapy.
References: 1. Tong RT, Boucher Y, Kozin SV, Winkler F, Hicklin DJ, Jain RK. Vascular Normalization by Vascular Endothelial Growth Factor Receptor 2 Blockade Induces a Pressure Gradient Across the Vasculature and Improves Drug Penetration in Tumors. Cancer Res. 2004 Jun 1;64(11):3731–6. 2. Mollard S, Ciccolini J, Imbs D-C, Cheikh RE, Barbolosi D, Benzekry S. Model driven optimization of antiangiogenics + cytotoxics combination: application to breast cancer mice treated with bevacizumab + paclitaxel doublet leads to reduced tumor growth and fewer metastasis. Oncotarget. 2017 Feb 18;8(14):23087–98. 3. Lavielle M. mlxR: Simulation of Longitudinal Data [Internet]. 2018 [cited 2019 Jan 22]. Available from: https://CRAN.R-project.org/package=mlxR
W-038
Pre-clinical Population PK Model of a Novel Antibacterial Lysin Exebacase (CF-301) for Evaluation of PK-
Efficacy Relationship
Joannellyn Chiu1, Parviz Ghahramani1, Tomefa Asempa2, Kamilia Abdelraouf2, David Nicolau2, Wessam Abdel
Hady3, Yan Xiong3, Arnold Bayer3, Teresa Carabeo4, Ray Schuch4, Cara Cassino4, Dario Lehoux4
1Inncelerex, Jersey City, NJ, 2Hartford Hosp., Hartford, CT, 3LA Biomed/UCLA Sch. of Med., Los Angeles, CA, 4ContraFect, Yonkers, NY
Introduction: Exebacase is a novel lysin with rapid S. aureus-specific bacteriolysis, potent anti-biofilm activity,
low propensity for resistance and pronounced synergy with antibiotics. Exebacase has been studied in Phase 1 and
Phase 2 trials, and demonstrated potential to improve clinical outcomes when used in addition to conventional
antibiotics.
Objectives: To develop a population pharmacokinetic (PPK) model for exebacase in animal species to simulate
exposures for PK-efficacy analysis.
Methods: Data was pooled from 15 PK studies in 4 animal species at various dosing regimens ranging from 0.03-50
mg/kg, which included q24h, q12h, and q8h fractionations. Data included intravenous (IV) and subcutaneous (SC)
injections. Injections in mice were IV or SC, and in other species, was IV. Various compartmental models with zero-
order and first-order absorption dose inputs were evaluated. Log of the observed data was modelled due to the large
range in concentrations. Body weight was used for allometric scaling in the model. Formulation and total daily dose
were examined as covariates. The PPK model was validated with goodness-of-fit plots, bootstrap, and visual
predictive check.
Results: Data included 592 animals (42 mice, 316 rats, 156 rabbits, and 78 dogs), with a total of 2,602 PK
observations. A 3-compartment linear model with first-order absorption followed by a delay input parsimoniously
described the PK data. Separate additive residual errors were estimated to account for differences in sensitivity of
various bioanalytical assays (LLOQ 0.1 ng/mL vs. 0.2 ng/mL) among studies. Dilution factor was included on
central clearance and central volume to account for errors in measurements of samples. Mean concentration-time
profile for exebacase remained detectable for 7-16 hours post dose across the species. Volume of distribution was
>16L in all species that is greater than the expected total body fluid. Bioavailability appeared to be dependent upon
route of administration and animal size. For IV injections, bioavailability was estimated to be 0.977 and 12% higher
in rabbits. Bioavailability following SC injection in mice was estimated to be about 7.5-fold higher than other
species with total daily dose as a predictor of bioavailability. There was no significant effect of formulation on PK.
Model validations (goodness-of-fit plots presented in Figure 1) support that the animal PPK model adequately
described the data. The PPK model was deemed suitable for simulation of exposures for various dosing regimens in
the efficacy studies.
Figure 1. Goodness-of-fit plots for final PPK model in 4 animal species (mouse, rat, rabbit, dog)
Conclusions: The PPK model was developed successfully for prediction of exposures in all species. The estimated
concentration-time profile of exebacase is consistent with single dose 2h IV-infusion recommended in humans. The
estimated volume of distribution indicates that exebacase penetrates well beyond vascular space and into peripheral
tissues. The results of this PPK analysis deemed suitable for PK/PD analysis.
W-039
Comparison of Sequential and Joint Nonlinear Mixed Effects Modeling of Tumor Kinetics and
Survival/Dropout following Durvalumab Treatment in Patients with Urothelial Carcinoma
Authors: Ting Chen1, Donald E. Mager1, Lorin Roskos2, Yanan Zheng2
1 Department of Pharmaceutical Sciences, University at Buffalo, SUNY, Buffalo, NY 2 Clinical Pharmacology and DMPK, AstraZeneca, South San Francisco, CA
Objectives: This study aims to evaluate the performance of sequential and joint modeling of tumor growth kinetics
and overall survival (OS)/dropout in urothelial carcinoma (UC) patients treated with Durvalumab, a monoclonal
antibody directed against PD-L1.
Methods: Longitudinal tumor size measurements and survival/dropout data from UC patients who received
durvalumab (10 mg/kg Q2W) were used for a comparison of the two modeling approaches. The base model was
developed around a prior modeling analysis from the same study [1]. For sequential modeling, the tumor kinetic
model was developed first and all parameters for individual patients were fixed in the subsequent survival and
dropout modeling. For joint modeling, tumor growth kinetics and survival/dropout data were modeled
simultaneously, which allows the survival/dropout data to inform the estimation/prediction of tumor kinetics. The
performance of these two approaches were assessed by comparing parameter estimates, standard goodness-of-fit
criteria, and predictability of a validation dataset.
Results: Both approaches performed equally well in predictions of tumor growth kinetics and survival/dropout rates.
However, the estimated individual parameters for the tumor kinetic model were different between the two
approaches: tumor growth rate constant (kg) was estimated to be greater for patients who progressed rapidly (OS≤16
weeks) compared to those who did not (kg=0.130 vs. 0.0551 week-1) with the joint approach, but similar for both
groups (kg =0.0624 vs.0.0563 week-1) with the sequential approach. The cell killing rate constant (kkill) was smaller
in patients with rapid progression with both approaches. Simulations showed that the joint approach predicted
greater tumor growth and substantial interpatient variability for patients with rapid progression when the model was
informed by survival/dropout data. In contrast, tumor growth simulations for these patients approached the
population predictions when sequential modeling was applied. Tumor size baseline (TBSL) and albumin
concentration (ALB) were identified as significant covariates for tumor kinetic parameters, and the neutrophil-to-
lymphocyte ratio (NLR), liver metastasis (LIVERBL), and Eastern Cooperative Oncology Group (ECOG)
performance status were added as the covariates for the hazard function of OS/dropout when the sequential approach
was used. With joint estimation, significant covariates (i.e., TBSL, ALB, LIVERBL, and NLR) were only added to
parameters in the tumor growth kinetic model. Sequential approach performed better in predicting the impact of
covariates on survival and dropout rate based on the visual predictive check that was stratified by covariate.
Conclusions: Both sequential and joint models of tumor growth kinetics and OS/dropout can characterize the
training and validation data reasonably well. Except for ECOG, the same covariates were identified by both
approaches, but the associated model parameters were different. The impact of covariate on survival and dropout
rate can be predicted better by using sequential approach.
Reference: 1. Zheng, Y. et al., Clin Pharmacol Ther, 2018. 103:643-652.
W-040
Pharmacokinetic (PK) analysis of weight-based and fixed dose cemiplimab in patients (pts) with advanced
malignancies
Anne Paccaly,1 Michael R. Migden,2 Kyriakos P. Papadopoulos,3 Feng Yang,1 John D. Davis,1 Ronda Rippley,1
Israel Lowy,1 Matthew Fury,1 Elizabeth Stankevich,4 Danny Rischin5
1Regeneron Pharmaceuticals, Inc., Tarrytown, NY, USA; 2Departments of Dermatology and Head and Neck
Surgery, University of Texas MD Anderson Cancer Center, Houston, TX, USA; 3START, San Antonio, TX, USA; 4Regeneron Pharmaceuticals, Inc., Basking Ridge, NJ, USA; 5Department of Medical Oncology, Peter MacCallum
Cancer Centre, Melbourne, Australia.
Background: Cemiplimab (cemiplimab-rwlc in the US), a human monoclonal anti–PD-1 antibody, has shown
substantial antitumour activity in pts with advanced malignancies and is FDA-approved for pts with advanced
(metastatic or locally advanced) cutaneous squamous cell carcinoma (CSCC). We sought to compare weight-based
and fixed dosing regimens of cemiplimab in pts with advanced malignancies.
Methods: This PK analysis included 505 pts with advanced tumours, including 135 pts with advanced CSCC, from
Phase 1 and Phase 2 studies of cemiplimab (NCT02383212 and NCT02760498, respectively). Pts received either
weight-based doses of cemiplimab (1, 3, 10 mg/kg every 2 weeks [Q2W] and 3 mg/kg every 3 weeks [Q3W]) or a
fixed dose regimen (200 mg Q2W). This analysis pooled data from 10,935 PK samples including 2,023 samples
from pts with advanced CSCC, to predict the Q3W fixed dose regiment with equivalent cemiplimab exposure as the
recommended Phase 2 dose (3 mg/kg Q2W) using a 2-compartment population PK (PopPK) model incorporating
covariates that improved goodness-of-fit statistics. The selected fixed dose regimen was confirmed by PK data of the
regimen in pts with advanced CSCC from the Phase 2 trial.
Results: In the Phase 1 study, dose-proportional kinetics were observed over the 1 to 10 mg/kg Q2W dose range.
The 350 mg Q3W fixed dose regimen was selected based on PopPK model simulations as it provided similar
cemiplimab exposure to 3 mg/kg Q2W at steady state (ss) over a 6-week (6wk) treatment period (e.g. mean
AUC6wk, ss [coefficient of variation (CV)]: 3,800 [37.2%] and 3,710 [35.9%] mg*day/mL for 350 mg Q3W and 3
mg/kg Q2W, respectively). The observed exposures in patients with advanced CSCC in the Phase 2 trial receiving
350 mg Q3W cemiplimab were comparable to the simulated exposures. In addition, the observed exposures at ss
were similar between 3 mg/kg Q2W and 350 mg Q3W.
Conclusions: The results from PopPK simulations support the use of a 350 mg Q3W fixed dose regimen, with
similar cemiplimab exposure to a weight-based 3 mg/kg Q2W regimen, in pts with advanced malignancies.
W-041
Interspecies Scaling in Pre-clinical Population Pharmacokinetics of PLG0206 – a Novel Antibacterial Peptide
Joannellyn Chiu1, Parviz Ghahramani1, Jonathan Steckbeck2 1Inncelerex, Jersey City, NJ, USA; 2Peptilogics, Pittsburgh, PA, USA.
Introduction: PLG0206 is a novel eCAP (engineered cationic antibiotic peptide) with potent activity in non-clinical
models against a broad spectrum of resistant bacteria and difficult-to-treat Gram-negative pathogens.
Objectives: Develop a population PK model for interspecies allometric scaling to extrapolate to human following
intravenous infusion to obtain human exposures for target attainment.
Methods: A population PK model was developed for data collected from various animal species in a total of 193
animals (64 monkeys, 92 rats, 34 mice, 3 dogs), with 1460 PK observations following intravenous administration.
1-, 2- and 3-compartmental models with zero-order absorption were evaluated to characterize the concentration-time
profiles of PLG0206 in different animal species with various infusion regimens. Body weight was included on
clearance and volume parameters for interspecies scaling. Total daily dose and sex were assessed as covariates. The
model was validated using bootstrap and pcVPC.
Results: Animal PK of PLG0206 from various species was best characterized with a 3-compartment linear model
with zero-order absorption. Diagnostic and goodness of fit plots supported the model is a good fit to the data. Total
daily dose (mg) is a predictor of both clearance (CL) and apparent bioavailability (F1). As total daily dose increases
from a 1 mg dose to a 2 mg dose, CL decreases by 33% and F1 increases by 7%. CL of IV infusions were
estimated to be 81% lower than CL of IV bolus injections. Sex was not statistically significant. Fixed-effects were
generally well estimated (%RSE < 20%). Bootstrap and pcVPCs (Figure 1) support that the model adequately
describes the pooled data across the species as well as each species separately.
Figure 1. pcVPC Stratified by Species
Conclusions: The final population PK model described PLG0206 animal PK data adequately for all species and was
deemed suitable for simulations to predict PK profiles for allometric extrapolation to human.
W-042
Developing Quantitative Methods to Compare Exposure-Response Relationships Between Pediatrics and
Adults to Support Pediatric Extrapolation
Qunshu Zhang1, James Travis2, Rebecca Rothwell2, Yaning Wang3, Jian Wang1
1Office of Drug Evaluation IV, OND, U.S Food and Drug Administration; 2Office of Biostatistics, OTS, U.S Food
and Drug Administration; 3Office of Clinical Pharmacology, OTS, U.S Food and Drug Administration;
Objectives: US FDA has allowed extrapolation of efficacy from adults and an abbreviated pediatric development
program where there is an expectation of a similar exposure-response (E-R) relationship in addition to a similar
disease progression and treatment response in children, when compared to adults. Historically, the assessment of E-
R similarity was based on visual inspection of two E-R curves. The objective of this study to develop a more
quantitative approach to compare exposure-response relationships between pediatric and adult populations to
inform decision making in pediatric drug development.
Methods: The methods were developed for both linear and logistic PK/PD models. The estimated treatment
response in pediatrics and adults was calculated at three points along the E-R curve (at the 10%, 50% and 90% adult
exposure quantiles), and the distribution of the estimated differences (pediatric – adult) at these points were
calculated using two methods, bootstrap and Bayesian methods. The E-R relationships were considered sufficiently
support efficacy extrapolation if the estimated probability of exceeding a non-inferiority margin based on the adult
and pediatric data at all three points was greater than a pre-defined non-inferiority threshold.
Results: Clinical trials of nine drugs were identified in which E-R similarity was concluded using visual inspection,
five drugs with linear E-R relationships (levetiracetam, oxcarbazepine, topiramate, lamotrigine, perampanel) and
four drugs with logistic E-R relationships (infliximab, golimumab, darunavir, esomeprazole). The bootstrap and
Bayesian posterior probability of exceeding the non-inferiority margin ranged from 53% to 100%.
Conclusions: This study developed and presented examples of a quantitative approach to assess the magnitude of
difference in exposure-response relationship between pediatric and adult patients. This method clarified that non-
inferiority instead of similarity is required to extrapolate adult efficacy to pediatric patients. It also provided reliable
objective criteria for non-inferiority assessment and inform pediatric trial design and decision making for efficacy
extrapolation in pediatric population.
W-043
A Semi-physiological Population Pharmacokinetic Model Developed Using Clinical Dose Escalation and Dose
Confirmation Data for an Oral Fixed-Dose Combination of CDA Inhibitor Cedazuridine with Decitabine
(ASTX727) in Subjects with Myelodysplastic Syndromes
Eric Burroughs1, Mohammad Azab, and Aram Oganesian2
1Metrum Research Group, Cambridge, MA, USA; 2Astex Pharmaceuticals, Inc., Pleasanton, CA, USA
Objectives: Cytidine deaminase (CDA) rapidly degrades decitabine (DAC), an approved treatment for
myelodysplastic syndromes, resulting in poor and variable bioavailability. Low doses of oral DAC co-administered
with a novel and potent CDA inhibitor, cedazuridine (E7727), have been shown in clinic to produce exposures
similar to IV DAC with acceptable inter-patient variability. The objective of this work was to further develop a
semi-physiological population PK model ([1]) to characterize the PK enhancement of oral DAC when co-
administered with cedazuridine and to identify potential covariates that impact the PK of DAC and/or cedazuridine.
Methods: Model development utilized serial cedazuridine and DAC plasma concentration observations of IV DAC,
oral DAC, and cedazuridine monotherapies and DAC+cedazuridine combinations. Observations from Phase 1/2
Study ASTX727-01 included dose escalation data (n=43; cohorts of 40:20, 60:20, 100:20, 100:40, and 100:30 mg
cedazuridine:DAC with n=6 per cohort), dose confirmation data (n=42; 35:100 mg cedazuridine:DAC), and an FDC
formulation extension (n=26). R was used for data processing, exploratory analysis, and visual predictive checks,
while model development and parameter estimation utilized NONMEM. Covariate effects were explored using a full
model approach.
Results: Mono- and combination therapy data were sequentially integrated into a semi-physiological population PK
model. Semi-physiological structural modeling elements included an IV DAC depot, oral DAC and cedazuridine
transit compartment absorption, and portal vein, liver, central, and peripheral compartments. CDA metabolism of
DAC primarily occurs in the liver compartment, with additional extra-hepatic metabolism. A maximum effect
(Emax) inhibition model, dependent on local cedazuridine concentrations, described the drug effect of the oral
ASTX727 combination therapy on CDA metabolism of DAC. IV DAC data were used to parameterize distribution
and metabolism of DAC, while oral DAC monotherapy data was used to parameterize oral absorption. cedazuridine
mono- and combination therapy data were used to parameterize cedazuridine PK parameters. Stratified individual-
level random effects did not demonstrate systematic biases for any covariates, including weight-based effects.
Conclusions: A semi-physiological population PK model was sequentially developed from mono- and combination
therapy observations of plasma concentrations from the ASTX727-01 dose escalation and confirmation study. The
analysis characterized the PK enhancement of oral DAC when co-administered with cedazuridine across a range of
dose regimens and found no significant covariate effects, including weight-based effects. The resulting model will
be used to interpret outcomes from an ongoing Phase 3 study (FDC ASTX727 of 35 mg DAC / 100 mg
cedazuridine), while simulations will quantitatively inform future clinical development of ASTX727.
[1] Burroughs, E., Oganesian, A., Zhang, X. and Hoke, F. Development of a Semi-Mechanistic PK/PD Model of an
Oral Fixed Dose Combination (FDC) of Cytidine Deaminase Inhibitor E7727 with Decitabine (ASTX727) in
Subjects with Myelodysplastic Syndromes. ASCPT Annual Meeting (2017)
W-044
Continuous learning in model-informed precision dosing: case study in pediatric dosing of vancomycin
Authors: Jasmine Hughes, Sirj Goswami, Ron Keizer
Institution: InsightRX, San Francisco, CA
Background: Model-informed precision dosing (MIPD) has the potential to optimize drug dosing for many narrow
therapeutic window drugs.[1-3] However, naively applying literature models into a new population often introduces
significant bias and/or imprecision.[4,5] Developing new models for each new patient population requires
considerable time and efforts, delaying potential optimal treatment of patients using MIPD. We have previously
proposed a “Continuous Learning” strategy [4,6], in which an initial model is implemented at the point of care, then
trained and updated continuously as new routinely collected TDM data becomes available.
Objectives: Evaluate the potential improvement in predictive performance of CL applied to MIPD vancomycin
dosing in a pediatric population.
Method: Models were trained and tested on de-identified patient data collected on the InsightRX platform during
routine care of pediatric patients (1 month – 20 years of age) treated with vancomycin at a single US hospital. A
popPK model was defined with a pre-specified model structure (serum creatinine and age affecting clearance, and
allometric scaling of clearance and volume parameters). The data supported only estimation of 1-compartment
models. This model was then trained on test datasets of varying sizes (n = 50, 100, 200, 300 patients), while
predictive performance was evaluated in a hold-out dataset of n=322 patients. Predictive performance, defined as the
ability of the tool to predict the next vancomycin trough level for the patient given all data available prior to the
collected level, was compared to that of three previously published parametric population PK (popPK) models for
vancomycin built on data from a general pediatric population [7-9]. Computation was automated using NONMEM
FOCEI and PsN proseval.
Results: CL improved predictive precision by 1-20% compared to the three literature models: RMSE for the CL
models was 5.1-5.3 mg/L while RMSE for literature models was 5.3-6.4 mg/L.
Conclusion: As demonstrated previously for vancomycin dosing in adults [10], CL allows for better predictive
performance compared to models from literature, even at low sample sizes. The benefit of training the model on
increasingly larger datasets appears limited in this study, but might allow further optimization of model structure, or
conditioning on more specific subpopulations. Further studies are ongoing to investigate the benefits of CL in other
drugs and populations.
References: 1. Chan D et al. Int J Pharmacokinet 2017 2. Gonzalez D et al. CTS 2017 3. Darwich AS et al. CPT
2017 4. Keizer RJ et al. CPT-PSP 2018 5. Bukkems LS et al. Int J Antimicrob Agents 2018 6. Keizer RJ et al.
PAGE 2018 7. Le et al. TDM 2014 8. Colin PJ et al. CPK 2019 9. Kloprogge F et al. AAC 2019 10. Keizer RJ et al.
PAGE 2019
W-045
A Novel Pathophysiological Drug-Survival Mathematical Modeling Framework for Ebola Virus Disease
Masood Khaksar Toroghi, Mohamed Kamal, Joel Kantrowitz, Thomas DiCioccio, Ronda Rippley
Regeneron Pharmaceuticals, Inc., 777 Old Saw Mill River Rd, Tarrytown, NY, USA
Objective: Ebola virus infection is a rare and deadly disease which remains a threat to global public health. Due to
the high mortality rate among Ebola patients, an effective treatment is urgently needed. In this work, we have
developed a mathematical modeling framework based on preclinical data that describes the antiviral effects of an
anti-Ebola antibody cocktail, REGN-EB3, and can be used to predict probability of survival in preclinical dose
ranging studies and guide dose selection for future investigations.
Method: To develop a modeling framework that connects drug concentrations in serum (i.e., exposure) to
probability of survival (i.e., response), first, a semi-mechanistic modular modeling approach was used to describe
key hallmarks of the disease (e.g., elevation in AST, ALT, BUN, and viral titer, as well as decrease in platelet count)
and drug mechanism of action (MOA). Then, a disease characteristic function (DCF) was introduced to describe
disease severity and integrated with the ordinary differential equations estimating the time course of clinical
biomarkers. Finally, the DCF, a survival analysis, and Monte Carlo simulation were combined and employed to
predict the probability of survival. Therefore, the proposed modeling framework comprises several ordinary
differential equations with some constraints (i.e., DCF, and survival analysis), and a simulation approach (i.e.,
Monte Carlo). To estimate the model parameters, data from several studies with REGN-EB-3 have been used. In
these studies, the drug was administrated at different dose levels and dosing regimens (i.e., multiple or single
dose(s)) using infected rhesus macaque as a non-human primate (NHP) model of disease.
Results: Simulation results indicate that the model can appropriately represent the dose-dependence of both the
magnitude and shape characteristics of the key hallmarks of Ebola infection, including the time course of viral load,
ALT and AST elevations, platelet count, and BUN. Also, utilization of the model to predict the probability of
survival for dose ranging studies illustrates the benefit of employing different dose levels in future investigations. In
addition, the exposure-response (E-R) analysis using the developed model demonstrates that the probability of
survival changes in a dose-dependent manner, eventually reaching a plateau (i.e., maximal efficacy).
Conclusion: The results from this analysis indicate the importance of the model-based data integration and design
to inform dose selection in future preclinical investigations. Also, the proposed framework has potential application
to guide clinical trial study design. Lastly, the model can be translated to humans to recommend an effective dose
for Ebola patients.
Acknowledgment: Authors thank the US Army Medical Research Institute of Infectious Diseases (USAMRIID)
for conducting animal studies and Biomedical Advanced Research and Development Authority (BARDA) for
funding the animal studies. Funding was provided by Regeneron Pharmaceuticals, Inc.
W-046
Pharmacokinetic-Pharmacodynamic Relationship of 2,4-Dinitrophenol in Mice and Human
Lyndsey F. Meyer1, Pooja Rajadhyaksha1, and Dhaval K. Shah1
1 Department of Pharmaceutical Sciences, School of Pharmacy and Pharmaceutical Sciences, The State University
of New York at Buffalo, Buffalo, NY 14214, USA
Objective: Oxidative phosphorylation uncouplers are promising molecules to combat the ongoing epidemic of
obesity. However, the pharmacokinetic-pharmacodynamic (PK-PD) relationship for these molecules is not well
understood in preclinical species and humans. We hypothesize that by understanding preclinical PK-PD relationship
for DNP, we can translate our findings to humans, and identify novel clinical strategies that can facilitate application
of uncouplers to combat obesity.
Methods: The PK-PD relationship of DNP was investigated in a diet-induced obese mouse model through IP and
oral administration as previously described [1]. We then employed mathematical modeling of previously developed
body composition models [2, 3] to investigate the validity of translating this relationship from mice to humans.
Results: Following continuous administration of DNP via drinking water, mice elicit an estimated 6% increase in
total energy expenditure compared to control mice (Fig 1). In addition, DNP treated mice gained on average 9% less
body weight compared to control mice without significant changes food consumption. The PK-PD relationship in
mice was translated to humans by assuming similar effect of plasma DNP concentrations on fold-change in human
energy expenditure. Simulations performed by translating the PK-PD model were compared with clinical data
obtained from DNP clinical trials performed in early 1930s [4].
Conclusions: Model simulations confirmed that reduction in body weight accumulation observed in mice translates
to body weight loss in humans. Here we have demonstrated that established mathematical models of body
composition can help in bridging the gap between different species and provide a useful tool for preclinical-to-
clinical translation of DNP efficacy.
Figure 1. Change in body weight over time for control (A-H) and DNP treated mice (I-P). Initial body weight and
time averaged food consumption were used as model inputs to describe body weight changes over time for
individual mice.
References: 1. Goldgof, M., et al., The Chemical Uncoupler 2,4-Dinitrophenol (DNP) Protects against Diet-
induced Obesity and Improves Energy Homeostasis in Mice at Thermoneutrality. The Journal of Biological
Chemistry, 2014. 289(28): p. 19341-19350. 2. Guo, J. and K.D. Hall, Predicting Changes of Body Weight, Body Fat,
Energy Expenditure and Metabolic Fuel Selection in C57BL/6 Mice. PLOS ONE, 2011. 6(1): p. e15961. 3. Hall,
K.D., Body fat and fat-free mass inter-relationships: Forbes's theory revisited. British Journal of Nutrition, 2007.
97(06): p. 1059-1063. 4. Dunlop, D.M., The use of 2,4-dinitrophenol as a metabolic stimulant. British medical
journal, 1934. 1(3820): p. 524-527.
W-047
Concentration-response modeling of ECG data from early-phase clinical studies to assess QT prolongation
risk of contezolid (MRX-I), an oxazolidinone antibacterial agent
Junzhen Wu1,2,3, Kun Wang5, Yuancheng Chen1,2,3,4, Hong Yuan1,2,3, Li Li1,2,3, Jing Zhang1,2,3,4*
1 Institute of Antibiotics, Huashan Hospital, Fudan University, Shanghai, China; 2 Key Laboratory of Clinical
Pharmacology of Antibiotics, National Health and Family Planning Commission, Shanghai, China; 3 National
Clinical Research Center for Aging and Medicine, Huashan Hospital, Fudan University, Shanghai, China; 4 Phase I
Unit, Huashan Hospital, Fudan University, Shanghai, China; 5 Certara Strategic Consulting China, Shanghai, China,
200122
Objectives: Contezolid (MRX-I) is an oxazolidinone antibacterial agent which is under development for the
treatment of skin and soft tissue infections. Relationship between concentration and QT of MRX-I is still unknown.
The aim of this study was to develop concentration-QT model to evaluate impact of MRX-I on cardiac
repolarization.
Methods: A population modeling approach was performed using data from Phase I study incorporating single
ascending dose, multiple ascending dose, and food effect assessments in healthy Chinese subjects. Linear mixed
effect models were used to assess the relationships between MRX-I plasma concentration and QT/QTc/ΔQTc
(baseline adjusted), in which different correction methods for heart rate have been included.
Results: The upper bound of the one-sided 95% confidence interval (CI) for predicted ΔΔQTc was <10 milliseconds
(ms) at therapeutic and supratherapeutic doses of MRX-I. Additionally, the upper bound of one-sided 95% CIs for
ΔQTcB and ΔQTcF model was higher than other three models. Model performance/suitability was determined using
diagnostic evaluations, which indicated rationality of one-stage concentration-QT model.
Conclusions: MRX-I may have no clinical effects on the QT interval. Concentration-QT model may be an
alternative to conventional thorough QT studies.
Population predicted values of QTc interval vs predicted concentrations of MRX-I. (a) female subjects. (b) male
subjects
W-048
The Use of a Mechanism-Based Model (MBM) to Better Characterize the Pharmacology of Combination
Colistin (COL) and Ceftazidime/Avibactam (CZA) Across Five Carbapenem-resistant Klebsiella Pneumoniae
Strains
Joseph Piscitelli1, Hongqiang Qiu1, Rajnikant Sharma1, Estefany Garcia1, Patrick Hanafin1, David van Duin2, Gauri
G. Rao1
1UNC Eshelman School of Pharmacy, Chapel Hill, NC, 2UNC School of Medicine, Chapel Hill, NC
Objectives: Carbapenem-resistant Klebsiella pneumoniae (CRKP) are resistant to multiple classes of antibiotics,
highlighting the need for combination therapy to ensure adequate sustained killing activity and suppress the
emergence of resistance. We aimed to develop a MBM using static concentration time-kill (SCTK) studies to best
characterize the pharmacodynamic (PD) activity of COL and CZA in monotherapy and combination across five
isolates, with varying susceptibilities.
Methods: SCTK were performed using five clinical blood isolates obtained from patients enrolled in The
Consortium on Resistance Against Carbapenems in Klebsiella and Other Enterobacteriacae (CRACKLE) study.
Both the COL and the CZA MICs ranged from 0.5 to 4 mg/L. SCTK were performed using colistin (0.5 to 16 mg/L)
& ceftazidime (16 to 128 mg/L)/avibactam (4 to 32 mg/L) alone & in combination (COL 0.5, 1, 2 mg/L; CZA 16/4,
32/8, 64/16 mg/L) against an initial bacterial inoculum of 106 CFU/mL. Bacterial quantification was done at 0, 1, 2,
4, 6, 8, and 24 hours. PD response data for the different isolates against COL and CZA were co-modeled using
maximum likelihood estimation in ADAPT5.
Results: The effect of each drug was described as enhancing the natural bacterial death rate by the Hill equation. A
proportionality constant was used to capture the shift of ceftazidime’s EC50 when given with avibactam, to better
describe its beta-lactamase properties. Across the five isolates that were evaluated, the constant was able to capture
the PD activity of CAZ in monotherapy and in combination, with its value ranging from 0.000828-0.0405 (median:
0.00913). Furthermore, the MBM was able to adequately capture the reduction in bacterial burden as well as the
minimal regrowth seen with each of the five isolates, even when the susceptibilities were varied. When comparing
the drug parameter estimates, the MBM was able to adequately predict a lower EC50, 0.246-10.4 mg/L (median:
1.31 mg/L), in the COL susceptible isolates and a higher EC50, 106-632 mg/L (median: 173 mg/L), for the more
resistant isolates. A similar trend was noted for CZA, with susceptible and resistant EC50’s ranging from 0.000601-
12.3 mg/L (median: 2.922 mg/L) and 7.90-319 mg/L (median: 219.6 mg/L), respectively. COL and CZA Emax
values ranged from 9.57-74.5 (median: 21.9) and 5.13-10.9 (median: 5.13).
Conclusions: The MBM allows us to compare the differences in the time course of bacterial killing across multiple
resistant clinical isolates. This modeling approach will help us optimize this antibiotic combination to further
maximize the killing profile with minimal emergence of resistance. We will validate these optimized combination
regimens in a dynamic in vitro infection model simulating humanized PK for this combination.
W-049
Establishing Translational Platform for Tuberculosis Drug Development in Murine Models with First-line
TB Drug Regimen
Authors: Nan Zhang 1, Sandeep Tyagi 2, Heena Soni 2, Eric L. Nuermberger 2, Rada Savic 1
Affiliations: 1. University of California San Francisco, San Francisco, CA, USA 2. Center for Tuberculosis
Research, Department of Medicine, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA
Objectives: The development of novel drug regimens to test in clinical trials for treatment of tuberculosis (TB)
relies heavily on results from preclinical murine models (1, 2). However, two clinical trials (TBTC study 29 and
29X) did not demonstrate the expected outcomes as shown in the mouse studies (3-5). This highlights a gap in
knowledge regarding the predictive accuracy of commonly used preclinical TB murine models (1). Translating the
results from commonly used murine models requires a full understanding of the relationship between various
factors, including pathology, immunity, mechanism of action for the drugs, pharmacokinetics (PK),
pharmacodynamics (PD), etc. Quantitative modeling and analysis to investigate the role of each factor using the data
of BALB/c mice with the treatment of the first-line TB drugs, including rifampin (RIF), isoniazid (INH) and
pyrazinamide (PZA), is important. It lays out the foundation of the translational platform for TB drugs or regimens
development. Most importantly, it enables more efficient and effective translation of the findings from animal
models to optimally inform the design and interpretation of clinical trials for TB drug development.
Methods: Bacterial infection model to describe bacterial number in immune-competent (BALB/c) with inhibitory
adaptive immune effect was developed using NONMEM (v7.4). The adaptive immune effect and drug effect of RIF
or INH monotherapy, were added into the baseline model either as an inhibition on bacterial growth or an
acceleration on bacterial death. Drug effect was characterized using different exposure-response relationships,
including linear, nonlinear, Emax and sigmoidal function. Goodness‐of‐fit plots and visual predictive checks were
used for model evaluation.
Results: A mechanistic model with the exponential growth and death of bacteria, sigmoidal adaptive immune effect
and drug effect of RIF and INH monotherapy in mouse model was established. Bacterial infection was described
and PK/PD relationship of RIF and INH monotherapy was characterized individually (Table 1). RIF (with higher
Emax, 554% vs 90.4%) and INH (with lower EC50, 0.225 vs 2.21 ug/mL) are both more active against replicating
bacteria than non-replicating bacteria.
Conclusions: The backbone of the translational platform for TB drug development with the first-line TB regimen
has been established. This work exemplifies the platform structure for evaluating new TB drugs or regimens in
preclinical studies and will eventually serve as an innovative solution to increase the accuracy with which preclinical
results are translated into clinical predictions in TB drug development.
References: 1. Microbiol Spectr. 2017 Jun;5(3). 2. Clin Transl Sci. 2017 Sep;10(5):366-379. 3. Antimicrob Agents
Chemother. 2012, 56 (8), 4331-40 4. J Infect Dis. 2012, 206 (7), 1030-40. 5. Am J Respir Crit Care Med.
2015, 191 (3), 333-43.
Table 1. Parameter Estimates of Translational Platform Model
W-050
Parameter Description Value (RSD, %)
Kg (day-1) Bacterial growth rate 1.22 (3%)
IIVKg (%) Interindividual variability of bacterial
growth rate 7.8 (13%)
Kd (day-1) Bacterial death rate 0.41 (FIX)
KB (%) Maximal inhibitory CFU-dependent
adaptive immune effect 22.4 (31%)
B50 (log10 CFU) CFU counts to reach half of KB 6.99 (2%)
B Steepness of the CFU-dependent adaptive
immune effect curve 2.84 (1%)
KT (%) Maximal inhibitory time-dependent
adaptive immune effect 66.9 (2%)
T Steepness of the time-dependent adaptive
immune effect curve 5.54 (17%)
T50 (days) Time to reach half of maximal time
covariate 19.2 (7%)
EC50INH_acute (ug/mL) Concentration to reach half of maximal
growth inhibitory effect of INH 0.225 (7%)
EMAXINH_acute (%) Maximal growth inhibitory effect of INH 69.8 (1%)
EC50INH_chronic (ug/mL) Concentration to reach half of maximal
growth inhibitory effect of INH 2.21 (124%)
EMAXINH_chronic (%) Maximal growth inhibitory effect of INH 62.7 (23%)
EC50RIF_acute (ug/mL) Concentration to reach half of maximal
death acceleratory effect of RIF 63.3 (72%)
EMAXRIF _acute (%) Maximal death acceleratory effect of RIF 554 (45%)
EC50RIF _chronic (ug/mL) Concentration to reach half of maximal
death acceleratory effect of RIF 46.5 (60%)
EMAXRIF_chronic (%) Maximal death acceleratory effect of RIF 90.4 (29%)
Population Pharmacokinetics Model of Sunitinib and its Active Metabolite (SU12662) in Cancer Patients in
Combination with Combination Antiretroviral Therapy (cART) for HIV: AIDS Malignancy Consortium trial
AMC 061.
Amir S. Youssef1,2,3, Johns Deeken4, David Aboulafia5, Vijay Ivaturi2, Jogarao Gobburu2, Michelle A. Rudek6
1Synteract Inc., Exton, PA, USA; 2Center for Translational Medicine, School of Pharmacy, University of Maryland
Baltimore, Baltimore, MD, USA; 3Department of Pharmaceutics, Faculty of Pharmacy, Cairo University, Cairo,
Egypt; 4Inova Comprehensive Cancer and Research Institute, Falls Church VA, USA; 5Division of Hematology and
Oncology, Virginia Mason Medical Center and University of Washington, Seattle, WA, USA; 6The Sidney Kimmel
Comprehensive Cancer Center at Johns Hopkins, and Department of Oncology and Medicine Johns Hopkins
University, Baltimore, Maryland, USA.
Objectives: Sunitinib is metabolized by CYP3A4 to its active metabolite (SU012662). Several cART drugs are
known to be substrates, inducers or inhibitors for CYP3A4. The objective of this analysis was to determine the
population pharmacokinetics of sunitinib and SU12662 in presence of different cART drugs.
Methods: Sunitinib maleate was given orally at dose of 25, 37.5 and 50 mg/day to 19 patients with solid or
hematologic malignancies who were HIV positive. Patients were stratified based on the cART therapy received.
Nine patients received ritonavir-based cART (a strong CYP3A4 inhibitor), and 10 patients received non-ritonavir
protease inhibitors cART therapy, including 2 patients who received efavirenz, a potent inducer of CYP3A4.
Sunitinib and SU012662 concentrations were measured in plasma for up to 24 hours after the first dose on cycle 1.
Weekly troughs were obtained on day 8, 15 and 22 with an additional 6 timepoints collected after the last dose of
sunitinib on days 29, 30, 32, 35, 38 and 42. Sunitinib and SU012662 concentration-time data were analyzed using
non-linear mixed-effects modeling. Age, race, ethnicity and concomitant cART therapy were tested as potential
covariates. Bootstrap and visual predictive check simulations were used to evaluate the final model.
Results: Sunitinib plasma data was best described by a 2-compartment model with first-order absorption and a lag-
time. The population estimates of clearance, central and peripheral volume of distribution were 46 L/hr, 1982 L and
658.4 L, respectively. None of the evaluated covariates were identified as statistically significant affecting sunitinib
pharmacokinetics. The metabolite data was fit sequentially using post-hoc parameters from the parent model to a 2-
compartment model. Fraction metabolized was fixed to 0.21 based on literature. Age (mean±SD 52.4±9.6 years),
race (47% white, 42% black and 11% unknow), ethnicity (21% Hispanic, 68% non-Hispanic and 11% unknown)
were not found to be significant covariates into the model. Concomitant cART therapy was identified as a
statistically significant covariate on the fraction metabolized and its inclusion in the model reduced the between-
subject variability in this parameter from 63.8% to 30.7%. Model validation techniques confirmed the stability and
robustness of the final model.
Conclusions: The pharmacokinetics of sunitinib and its active metabolite were adequately described by a two-
compartment model for each compound. Co-administration of certain cART drugs with sunitinib was confirmed as a
significant covariate on the fraction metabolized leading to marked changes in SU12662 exposure. The current
model will be utilized to conduct exposure-response analysis to support dosing recommendations. This study shows
how combined efforts of clinical pharmacologists, pharmacometricians and clinicians can assist to generate
important dosing related questions. Future work will involve the development of a PBPK model to ascertain the
impact of this drug-drug interaction.
W-051
Using Medical Claim Data and Modified Wald’s Approximated Covariate Selection Method to Develop a
Population Disease Progression Model for Leuprorelin-treated Subjects with Hormone-sensitive Prostate
Cancer
Yixuan Zou1, Fei Tang1, Chee M. Ng1 1University of Kentucky College of Pharmacy, Lexington, KY
OBJECTIVE: Leuprorelin is widely used as initial treatment for patients with hormone-sensitive prostate cancer
(PCa). However, the treatment effects of leuprorelin and patient-specific factor on disease progression in PCa
patients have never been characterized. Most of the published disease progression models are developed using
randomized controlled studies (RCT) data. The strict inclusion/exclusion criteria used in the RCT may limit the
study results in addressing clinical questions in a real-life patient population. This RCT limitation has led to an
increased reliance on using medical claim data in designing healthcare policies related to real-life clinical practice.
However, medical claim data have never been used to develop disease progression model. Therefore, the objective
of this study was to use a medical claims data to develop the first disease progression model for leuprorelin-treated
PCa patients.
METHOD: 1113 prostate cancer antigen (PSA) observations from 264 PCa subjects in Humana clinical database
were used in the analysis. Two mathematical models with either constitutive or inducible drug resistance were
developed using MCPEM (Method=IMP) in the NONMEM 7.3. The PSA with BQL values were included and
handled as fixed censored observations in the analysis. Eleven covariates were tested in the analysis. A modified
Montel-Carlo Parametric Expectation (MPEM)-based Wald’s approximated Method (M-WAM) developed by our
group was used as an efficient covariate model selection method to construct the final model. In this approach, a
single full model that included all the covariate effects on model parameters was developed to generate a covariate
matrix for WAM calculation. Then a backward elimination (BE) using WAM-derived likelihood from the full
model was used to efficiently eliminate insignificant covariates, followed by BE with an actual NONMEM run to
yield the final model.
RESULTS: The model with constitutive drug resistance was selected as the final model. PSA kinetics in
leuprorelin-treated PCa patients were well described by the model and parameters were estimated with good
precision (%CV<50)(Table 1). Population average of RP was 3.94, indicating that 1.94% (R=e-RP) of the original
PSA-producing cancer cell population was inherently resistant to leuprorelin treatment. Lower hemoglobin (HGB)
level was associated with lower RP and reduced response to leuprorelin treatment. Higher PSA levels was related to
shorter drug resistance development time. These findings were consistent with the clinical observations.
Furthermore, the model-based simulation provided the first mechanism insights for the complex interaction between
HGB/baseline PSA and observed PSA kinetic parameters (nadir levels and median time to nadir levels) related to
clinical outcomes.
CONCLUSIONS: First disease progression model for leuprorelin-treated PCa patients were successfully developed
using routine medical claim data, and provided a mechanistic insights for treatment effect and patients-specific
factors on drug resistance development and disease progression in leuprorelin-treated PCa patients.
REFERENCES: [1] Zou YX, et al (2017). J Pharmacokinet Pharmacodyn, 44,S75
Table 1. Parameter estimates of the final Disease Progression model for Leuprorelin-treated PCa Patients.
Parameter Estimate %CV
Structural Model aDS (day-1) 3.78 x 10-2 6.19 bGs (day-1) 1.96 x 10-3 22.5 cRP 3.94 7.44 dGR (day-1) 6.54 x 10-4 28.4
Interindividual Variability
𝜔𝐷𝑆2 0.453 14.8
𝜔𝐺𝑆2 2.59 19.7
𝜔𝑅𝑃2 0.944 16.4
𝜔𝐷𝑅2 3.76 21.1
Covariate Model eHGB on RP (𝜃𝐻𝐺𝐵_𝑅𝑃) 2.30 24.7 fBAS on DS (𝜃𝐵𝐴𝑆_𝐷𝑆) 0.174 24.5 gAND on DS (𝜃𝐴𝑁𝐷_𝐷𝑆) 0.677 30.0
Residual Variability
Additive error (𝜎𝑎𝑑𝑑) 2.01x10-1 3.49
%CV = percent coefficient of variation. adrug effect on drug-sensitive tumor cells, bgrowth rate of drug-sensitive
cells, exp(-RP) represents the fraction of drug-sensitive tumor cells in the original tumor, dgrowth rate of drug-
resistant cells, eeffect of hemoglobin level on RP, feffect of baseline prostate-specific antigen level on Ds, geffect of
antiandrogen use on Ds
W-052
Population Pharmacokinetics of Cabotegravir in Adult Healthy Subjects and HIV-1 Infected Patients
Following Administration of Oral Tablet and Long Acting Intramuscular Injection
Authors: Kelong Han, Mark Baker, Mark Lovern, Prokash Paul, Yuan Xiong, William Spreen, Katy Moore, Susan
L Ford
Affiliations: GSK, ViiV Healthcare, Certara
Objectives: Cabotegravir is an integrase inhibitor currently in Phase 3 development as an oral tablet and an
intramuscular long-acting injection (LA) for HIV treatment and prevention. The aim of this analysis was to
characterize cabotegravir population pharmacokinetics (PopPK) using data from Phase 1, 2 and 3 studies, evaluate
the association of intrinsic and extrinsic factors with the variability of cabotegravir PopPK, and perform simulations
to inform dosing strategies.
Methods: All analyses were implemented in NONMEM 7.3 and R. The M3 method was used to model
concentrations below the quantitation limit. Likely covariate relationships were evaluated using a forward addition
(p<0.01) and backward elimination (p<0.001) approach. Model adequacy and predictive performance were assessed
using bootstrapping and visual predictive checks. The relative contributions of covariates on trough concentration
following the first injection (Cmin-LD), steady-state trough (Cmin-SS) and peak concentration (Cmax-SS) were
evaluated through simulations.
Results: A total of 23,926 cabotegravir plasma concentrations were collected from 1647 healthy (28%) and HIV
infected (72%) adult subjects (age 18 to 74) in 16 studies at 7 dose levels (10 mg to 60 mg for oral tablet; 100 mg to
800 mg for LA). LA was administered in >75% of the subjects. A two-compartment model with first-order oral and
intramuscular absorption and elimination adequately described the data. The M3 method significantly improved the
fitting. Clearances and volumes were scaled to body size. Relative bioavailability of the oral to LA formulation was
75.6%. Race and age (adult) were not significant covariates. LA absorption rate constant (KALA) was 50.9% lower
in females than males, and 47.8% higher if the LA dose was given as two split injections instead of one single
injection. KALA decreased with increasing BMI and decreasing needle length. Clearance was 17.4% higher in
current smokers. Gender and BMI had the largest contribution on Cmin-LD. Median Cmin-LD was 31% lower in
females compared to males and 31% lower in subjects with BMI ≥30 kg/m2 than subjects with BMI <30 kg/m2.
Body weight, split injection, needle length and smoker status had relatively small contribution of <20% on Cmin-LD.
All covariate effects on Cmin-LD were still within the concentration ranges observed to be effective in the Phase 3
studies. All covariates had relatively small contribution of ≤15% on Cmin-SS and ≤21% on Cmax-SS.
Conclusions: A robust PopPK model for cabotegravir was developed that can be used for simulations to support
dosing strategies and future studies. Gender, BMI, split LA injection, and needle length were significant covariates
on cabotegravir LA absorption, and smoker status was a significant covariate on clearance. However, the magnitude
of their relative contribution on cabotegravir exposures was relatively small. No dose adjustment of cabotegravir
based on these covariates is recommended.
W-053
NKTR-262 Released Below Quantifiable Levels of TLR 7/8 Agonist in Human Plasma in Phase 1b/2 Clinical
Study as Predicted A-Priori by PK Modeling and Scaling to Humans
Kavitha Bhasi1,a, Nathan Hanan2,a, Anh Nguyen2, Ute Hoch2, and Werner Rubas1
1Non-Clinical PK, 2Clin Pharm and Quant. Sciences, Nektar Therapeutics, San Francisco, CA aBoth authors contributed equally to this work.
Background: NKTR-262 is a novel polymer-modified prodrug of a TRL7/8 agonist designed to provide sustained
intratumoral engagement of the TLR7/8 pathway, promoting an immune stimulatory environment and tumor antigen
release. When administered with bempegaldesleukin (NKTR-214), a CD122-preferential IL-2 pathway agonist, the
combined effect of innate immune stimulation and enhanced antigen presentation with sustained T cell activation
leads to systemic tumor immunity. A Phase 1b/2 trial was initiated to evaluate the safety, efficacy and
pharmacokinetics (PK) of intratumorally (IT) administered NKTR-262 in combination with bempegaldesleukin.
Here we present initial plasma pharmacokinetics (PK) of NKTR-262 and released active TLR 7/8 agonist in
comparison to model predicted concentration values derived from in-vitro studies, non-clinical animal PK, and
clinical literature values.
Methods: Allometric scaling from mouse, rat and dog PK studies was used to predict Volume (V) and clearance
(CL) of NKTR-262 in humans. These parameters were applied to predict both 1- and 2-compartmental dispositions
of NKTR-262 in humans. The absorption rate for NKTR-262 following IT administration was fixed to that from
nonclinical species after IT (0.643 1/hr) and subcutaneous (SC) administrations (0.0655 1/hr). The in vitro release
rate of the TLR7/8 agonist from NKTR-262 was determined in human plasma. PK parameters for the TLR7/8
agonist in humans were fixed to those reported in the literature. Oral bioavailability (F) of TLR7/8 in humans was
estimated between 0.5-1 using PBPK modeling in GastroPlus®. A value of 0.5 was used for F to estimate V and CL
parameters for the TLR 7/8 agonist in humans. Modeling was conducted using NONMEM (ver 7.4).
Results: As of Feb 19, 2019, plasma PK of NKTR-262 and the active TLR 7/8 agonist were available for
evaluation from NKTR-262 IT doses ranging from 30-120 g. The majority of observed concentrations fell
between the limit of detection (NKTR-262: 0.7 ng/mL; TRL 7/8 agonist: 0.007 ng/mL) and the lower limit of
quantitation (NKTR-262: 4.149 ng/mL; TRL7/8 agonist: 0.1 ng/mL), but appeared to increase with dose. The 2-
compartment disposition model using the IT absorption rate predicted the observed NKTR-262 concentrations well.
Observed plasma concentrations of the TLR7/8 agonist were higher than those predicted by the model. While the
modeled in vivo release rate in mice agreed with those measured from in vitro mouse plasma incubations, in vivo
release in humans was faster than predicted from the model based on incubation in human plasma, thereby
contributing to the observed differences in TLR7/8 agonist concentrations in humans. Additional data will be
collected to refine the model.
Conclusion: A PK model developed by allometric scaling of nonclinically derived PK parameter estimates
accurately predicted NKTR-262 concentrations and observed data confirmed low systemic exposure of TLR7/8
agonist in a first in human study of IT administered NKTR-262.
W-054
POPULATION PHARMACOKINETICS AND EXPOSURE-RESPONSE ANALYSES FOR VENETOCLAX
IN COMBINATION WITH R-CHOP IN PREVIOUSLY UNTREATED DLBCL PATIENTS
Authors: Divya Samineni,1 Hao Ding,1 Rong Zhang,1 Chunze Li,1 Sandhya Girish,1 Arijit Sinha,2 Richa Rajwanshi,1
Jessie Randhawa,2 Kathryn Humphrey,2 Alexandra Bazeos,2 Nathalie Spielewoy,3 Ahmed H Salem,4 Dale Miles1
Affiliations: 1Genentech, Inc., South San Francisco, CA, USA; 2Roche Products Ltd, Welwyn Garden City, UK; 3F.
Hoffmann-La Roche Ltd, Basel, Switzerland; 4AbbVie Inc., North Chicago, IL, USA
OBJECTIVES: Venetoclax (VEN) is a selective, potent BCL-2 inhibitor with clinical activity in non-Hodgkin
Lymphoma (NHL). We aimed to use population pharmacokinetics (PopPK) and exposure-response (ER) analyses
from the Phase Ib/II CAVALLI study (NCT02055820) to confirm the dose selection of VEN in combination with
R-CHOP for future studies.
METHODS: CAVALLI included 228 patients with relapsed or refractory (R/R) or previously untreated (1L) NHL
(208 1L diffuse large B-cell lymphoma [DLBCL]) treated for 8 cycles with 400–800 mg VEN (dosing regimen of
10 doses/21-day cycle) and R-CHOP. A previously developed PopPK model for VEN from patients with R/R
CLL/small lymphocytic lymphoma and R/R NHL was used to describe the observed pharmacokinetic data of VEN
in combination with R-CHOP and provide post-hoc predicted pharmacokinetic parameters. In this interim evaluation
of 223 patients, VEN steady-state exposure (AUCss) was used as the exposure metric for predicting clinical
response (efficacy, safety, or tolerability). To isolate the VEN effect, ER analyses referenced data from the R-CHOP
arm of a historical control study in 1L DLBCL (BO21005, N=571 with IPI 2–5). Logistic regression and Cox
proportional hazard models were used to assess impact of exposure on key efficacy endpoints, i.e.
PET-complete response (CR)/overall response (OR) and investigator-assessed progression-free survival (PFS).
Logistic regression was used to assess impact of exposure on key safety endpoints, i.e. selected grade ≥3 adverse
events in the intent-to-treat (ITT) population and a high-risk BCL-2+ sub-population (N=30). Covariates
(demographics and baseline disease characteristics) were evaluated for impact on efficacy and safety. The effect of
VEN exposure on dose intensity of individual R-CHOP components was also evaluated.
RESULTS: The observed pharmacokinetic data for VEN given with R-CHOP were consistent with the original
PopPK model predictions; the pharmacokinetic parameters were comparable with those reported previously for
VEN when administered as monotherapy or with rituximab. No significant trends were observed between VEN
exposures and PFS (p-value range: 0.56–0.79) and the probability of CR/OR (p-value range: 0.27–0.85) for both the
ITT and BCL-2+ populations. Statistically significant increases in grade ≥3 neutropenia (p=0.02) and infections
(p=0.004) were observed with VEN exposure; no significant trends were observed with grade ≥3 thrombocytopenia
(p=0.05) or febrile neutropenia (p=0.23). Similar dose intensities were observed for VEN and R-CHOP components
across VEN exposures, suggesting VEN did not impact delivery of the R-CHOP backbone. In addition, a positive
benefit–risk profile with clinical efficacy and safety data was observed at the 800 mg VEN dose given with
R-CHOP in both the ITT and BCL-2+ populations.
CONCLUSIONS: The PopPK and ER analyses, in addition to the positive benefit–risk observed in the clinical data,
support the selection of 800 mg VEN in combination with R-CHOP in 1L DLBCL patients for future studies.
Disclosures: Venetoclax is being developed in collaboration between Genentech and AbbVie. Genentech and
AbbVie provided financial support for the study and participated in the design, study conduct, analysis, and
interpretation of data as well as the writing, review, and approval of this publication. Editorial assistance was
provided by Lynda McEvoy of Gardiner-Caldwell Communications, UK, and was funded by F. Hoffmann-La Roche
Ltd.
W-055
Efficacious dose selection of a phosphoinositide 3-kinase inhibitor using pharmacokinetic-pharmacodynamic
analysis
Suein Choi1,3, Sangil Jeon2, Seunghoon Han1,3, Dong-seok Yim1,3
1PIPET (Pharmacometrics Institute for Practical Education and Training), College of Medicine, The Catholic
University of Korea, Seoul 06591, Korea; 2Q-fitter, Inc.; 3Department of Pharmacology, College of Medicine, the
Catholic University of Korea, Seoul 06591, Korea;
Objectives: Phophoinositide-3-kinase (PI3K) inhibitor induces apoptosis and prevents proliferation in cell lines
derived from malignant B-cells and primary tumor cells showing potential as second-line drug for relapsed chronic
lymphocytic leukaemia patients. The aim of this study was to predict the appropriate efficacious dose of PI3K
inhibitor for humans in terms of the inhibition of tumor size by developing PI3K
pharmacokinetic/pharmacodynamic (PK/PD) model based on data from preclinical studies and information of
comparator drug when the preclinical data is not sufficient.
Methods: PI3K inhibitor was orally administered to rats (1, 3mg/kg, n=6) for PK assessment and mouses (3, 10, 30,
100mg/kg, n=50) for PD assessment. Animal PK and PK/PD model of all dose levels in each species were
developed using mixed effect modeling analysis by NONMEM (Version 7.4). As rat PK data is not enough to
predict human PK parameter, to gather more information, rat and dog PK raw data of comparator drug was drawn
from FDA approval document by using digitizer and PK parameters were estimated by developing PK models.
Allometric approach with various physiological factors (e.g. Brain weight, MLP) and information of comparator
drug were used to predict human PK parameter. Human PK/PD profile was simulated to estimate the potential
efficacious dose in terms of tumor size and known effective Ctrough level of comparator drug.
Results: Two-compartment first-order elimination model with first-order absorption followed by zero-order
absorption model was selected as final PK model, and logistic model was selected as final PD model to describe the
tumor volume (CL/F = 1.622L/hr, Vc/F = 3.68L, F1 = 0.423, Q/F = 1.934L/hr, Vp/F = 15.4L, ka = 3.68/hr, D2 =
0.27hr, ALAG2 = 1.97hr, Kin = 0.00642/hr, Kout = 0.000316, TVmax = 11900 mm3, KC50 = 8.83ng/mL, Kmax =
0.0665, = 2.26). Human PK parameters were predicted using brain weight applied allometric approach showing
fittest in animal PK parameters (CL/F = 23.45L/hr, Vc/F = 47.74L, ka = 3.68/hr, Q/F = 144.24L/hr, Vp/F = 364.85L).
Human PK/PD profile was simulated and 85mg QD was efficacious dose in terms of tumor size. Also, by applying
in vitro potency ratio (IC50 = 0.5nM vs 1.39nM) to known effective Ctrough level in human of comparator drug
(125ng/mL), 50mg QD was expected to be effective showing Ctrough level over 50ng/mL.
Conclusion: From the human simulation with predicted PK/PD parameter using preclinical data and known
parameters of comparator drug, 50mg QD and 85mg QD were predicted as efficacious dose for human despite of
limited animal PK data.
Reference: Food and drug administration center for drug evaluation and research review report for idelalisib (2014)
Ramanathan S et al. Clinical Pharmacokinetic and Pharmacodynamic Profile of Idelalisib. Clin. Pharmacokinet.;
2016;55:33–45.
W-056
Application of a Mechanistic Weight Loss Model to Predict Drug-Induced Body Weight Change in the Obese
Population
Authors: Sumit Basu, Brian Maas, Akshita Chawla, Matthew L. Rizk, Larissa Wenning
Affiliations: Merck & Co., Inc., Kenilworth, NJ, USA
Objectives: Weight loss is defined by the imbalance between metabolizable energy intake and energy expenditure,
which usually changes dynamically with time. Clinical studies have indicated that most anti-obesity drugs induce
body weight loss in humans by reducing net daily energy intake through various mechanisms such as reduction of
food consumption, decrease of macronutrient absorption, or decline in metabolizable energy intake, with minor
effects on energy expenditure. Previously, Gobel and collaborators1 developed a dynamic mechanistic weight loss
model based on the principle of energy balance to quantify the metabolizable energy intake changes during long-
term obesity pharmacotherapy. The aim of this work to apply this model to predict the drug-induced body weight
change from early calorie intake reduction data induced by various anti-obesity drugs.
Methods: The mean body weight time course data of various anti-obesity drugs reported in published literature were
provided as an input to the mechanistic weight loss model parameterized by three key parameters, namely, initial
calorie intake reduction (Pearly), late calorie intake reduction (Plate) and the time-constant from Pearly to Plate (τ).
Parameters from model fits to body weight time course data for 14 drugs in 28 trial-arms1 and a dose ranging phase
2 clinical data from a multiple-dose semaglutide 52-week trial2 were separately fit to two linear models of Pearly vs.
Plate.
After the determination of the model parameters, drug-induced body weight was predicted at various time intervals
and were compared to a linear regression meta-analyses of body weight at the same fixed timepoints from previous
reports3. Model development, evaluation and simulations were performed in R 3.5.1.
Results: The goodness of fit of the model parameters was well described by the relatively high coefficients of
determination (R2 > 0.8). The model indicated that there is a linear relationship between the Pearly and Plate. In
addition, the model can predict body weight loss over 52 weeks of treatment.
Conclusions: The mechanistic weight loss model can be utilized as a quantitative modeling platform to predict the
drug-induced body weight change for new investigational anti-obesity compounds in the early clinical development
phase as well as can be compared with the results of the existing marketed compounds.
References: [1] Göbel et al. Obesity. 2014;22(10):2105-8. [2] O’Neil et al. Lancet. 2018;392(10148):637-649. [3]
Plock et al. J. Clin. Pharmacol. 2017;57(1):52-63.
W-057
An Exemplar of Model-Informed Drug Discovery and Development towards Discovering Promising
Combination Therapies for Mycobacterium tuberculosis
Sarah Kim1, Jenny Myrick2, Jocelyn Nole2, Michael Maynard2, Brandon Duncanson2, Arnold Louie2, Stephan
Schmidt1,*, George L. Drusano2
1 Center for Pharmacometrics and Systems Pharmacology, Department of Pharmaceutics, College of Pharmacy,
University of Florida, Orlando, FL, USA 2 Institute for Therapeutic Innovation, College of Medicine, University of
Florida, Orlando, FL, USA * Corresponding author’s email: [email protected]
Objectives: Tuberculosis (TB) is one of the top 10 causes of death worldwide, causing an estimated 1.6 million
deaths in 20171. Discovering promising combination therapies for TB are needed in order to combat antimicrobial
resistance as well as to shorten the therapy duration. Due to a huge number of possible drug combinations, a model-
informed strategy will help to accelerate the discovery. The objective of this research was to inform experimental
trial design of studies to investigate effects of two combination therapies of 1) bedaquiline (BDQ) with pretomanid
(PMD) and 2) clofazimine (CFZ) with PMD against Mycobacterium tuberculosis (M. tuberculosis) in Log-, Acid-,
and Nonreplicating-Persister (NRP)-metabolic phases through mathematical modeling approach.
Methods: M. tuberculosis bacterial counts were obtained in duplicate from the combination regimens for each
metabolic phases through in vitro checkerboard assay. The use of the Greco model2 allows us to characterize the
effect (bacterial killing) of the different concentrations of the drugs in combination in a quantitative manner by
evaluating the interaction parameter α and its confidence interval (CI). Synergy was declared if α had a positive
value and the lower bound of the 95% CI does not cross zero. The interaction was antagonistic if α was negative
and the upper bound does not cross zero. If the 95% CI of α contains zero, then additivity is declared for the effect
of the drug combination. Model parameters were estimated in ADAPT 5 using either maximum likelihood or
weighted least squares estimation methods. R (version 3.5.1) was used for data management and Mathematica
(version 11.3) was used for creating 3-dimensional goodness-of-fit plots.
Results: The combination of BDQ and PMD showed additive effects against M. tuberculosis in all three metabolic
phases. The combination of CFZ and PMD showed additive effects with a trend towards synergy in both Acid- and
NRP-metabolic phases while it showed additive effects in Log-metabolic phase. The 3-dimensional diagnostic plots
using the final parameter estimates suggest that the model was able to capture the observed checkerboard data
reasonably well. The presence of mosaic surfaces suggests the existence of less-susceptible bacterial clones, which
cannot be conclusively evaluated with the Greco model on the basis of checkerboard data.
Conclusions: The analyzed combination therapies were identified as promising options to promote M. tuberculosis
bacterial killing in Log-, Acid-, and NRP-metabolic phases. Thus, these drug combinations will be further evaluated
for resistance suppression in an in vitro hollow fiber infection model and in animal models. This study provides an
exemplar for model-informed drug discovery and development towards discovering promising combination
therapies for M. tuberculosis.
References: 1. Global tuberculosis report 2018. https://www.who.int/tb/publications/global_report/en/. 2. Greco, W.
R.; Bravo, G.; Parsons, J. C., The search for synergy: a critical review from a response surface perspective.
Pharmacol Rev 1995, 47 (2), 331-85.
W-058
A Novel Approach for the Parameter Estimation of Dynamically-Constrained, Differential Algebraic
Equation Models of Biophysical Systems
Florencio Serrano Castillo1, Timothy E. Corcoran1,2, Carol A. Bertrand3, Monica E. Shapiro1, Robert S. Parker1
1Department of Chemical and Petroleum Engineering, Swanson School of Engineering, University of Pittsburgh,
Pittsburgh, PA, USA; 2Division of Pulmonary, Allergy, and Critical Care Medicine, School of Medicine, University
of Pittsburgh, Pittsburgh, PA, USA; 3 Department of Pediatrics, School of Medicine, University of Pittsburgh,
Pittsburgh, PA, USA.
Objectives: The advent of systems pharmacology has driven the development of more complex models of disease
systems that aim not only to describe a dataset, but also simulate the biophysics of the system. These models provide
invaluable biological insight that is often infeasible to obtain from either the lab or clinic. However, the same
characteristics that make these models useful often result in numerical issues that significantly slow their
development. Specifically, this project aims to address the largely underexplored issue of discontinuous parameter
hyperspaces, which significantly complicates the accurate estimation of model parameters through traditional
deterministic algorithms, and often requires the use of brute-force exploration.
Methods: We developed an algorithm that uses a combination of an adaptive Latin Hypercube (LHC) sampling
scheme, and an a priori feasibility check to pre-screen potential initial parametric guesses and determine if they are
located within a physiologically plausible region of the parametric hyperspace. The algorithm verifies the numerical
feasibility of these guesses and excludes those that either fail to converge (based on a goodness of fit measure) or
produce biophysically invalid results (based on a feasibility criterion). The algorithm was implemented in Pyomo, an
open source, Python-based, modeling framework designed for the formulation of mathematical models for complex
optimization problems (http://www.pyomo.org/installation/). Pyomo combines the advantages of a full-features
scripting language like Python with the optimization capabilities of traditional algebraic modeling languages and
extends them to include features important for the development of systems pharmacology models, such as the
handling of generalized disjunctive programming and integration of dynamic systems.
Results: We have successfully applied this algorithm to the development of a biophysically inspired, differential-
algebraic model of transepithelial airway electrophysiology. The complex nature of the ion and liquid transport
network across these membranes presents a numerically challenging system to model. The model requires the
inclusion of both differential states to describe ion and water flux dynamics and algebraic states to dynamically
constrain the electrical component of transport within a physiologically meaningful space. Our algorithm facilitated
the estimation of ionic permeabilities for primary human nasal epithelial cell cultures harvested from three subject
groups (healthy controls, patients with Cystic Fibrosis (CF), and carriers of a single CF-causing mutation). Model
predictions accurately described the dataset, appropriately predicted unobserved states within biophysically
appropriate ranges, and recapitulated experimental data (not used for fitting) of the relative changes observed in ion
channel functionality in healthy and CF epithelium.
Conclusions: We developed a novel approach to systematically address parameter space discontinuities and
facilitate the development of highly constrained, biophysically inspired models. This tool can be used as part of a
simple, iterable workflow for parameter estimation within complex discontinuous hyperspaces.
W-059
Bayesian-Koopman Techniques for Optimization of Intervention With Respect to Uncertainty in PuMaS.jl
Christopher Rackauckas1,2, Vijay Ivaturi1
1Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD, USA; 2Massachusetts Institute of Technology, Boston MA
Objectives: The goal of precision medicine is to improve patient outcomes by accurately identifying the benefit-risk
ratio and maximize the probability of therapeutic success. While Bayesian estimation methods have allowed for the
quantification of parameter uncertainties, the ability to subsequently optimize an individual's therapeutic strategy
given this information has seen little innovation. In this work, we develop a method to minimize cost functions on
the probability of therapeutic success using an implementation in PuMaS.jl
Methods: We describe a method which utilizes Koopman operator theory [1] to directly calculate the therapeutic
risk of intervention schemes given the probability distributions from Bayesian posterior distributions. This avoids
parameter sampling to drastically reduce the computational cost versus traditional techniques. The Koopman
operator is derived via its adjoint, the Frobenius-Perron operator, and results in a backpropagation relation on
uncertainty by its action against a chosen cost function. These cost functions are used to define therapeutic success
targets, such as a therapeutic window for drug concentrations or probability thresholds for patient-reported
outcomes. A direction computation of the action then leads to a high-dimensional integral which is solved by h-
cubature and VEGAS importance sampling techniques. To parallelize the calculations, a GPU-based adaptive ODE
solver was developed and a Julia implementation of parallelized high-dimensional quadrature was made GPU-
compatible. Using this uncertainty-based cost, an optimization routine is run to optimize dosage regimens and
intervention parameters to maximize therapeutic success.
Results: This method on a multiple response Pk/Pd model demonstrate allows for a 1-2 order of magnitude
reduction in the number of ODE solves required to compute the action of the uncertainty against the cost function
(Figure 1). The GPU-based stiff and non-stiff ODE solvers successfully simultaneously solve nearly 4000 ODEs on
an NVIDIA Quadro P6000 GPU, and demonstrate the cost reduction given by the high-dimensionality optimized
quadrature schemes.
Conclusions: Further discussions will focus on the scalability of the technique and the computational challenges
needed to be addressed for intervention optimization with respect to uncertainty on QsP models. Our tests were
effective but showed that large parameter QsP models required the high memory capacity of the state-of-the-art
Quadro P6000 GPU. In order to scale these methods on typical user hardware, early results with alternative GPU-
based ODE solvers based on low-memory ROCK schemes for stiff equations will be discussed.
Figure 1: Koopman efficiency. Shown is the convergence of the expected therapeutic risk calculated with respect to
the uncertainty distributions of the parameters obtained by Bayesian posteriors. Blue: Koopman operator method
converges in using less than 100 ODE solver calls. Orange parameter sampling does not converge after 500 ODE
solver calls.
References: [1] Fabbri, Giorgio, Fausto Gozzi, and Andrzej Swiech. "Stochastic optimal control in infinite
dimension." Probability and Stochastic Modelling. Springer (2017).
W-060
Neural-embedded nonlinear mixed effects models (NENLME) in PuMaS.jl
Christopher Rackauckas1,2, Vijay Ivaturi1
1Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD, USA; 2
Massachusetts Institute of Technology, Boston MA
Objectives: The ability for pharmacometric models to predict drug response differences between individuals is
dependent on the ability to discern meaningful relationships between model parameters and covariates. However,
the modern era of “Big Data” has led to a flood of potentially useful data, such as dense time series from wearables
and easily available genomic/transcriptomic assays due to next-generation sequencing techniques. In order to
incorporate high dimensional data sources into pharmacometric models, we demonstrate a technique for embedding
neural networks into nonlinear mixed effects models using PuMaS.jl.
Methods: We define a neural-embedded nonlinear mixed-effects model (denoted NENLME) as a nonlinear mixed-
effects model where some or all nonlinearities are prescribed by an embedded neural network with learnable
parameters. By building a pure-Julia NLME estimation stack, we utilize the differentiable programming capabilities
of the Julia language [1] to perform reverse-mode automatic differentiation (AD) in a way that mixes the neural
backpropagation pass with the gradient calculation of the maximum likelihood on a pure-Julia NLME software
stack. Through this gradient, performing a parameter estimation against time series data using common marginal
log-likelihood schemes (such as LaplaceI and FOCE) allows for the embedded neural network parameters to be
trained simultaneously to the population parameters.
Results: We demonstrate how this method can be used to train NENLME models with Pk/Pd and QsP dynamical
models. We showcase how this technique can allow for a neural network to learn relationships between the data and
dynamical parameters. We show how this could be potentially useful in pharmacometric research as a device for
hypothesis discovery by analyzing the sensitivities of the trained networks to predict possible covariates. We
additionally show how NENLME models can learn partially prescribed dynamics: utilizing an embedded neural
network to learn the dynamical Pd equations with a specified Pk portion and discuss how this can allow for
automated model identification and quantification of model accuracy (Figure 1).
Conclusions: Embedding neural networks into traditional pharmacometric practice allows for the data generality
advantages of machine learning while keeping the explanatory power of mechanistic models. As a tool within a
pharmacometric framework, the predictions of the neural networks directly give rise to hypotheses about disease
mechanisms which can be externally validated. Utilizing prescribed dynamical models with neural networks lets the
modeler directly specify which portions of the model are known and let the unknown portions be learned through
data-driven discovery.
References: [1] Rackauckas, Chris, et al. "DiffEqFlux.jl - A Julia Library for Neural Differential Equations." arXiv
preprint arXiv:1902.02376(2019).
Figure 1: Neural network defined dynamics. Click for animation (wait 3 seconds). Shown is the Julia code for a
differential equation system defined by a neural network (x’=NeuralNetwork(p,x)) which, when trained against the
time series data, learns the underlying dynamical equations.
W-061
Handling Steady-state Dosing in Stiff State-Dependent Delay Differential Equations in PuMaS.jl
David Widmann3, Vijay Ivaturi1, Chris Rackauckas1,2
1Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD, USA;2
Massachusetts Institute of Technology, Boston MA; 3Uppsala University
Objectives: Chemical signals in living organisms are not instantaneous, and these delays can have a profound effect
on the system's resulting behavior. Recently, there has been a trend to model these delays explicitly using delay
differential equations (DDEs). However, no pharmacometric software has been capable of handling all of the ODE
features in the generalized delay context, such as stiffness, steady-state dosing, and state-dependent delays. Here we
demonstrate the ability to handle all of these components together inside of PuMaS.jl.
Methods: Our approach utilizes a metaprogramming-based method of steps implementation to compile a version of
the Julia-based ODE solvers within themselves and handle the resulting implicit system by decoupled stepping. This
approach computationally generates high order integrators for stiff delay differential equations based on Rosenbrock
and ESDIRK tableaus. State-dependent delays are to be handled through the event handling (mtime) framework to
detect implicit propagation of discontinuities. Further, a history purging technique is employed to modify the
interpolation schemes of the embedded ODE solver to accurately account for the repeat interval dynamics of steady-
state calculations.
Results: We demonstrate the integration software on a Pk/Pd problem with steady state dosing and state-dependent
delays (Figure 1). The scalability of the stiff delay differential equation integrators is demonstrated on a stiff quorum
sensing application.
Conclusion: We implemented steady-state dosing in stiff state-dependent DDE problems in PuMaS.jl. This is the
first such implementation in a pharmacometrics software that the authors are aware of.
Figure 1: Steady-state dosing in delay differential equations. Shown is the result of a steady-state dose at every 12
hours on one-compartment model with delayed reaction. (A) Results without history purging. (B) Results with
history purging.
W-062
Optimizing sampling strategy in vaccine efficacy trials for self-limiting infections: Balancing cost and
likelihood of success
Daniel I. S. Rosenbloom1, Jeffrey R. Sachs1, Oliver Bautista1, Anushua Sinha1, Huaping Tang1, Beth Arnold1, Nitin
Mehrotra1
1 Merck & Co., Inc., Kenilworth, NJ, USA
Objectives: Assay-based endpoints are often used to count infection cases in clinical trials of vaccine efficacy. If an
infection is self-limiting, however, frequent specimen assays may be required to ensure case detection, increasing
trial costs. Vaccine efficacy is defined as reduction in infection incidence in the active group (relative to placebo), so
if the placebo rate of disease is small, even a small false positive rate in the assay can substantially dilute estimates
of efficacy and, hence, trial power. Studies requiring frequent assay measurements may be particularly vulnerable to
false positives. Optimizing specimen sampling is therefore essential to balancing trial cost and likelihood of success.
Methods: Simulation of a vaccine efficacy trial involving 1400 participants (700 active, 700 placebo) was used to
calculate effect of different sampling strategies on (1) risk of missed cases causing delay in trial completion; (2)
effect of assay false positives on trial power; and (3) potential cost-saving versus impact on outcomes. Six infection
sampling strategies were simulated: sampling of one or two different kinds of biological specimens at a time
(declaring an infection case if either specimen is positive); sampling every one, two, or three months. An exact
Poisson rate ratio test, requiring that the 95% CI include only positive efficacy, was used to determine trial power
(5000 simulations per scenario). Simulated detection probability decayed exponentially over time, according to rates
suggested by clinical infection assay data.
Results: If all cases are detected with no false positives, the trial takes 2-3 years to accumulate the 24 cases
required, and trial power is 95% (assuming 2% annual placebo incidence and true vaccine efficacy of 85%). Both
trial time and power can change either if the infection resolves quickly to become undetectable or if false positives
occur (see Figure). If the infection is detectable for only 1 month, then infrequent sampling (one specimen every
three months) increases trial time to ~4 years. On the other hand, if frequent sampling is used (two specimens
monthly) and the infection is detectable for 9 months, then even a small false positive rate (1 per ~1800 assays)
dilutes observed efficacy to 57%, yielding 66% trial power. Power loss is mitigated by reducing sampling to once
every three months (92% power) or by reducing false positives 10-fold via assay improvement or confirmatory
measurement (93% power).
Conclusions: Clinical trial simulation can help optimize sampling strategy in vaccine efficacy trials. In trials
involving low-incidence, self-limiting infections, frequent sampling (multiple biological specimens each month) can
ensure detection of nearly all infection cases, shortening trial times. If a frequent sampling strategy is used, however,
even rare false positives may dilute observed trial efficacy and power, leading to unwarranted no-go decisions.
Simulations quantify this trade-off.
W-063
Implementation of Markov Modulated Poisson Process to Model Time-Variant Pharmacokinetic Effects on
Categorical Responses
Po-Wei Chen, Fiona Chandra, Sameer Doshi, Sandeep Dutta, Chih-Wei Lin
Clinical Pharmacology Modeling and Simulation, Amgen, Thousand Oaks, CA, USA
Objective: Modeling categorical responses is critical in clinical Pharmacokinetics (PK) and Pharmacodynamics
(PD) to understand exposure-response relationships and to further support model-informed drug development.
However, modeling categorical responses becomes challenging when the response and predictor variables are time-
variant. One approach to overcome this challenge is utilizing Markovian modulated Poisson process (MMPP) for
continuous-time discrete-state systems (1,2). The MMPP can be implemented to analyze clinical PKPD studies
where clinical responses can be a number of events with different magnitudes over a period of time, such as,
worsening of an adverse event (AE) from one grade to another over the treatment period. In this work, an
implementation of a MMPP to model time-variant PK effects on categorical responses is presented.
Methods: The MMPP approach provides a tool to describe time-variant effect (e.g., PK) on the rates of changing
from one state to another (e.g., from no-event to event, or grade 2 AE to Grade 3 AE) through the Markovian
process. The rates of change were described by a non-homogenous Poisson process, which utilize a thinning
process to reshape the rates to accommodate time-variant effects. An illustration of the method is described in
Figure 1. The changes in different states for the categorical response was described by a Markov jump process with
a forward filtering-backward sampling algorithm. Given the complexity of the likelihood function, statistical
inference was via a Markov chain Monte Carlo based method with an auxiliary variable Gibbs sampler.
Results: The MMPP was implemented in R to describe the PK effect on a hypothetical categorical response with 3
levels. The PK was assumed to have first-order absorption and elimination and had an effect on the rate of
worsening from low level response(s) to high level response(s) following an Emax-relationship over the treatment
period. MMPP was able to describe the time-variant effect on the categorical response over time. In this
simulation-based analysis, the inference was through a Bayesian MCMC approach using a Gibbs sampler. The
simulation-based analysis demonstrated the MMPP provides an appropriate approach to describe time-variant
categorical response.
Conclusions: This work provides a MMPP workflow containing model construct, estimation and visualization, in
which model estimation is based on a Bayesian MCMC approach to adopt prior knowledge of inter-subject
variability. The major purpose/expectation is to bridge the well-applied MMPP approach to the PKPD environment,
and to provide a platform for easier access to the method.
References: 1. V Rao, YW Teh. MCMC for continuous-time discrete-state systems. The Journal of Machine
Learning Research 2012 2. V Rao, YW Teh. Fast MCMC Sampling for Markov Jump Processes and Extensions.
The Journal of Machine Learning Research 2013
W-064
Extended Top-down Approach of Physiologically-Based Pharmacokinetics (PBPK) Modeling: An Application
for Effects of Obesity on Clinical Pharmacokinetics
Toshimichi Nakamura (1, 2), Emilie AG Molins (2), Melissa Natavio (3), Frank Z Stanczyk (3), William J Jusko (2)
1. DMPK Research Department, Teijin Pharma Limited, Hino, Tokyo, Japan 2. Department of Pharmaceutical
Sciences, State University of New York at Buffalo, Buffalo NY 3. Department of Obstetrics and Gynecology, Keck
School of Medicine, University of Southern California, Los Angeles, CA
Objectives: As a top-down approach, Population PK (“PopPK”) analysis is a useful tool, but this approach often
needs improved physiological components as covariates in diverse physiological state such as obesity. Suitable data
sets to build adequate models are needed to predict clinical PK by PopPK. In contrast, Physiologically Based PK
(PBPK) models can be assembled by extending average subject data from a limited population (“Conventional
PBPK”). As a bottom-up approach, Conventional PBPK is used to simulate clinical PK with a virtually generated
population adjusted with physiological components. The inter-individual variability (IIV) can be provided as
physiological variabilities of virtual subjects generated by an in silico method. Therefore, predicting variance of
clinical PK using Conventional PBPK has limitations. An extended top down PBPK modeling approach (“Extended
PBPK”) for clinical PK prediction was developed. Our new approach involves: [1] using individual subject data to
estimate parameters, [2] defining IIV and residual variability (RV) from observed population data, and [3] building
in diverse physiologic changes produced by obesity in the structural model. The present study demonstrates the
usefulness of an extended PBPK modeling approach for a clinical study that includes obese subjects with
comparisons of conventional PBPK and PopPK methods.
Methods: Individual plasma concentration-time profiles of 26 subjects including normal, obese, and extremely
obese women who received 1.5 mg oral levonorgestrel (LNG) were utilized (1). Relationships for organ/tissue
weights, composition, and blood flows in relation to these factors for adult humans were found and/or adapted from
literature sources. Specific PBPK body components were then generated from the physiologic metrics for individual
subjects. Tissue/plasma partition coefficients (Kp) were calculated using published methods. All analyses and
parameter estimates were conducted using Phoenix NLME.
Results: The plasma concentration-time profiles of LNG showed considerable IIV. Without a covariate of BMI,
PopPK analysis could not describe the large IIV. Parameter estimates of extended PBPK model were fitted using
individual subject data and then typical values and random effects were obtained using nonlinear mixed effect
(NLME) models. Three of four IIVs did not require BMI as a covariate, but the IIV of absorption rate (ka) that had
no physiological basis was slightly affected. When parameters only from normal subjects were used for simulation
of a virtual population, the effect of obesity on PK was well reproduced. The extended PBPK model well described
the general PK variance of LNG without any covariates such as BMI despite including extremely obese subjects
(Figure).
Conclusions: Unlike conventional PBPK using average subject physiologic data, the extended PBPK approach for
individual subjects includes realistic parameters for disturbed physiology. As top-down approach, extended PBPK
model may be more informative and extensive approach than PopPK.
References: Natavio M. et al., Contraception 2019 (in press)
Figure
Virtual clinical PK prediction using parameters from normal subjects.
W-065
Identification of Potential Formation Causes of Adverse Effects Caused by Brand Swapping Using PBPK/PD
Modeling
Authors: Fudan Zheng, Tonglei Li
Institutions: Department of Industrial and Physical Pharmacy, Purdue University, 575 Stadium Mall Drive, West
Lafayette, Indiana, 47907, United States
Objectives: Warfarin sodium tablet is a commonly prescribed anticoagulant with narrow therapeutic index (NTI).
Substantial concerns over its efficacy and safety were raised regarding the interchangeability between brand-name
and generic warfarin products.1 A physiologically-based pharmacokinetic/pharmacodynamic (PBPK/PD) modelling
strategy was applied herein to delineate the impacts of formulation variations among warfarin products on the oral
absorption process and subsequently therapeutic responses.
Methods: Advanced compartmental and transit (ACAT) model in account for the phase transition of warfarin salt
was integrated with a compartmental PK model and a population-based pharmacodynamic model (popPD) of
warfarin sodium tablets.2
Results: The oral absorption of a single 10-mg dose was completed in GI tract within seven hours, of which jejunum
contributed 50%. The stomach pH in fed state population and bile salt content in small intestine among fasted and
fed state groups affected the oral absorption fraction. The predicted AUC, Cmax and INR of warfarin sodium tablet
are within 15% deviations of the clinically measured data. The high accuracy of the predicted PK and PD data
allows the use of our model to assess the role of single formulation variable played on the in vivo performance of
warfarin sodium tablet. The fraction of (S)-warfarin and initial particle size are identified as two critical attributes
that lead to significant fluctuation of the steady-state INR among various patient groups. Patients are more
susceptible to thrombotic or hemorrhagic event when (S)-warfarin fraction is out of the range of 40%-80%, which is
not affected much by the genotypes, age and food effects. Hemorrhagic risk increases largely when the mean initial
particle size is below 140 µm. Preprandially administered group is more susceptible to warfarin dose than
postprandially administered group. The diffusion coefficient and solubility of warfarin sodium salt both influence
the dissolution rate and are predicted to fluctuate the INR moderately.
Conclusions: The chiral ratio and mean initial particle size were identified as two critical attributes that may be the
culprits when switching among warfarin brand products. Given the fact that the chiral ratio is most likely even
between R and S form of warfarin, difference in particle size among warfarin tablet products may be the main cause.
References: 1. Ghate SR, Biskupiak JE, Ye X, Hagan M, Kwong WJ, Fox ES, Brixner DI 2011. Hemorrhagic and
thrombotic events associated with generic substitution of warfarin in patients with atrial fibrillation: a retrospective
analysis. The Annals of pharmacotherapy 45(6):701-712. 2. Sjogren E, Westergren J, Grant I, Hanisch G, Lindfors
L, Lennernas H, Abrahamsson B, Tannergren C 2013. In silico predictions of gastrointestinal drug absorption in
pharmaceutical product development: application of the mechanistic absorption model GI-Sim. European journal of
pharmaceutical sciences : official journal of the European Federation for Pharmaceutical Sciences 49(4):679-698.
W-066
PBPK modeling for therapeutic nanoparticles loaded with drug: distribution and release
Evgeny Metelkin1, Oleg Demin1
1InSysBio LLC, Moscow, Russia
Objectives: Nanoparticles are promising vehicles for drug delivery which formulations differ by release kinetics of
the encapsulated drug. We have proposed an approach to develop mathematical model describing drug release from
nanoparticles and distribution for both. Model of nanoparticles distribution in tumor bearing mice with loaded
vincristine (VCR) was developed to illustrate the approach.
Methods: The model of nanoparticle distribution was based on the diffusion-limited model of the distribution of
macromolecules. The release of the drug from the nanoparticles was described as reported in [1]. The model of drug
distribution was based on the standard distribution model for small molecules. Parameters of the model
nanoparticles loaded with VCR were taken from literature or fitted against in vivo preclinical data [1, 2].
Results: Mathematical model of VCR distribution calibrated on the data for inactive nanoparticles and conventional
VCR PK was able to correctly simulate the distribution of the model therapeutic formulation. The results in this
abstract have been previously presented at [ACSPT, Orlando, March 21-24, 2018] and published as paper [1].
Conclusion: The approach can be applied for mathematical modeling of therapeutic nanoparticles with different
release profiles.
References: [1] Shalgunov et al, J Contr Release 261 (2017) 31–42 [2] Fredenberg et al, Int J Pharm, 415 (2011)
34–52
W-067
A Multi-Scale PBPK-Toxicodynamic Model for Cisplatin-induced Nephrotoxicity and Interaction with
Cimetidine, a Reno-protective Agent Inhibitor of OCT2 Transporter
Hardik Mody1, Tanaya R. Vaidya1, Sreenath Nair2, Adrian Rump1, Timothy Garrett3, Lawrence Lesko1, Sihem Ait-
Oudhia1
1Center for Pharmacometrics and Systems Pharmacology, Department of Pharmaceutics, University of Florida,
Orlando, FL, USA 2Pharmaceutical Sciences Department, St. Jude Children’s Research Hospital, Memphis, TN,
USA 3Southeast Center for Integrated Metabolomics, University of Florida, Gainesville, FL, USA
Objectives: To develop a physiologically based pharmacokinetic (PBPK) and toxicodynamic (TD) model for
cisplatin-induced nephrotoxicity (CINT), and to integrate the reno-protective action of cimetidine, an Organic
Cationic Transporter 2 (OCT2) inhibitor, on CINT.
Methods: The overall approach for model development and qualification is summarized in Fig 1A. First, cisplatin
physicochemical properties, mice and human biological properties, and OCT2 kinetics were incorporated into a
whole-body PBPK model for cisplatin using GastroPlus® and SimCYP® software platforms. The time-course of
plasma and kidney total cisplatin concentrations in mice and humans after intravenous administration of cisplatin
were extracted from literature [1, 2], and used to develop the model. Inclusion of cimetidine interaction with
cisplatin was carried out using a published PBPK model for cimetidine [3]. Second, CINT was assessed
experimentally on human kidney proximal tubular epithelial cells (SA7K), and computationally with TD models for
cisplatin/cimetidine actions on SA7K. The observed TD data comprised: 1) SA7K viability, and 2) novel early
biomarkers for CINT identified by LCMS-MS using untargeted global pharmaco-metabolomics approach. The
developed TD models included a combination of simple Hill functions for SA7K responses to cisplatin/cimetidine
exposures, and transit compartments with precursor pool indirect response models for metabolite-based biomarkers.
All mathematical modeling was performed using Monolix® software and parameters were estimated as
mean±%RSE.
Results: Total cisplatin plasma concentration time profiles in mice and humans were characterized well with the
developed whole-PBPK models (Fig 1B). Predicted area under cisplatin plasma concentration curves (AUCCis) were
within 1.5-fold of observed data. Additionally, AUCCis was unchanged after co-administration with cimetidine (800
mg), indicating that cimetidine did not alter cisplatin disposition as reported in published clinical studies. The
developed TD models (Fig 1C-D, top) described well the CINT responses as well as reno-protective effects of
cimetidine (Fig. 1C-D, bottom). The first-order death rate constant for SA7K cells was estimated at 0.00523±6.5% h-
1. The maximal effect (Kmax) of cisplatin stimulation of SA7K cell death was determined at 0.082±10% h-1, while the
cisplatin concentration inducing 50% (KC50) of maximal effect was estimated at 82.3±16.9% µM. The concentration
of cimetidine leading to 50% of Kmax was determined at 608.5 µM. Seven novel potential metabolite-based early
biomarkers for CINT were identified, of which TD models for glutathione and glutathione disulfide (Fig.1E-F, top)
captured well the time-course of observed data (Fig.1E-F, bottom).
Conclusions: The final whole-body PBPK-TD model for cisplatin and interaction with cimetidine described well all
observed data. The CINT responses were captured well with the developed TD models. Model-based simulations for
optimization of clinical dosing regimens to maximize cisplatin action and minimize CINT are ongoing.
References: 1] Himmelstein KJ, et al. Clin Pharmacol Ther 1981; 29: 658-664 2] Zamboni WC, et al. Clin Cancer
Res 2002; 8: 2992-2999 3] Burt HJ, et al. Eur J Pharm Sci 2016; 88:70-82
W-068
Mechanistic whole-body physiologically based modeling to predict the clinical pharmacokinetics of
monoclonal antibodies after intravenous and subcutaneous administration
Shihao Hu and David Z. D’Argenio
Department of Biomedical Engineering, University of Southern California, Los Angeles, CA, USA
Objectives: Monoclonal antibodies (mAbs) are a primary class of treatment for chronic conditions, including
cancers, autoimmune and inflammatory diseases. Since 2008, 64% of the mAbs approved by U.S. Food and Drug
Administration have involved intravenous infusion (IV), with 36% for subcutaneous injection (SC). However, the
multifactor complexities of the SC absorption process together with the characteristics of mAbs have presented a
challenge to the development of models for SC delivery of mAbs. To better predict clinical PK of mAbs following
SC administration, we modified and extended an established mechanistic physiologically-based PK (PBPK) model
for IV mAbs [1].
Methods: The previously reported PBPK model for IV mAbs included different organs, in which they described the
interaction between neonatal Fc receptors (FcRn) and mAbs, target turnover and target-mediated drug disposition
[1]. We adapted this model according to available physiological information: First, we included a separate peripheral
sampling site with contributions from surrounding skin and muscle, based on the reported deviation of peripheral
sampling plasma concentration from central venous plasma concentration [2]. We also optimized the production rate
of FcRn. In addition, we updated some physiological parameters, including venous plasma volume and lymph node
transit time. The SC absorption model was added to the base PBPK model, assuming the major absorption pathway
was from interstitial space to lymphatics system [3]. Model parameters were fixed based on physiological
information in published literature as done in [1], except two drug-specific coefficients controlling extravasation
process (pinocytosis and convection) of mAbs and SC absorption model parameters. Plasma concentration data of
16 mAbs in humans were digitized from 18 published studies. The modified PBPK model was fit (ADAPT Version
5) to the pooled multiple-dose data from each of the IV mAbs to estimate the two parameters affecting mAb
extravasation. Based on the resulting IV model, absorption models were investigated using the corresponding SC
data.
Results: The base PBPK model adequately captured the pharmacokinetics of 16 mAbs in humans after IV delivery
at multiple doses. Two coefficients estimated from IV data explained the inter-antibody variability reliably (standard
error less than 13%). Coefficients controlling pinocytosis and convection range from 0.05 to 0.20 and 3.78 to 13.20,
respectively, except one mAb. The fittings of extended PBPK model to 7 mAb SC datasets identified the entry of
mAb into lymphatics system as the rate-limiting step, which is consistent with some previous experimental studies.
Conclusions: Our study presents a mechanistic, whole-body PBPK model for predicting clinical PK of mAbs after
IV or SC administration.
References: [1] Patrick Glassman et al, J Pharmacokinet Pharmacodyn (2016) 43:427–446. [2] Helen Musther et al,
The AAPS Journal (2015) 17: 5. [3] Margarida Viola et al, Journal of Controlled Release (2018) 286: 301-314.
W-069
Evaluation of Drug-Drug Interaction Potential of Probuphine® Implants using a Physiologically-Based
Pharmacokinetic Model
Authors: Biju Benjamin1, Sunil Sreedharan2, Jogarao Gobburu1
Institutions: 1Center for Translational Medicine, University of Maryland, Baltimore, MD USA; 2Titan
Pharmaceuticals, South San Francisco, CA, USA.
Objectives: Probuphine® subdermal implants, each containing 80 mg buprenorphine hydrochloride (equivalent to
74.2 mg of buprenorphine free base) is approved in the United States of America and in Canada for the maintenance
treatment of opioid dependence in patients who have achieved and sustained prolonged clinical stability on low to
moderate doses of a trans-mucosal buprenorphine-containing product. The goal of this research was to develop a
physiologically-based pharmacokinetic (PBPK) model for buprenorphine released from Probuphine implants in
order to predict potential drug-drug interactions (DDI) on buprenorphine exposure from the implants.
Methods: The rich knowledge base on the absorption, distribution, metabolism and excretion of buprenorphine
were drawn on in order to build a reliable PBPK model. Phase-1 metabolism of buprenorphine is largely mediated
by cytochrome P450 (CYP450) enzymes with CYP3A4 as the major- and CYP2C8 as a minor-oxidation pathway,
respectively. The absorption model for Probuphine was derived by deconvolution of the PK profile relative to the
published intravenously-delivered (IV) buprenorphine profile. The model was qualified using PK data from
Probuphine phase 3 clinical studies. PBPK models for the strong CYP3A4 inhibitor, ketoconazole, and CYP3A4
inducer, rifampicin, were also developed. The final PBPK model was employed to predict DDI of buprenorphine at
the clinical dose of 4 Probuphine implants (80 mg buprenorphine hydrochloride per 6-month implant) with repeat
doses of ketoconazole (200 mg BID), or rifampicin (600mg QD).
Results: Ketoconazole has a negligible effect (about 10%) on buprenorphine exposure after treatment with
Probuphine implants. Rifampicin reduced the buprenorphine exposure from Probuphine implants by about 25-30%.
The estimated fraction of buprenorphine metabolized via the uridine glucuronyl transferases (UGT) pathway has
been reported in the range of 8%-37%. The in-vivo clearances for the CYP3A4 and UGT pathways are 22.6
mL/min/kg and 13.6 mL/min/kg, respectively. A published Km of 22.6 uM for UGT1A1, the major contributor of
buprenorphine metabolism in the UGT pathway, translates into a very high in-vivo buprenophine concentration
(~10,000 ng/mL) relative to the average sustained plasma buprenorphine concentration of 0.82 ng/mL after
treatment with 4 Probuphine implants; making any DDI with the UGT pathway unlikely.
Conclusion: A reliable PBPK model for buprenorphine exposure following treatment with Probuphine implants has
been developed. This model was employed to predict DDI on buprenorphine exposure following treatment with
Probuphine. Strong CYP3A4 inhibition negligibly increases the buprenorphine exposure from Probuphine; while
strong CYP3A4 induction is predicted to decrease the exposure by 25-30%. The role of the UGT metabolic pathway
on buprenorphine released from Probuphine was assessed to be negligible.
W-070
Physiologically-based Pharmacokinetic (PBPK) Modeling Workflow in PuMaS
Praneeth Jarugula1, Vaibhav Dixit1,3, Christopher Rackauckas1,2, Joga Gobburu1 and Vijay Ivaturi1
1Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD, USA; 2Massachusetts Institute of Technology, Boston MA; 3Department of Mathematical Sciences, Indian Institute of
Technology (B.H.U.), Varanasi, Uttar Pradesh 221005 India
Objectives: Physiologically-based pharmacokinetic (PBPK) modeling is a valuable and widely-used tool in drug
development with several available software platforms. The flexibility to model an organ or cellular system by
modifying differential equations is a coveted feature for PBPK software. PuMaS is a new Julia-based platform
which can be used for PBPK model development with flexible control over model equations. Our objective was to
demonstrate the complete workflow of a PBPK model in PuMaS.
Methods: The workflow of implementing a PBPK model in PuMaS was demonstrated using a previously published
adult and pediatric voriconazole PBPK model by Zane and Thakker1 which was also previously implemented in
mrgsolve2. The general approach to develop PBPK models suggested by the FDA was followed.3 The modeling was
performed in three steps: model development, model qualification and application. In each of these steps, we
demonstrate the enabling features provided in PuMaS that ensure efficiency and reproducibility.
Results: The results of Zane and Thakker as well as the results of further optimization done by Elmokadem et al in
mrgsolve were replicated in PuMaS. The two most influential parameters highlighted in Elmokadem’s sensitivity
analysis, muscle:plasma partition coefficient (KpMu) and the blood:plasma concentration ratio (BP), showed similar
sensitivity after analysis done in PuMaS (Figure 1). Further, we showcase features that allow parameter optimization
via maximum likelihood or Bayesian methods and tools for global and local sensitivity analysis (SA) that together
provide a powerful toolkit for PBPK modeling. The global SA tools available in PuMaS include Sobol, Morris,
GLM Regression Based SA, Derivative Based SA, Owen SA and Gaussian Process Regression SA.
Conclusion: The previously published voriconazole PBPK model was successfully implemented in PuMaS. While
PuMaS is able to implement models, it is currently lacking the parameter databases present in other commercial
software. Future work will compare the performance of these different SA tools.
Reference: 1. Zane NR, Thakker DR. A Physiologically Based Pharmacokinetic Model for Voriconazole
Disposition Predicts Intestinal First-pass Metabolism in Children. Clin Pharmacokinet. 2014; 53: 1171-82. 2.
Elmokadem A, Gastonguay MR, Baron KT, et al. A Physiologically-based Pharmacokinetic Model for Voriconazole
Explores Differences in Pharmacokinetics between Adults and Children. J Pharmacokinet Pharmacodyn. 45: S3-
S134. 3. Food and Drug Administration. Guidance for Industry: Physiologically Based Pharmacokinetic Analyses –
Format and Content (2018). 4. Poulin P, Theil FP. Prediction of Pharmacokinetics Prior to In Vivo Studies. 1.
Mechanism-Based Prediction of Volume of Distribution. J Pharm Sci. 2002; 91: 129-15
Figure 1 Sensitivity analysis plots of muscle:plasma partition coefficient (Kpmu) and blood:plasma
concentration ratio (BP). The values used to analyze sensitivity were based on calculation using the Poulin and
Theil method4 and the remaining two values were half and twice that calculated value, as was done in Elmokadem et
al.
W-071
Using PBPK Modeling to Predict Maternal Exposure to ARV Drugs in Pregnant Women Living With HIV
Authors: Mary Gockenbach1, Manuela Grimstein2, Jeremiah Momper3, Mark Mirochnick4, Edmund Capparelli3,
Kimberly Struble5, Jian Wang6, Yaning Wang2, Tamara Johnson1, Hari Cheryl Sachs1, for the Antiretrovirals in
Pregnancy PBPK Modeling Group
Institutions: 1Division of Pediatric and Maternal Health, Office of New Drugs, CDER, FDA, Silver Spring, MD, 2Divison of Pharmacometrics, Office of Clinical Pharmacology, CDER, FDA, Silver Spring, MD, 3University of
California, San Diego, CA, 4Boston Medical Center, Boston, MA, 5Division of Antiviral Products, Office of
Antimicrobial Products, CDER, FDA, Silver Spring, MD, 6Office of New Drugs, CDER, FDA, Silver Spring, MD
Objectives: The pharmacokinetics (PK) of antiretroviral (ARV) drugs in pregnant women living with HIV (PWLH)
are affected by both physiological changes of pregnancy, and the HIV infection itself. Physiologically-based
pharmacokinetic (PBPK) modeling provides an approach for predicting maternal exposure while accounting for both
conditions. A PBPK compound model of rilpivirine (RPV) was developed for the existing nonpregnant and pregnant
populations in Simcyp®. Then, laboratory data of PWLH from the IMPAACT Network Protocol P1026s were used
to modify the Simcyp® pregnancy population model to reflect the population of interest.
Methods: We developed a PBPK compound model for RPV in Simcyp® V17 using physicochemical, in vitro, and
clinical parameters, and verified it with clinical PK data in the nonpregnant population. Next, the RPV PBPK model
was applied to the Simcyp pregnancy model. Simulated PK profiles were compared with observed data from three
clinical trials of PWLH. We then modified the pregnancy model with time dependent functions describing
hematocrit (HCT), albumin (ALB), and serum creatinine (SCR) values collected from 500 women enrolled in
P1026s. PK predictions using the updated pregnancy model were compared against original results obtained using
the existing pregnancy model.
Results: The PBPK model for the nonpregnant population predicted RPV PK within a two-fold range of mean
observed values in nonpregnant adults. In the pregnant population, the predicted values were within ± 50% of the
mean (n=30 in second trimester, n=57 in third trimester) observed values for Cmax, Tmax, AUC, and Cmin. HCT, ALB,
and SCR in PWLH were significantly lower than the Simcyp reference population during the second and third
trimesters. Second degree polynomial equations were generated to describe change of HCT, ALB, and SCR through
pregnancy. Updating the pregnancy model with these functions resulted in a predicted decrease in RPV exposure of
11–16% for Cmax, 12–18% for AUC, and 13–18% for Cmin, reducing the overprediction of exposure obtained with
the original pregnancy model. The results for RPV predictions using the Simcyp® pregnancy population were
presented at the International Workshop on Clinical Pharmacology of HIV, Hepatitis and Other Antiviral Drugs,
Noordwijk, Netherlands, on 5/16/19.
Conclusions: Our PBPK model for RPV captured the effects of pregnancy on maternal exposure from 20–38 weeks
gestation. The update model accounted for progressive changes in pregnancy including CYP3A4 enzyme activity
and glomerular filtration rate, and differences in HCT, ALB, and SCR between PWLH and pregnant women without
HIV. Incorporating these differences improved prediction of PK parameters for RPV, which may help to inform
dosing for PWLH.
References: Abduljalil, K., Furness, P., Johnson, T. N., et al. (2012). Anatomical, Physiological and Metabolic
Changes with Gestational Age During Normal Pregnancy. Clinical Pharmacokinetics, 51(6), 365-396.
US FDA NDA 202022: Clinical Pharmacology Review, 5/20/2011.
https://www.accessdata.fda.gov/drugsatfda_docs/nda/2011/202022Orig1s000ClinPharmR.pdf
W-072
In vitro to in vivo extrapolation of CD4 T cell polarization in lymph nodes using QSP Modeling
Oleg Demin1, Svetlana Rubina1, Dmitry Shchelokov1, Oleg Demin Jr1
1InSysBio, Moscow, Russia
Objectives: There are a lot of data describing in vitro CD4 T cells polarization available in literature. Either
cytokine positive cell counts, or cytokine concentrations are typically measured under specific polarization
conditions (Th1, Th2, Th17 etc) provided by administration of anti-CD3/anti-CD28 stimuli and specific cytokines
shifting polarization to the particular direction. In vivo conditions in lymph nodes (LN) are different from those of in
vitro T cell polarization experiments. The difference includes (i) simultaneous presence of multiple cytokines in LN
vs few selected cytokines in in vitro experiments, (ii) interaction of naïve and polarized T cells with stromal cells
and extracellular matrix infrastructure of LN, (iii) interaction of naïve T cells with antigen presenting cell in LN vs
stimulation of T cell response by anti-CD3/anti-CD28 antibodies in vitro, (iv) influx/efflux of T cells with lymph
flow. We aimed to develop mathematical model of T cell polarization in vitro and translate it to the in vivo LN. The
model was applied to predict response of T cell polarization in LN to anti-TNFa, anti-IL6R, anti-TGFb, anti-IL4Ra,
anti-IL12p40 antibodies and pomalidomide and to explore possible differences in responses between in vitro and in
vivo conditions.
Methods: Sub-models describing proliferation, differentiation, death and cytokine production of T cells were taken
from Immune Response Template [IRT2.1.0, https://irt.insysbio.com], upgraded to take into account cytokine
regulatory effects not included in the version of the IRT and assembled to form T cell polarization model describing
in vitro experiments. Parameters values describing surface molecules mediated processes in immunological synapse,
T cell apoptosis and inactivation were taken from IRT. Values of remaining parameters were either taken from
literature or fitted against literature data from multiple sources.
To translate model from in vitro to in vivo in LN we (1) took into account volume of LN, (2) added cytokine
degradation rates, influx of Th0 and effluxes of T cells with lymph flow, (3) fitted rate constants of cytokine
degradation to describe literature data on cytokine concentrations measured in vivo in plasma of healthy subjects
(HS).
Results: The model based on in vitro data and fitted against in vivo cytokine concentrations allows to satisfactory
predict cell counts of CD4 T cells measured in LN of HS (Fig. 1). Simulation of different therapies allows to predict
changes in T cell counts in LN. For example, Dupilumab decreases Th2 count by about 30% and increases Th1
count due to IL4 decrease, Ustekinumab decreases Th1 by about 4 folds and increases Th2 and Th9 by 2-3 times due
to IL12 decrease.
Conclusions: Proposed method of in vitro to in vivo extrapolation demonstrates satisfactory quality of prediction of
cell counts in LN and can be applied for evaluation of novel therapies.
W-073
An integrative quantitative systems pharmacology model of asthma inflammation and constricted respiratory
airflow
Authors: Justin Feigelman, Siddharth Sukumaran, Fang Cai, Rebecca Bauer, Tracy Staton, Cynthia Stokes, Heleen
Scheerens, Saroja Ramanujan Kapil Gadkar
Institutions: Genentech, Inc., South San Francisco, CA
Objectives: Using a quantitative systems pharmacology approach, we aim to provide a broad mathematical
framework capturing the impact of type 2 biology on airflow during acute or chronic inflammation in asthma.
Methods: We extended our previous QSP platform model of type 2 biology by adding in a mechanical model
module which captures acute bronchoconstriction dynamics. The included model is based on the model from
Lambert and Wilson [1] and captures several of the relevant forces in the airways including muscle contraction due
to spasmogenic regulation, elastic deformation of the airway wall, and compressive forces from the surrounding
parenchymal tissue. The model is used to obtain functional output of FEV1 with modulation of four key parameters
(namely, maximal airway smooth muscle force generation, airway wall thickness, total lung capacity and vital
capacity) that represent inter-subject variability, and which are impacted by various therapeutic interventions.
Results: The integrated QSP/mechanical model captures essential type 2 biology including Th2, dendritic cells,
mast cells, basophils, eosinophils and progenitors, epithelium, and airway smooth muscle cells. It includes
associated essential mediators such as IL-4, IL-13, IL-5, and chemokines such as CCL13, CCL17, and eotaxin. The
model also contains bronchoconstrictors including histamine and leukotriene. The mechanical model module
extends the QSP model by providing accurate predictions for FEV1 and other spirometric features as a function of
airway bronchoconstrictor concentrations. The integrated model has been used to represent FEV1 changes in
different patient populations with different classes of interventions including LABA, LAMA, corticosteroids,
Dupilumab and Tezepelumab. Wherever available, model predictions have been verified against clinical data.
Various challenge studies such as methacholine, allergen and mannitol challenges, which are commonly performed
tests to inform airway sensitivity or allergic status, are also included in this integrated model.
Conclusions: The integrated model provides a mechanistic link between underlying pathophysiology and impact on
airflow. Further, it is a robust quantitative tool to evaluate interventions for changes in the underlying
pathophysiology and subsequently airflow improvements (measured as FEV1). Inclusion of challenge studies
utilized for evaluating new therapeutics in this framework aids clinical study design and analyses.
References: Lambert, R. K., & Wilson, T. A. (2005). Smooth muscle dynamics and maximal expiratory flow in
asthma. Journal of Applied Physiology, 99(5), 1885–1890. http://doi.org/10.1152/japplphysiol.00450.2005
W-074
Understanding multilineage hematopoietic toxicity using a QSP modeling approach to bridge in vitro and
clinical data of anti-cancer agents
Jennifer L. Wilson1, Dan Lu1, Nick Corr2, Aaron Fullerton2, James Lu1
1Department of Clinical Pharmacology, Genentech, Inc., South San Francisco, CA, USA; 2Department of Safety
Assessment, Genentech, Inc., South San Francisco, CA, USA
Objectives: Cytopenias are a source of late-stage drug attrition for anti-cancer therapies and the persistence of these
toxicities necessitates further early-stage assessment tools1. For single cytopenias (e.g. neutropenia) it is possible to
model and select dose schedules to optimize the neutrophil nadir2; however, these decisions require sufficient
clinical data from late-stage development. Additionally, anti-cancer therapies differentially affect multiple
hematopoietic cell lineages (including neutrophils, platelets and red blood cells) due to the cell populations’ varying
sensitivities. Species-specific kinetics and half-lives, and the lack of sufficient patient data complicate the
optimization of dosing parameters for multilineage toxicities early in development. Our objective was to reduce drug
attrition by predicting multilineage cytopenias without late-stage data and hypothesize dosing schedules optimized
for these toxicities.
Methods: We used an in vitro QSP model describing drug-induced multilineage effects3 and in vitro multilineage
assay data to estimate Emax and EC50 parameters that quantified drug effects on multi-species proliferation and cell
killing and used a consensus, clustering technique to group compounds based on these mechanistic, kinetic effects.
We applied these learned parameters to an in vivo QSP model4 and related drug effects on multilineage toxicities to
clinically observed rates of thrombocytopenia, anemia, and neutropenia. We used this approach to assess dose
schedules optimal for multilineage toxicities.
Results: Using the in vitro QSP model, we described the mechanistic toxicity of ~50 drugs in terms of anti-
proliferative and cell killing Emax and EC50 effects on 12 cell species. These therapies included control compounds
and multi-class, anti-cancer compounds (e.g. CDK, BCL, and HDAC inhibitors) without/with known hematopoietic
toxicities. We demonstrated that within-class compounds exhibit varying effects on cell proliferation and cell
killing. Taking CDK inhibitors as a case study, palbociclib demonstrated greater anti-proliferative effects compared
to dinaciclib, which had stronger cell killing effects. We demonstrated the ability to bridge in vitro to in vivo
myelosuppression profiles; the antiproliferative effect of palbociclib was consistent with clinically reported
cytopenias5. For select compounds, we discovered optimal schedules based on multi-lineage toxicities.
Conclusions: We deepened understanding of multilineage toxicity mechanisms in terms of quantitative anti-
proliferative and cell-killing effects using a combined in vitro assay and QSP model for multiple classes of anti-
cancer therapies. We bridged in vitro and clinical data and used modeling to select dosing schedules that optimized
multilineage toxic effects. Our results suggest that it may be possible to predict hematopoietic toxicities in early-
stage development, and this approach is relevant to complex development paradigms, including combination
therapies.
1Kurtin S. (2012). J Adv Pract Oncol, 3(4):209-24. 2Sun W et al. (2017). The Journal of Clinical Pharmacology, 57(9) 1159–1173. 3Lu J et al (2019) ASCPT Annual Meeting. 4Lu J et al (2019). PAGE Meeting. 5Hu W et al (2016).
Clinical Cancer Research, 22(8), 2000-2008.
W-075
Assessing the Risk of Arrhythmias in Healthy and Failing Cardiomyocytes
Authors: Anna Sher1, Anna Kirpichnikova2, Anna James-Bott3, Michael Clerx3, CJ Musante1, David Gavaghan3
Affiliations: 1 Pfizer, Cambridge, USA 2 University of Stirling, UK 3 University of Oxford, UK
Objectives: Chronic heart failure (HF) is a cardiovascular disease with a leading and growing hospitalization
burden, affecting at least 26 million people worldwide [1]. HF is associated with a higher risk of onset of ventricular
arrhythmias. Markers based on the traditional S1-S2 [2] and dynamic [3] restitution protocols including the slope
of >1 of the restitution curves and the point where 1:1 behavior ends have been found to give false negative results
and to be unreliable in predicting the onset of arrhythmic behavior [4]. In this work we compare different restitution
protocols in silico and evaluate novel restitution protocols for predicting arrhythmogenic behavior in failing
cardiomyocytes.
Methods: We use existing mathematical models of human ventricular myocytes such as Ten Tuscher 2006 [5] and
Ohara-Rudy 2011 [6] modified to represent HF conditions to demonstrate novel restitution protocols for the
prediction of onset of arrhythmias in failing cardiomyocytes. Simulations are performed using Matlab as well as
Myokit [7] software.
Results: Novel restitution protocols identify the onset of the multistability region (where 1:1 behavior coexists with
non-1:1 responses). Depending on the width of the multistability region that includes 1:1 behavior, we show the
onset of arrhythmias to be detected ~30ms and ~50ms earlier in healthy and failing cardiomyocytes respectively.
Conclusions: This work demonstrates that while the underlying pathophysiology and mechanisms of HF are still an
area of active research and debate, we can evaluate the onset of arrhythmias in HF cardiomyopathies by using novel
restitution protocols.
References: 1. P Ponikowski, SD Anker, KF AlHabib, MR Cowie, TL Force, S Hu, T Jaarsma, H Krum, V Rastogi,
LE Rohde, UC Samal, H Shimokawa, B Budi Siswanto, K Silwa, G Flippatos. Heart failure: preventing disease and
death worldwide. ESC Heart Fail. 1(1): 4-25. 2014. 2. BG Bass. Restitution of the action potential in cat papillary
muscle. Am J Physiol. 228: 1717–1724. 1975. 3. ML Koller, ML Riccio, RF Gilmour Jr. Dynamic restitution of
action potential duration during electrical alternans and ventricular fibrillation. Am J of Physio. 275(5):H1635–
H1642. 1998. 4. JI Goldhaber, LH Xie, Duong T, C Motter, K Khuu, JN Weiss. Action potential duration restitution
and alternans in rabbit ventricular myocytes. Circ Res. 96:459-466. 2005. 5. KHWJ Ten Tusscher, AV Panfilov.
Alternans and spiral breakup in a human ventricular tissue model. Am J of Physio. 291(3):H1088–H1100. 2006. 6. T
O’Hara, L Virag, A Varro, Y Rudy. Simulation of the undiseased human cardiac ventricular action potential: model
formulation and experimental validation. PLoS Comp Bio. 7(5):e1002061. 2011. 7. M Clerx, P Collins, E Lange,
PGA Volders. Myokit: A simple interface to cardiac cellular electrophysiology. PBMB. 120(1–3):100–114. 2016.
W-076
Development of Rheumatoid arthritis QSP model capturing mechanistic pathways and clinical response to
anti-TNF and methotrexate therapies
Author: Madhav Channavazzala (1), Dinesh Bedathuru (1), Maithreye Rengaswamy (1), Tamara Ray (1), Mrittika
Roy (1), Rukmini Kumar (1)
Institution: Vantage Research, India (1)
Objectives: * Develop QSP model of appropriate physiological detail to integrate mechanistic and population scales
in RA * Incorporate appropriate physiology and clinical outcomes: i. Key cellular and cytokine interactions in the
localized joint compartment that can be perturbed to implement the effect of therapies ii. Key clinical read-outs at
the population level such as ACR and DAS-28CRP that can evaluate the impact of novel therapies and trial designs
* Use model to visualize trial outcome and predict patient characteristics in a population level such as outcomes for
novel therapies, combinations of existing therapies, identifying sub-populations with greater response to therapies,
and simulate and op timize novel trial designs
Methods: Model design, engineering, survey of published physiological and clinical data was carried out in
accordance with standard QSP approaches1, 4. An average, inflamed joint capturing the disease at steady-state (i.e.,
with no disease progression or episodic inflammation) is modelled.
Using Ordinary Differential Equations (ODEs), the model captures cellular lifecycle and interactions of Fibroblast
like Synoviocytes (FLS), B cells, T cells and macrophages among other relevant cell types and relevant pro and anti-
inflammatory cytokines (e.g., Il-6, TNF-alpha, TGF-beta). Reference virtual subjects are generated and calibrated to
be average responders/non-responders to methotrexate (standard of care) and to anti-TNF-α therapy. Model
parameters are constrained by clinical trial data as well as by data from basic science literature.
Results: A QSP model capturing multiple physiological pathways of interest and response to specified therapies in
RA was developed that can be used for clinical trial visualization, trial optimization, responder/non-responder
identification etc. A Virtual Population has been developed that captures DAS-28CRP entry range at baseline and
shows a comparable response (reduction of disease severity) to the two therapies studied in the BEAM2 and
Premier3 trials. Reference subjects corresponding to responders and non-responders to Methotrexate and anti-TNF-
alpha therapies were also generated consistent with the aggregate results.
Conclusions: The Vantage RA-QSP model captures the physiology and clinical outcomes of RA, including
response to Methotrexate and anti-TNF-alpha therapies. Future efforts will add therapeutic pathways including anti-
IL-6, anti-IL-17, JAK-inhibitors and anti-IL-23.
Additional Note: A poster based on the Vantage RA model but focused on creation of Reference Virtual Patient has
been approved as a poster for PAGE 2019. However, this ACoP poster abstract focuses on Virtual Population and
model applications in QSP research.
References: 1. Gadkar K, et al., CPT pharmacometrics Syst Pharmacol. 2016;5(5):235-249.
doi:10.1002/psp4.12071 2. Breedveld FC, et al., Arthritis Rheum. 2006;54(1):26-37. doi:10.1002/art.21519 3.
Beattie S et al., N Engl J Med. 2017;376(7):652-662. doi:10.1056/nejmoa1608345 4. Greef J van der et al, IEE Proc
Syst Biol. 2005;152(4). doi:10.1049/ip-syb
W-077
Towards a multiscale QSP model of allergic airway inflammation in chronic asthma
Kamau Pierre1, Randolph J. Leiser1, Mengdi Tao1, Yui-Hsi Wang2, Christine Xu3, Haobin Luo4, John Pappas4, Zhiwei
Zhang4, Spyros Stamatelos1, Susana Zaph1 and Panagiota Foteinou1
1 Systems Pharmacology US, Sanofi, Bridgewater, NJ, 2 Type 2 Immunology & Fibrosis Cluster, Sanofi, Cambridge,
MA 3 Pharmacokinetics, Dynamics and Metabolism US, Sanofi, Bridgewater, NJ 4 RES Group, Inc, Needham, MA
Objectives: Asthma, a chronic inflammation of the bronchial tubes, involves numerous mediators and underlying
immunological pathways, such as atopy (IgE-mediated response) and Type 2 (Th2, eosinophilic) inflammation.
Despite the availability of several biologic products to treat moderate-to-severe asthma, some patients fail to reach
sufficient control of the disease, highlighting the need for novel approaches for effective treatment of this
population. Integrative initiatives such as those offered by Quantitative Systems Pharmacology (QSP) modeling are
identified as valuable in understanding how the complex interactions between drugs and their cellular targets
contribute to drug efficacy. In order to support evaluation of novel targets, a QSP modeling strategy has been
developed for chronic (Th2-driven) asthma. A unique aspect of this modeling framework is the mechanistic
quantification of emergent clinical and cellular functional responses as a result of interacting network modules
spanning multiple scales. In terms of application, such a modeling framework is expected to impact both Th2
research and early development by addressing key translational questions.
Methods: The current asthma QSP model is modular and multiscale in design, allowing the mechanistic
representation of pathological processes underlying asthma in their appropriate organ compartments. Core modules
have been constructed using an iterative process based on published information on key inflammatory mechanisms
in the lymph node and lung. In the context of a pharmacologic perturbation, these modules combine a detailed
representation of drug action at the molecular-level to the therapeutic effects at the cellular- and organ-levels
ultimately resulting in changes to clinical endpoints. To facilitate translational research, the integrated QSP/asthma
model also mechanistically incorporates key clinical biomarkers predictive of target/pathway engagement or
treatment response. These biomarkers (serum eotaxin, periostin, TARC, IgE and eosinophils) are connected to
clinical endpoints of interest represented in the model: FEV1 and exacerbation index. During the model
development phase, datasets from diverse sources (experimental and clinical) have been used to constrain and
parameterize the model, with the majority of data derived exclusively from the published human studies. These
include quantitative in-vitro (e.g. dose-dependent) data derived from human cells, as well as receptor and cytokine
expression levels, tissue cell numbers, half-lives, and cytokine-receptor affinities.
Results: All modules comprising the QSP asthma model have been constructed and calibrated successfully to
relevant experimental data. When integrated, the proposed model connects mechanism of action of several biologics
to changes to biomarkers and clinical endpoints, making it a good example of the value of using mathematical
modeling to gain mechanistic understanding of the system and a useful platform for predicting response of novel
therapeutics.
Conclusions: A robust modular and multiscale QSP strategy has been successfully implemented to capture key
aspects of Th2 pathophysiology in asthma.
W-078
Adapting a quantitative systems toxicology model of mitochondrial dysfunction in liver to kidney
Shailendra B. Tallapaka, Yeshitila Gebremichael, Scott Q. Siler, Brett A. Howell, Jeffrey L. Woodhead
DILIsym Services, Inc., a Simulations Plus company, Research Triangle Park, NC, USA
Objective: Kidney, as a major excretory organ, is exposed to high levels of drugs and their metabolites. Therefore,
kidney toxicity is an important part of drug safety assessment in clinical trials. RENAsym® is a quantitative systems
toxicology (QST) model of drug induced acute kidney injury (AKI) currently under development. In its current
form, the model includes representations of proximal tubule cell (PTC) lifecycle, bioenergetics, cellular injury and
death pathways. Our objective is to develop a mechanistic mathematical model of mitochondrial dysfunction in
proximal tubule cells to predict drug induced AKI.
Methods: We adapted the mitochondrial dysfunction model existing in DILIsym®, a QST model of drug induced
liver injury, by modifying the equations to accommodate the physiological differences between kidney and liver.
Changes made in order to translate the model to the kidney include (but are not limited to) eliminating de novo
lipogenesis and glycogen storage, refining PTC bioenergetics, and changing mitochondrial substrate utilization, for
example glucose oxidation was removed during homeostasis as little glucose oxidation was observed in rat proximal
convoluted tubules [1]. We then simulated gentamicin as an exemplar compound to qualitatively validate the model.
Gentamicin in vitro mitochondrial toxicity was measured in HepG2 cells and converted to RENAsym® parameters
using MITOsym®. Parameters for oxidative stress and kidney exposure were obtained from literature.
Results: Simulations in rat at 100 mg/kg QD dosing predicted significant toxicity within 24h (Fig 1), with the
fraction of viable PTCs reduced to ~16% when mitochondrial toxicity and oxidative stress mechanisms were turned
on. Interestingly, fraction of viable PTCs reduced to only ~68% when mitochondrial toxicity mechanism was turned
off, suggesting that the majority of the toxicity is due to mitochondrial toxicity.
Conclusions: A mitochondrial dysfunction model originally constructed for the liver has been adapted to the kidney
and reasonably predicts gentamicin-induced AKI. Simulations show that mitochondrial dysfunction causes more
toxicity than oxidative stress.
Figure 1: RENAsym ® simulations showing predictions of
gentamicin-induced toxicity in rats with and without the model of
mitochondrial dysfunction.
1. Klein, K.L., Wang, M.S., Torikai, S., Davidson, W.D. and
Kurokawa, K., 1981. Substrate oxidation by isolated single
nephron segments of the rat. Kidney international, 20(1), pp.29-35.
W-079
Dynamic-TMDD: A Pharmacokinetic Modeling framework for Immuno-oncology
Authors: Daniel C. Kirouac, David Flowers, Fei Hua, John M. Burke, Joshua F. Apgar
Institution: Applied BioMath, LLC, Concord, MA, USA
Objectives: Clearance of biologics is mediated by two routes: non-specific (linear) protein catabolism and target-
mediated (saturable) endocytosis. The hallmark of such target-mediated drug disposition (TMDD) is non-linear
clearance, wherein drug clearance and resulting exposure is concentration-dependent. Pharmacokinetic (PK) models
incorporating TMDD in common use presume target synthesis rate to be constant. While this is generally a
reasonable assumption, there are documented cases where it does not hold. If target synthesis changes over time due
to pharmacological activity of the drug, the rate of TMDD and clearance will thus change with successive doses,
affecting drug concentration profiles and activity. This creates a nonintuitive phenomenon, wherein
pharmacodynamics (PD) affect pharmacokinetics – a PD/PK model. This phenomenon is especially important in
oncology/immuno-oncology.
For drugs which deplete target expressing cells, target expression and drug clearance decreases over time (e.g. Anti-
CD20). For biologics which induce expansion of target expressing cells, target expression and drug clearance
increases over time (e.g. immuno-cytokines such as IL-2). A modelling framework which quatifies drug-induced
changes in target synthesis may be important to capture pharmacokinetic profiles of such agents.
Methods:We have developed a novel pharmacokinetic model, termed dynamic TMDD, to mechanistically describe
the phenomenon of drug induced changes in target expression and resultant TMDD. The model is based on first
principle reaction kinetics and contains easily interpretable and biology-based parameters, linking target engagement
to changes in synthesis.
Results:We demonstrate how tuning a single parameter can convert between a static-, target-depleting-, and target
inducing-TMDD. Ease of utility is demonstrated by reproducing PK profiles of two drugs: Rituximab (anti-CD20)
for target-depleting TMDD, wherein clearance decreases over successive cycles, and CEA-IL2 for target-inducing
TMDD, wherein clearance increases over time. We explore properties of these differing classes of molecules
through simulations of the effect of dose and affinity on PK/PD profiles.
Conclusions: For molecules exhibiting target-depleting TMDD, higher affinity always results in greater biological
activity. However, for the case of target-inducing TMDD, a lower affinity drug can provide improved biological
activity by balancing TMDD-mediated clearance vs target engagement. Non-linear clearance profiles typical of
TMDD are however not produced in the latter case, which may mask this phenomenon when examining single
ascending dose-PK profiles. In both cases, doses may need to be adjusted over successive cycles to maintain
consistent exposure and target engagement.
W-080
QSP model of human granulopoiesis and neutrophil homeostasis for assessment of pharmacological
interventions
Galina Lebedeva1, Alexander Stepanov2
1InSysBio UK Ltd, Edinburgh, UK; 2InSysBio LLC, Moscow, Russia
Objectives: Neutrophils play an essential role in the innate immunity. Homeostasis of neutrophil counts in blood is
tightly controlled by a complex regulatory network that maintains the balance between neutrophil production in
bone marrow, their release to blood circulation, migration to peripheral tissue and apoptosis/degradation.
Perturbation of any of these processes due to pathology or by pharmacological interventions can cause significant
changes in neutrophil levels (e.g. neutropenia) thus compromising the immunity. We sought to develop a
comprehensive model of granulopoiesis in healthy human based on modern understanding of neutrophil regulation
by growth factors, cytokines and chemokines. We aimed to apply the model to elucidate the mechanisms of
neutrophil homeostasis and analyze/predict effects of standard therapies.
Methods: The ODE-based model was constructed that consists of the following modules:
1) Neutrophil cell life cycle model for neutrophil cell dynamics in bone marrow (BM), blood plasma (PL) and
tissue (TIS) compartments. This includes maturation and proliferation of the neutrophil linage cells in BM,
their egress from BM to circulation, margination, migration to tissue and apoptosis/degradation. Known
regulatory effects of growth factors (G-CSF, GM-CSF), cytokines (IL-6, IL-3) and chemokines (CXCL12,
CXCL1, etc) on relevant stages of neutrophil development and trafficking were incorporated.
2) Production of biomolecular mediators (cytokines, chemokines) in BM, PL and TIS compartments by
relevant cell types, including neutrophils, endothelial cells, BM stromal cells and osteoblasts.
3) Distribution of biomolecular species between compartments and their degradation.
4) Drug effect module, that describes the impact of selected therapies (e.g. recombinant G-CSF and
chemotherapeutic agents) on target processes and includes implementation of relevant PK models.
The model accounts for changes in expression level of cell surface receptor molecules (CXCR4, CXCR2, GCSFR)
during neutrophil maturation via updating specific regulatory terms controlling granulocyte retention in BM and
their egress to circulation.
The neutrophil life cycle and biomolecular mediator production/degradation were partially calibrated based on
public in vitro/ex vivo data. The remaining unknown parameters were fitted to the in vivo data on the baseline cell
numbers in BM and blood and cytokine levels in plasma.
Results: The model allows for successful reproduction of clinical data on neutrophil homeostasis in healthy human,
including baseline levels of various neutrophil precursors in BM, neutrophil production rate in BM and neutrophil
blood counts. It predicts well the dynamics of neutrophil count in blood in response to administration of growth
factors and cytotoxic therapies. The model demonstrates the role of BM microenvironment in controlling neutrophil
maturation and egress to blood.
Conclusions: The model of neutrophil development and homeostasis aids in understanding of the complex
regulation of granulopoiesis, predicts the effect of pharmacological interventions on neutrophil levels and can serve
as a building block for establishing a comprehensive model of hematopoiesis.
W-081
Investigating the Mechanism of Action for Favipiravir against ZIKV Infection by Studying Target Site
Penetration and Mathematic Modeling
Authors: Xun Tao (1), Kaley C. Hanrahan (2), Jieqiang Zhou (1), Jürgen B. Bulitta (1), Ashley N. Brown# (2)
Institutions: (1) Center for Pharmacometrics and Systems Pharmacology, Department of Pharmaceutics, College of
Pharmacy, University of Florida, Orlando, Florida, USA, (2) Institute for Therapeutic Innovation, Department of
Medicine, College of Medicine, University of Florida, Orlando, Florida, USA
Objectives: Human ZIKV infection exhibits a broad cell and tissue tropism. Our previous research (Pires de Mello,
Camilly P. et al, AAC, 62.9 (2018)) revealed the promising anti-ZIKV activity of favipiravir (FAV) in HUH-7 cells.
Nevertheless, it remains unknown whether FAV can exert activity in wide range of cell types. Here we study FAV
effectiveness against ZIKV in two different cell lines of human origin, HELA and LNCAP. We also monitored the
intracellular triphosphate formation of the active FAV metabolite (RTP) to understand the degree of FAV activity in
each cell type. Furthermore, we established a mathematic model to identify the mechanism of action for FAV
activity.
Methods: ZIKV-infected HELA and LNCAP cells were seeded into 6-well tissue culture plates and exposed to
serial two-fold dilutions of FAV ranging from 62.5 µM to 1000 µM. An untreated control was also included.
Intracellular FAV and RTP concentrations were quantified by LC-MS/MS. A mechanism based model (MBM)
utilizing SADAPT-TRAN was developed to characterize the time course of ZIKV burden in both cell lines. The post
hoc analysis was utilized to diagnose the missing part during model development.
Results: ZIKV replication was substantially inhibited by FAV in HELA cells, whereas such activity was absent in
the LNCAP cell line (Fig.A&B). LC-MS/MS analysis revealed that active RTP was only present in HELA cells
(Fig.C). ZIKV egress from HELA cells was rapid and sustained; indicating that a high proportion of infected HELA
cells survive and support further rounds of viral replication post-cellular exit (recycling process). Our initial MBM
model yielded a significant bias of population fitting. The post hoc analysis suggested that recycled HELA cells
decreased with increasing concentrations of FAV (Fig.D). We excluded the cytotoxicity hypothesis, as cellular
toxicity was not observed with FAV treatment. Instead, we hypothesized that RTP resulted in the production of
inactive virus or defective interfering viruses (DIV). DIVs may compete with infectious ZIKV for cellular surface
receptors. This mechanism protects the uninfected host cell from becoming infected. The optimized MBM (Fig.E)
successfully described viral burden for all FAV treatment arms (r=0.985, population fits).
Conclusions: MCPEM algorithm with full Omega block model facilitates the post hoc analysis, which is
substantially useful for hypothesis inspiration. RTP is a well-known nucleoside analogue which inhibits intracellular
virus replication. Our model suggests an additional mechanism of action in that RTP can result in the formation of
DIVs that compete with ZIKV for cellular receptors. This additional mechanism may further enhance the antiviral
activity of FAV in HELA cells. Nevertheless, the FAV activity is absent in LNCAP cells. The addition of a second
anti-ZIKV agent may be required together with FAV to cover for the varying degrees of FAV activity in different
cell types.
W-082
Model Predictive Control as a Framework for Patient-Tailored (sub)Optimal Dosing under Constraints
Robert S. Parker, PhD1,2, and Timothy Knab, PhD1,3
1 Dept. of Chemical and Petroleum Eng., University of Pittsburgh, Pittsburgh, PA; 2 University of Pittsburgh Cancer
Institute, UPMC, Pittsburgh, PA; 3 currently with Metrum Research Group, Tariffville, CT
Objectives: To design a framework for dosing agents under input (e.g., dose magnitude, dose rate, cumulative
exposure, etc.) and output (e.g., measured variable peak/trough levels, time above/below a specified value,
trajectory) constraints. This model-based optimization problem must respond to deviations between predicted and
measured patient characteristics (e.g., cell counts, other biomarkers) and alter treatment accordingly, while
satisfying applied constraints.
Methods: We employ a docetaxel administration / neutrophil toxicity case study, drawing upon our physiologically
based PK (PhysPK) model of docetaxel [1] and our nonlinear dynamic model of neutrophil toxicity [2]. Models are
built in Pyomo (Python Optimization and Modeling Objects), a script-based language. The treatment objective
function over a user-specified number of weeks penalizes tumor volume and small drug doses, which increase
toxicity without improving antitumor efficacy. Dosing is limited to once per week, during working hours, Monday
to Friday, and the maximum quantity of drug allowed over a 3-week cycle is 105 mg/m2 (similar to existing 100
mg/m2 q3w and 35 mg/m2 qw x 3 every 4 weeks schedules). Neutrophil count constraints are set by toxicity level,
thereby limiting acute (1-day) and extended (week duration) toxicity exposure. Problem formulation discretizes
time, acceptable dose ranges, and dosing days; solutions via interior point optimization (via IPOPT) and direct
simulation leveraging GPU computing are compared for performance and solution time. We reoptimize every 21
days, when the algorithm receives a new neutrophil measurement from the simulated patient.
Results: Toxicity constraints stipulated no grade 3 neutropenia, and grade 2 neutropenia could last up to 7 days in
any 3-week cycle. Simulated patients satisfy the toxicity constraints, which are violated for 100 mg/m2 dosing on a
q3w schedule. Resolving every 21 days with an updated neutrophil measurement from the simulated patient, the
algorithm returns the dosing profile and neutrophil trajectory shown in Figure 1. Changes in patient characteristics
(e.g., PK, drug sensitivity) can be incorporated in the algorithm’s model; some changes can also be adaptively
corrected for over time.
Conclusions: Model-based receding-horizon control formulations, such as model predictive control, can be used as
a framework to design treatment schedules under drug- and clinically-relevant constraints. The framework is
extensible to other model types and applications beyond cancer.
Figure 5: Patient-specific optimal treatment profile under constraints. Left: docetaxel administration schedule, with
21-day cyclical structure. Right: Neutrophil profile after docetaxel infusion at left. Horizontal lines are toxicity
grades 2 (black), 3 (pink), and 4 (red).
References: [1] J.A. Florian, Jr. Modeling and Dose Schedule Design for Cycle-Specific Chemotherapeutics. PhD.
Dissertation. University of Pittsburgh, 2008. [2] T. Ho, G. Clermont, and R.S. Parker. “A Model of Neutrophil
Dynamics in Response to Inflammatory and Cancer Chemotherapy Challenges.” Computers & Chemical
Engineering 51, 187-196, 2013.
W-083
Combined application of Quantitative Systems Pharmacology (QSP) and PK-PD modeling to guide dose
selection for a Phase 3 study of anti-TFPI therapy in patients with hemophilia
Authors: Satyaprakash Nayak1, Chay Ngee Lim1, Joan Korth-Bradley2, Pinky Dua3, Tong Zhu1 and Lutz O.
Harnisch3
Affiliations: 1Pfizer Inc, Cambridge, MA, USA; 2Pfizer Inc, Collegeville, PA, 3Pfizer Ltd., UK.
Objectives: Inhibition of Tissue Factor Pathway Inhibitor (TFPI) is an effective strategy to increase clot formation
in patients with hemophilia, who are at risk of blood loss and long-term joint damage due to decreased blood
coagulation. The objective of this analysis was to determine the dose regimen for PF-06741086 (an anti-TFPI
monoclonal antibody) for a Phase 3 study based on a combined population PK-PD and QSP model of blood
coagulation, which described key clinical markers, including peak TGA and dilute Prothrombin Time (dPT) as
observed in Phase 1 and Phase 2 studies.
Methods: A Target Mediated Drug Disposition (TMDD) model was developed to characterize total TFPI (bound to
drug) and total drug concentration profiles, allowing prediction of free TFPI levels, which were subsequently used
as an input to the QSP model. The QSP model1 is a mechanistic model of coagulation, which includes the extrinsic
(or initiation), the intrinsic (or amplification) and the common pathway, as well as positive and negative feedback
loops present in the coagulation system. It can simulate in vitro, ex vivo and in vivo biomarkers (e.g., thrombin
generation assay and various clot-times) and has been validated with in-house and literature data1,2.
Results: Simulations of PD responses (peak and lag time from TGA, dPT) following 300 mg subcutaneously (SC)
and 150 mg SC once-weekly (with a loading dose of 300 mg) dosing regimens were generated in virtual populations
of 200 subjects for each cohort, representative of the demography (mean and standard deviation) observed in the
Phase 2 study. Peak TGA as observed in the study was accurately predicted from this model with the lower limit of
the 90% prediction interval being above the pre-defined threshold throughout the dosing period for both 150 mg and
300 mg dosing regimens. Simulations also showed that a 75 mg SC once-weekly dose would not be sufficient to
achieve the pre-defined criteria.
Conclusions: A combined population PK-PD and QSP modeling approach was implemented to describe the drug-
target kinetics and biomarker behavior following anti-TFPI therapy in patients with hemophilia, to facilitate dose
regimen selection for a Phase 3 study. Based on the observed Phase 2 efficacy results and predictions from this
combined PK-PD and QSP modeling approach, a 150 mg SC (300 mg loading) dose regimen was chosen as the
potential regimen for Phase 3.
References: 1. S. Nayak et al., Using a systems pharmacology model of the blood coagulation network to predict
the effects of various therapies on biomarkers, CPT:PSP, 4 (7):396 – 405, 2015 2. D. Lee et al, A quantitative
systems pharmacology model of blood coagulation network describes in vivo biomarker changes in non‐bleeding
subjects, Journal of Thrombosis and Hemostasis, 14 (12):2430-2445, 2016
Figure
Figure 1: The figure shows peak TGA as predicted by the model. Panel A describes the peak TGA for a dosing
regimen of 150 mg QW dosing with a 300 mg loading dose. Panel B shows the same for a 300 mg QW dose
regimen. The red band shows 5th – 95th percentile of peak TGA values. Both dosing regimens met the pre-specified
criteria of lower 5th percentile being above the threshold value.
W-084
Network-guided Evaluation of Targeted Combination Chemotherapy in Diffuse Large B-Cell Lymphoma
Authors: Van Anh Nguyen1 and Donald E. Mager1
Affiliations: 1Department of Pharmaceutical Sciences, University at Buffalo, The State University of New York,
Buffalo, NY 14214, USA
Objectives: To (i) evaluate how distinctive protein expression patterns in two subtypes of diffuse large B-cell
lymphoma (DLBCL) affect B-cell signaling dynamics and proliferation in silico in the absence and presence of
pharmacological treatment; and (ii) assess network predictions for targeted interventions in vitro.
Methods: Network analyses of a B-cell signaling Boolean model were performed to examine intracellular
determinants of inter-disease variability in DLBCL. Minimal intervention analysis (MIA) was conducted in
CellNetAnalyzer to identify single and combination treatment strategies that can induce cell death and prevent cell
growth in activated B-cell (ABC) and germinal center B-cell (GCB) subtypes of DLBCL. Attractor analysis was
conducted with BoolNet to assess the effectiveness of potential single agent and multidrug therapeutic interventions.
Simulations with 106 random initial conditions were performed with a synchronous updating scheme, and the
activation frequency of apoptosis was evaluated across suggested treatments. Cell cytotoxicity studies were
performed at 48 hours to assess network predictions for single-agent and combination interventions. SU-DHL2 and
SU-DHL4 cell lines were utilized as model systems for studying ABC and GCB DLBCL. Bortezomib, vorinostat,
and rituximab were used as model compounds for evaluating the effects of proteasome inhibition, histone
deacetylase (HDAC) inhibition, and CD20 activation (aCD20). Combination drug effect was modeled using the
noncompetitive interaction equation, and a Ψ parameter was estimated to determine the nature of bortezomib-
vorinostat interaction. The equation was modified to incorporate a multiplicative interaction parameter ΨR to
quantify any changes in drug combination efficacy due to addition of rituximab.
Results: MIA suggested that no single-agent targeted intervention was capable of achieving cell death in ABC
DLBCL. In contrast, GCB DLBCL was found to be more susceptible to targeted monotherapies. Attractor analysis
suggested that proteasome and HDAC inhibitors may have differential effects in two subtypes of DLBCL. However,
the difference in the distribution of outcome (probability of achieving cell death) decreases when these agents are
combined and becomes nonexistent with addition of aCD20 to the combination. Bortezomib and vorinostat were
found to have different potencies across two cell lines (bortezomib IC50: 6.01 and 2.98 nM, vorinostat IC50: 0.76 and
1.27 μM in SU-DHL2 and SU-DHL4). The combination of bortezomib and vorinostat was synergistic in both cell
lines (Ψ estimated at 0.792 and 0.545 in SU-DHL2 and SU-DHL4). Triple combination study in SU-DHL2
suggested that addition of rituximab to the combination of bortezomib and vorinostat resulted in enhanced cell
death, as reflected by ΨR of 0.590.
Conclusions: The analysis of a network-based systems pharmacology model of B-cell signaling led to the
identification of differences in signal propagation in two subtypes of DLBCL. Boolean network modeling can aid in
qualifying potential therapies within the context of molecular complexity and heterogeneity of DLBCL.
W-085
Visualizing Parameter Source Reliability and Sensitivity for QSP Models
Sietse Braakman1, Ricardo Paxson1, Stacey Tannenbaum2, Abhishek Gulati2
1MathWorks Inc., Natick, Massachusetts 2Clinical Pharmacology & Exploratory Development, Astellas Pharma,
Northbrook, Illinois
Objectives: Published quantitative systems pharmacology (QSP) models usually present tables with initial
parameter values including their sources as references. Reliability of the sources is rarely evident, and evaluating
each reference individually can be an enormous undertaking for models that have hundreds of parameters. The
present work builds on a previously introduced methodology (1) and aims to evaluate the reliability of initial
parameter values in QSP models, as well as the sensitivity of model outputs to each parameter, thus helping to build
confidence in the model predictions.
Methods: A published QSP model of PCSK9 therapy (2) to reduce LDL-cholesterol (LDLc) was used to illustrate
the method. The publication includes a SimBiology model (8 ODEs, 42 parameters) and sources of parameter
values. As part of this method, sources were categorized into nine “reliability categories” based on their reliability
and relevance to the research question [see Figure legend]. In the model, each parameter (𝒑𝒊) is assigned a reliability
category, and local sensitivity scores (𝜎𝑖,𝑗 = ∑ (|∆𝑅𝑗(𝑡)
∆𝑝𝑖| ∗ ∆𝑡) 𝑇𝑙𝑎𝑠𝑡
0 ) were calculated for each response (𝑹𝒋(𝑡)),
where 𝑇𝑙𝑎𝑠𝑡 signifies the last simulation time. In the proposed visualization [see Figure], the height of each bar
represents the sensitivity score for that parameter, while its color represents its reliability category. LDLc is used as
the only model response in the Figure because it represents the clinically relevant outcome of PCSK9 therapy.
Sensitivity score calculations, categorization and visualization were performed with MATLAB and SimBiology
R2019a.
Results: The figure shows a significant number of sensitive parameters (green, turquoise) that seem to have reliable
sources, providing confidence in the model. However, the tall red and orange bars represent sensitive parameters
that had unknown or unreliable sources and warrant further investigation to increase confidence in the model. The
shorter red and orange bars in the Figure indicate that even if some parameters had less reliable sources, the outcome
is less sensitive to their values.
Conclusions: This work introduces a visualization method that provides a bird’s eye view of the reliability (based
on the assignment by the modeler of “reliability categories”) and importance of sources of initial parameter values in
a QSP model. We propose that this method can become an integral part of the model building process to help
establish confidence in a QSP model: for regulatory purposes, for internal cross-functional discussions, and as a
standard format for publishing QSP models. The visualization method can be extended to include multiple model
responses and can be implemented in any modeling environment of choice.
References: Gulati A, Tannenbaum S. ACoP Annual Meeting, San Diego, CA; Oct 7-10, 2018.
Gadkar K, et al. CPT Pharmacometrics Syst Pharmacol. 2014;3(11):e149.
W-086
Building a Mechanistic Modeling Platform for HIV Cure Drug Development
Authors: Youfang Cao1, Malidi Ahamadi1, Daniel Rosenbloom1, SoHyun (Irene) Bae1, Ryan Vargo1
Institutions: 1Pharmacokinetics, Pharmacodynamics and Drug Metabolism, Merck & Co., Inc., Kenilworth, NJ
Objectives: Current antiretroviral therapy (ART) effectively controls HIV in most patients but does not cure it. Side
effects from life-long ART and risks of drug resistance become long-term burdens for patients. To develop drugs
towards a HIV cure, novel approaches – such as reactivation of latent provirus (“shock”) and immunotherapies – are
being explored. However, it is an open question whether these approaches – either alone or together – might lead to
a cure. Mechanistic mathematical models that describe both within-host viral load dynamics and immunologic
control of HIV infection are essential to integrate clinical data, assess therapeutic response, and generate hypotheses
in support of HIV cure drug development.
Methods: We built the Latent Viral Dynamics Modeling (LVDM) platform, based on recently developed
mathematical models (1, 2) which integrate potential mechanisms that may lead to a cure (Figure 1). The LVDM
with large viral load (VL) dataset from separate clinical studies covering both long-term viral suppression during
ART, and analytical treatment interruption (ATI) were fit using Monolix (N=896 subjects). Despite the high
nonlinearity of the LVDM, five important parameters along with their inter-individual variabilities were estimated
with high accuracy (Figure 1). From estimated parameters, we sampled a virtual population (N=2000) to run clinical
trial simulations (CTS) for potential curative interventions.
Results: The results showed high-quality fits of LVDM to the large VL dataset (Figure 1), and the LVDM can
accurately capture the complex individual viral load dynamics during both ART and ATI phases. The CTS based on
estimated parameters illustrated that the LVDM can reliably recapitulate and predict viral load responses during all
phases of HIV infection. To explore the effects of potential curative interventions, the viral suppression off therapy
(VSOT) time after ATI from large-scale CTS was calculated as the predicted responses of cure trials. The results
show that long-term ART will increase the VSOT time, but this effect reaches a plateau after a few years (Figure 1),
and early start of ART does not help achieve longer VSOT. Additionally, we found that a “shock” mechanism on
top of ART will significantly increase the VSOT (Figure 1), and it does not reach a plateau even after 10 years of
continuous treatment. Furthermore, it was observed that the VSOT time has a large variability across individuals
regardless of the cure treatment. These initial simulations for long-term ART and “shock” mechanisms suggest a
potential role for these mechanisms in HIV cure.
Conclusions: The LVDM platform will be continuously developed by incorporating new mechanisms as a platform
to identify new targets and generate hypotheses about combination therapies, and design HIV cure trials through
large-scale CTS.
References: 1. Conway and Perelson, (2015) PNAS 112(17) 5467-5472. 2. Cao, Cartwright, Silvestri, Perelson
(2018) PLoS Pathog 14(10): e1007350.
W-087
A quantitative systems pharmacology (QSP) model for Abeta pathway to predict effect of aducanumab on
Abeta SUVR change
Authors: Lin Lin2, Cristian Salinas1, Carissa Young2,3, Thierry Bussiere1, Jennifer Park2, Alvydas Mikulskis 1,
Yaming Hang1,3, Kumar Muralidharan1, Kubra Kamisoglu1, Ping Chiao1, John M Burke2, Joshua F Apgar2, Paul H.
Weinreb1, Fei Hua2, Ivan Nestorov1
Institutions: 1 Biogen Inc. 225 Binney Street, Cambridge, MA USA 2Applied BioMath, LLC; 561 Virginia Road,
Concord, MA, USA 3Current address: Takeda Pharmaceuticals, Cambridge, MA, USA
Objectives: Aducanumab is a human monoclonal antibody selective for aggregated forms of beta-amyloid peptide
(Abeta), with weak binding to monomers. During phase 1 study, aducanumab reduced Abeta plaque, as measured by
standard uptake value ratio (SUVR) of amyloid PET imaging. The goal of this work was to use a quantitative
systems pharmacology (QSP) model to understand the relationship between the aducanumab dosing regimen and
SUVR changes.
Methods: An ordinary differential equation based QSP model was implemented using KroneckerBio v0.5
(https://github.com/kroneckerbio). The model describes Abeta production through Amyloid Precursor Protein (APP)
cleavage by gamma-secretase and beta-secretase, Abeta aggregation into soluble oligomer, then insoluble plaque as
well as protein synthesis and turnover. In addition, the model also describes aducanumab clearance, distribution into
peripheral tissue, brain interstitial fluid (ISF) and cerebrospinal fluid (CSF), aducanumab binding to different forms
of Abeta as well as antibody dependent cellular phagocytosis (ADCP). The change of brain insoluble plaque in the
model was used to correlate to the SUVR change. The model was first calibrated against the pharmacokinetics (PK),
plasma Abeta and SUVR data with 1 year of aducanumab treatment from the phase 1b study. The model was used to
predict the SUVR change with 2-year aducanumab treatment. The developed model was also analyzed to understand
the parameters that control the dynamics of plaque reduction with aducanumab treatment.
Results: After model calibration, the model adequately captured single dose PK and plasma Abeta as well as
multiple dose PK data for aducanumab. Moreover, the rate of insoluble plaque turnover was identified as a sensitive
parameter for insoluble plaque reduction, yet not well constrained by the 1-year SUVR data. Two sets of parameters
were identified that can fit to the existing data equally well. However, when the model predictions were compared to
the 2-year SUVR data, the set of parameters based on slower plaque turnover was able to match the data much better
than the set with faster plaque turnover rate. The parameter set with slow plaque turnover was chosen as the final
parameter set. The model with the final parameter set was also able to predict the SUVR changes from a titration
group. Model analysis predicted that as the dose level of aducanumab increases, within the dose ranges tested, the
rates of plaque reduction increases and the maximal plaque reduction increase. In addition, model simulation
suggests that the same total dose of aducanumab with different dosing frequency is likely to lead to similar plaque
reduction.
Conclusions: A QSP model integrating biological understanding as well as dynamic data provided insights of the
mechanism of action of the drug and identified key parameters influencing aducanumab’s effect on plaque
reduction.
W-088
Three-Dimensional Heart Model–Based Screening of Proarrhythmic Potential by In Silico Simulation of
Action Potential and Electrocardiograms: Verapamil and Ranolazine vs. Dofetilide
Dong-Seok Yim1,2, Minki Hwang3, Seunghoon Han1,2, Min Cheol Park4, Eun Bo Shim4
1Department of Clinical Pharmacology and Therapeutics, Seoul St. Mary’s Hospital, 222 Banpo-daero, Seocho-gu,
Seoul 06591, Korea 2PIPET (Pharmacometrics Institute for Practical Education and Training), College of Medicine,
The Catholic University of Korea, 222 Banpo-daero, Seocho-gu, Seoul 06591, Korea 3SiliconSapiens Inc., Seoul,
Republic of Korea 4Department of Mechanical and Biomedical Engineering, Kangwon National University,
Chuncheon 200-301, Republic of Korea
Objectives: The proarrhythmic risk is a major concern in drug development. The Comprehensive in vitro
Proarrhythmia Assay (CiPA) initiative has proposed the JTpeak interval on electrocardiograms (ECGs) and qNet, an
in-silico metric, as new biomarkers that may overcome the limitations of the hERG assay and QT interval [1]. In this
study, we simulated body-surface ECGs from patch-clamp data using realistic models of the ventricles and torso to
explore their suitability as new in silico biomarkers for cardiac safety.
Methods: We tested three drugs in this study: dofetilide (high proarrhythmic risk), ranolazine, and verapamil (QT
increasing, but safe). Human ventricular geometry was reconstructed from computed tomography (CT) images, and
a Purkinje fiber network was mapped onto the endocardial surface. The electrical wave propagation in the ventricles
was obtained by solving a reaction-diffusion equation using finite-element methods. The body-surface ECG data
were calculated using a torso model that included the ventricles. The effects of the drugs were incorporated in the
model by partly blocking the appropriate ion channels. The effects of the drugs on single-cell action potential were
examined first, and three-dimensional (3D) body-surface ECG simulations were performed at free Cmax values of
1x, 5x, and 10x.
Results: In the single-cell and ECG simulations at 5x Cmax, dofetilide, but not verapamil or ranolazine, caused
arrhythmia (Figure). However, the non-increasing JTpeak caused by verapamil and ranolazine that has been
observed in humans was not reproduced in our simulation.
Conclusions: Our results demonstrate the potential of 3D body-surface ECG simulation as a biomarker for
evaluation of the proarrhythmic risk of candidate drugs.
References: Vicente, J., Zusterzeel, R., Johannesen, L., Mason, J., Sager, P., Patel, V., Matta, M.K., Li, Z., Liu, J.,
Garnett, C., Stockbridge, N., Zineh, I., and Strauss, D.G. (2018). Mechanistic model-informed proarrhythmic risk
assessment of drugs: Review of the "CiPA" initiative and design of a prospective clinical validation study. Clin
Pharmacol Ther 103, 54-66.
Figure. Simulated effects on single-cell action potential in the 3D heard model. Action potential curves are shown
for dofetilide, ranolazine, and verapamil at Cmax values of 1x, 5x, and 10x. The action potentials of endocardial, M,
and epicardial cells are shown for each drug and Cmax value. Dofetilide at Cmax values of 5x and 10x induced
tachycardia.
W-089
Evaluating Relationships Between Biomarkers Variability and responders and non-responders in response to
mineralocorticoid receptor (MR) antagonist therapy: A QSP Approach
Authors: Md Fazlur Rahman1, K. Melissa Hallow1
Affiliations: 1School of Chemical, Materials, and Biomedical Engineering, University of Georgia, Athens, GA,
USA
Objectives: The prevalence of primary aldosteronism accounts for approximately 10% of hypertensive patients,
with recent evidence suggesting that this number may be significantly higher [1]. MR antagonists are standard of
care for the treatment of primary aldosteronism [1]. However, there is wide variability in response to MR
antagonists, especially within essential hypertension populations. The aldosterone-renin-ratio (ARR) is used as a
diagnostic measure for aldosteronism, but neither this ratio or its individual components have been consistently
shown to predict therapy responsiveness in essential hypertension [2, 3]. This may be due in part to the impact of
additional sources of variability, such as sodium intake, on renin and aldosterone secretion and sensitivity. The
objective of this analysis was to evaluate physiological sources of variability in response to MR antagonists, with the
aim of evaluating the true relationship between aldosterone, renin, and blood pressure reduction with MR
antagonists when additional soruces of variability are removed.
Methods: We fit a systems pharmacology model of the cardiorenal system to a published data for different doses of
spironolactone therapy in healthy volunteers [4] to determine the dose-response relationship. The pharmacodynamic
parameters (Emax and ID50) were optimized by fitting the observed Na+ excretion data in healthy volunteers and
ability of the model in response to spironolactone therapy to produce the reduction in mean arterial pressure (MAP)
was simulated in different hypertensive virtual patients. We then evaluated the blood pressure response to variability
in model parameters, including renin secretion, aldosterone secretion, and sodium intake.
Results: The model produced a wide variability in MAP reduction in response to spironolactone therapy in different
virtual patients consistent with the observed data [2, 3]. Simulation indicated differential expression of different
biomarkers singly or in combination can regulate MAP reduction in response to spironolactone therapy.
Conclusions: A systems pharmacology model can be used to predict the sources of variability in response to MR
antagonist therapy.
References: 1. Funder, J.W., Mineralocorticoid receptor antagonists: emerging roles in cardiovascular medicine.
Integr Blood Press Control, 2013. 6: p. 129-38. 2. Levy, D.G., R. Rocha, and J.W. Funder, Distinguishing the
antihypertensive and electrolyte effects of eplerenone. J Clin Endocrinol Metab, 2004. 89(6): p. 2736-40.
3. Schersten, B., et al., Clinical and biochemical effects of spironolactone administered once daily in primary
hypertension. Multicenter Sweden study. Hypertension, 1980. 2(5): p. 672-9. 4. Hallow, K.M. and Y. Gebremichael,
A quantitative systems physiology model of renal function and blood pressure regulation: Model description. CPT
Pharmacometrics Syst Pharmacol, 2017. 6(6): p. 383-392.
Figure. Variability in Na+ intake obscures the relationship between renin, aldosterone, and mean arterial pressure.
Removing Na+ intake variability makes this relationship apparent.
W-090
When Spatial Effects Matter: A Quantitative Systems Pharmacology (QSP) Model to Describe Drug
Concentration Gradients in Realistic Tumor Vasculature with Applications to Antibody-Drug Conjugates
Jackson K. Burton1, Dean Bottino2, and Timothy W. Secomb3,4
1Critical Path Institute, Tucson, Arizona, USA; 2Takeda Pharmaceuticals, Cambridge, Massachusetts, USA; 3Program in Applied Mathematics, University of Arizona, Tucson, Arizona, USA; 4Department of Physiology,
University of Arizona, Tucson, Arizona, USA
Objectives: Pharmacokinetic-based models that quantify the time-course of drug concentration often assume the
drug is well-mixed in various compartments. For several classes of drugs, however, significant gradients in
concentration are observed on the microscopic scale in solid tumors. A novel QSP model is used to simulate time-
dependent spatial gradients in drug concentration within a three-dimensional network of tumor blood vessels with
realistic geometry. The objective is to utilize this model to describe the transport and kinetics of
antibody-drug conjugates (ADCs) in solid tumors to provide insights for drug development.
Methods: A time-dependent Green’s function is used to solve reaction-diffusion equations that describe the
transport and kinetics of ADCs in a tumor microvessel network previously digitized from an imaging study.
Governing equations for the time course of the ADC in plasma are specified. The model was implemented
numerically in C, with parallel processing using graphical processing units. All parameter values were derived from
the literature and no parameter fits were used for the reaction-diffusion model.
Results: The resulting model simulates time-dependent drug concentration in arbitrarily shaped 3-D vessel
networks. Three previous experimental studies that quantified the time course and spatial distribution of ADCs in
solid tumors were chosen as case studies for model validation. Each study showed excellent agreement with model
simulations with some explainable exceptions. In one example, model simulations were used to predict mean
fluorescent intensity as a function of distance to nearest vessel in mice xenografts and compared with experimental
data (1) (Figure 1).
Conclusions: The model and solution method serve as a novel mechanistic framework to address challenges in
describing drugs that induce gradients in concentration on the microscopic scale. The efficiency of the solution
method presents opportunities for modeling in which traditional mechanistic QSP modeling is not tractable. This
tool is potentially customizable for pre-clinical and translational drug development, with applications to ADCs and
other drugs that induce steep gradients in concentration
References: 1. Rhoden JJ, Wittrup KD. Dose Dependence of Intratumoral Perivascular Distribution of Monoclonal
Antibodies. Journal of Pharmaceutical Sciences. 2012;101(2):860-7. doi: 10.1002/jps.22801.
W-091
Vaccine Research and Development: A Fertile Field for PMx Innovation and Impact
Jeffrey P. Perley1, Nitin Mehrotra1, Paula Annunziato2, Daria Hazuda3, Amy Espeseth3, Long Wang4, Eseng Lai5,
Antonios Aliprantis5, Jon Hartzel6, Vikram Sinha1, Jeffrey R. Sachs1
1 Pharmacokinetics, Pharmacodynamics, and Drug Metabolism – Quantitative Pharmacology and Pharmacometrics,
Merck & Co., Inc., Kenilworth, NJ USA; 2 Clinical Research – Vaccines, Merck & Co., Inc., Kenilworth, NJ USA; 3 Biology/Discovery – Infectious Diseases/Vaccines, Merck & Co., Inc., Kenilworth, NJ USA; 4 Regulatory Affairs
– Vaccines/Infectious Diseases, Merck & Co., Inc., Kenilworth, NJ USA; 5 Clinical Research – Translational
Pharmacology, Merck & Co., Inc., Kenilworth, NJ USA; 6 Biostatistics – Vaccines, Merck & Co., Inc., Kenilworth,
NJ USA
Objectives: Pharmacometrics (PMx) has only recently been introduced to vaccine research and development (R&D)
[1] and has already shown great potential to impact decision-making. The objective of this work is to (1)
demonstrate, using examples, how PMx has influenced vaccine R&D decisions, and (2) to highlight further
opportunities for PMx to impact vaccine R&D.
Methods: We have used Quantitative Systems Pharmacology (QSP) modeling, mixed-effects modeling and
regression, Bayesian inference, trial simulation, machine learning, and model-based meta-analyses (“comparator
modeling”).
Results: Integrating PMx into the decision-making process has already impacted vaccine R&D. For example:
A. Preclinical Translation: We are supporting selection of candidates for clinical trials by determining
whether nonhuman primates (NHP) are predictive of the human immunogenicity response under
corresponding conditions for a multivalent vaccine candidate. A linear mixed-effects model across multiple
antigen types and time points was used to predict clinical neutralizing antibody titer from preclinical data
(Fig. 1A).
B. Design of Clinical Trials: Clinical trial simulation (CTS) was used to investigate how (dose-level) of
antigen, formulation, and number of animals/subjects influence the outcome of a study. This enabled
efficient optimization of the trial design to increase the probability of success (Fig. 1B).
C. Biomarker Response Prediction: Existing biomarker data were integrated with a mechanistic
understanding of the underlying immunology to generate a QSP model of vaccine-induced
immunogenicity. The model enabled the prediction of longitudinal biomarker responses to formulations
and regimens untested in the lab or clinic (Fig. 1C), suggesting that inclusion of regimen (“A”) could help
validate our understanding of a minimal dose-level/number of doses.
Other important examples that were not illustrated here for brevity include ►quantifying relationships between
biomarkers (e.g., neutralizing antibody titers) and efficacy, ►understanding how to stratify vaccine responders
versus non-responders based on demographic, phenotypic, and genetic factors [2], ►streamlining sampling
strategies for immune response biomarkers and disease state assessment, ►making quantitative predictions to
support objective go/no-go decisions, and ►integrating data across multiple trials for more informed decision-
making.
Conclusions: PMx has already had a substantial impact on vaccine R&D despite its recent introduction. Fully
integrating it into the decision-making process can boost probability of success and lead to even better scientific
understanding, savings in time and resources, and additional benefits that have been seen in the other therapeutic
areas that have adopted PMx.
References: [1] Heaton, Penny, comments during ASCPT 2019 presentation From Molecule to Patient: A Global
Health Perspective, Washington, DC. https://www.ascpt.org/Meetings/Annual-Meeting/Program-Highlights
[2] Hayes, S., Swaminathan, G., White, C., Cristescu, R., Citron, M., Sachs, J., … Cho, C. (2019, March).
Understanding the Role of the Microbiome in Vaccine Hyporesponse in the Elderly using Machine Learning and
Quantitative Systems Pharmacology. Poster presented at ASCPT 2019 Annual Meeting, Washington DC.
https://www.ascpt.org/Meetings/Annual-Meeting/Program-Highlights
W-092
Mechanistic model of a pathway regulated by protein sequestration helps define optimal therapeutic target
Jangir Selimkhanov1, Erica L. Bradshaw1
1Quantitative Solutions, Quantitative Translational Sciences, Takeda Pharmaceuticals, San Diego, CA, USA
Introduction/Objective: In certain biological systems, protein sequestration by a native inhibitor can provide
regulation that is necessary for an ultrasensitive pathway response [1]. In such a system, the sequestered protein
often activates a particular biological pathway, while the sequestering protein serves to inhibit it. In a
pathophysiological environment, the activity of such a pathway can be dysregulated which can further exacerbate
the pathological response. Restoration of normal pathway activity may provide some therapeutic relief or counteract
the pathophysiology altogether. From the drug development standpoint, it is not always clear what the optimal target
in a pathway with such regulation should be. A mechanistic mathematical model provides a quantitative framework
to systematically evaluate complex system parameters and identify which are most relevant for target selection and
to help with the drug development process.
Methods and Results: Here, we derive a mathematical model of a pathway consisting of a protein agonist, its
receptor protein and a sequestering protein. In this simplified system, the protein agonist binds to the receptor
resulting in pathway activation, while the sequestering protein inhibits this pathway by binding to the protein
agonist, which prevents it from binding to the receptor. In the state of pathway inactivation due to increased
concentration of sequestration protein, we evaluate the receptor and the sequestering protein as two potential drug
targets to increase pathway activity. We consider various protein and drug binding affinities and protein turnover
rates as potential differentiators for target selection. Using a model sensitivity analysis, we identify regions of
protein binding affinities where developing a drug that prevents sequestration of the protein agonist can be more
effective in activating the pathway than a drug receptor agonist. Additionally, we show that the protein turnover rate
differences between the protein agonist and the sequestering protein can further inform the drug target selection
process, while receptor turnover rate is less informative. Finally, we show that a drug that binds the receptor and the
sequestering protein can further boost pathway activation over a drug that can only accomplish one of those
functions.
Conclusions: These analyses illustrate how a simple mechanistic model can be used for target selection and to
identify drug properties required to achieve appropriate level of efficacy based on characterization of drug’s
intended target and target pathway.
References: Buchler, N. E. and F. R. Cross (2009). "Protein sequestration generates a flexible ultrasensitive
response in a genetic network." Molecular systems biology 5(1): 272.
W-093
Comparison of Phase I combination therapy designs by clinical trial simulations to evaluate early tumor
shrinkage
Authors: Jérémy Seurat (1), Pascal Girard (2), Vishnu Dutt Sharma (3), Kosalaram Goteti (3), France Mentré (1)
Institutions: (1) IAME, INSERM, UMR 1137, University Paris Diderot, Paris, France (2)Merck Institute for
Pharmacometrics, Merck Serono S.A., Lausanne, Switzerland (3)EMD Serono R&D Institute, Billerica, MA, USA
Objectives: In oncology, there is a growing interest in the use of combination therapies in early clinical trials. The
aim of this analysis was to compare in silico several combination designs of drug M with cetuximab (C) in the
treatment of solid tumors and to define the appropriate dose of C, assuming dose M is fixed. The performances, type
I error (α) and power, of several one-stage designs were compared to test the superiority of the combination C+M to
C alone using modeling and simulation of exposure-tumor growth inhibition (TGI).
Methods: Clinical trial simulations were performed, using an exposure-TGI model [1]. Different combination
effects of M and C (no effect/additive/synergistic) were explored. 1, 2 and 4 arm designs, were evaluated. In the 1
arm design, 60 patients received C+M. In the 2 arms design, 30 patients received C alone and 30 patients received
C+M. In the 4 arms design, in addition to C and C+M at standard doses, combination arms with lower doses of C
were evaluated (15 patients/arm). The data for all arms were fitted using SAEM algorithm in Monolix 2018R2 to
obtain individual ETS predictions at week 8 (ETS8). Comparison test were performed on predicted and observed
ETS8 (with residual variability) between the different arms.
Results: With the same number of individuals, the 1 arm design shows a better power of tests than 2 or 4 arms, but
implies strong assumptions on the historical reference value, leading to strong inflation of type I error in case of
misestimated reference: for instance, α is 34% if reference ETS8 is 14% lower than true one. Choosing a 2 or a 4
arms depends on the objective of the study. A 2 arms design is preferable than 4 arms to reach a good power of
statistical tests. Nevertheless, a 4 arms design allows a better understanding of the dose-exposure relationship and
thus a better dose selection for C.
Conclusions: This work highlights the strengths and weaknesses of the different early clinical combination designs.
An extension of this work is to perform model-based adaptive two-stage designs using the Fisher Information
Matrix to optimize the second stage of the study, where arms could also be added or dropped at the end of first
stage.
The results in this abstract have been previously presented in part at PAGE, Stockholm, 11-14th June, 2019 and
published in the conference proceedings as abstract PAGE 28 (2019) Abstr 9046.
References: 1. Claret L, Girard P, Hoff PM, Van Cutsem E, Zuideveld KP, Jorga K, et al. Model-based prediction
of phase III overall survival in colorectal cancer on the basis of phase II tumor dynamics. J Clin Oncol Off J Am Soc
Clin Oncol. 2009;27:4103–8.
W-094
Mechanistic model of a pathway regulated by protein sequestration helps define optimal therapeutic target
Jangir Selimkhanov1, Erica L. Bradshaw1
1Quantitative Solutions, Quantitative Translational Sciences, Takeda Pharmaceuticals, San Diego, CA, USA
Introduction/Objective: In certain biological systems, protein sequestration by a native inhibitor can provide
regulation that is necessary for an ultrasensitive pathway response [1]. In such a system, the sequestered protein often
activates a particular biological pathway, while the sequestering protein serves to inhibit it. In a pathophysiological
environment, the activity of such a pathway can be dysregulated which can further exacerbate the pathological
response. Restoration of normal pathway activity may provide some therapeutic relief or counteract the
pathophysiology altogether. From the drug development standpoint, it is not always clear what the optimal target in a
pathway with such regulation should be. A mechanistic mathematical model provides a quantitative framework to
systematically evaluate complex system parameters and identify which are most relevant for target selection and to
help with the drug development process.
Methods and Results: Here, we derive a mathematical model of a pathway consisting of a protein agonist, its receptor
protein and a sequestering protein. In this simplified system, the protein agonist binds to the receptor resulting in
pathway activation, while the sequestering protein inhibits this pathway by binding to the protein agonist, which
prevents it from binding to the receptor. In the state of pathway inactivation due to increased concentration of
sequestration protein, we evaluate the receptor and the sequestering protein as two potential drug targets to increase
pathway activity. We consider various protein and drug binding affinities and protein turnover rates as potential
differentiators for target selection. Using a model sensitivity analysis, we identify regions of protein binding affinities
where developing a drug that prevents sequestration of the protein agonist can be more effective in activating the
pathway than a drug receptor agonist. Additionally, we show that the protein turnover rate differences between the
protein agonist and the sequestering protein can further inform the drug target selection process, while receptor
turnover rate is less informative. Finally, we show that a drug that binds the receptor and the sequestering protein can
further boost pathway activation over a drug that can only accomplish one of those functions.
Conclusions: These analyses illustrate how a simple mechanistic model can be used for target selection and to identify
drug properties required to achieve appropriate level of efficacy based on characterization of drug’s intended target
and target pathway.
References: Buchler, N. E. and F. R. Cross (2009). "Protein sequestration generates a flexible ultrasensitive response
in a genetic network." Molecular systems biology 5(1): 272.
W-095
Machine Learning Methods can Inform Hypercoagulation Risk and Patient-specific Model Parameterization
M.A. Pressly1, G. Clermont1,2, M.D. Neal2,3, J.H. Waters4, A. Jeyabalan5, S. Beck5, R.S. Parker1,2;
Departments of 1 Chemical Engineering; 2 Critical Care Medicine; 3 Surgery; 4 Anesthesiology, 5 Obstetrics,
Gynecology & Reproductive Sciences; University of Pittsburgh, Pittsburgh, PA
Objectives: Venous thromboembolism (VTE) along with adverse hypercoagulation events during pregnancy and in
the postpartum period lead to more severe outcomes for mothers, making up 9 percent of maternal mortality [1].
Developing a statistical model to identify risk factors for postpartum VTE and assisting the synthesis and
parametrization of a compartmental dynamic model of coagulation may lead to improved maternal outcomes.
Methods: The Magee Obstetric Maternal and Infant Database is an extensive obstetric database of over 300
variables for each of 190,000 deliveries [2]. The subset of deliveries performed between 2009 and 2018 includes
55,587 deliveries with complete demographic and process variables. There were 94 VTE events recorded (incidence
1.7 in 1000). Preliminary risk factors were identified using univariate logistic regression (LR). A multivariate LR
model with split sample validation was used to derive the final risk model. A separate model using only variables
present pre-delivery was developed. Model performance was evaluated using area under the receiver operating curve
(AUC) on the test set. The LR prediction of risk informs parameters used in a compartmental model that includes
the uterus and vascular volumes, along with the activated coagulation factor (ACF) concentrations. The dynamic
model accounts for blood flowrates, blood loss during delivery, and activation and deactivation rates of clotting
factors.
Results: Risks factors’ odds ratio and p-values are shown in the Feature Table. Overall accuracy in predicting
thromboembolic events yield an AUC of 0.701. The model using pre-delivery features results in an AUC of 0.659.
From the compartmental model simulation of delivery and the postpartum period, the effect of increasing the
activation rate of ACFs results in increased ACF concentrations and duration above baseline ACF concentrations in
both compartments. This work presents a method of taking patient-specific estimates of apparent risk to inform
simulations of ACF concentrations.
Conclusions: Identifying patients at high risk of thromboembolism after delivery may assist targeting interventions
that mitigate this risk. Identifying relevant risk factors guides the development and parametrization of a dynamic
model of coagulation, which can be used to formulate time-dependent risk and therapeutics.
References: [1] Creanga AA, Berg CJ, Syverson C, Seed K, Bruce FC, Callaghan WM. Pregnancy-Related
Mortality in the United States, 2006–2010. Obstet Gynecol. 2014;125(1): 5–12.
[2] Catov JM. Magee-Womens Research Institute [Internet]. 2017. https://mageewomens.org/
Feature Table: Factors utilized in the predictive model of VTE risk. Italicized features are removed for the pre-
labor prediction.
Feature Odds Ratio P-value
Sickle Cell Crisis 60.1 9.18e-05
Sickle Cell Trait 12.7 0.000439
Uterine Rupture 7.66 0.00468
Structural Heart Disease 5.71 0.000725
Placental Abruption 4.84 0.000675
Hemorrhage 3.95 9.94e-07
Placenta Accreta/Increta/Percreta 3.90 0.0212
Preterm Delivery 3.12 4.64e-06
Hypertension 2.76 0.0286
Previous Cesarean Section 2.42 0.000464
Multiple Gestation 2.01 0.018
W-096
Approximation of Transition Probability to Hazard Rate Can Simplify Simulations, Diagnostic Plots, and
Parametrization of Discrete 2-State Markov Models
Authors: Varun Goel
Institutions: Alnylam Pharmaceuticals, Cambridge, MA
Introduction: Recurrent event data are routine in clinical trials, examples include bleed events in hemophilia and
attacks in acute hepatic porphyria. Statistical methods for handling recurring events include Markov models,
repeated time to event, Poisson model, or negative binomial models. In current example we used mixed effects 2-
state Markov model to characterize recurrent event of attack status, i.e., in attack versus no-attacks, as a function of
biomarker levels and other disease specific predictors [1]. The objective of this work is to show that approximation
of transition probability to hazard yielded closed form solution of data statistics, e.g., mean time to event, mean time
in event, annualized rate of attacks (AAR), that can be used to make diagnostic plots e.g., (IPRED, PRED vs DV for
these statistics). Further, it simplified interpretation from the parameter estimates of such model.
Methods: Mathematically, one can show that transition probability from no-attack(N) to attack (A) (pN→A) estimates
from a Markov model for small delta discrete time interval t can be approximated as a hazard rate. Such
approximation leads to close form solution of important data statistics: mean time to next attack (MTA) = 1/pN→A,
mean duration of attack (MDA) = 1/pA→N, attack cycle (AC) = MTA + MDA; AAR = 365.25 /AC. Probability of no-
attack over an observation period t = exp(-pN→A* t). As these statistics are of clinical interest, one can compute
IPRED, PRED versus DV plots for these variables to assess model fit. Simulations were conducted from the Markov
model to compute individual AAR and compared with closed form solution. Estimation and simulations were done
in NONMEM and R.
Results: Simulations show that long term AAR computed from closed form solution matches closed form solution
AAR greatly simplifying model simulations and interpretation. Alternative parameterization of the model as 1/p
compared to logit(p) yields interpretable estimates as mean tmes.
Conclusions: There is sparse literature on diagnostic plots for Markov models and interpretation from the model
requires simulation from the model which could be time consuming. Approximation of transition probability to
hazard rate provides closed form solution to important data statistics and may also allow handling situations of
censoring and varying discrete observation windows. Approximation of Markov models to other statistical models
for handling recurrent data will be shown.
[1] Varun Goel, Michael Dodds, Sagar Agarwal, Amy Simon, and Gabriel Robbie. Relationship Between Urinary
Aminolevulinic Acid (ALA) Levels and Porphyria Attacks in Acute Hepatic Porphyria Patients in Clinical Trials
with Givosiran; Submitted Abstract at ACOP 2019
W-097
An Online Solution to Match Control Subjects to Renal and/or Hepatic Impaired Patients in
Pharmacokinetic Studies
Guillaume Bonnefois1, Raphaël Vlavonou1, Julie Desrochers1, Pierre-Olivier Tremblay1, Mario Tanguay1, 2
1Syneos Health, Montreal, QC, Canada; 2Université de Montréal, Montreal, QC, Canada.
Objectives: Pharmacokinetic (PK) studies are conducted to assess the effects of renal and/or hepatic impairment on
the drug’s PK and to provide appropriate dosing recommendations, if necessary. These studies should include, for
comparison purposes, a group of control subjects who should be comparable to renal/hepatic patients in terms of
demographics. Various matching strategies can be applied such as “mean matching” or “one-to-one pairing” [1].
Based on the regulatory guidance documents and literature, no well-established “matching” methodology was
defined to ensure appropriate comparability in terms of demographics (e.g. age, body mass index: BMI…) [2-3].
Arbitrary limits were introduced to enable flexibility in the selection of control subjects (e.g. mean age±10 years for
mean matching). To ensure a similar demographics distribution, a computational approach was therefore developed
to propose a more robust and quantitative matching procedure via a web-based application.
Methods: The two commonly used matching strategies were adapted. For mean matching, the mean and standard
deviation (SD) were calculated for each cohort (i.e. mild, moderate, severe). A statistically significant difference
between cohorts’ means implies the control subjects should be matched according to each of the three separate
cohorts. Otherwise, the control subjects should be matched to the pooled cohorts. In that case, the variables of the
control group were then distributed according to an empirical rule, i.e. proportion of patients within 1-, 2-, and 3 SD
of the mean or 68%, 95%, 99%, respectively. For one-to-one pairing, the empirical distribution of demographics was
firstly estimated by the kernel density estimation. The associated probability density function was divided in two
parts or more defined by corresponding intervals and densities. The latter was used to calculate the number of
control subjects in each part.
Results: A computational approach was developed to be applied for both matching strategies through the Shiny
application for renal and/or hepatic impaired patients. Demographics variables can be tested (e.g. age, BMI...) to
develop an integrated tool that would be convenient for clinicians (see Figure).
Conclusions: An R-Shiny application was developed through a user-friendly interface to provide a quantitative
approach to facilitate and guide the selection of control subjects. The tool serves as proof-of-concept and may
require further validation to confirm its practical application to renal and/or hepatic impairment studies using real
patient datasets.
References:[1] Paglialunga S et al. Update and trends on pharmacokinetic studies in patients with impaired renal
function: practical insight into application of the FDA and EMA guidelines. Expert Review of Clinical
Pharmacology. 2017 10(3): 273-283 [2] Guidance for Industry “Pharmacokinetics in Patients with Impaired Renal
Function”, FDA, March 2010. Available at: https://www.fda.gov/downloads/drugs/guidances/ucm204959.pdf [3]
Guidance for Industry “Pharmacokinetics in Patients with Impaired Hepatic Function”, FDA, May 2003. Available
at: https://www.fda.gov/downloads/drugs/guidancecomplianceregulatoryinformation/guidances/ucm072123.pdf.
W-100
Comparison of Bayesian Functionality between PuMaS and NONMEM
Authors: Simon Byrne1, Andreas Noack1, Joakim Nyberg2, Christopher Rackuackas3,4, and Vijay Ivaturi4
Affiliations: 1Julia Computing; 2Pharmetheus, Uppsala, Sweden; 3Massachusetts Institute of Technology,
Cambridge MA; 4Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD,
USA
Objectives: PuMaS.jl, a Julia Language based dynamic simulation and estimation engine for non-linear mixed
effects modeling, provides a Bayesian inference engine. This uses HMC simulation from the DynamicHMC.jl[1]
package for full Bayesian data analysis through its seamless and easy to use Domain Specific Language (DSL). The
objective of this work to compare the results of the Bayesian analysis from PuMaS with that from NONMEM
BAYES.
Methods: The commonly used Theophylline dataset and model available online was used to compare the results
between NONMEM and PuMaS. The tests were conducted on the same system.
The Distributions.jl package is used for both specifying the prior, random effects, and sampling distributions, as well
as evaluating their densities [2]. HMC requires the computation of gradients of the posterior log-density, which we
compute using forward-mode automatic differentiation via ForwardDiff.jl package [3], that works directly with both
the ODE solvers and the distributions objects. In order to improve performance, we exploit the conditional
independence structure of the model by computing the gradient of each subject separately. The posterior sampling is
performed via the DynamicHMC.jl package, which implements the No-U-turn sampler [4] along with automatic
step-size adaptation during burn-in.
Results: There results of the posterior distributions of fixed and random effects matched between NONMEM
BAYES and PuMaS. Figure 1 shows the trace plots of the model parameters. A slight divergence of the results was
noticed in the estimate of one of the omega parameters that is being evaluated. Here we report the results from a
single chain to compare with NONMEM. However, PuMaS has the capability to run multiple chains in parallel
mode without any user overhead of setting up the parallelization.
Discussion: Bayesian inference engine is of one of the many estimation routines for non-linear mixed effects model
available in PuMaS in addition to maximum likelihood and expectation maximization methods. It works out of the
box without any additional changes required with ODE based and analytical models. This work represents our first
demonstration of the results matching existing tools.
References: [1] https://github.com/tpapp/DynamicHMC.jl [2] Mathieu Besançon, David Anthoff, Alex Arslan,
Simon Byrne, Dahua Lin, Theodore Papamarkou, John Pearson (2019) “Definition and modeling of probability
distributions using the JuliaStats ecosystem”, in submission [3] Revels, J., Lubin, M., Papamarkou, T. (2016)
“Forward-Mode Automatic Differentiation in Julia”, arXiv:1607.07892 https://arxiv.org/abs/1607.07892
[4] Hoffman, M. D., & Gelman, A. (2014). The No-U-turn sampler: adaptively setting path lengths in Hamiltonian
Monte Carlo. Journal of Machine Learning Research, 15(1), 1593-1623.
Figure 1: Traceplots and corresponding posterior density plots of theophylline model parameter estimates that are
easily obtained in PuMaS by building a wrapper around MCMCdiagnostics.jl
W-098
Comparison of Bayesian Functionality between PuMaS and NONMEM
Authors: Simon Byrne1, Andreas Noack1, Joakim Nyberg2, Christopher Rackuackas3,4, and Vijay Ivaturi4
Affiliations: 1Julia Computing; 2Pharmetheus, Uppsala, Sweden; 3Massachusetts Institute of Technology,
Cambridge MA; 4Center for Translational Medicine, University of Maryland School of Pharmacy, Baltimore, MD,
USA
Objectives: PuMaS.jl, a Julia Language based dynamic simulation and estimation engine for non-linear mixed
effects modeling, provides a Bayesian inference engine. This uses HMC simulation from the DynamicHMC.jl[1]
package for full Bayesian data analysis through its seamless and easy to use Domain Specific Language (DSL). The
objective of this work to compare the results of the Bayesian analysis from PuMaS with that from NONMEM
BAYES.
Methods: The commonly used Theophylline dataset and model available online was used to compare the results
between NONMEM and PuMaS. The tests were conducted on the same system.
The Distributions.jl package is used for both specifying the prior, random effects, and sampling distributions, as well
as evaluating their densities [2]. HMC requires the computation of gradients of the posterior log-density, which we
compute using forward-mode automatic differentiation via ForwardDiff.jl package [3], that works directly with both
the ODE solvers and the distributions objects. In order to improve performance, we exploit the conditional
independence structure of the model by computing the gradient of each subject separately. The posterior sampling is
performed via the DynamicHMC.jl package, which implements the No-U-turn sampler [4] along with automatic
step-size adaptation during burn-in.
Results: There results of the posterior distributions of fixed and random effects matched between NONMEM
BAYES and PuMaS. Figure 1 shows the trace plots of the model parameters. A slight divergence of the results was
noticed in the estimate of one of the omega parameters that is being evaluated. Here we report the results from a
single chain to compare with NONMEM. However, PuMaS has the capability to run multiple chains in parallel
mode without any user overhead of setting up the parallelization.
Discussion: Bayesian inference engine is of one of the many estimation routines for non-linear mixed effects model
available in PuMaS in addition to maximum likelihood and expectation maximization methods. It works out of the
box without any additional changes required with ODE based and analytical models. This work represents our first
demonstration of the results matching existing tools.
References: [1] https://github.com/tpapp/DynamicHMC.jl [2] Mathieu Besançon, David Anthoff, Alex Arslan,
Simon Byrne, Dahua Lin, Theodore Papamarkou, John Pearson (2019) “Definition and modeling of probability
distributions using the JuliaStats ecosystem”, in submission[3] Revels, J., Lubin, M., Papamarkou, T. (2016)
“Forward-Mode Automatic Differentiation in Julia”, arXiv:1607.07892 https://arxiv.org/abs/1607.07892 [4]
Hoffman, M. D., & Gelman, A. (2014). The No-U-turn sampler: adaptively setting path lengths in Hamiltonian
Monte Carlo. Journal of Machine Learning Research, 15(1), 1593-1623.
Figure 1: Traceplots and corresponding posterior density plots of theophylline model parameter estimates that are
easily obtained in PuMaS by building a wrapper around MCMCdiagnostics.jl
W-099
Using Hybrid OpenMP/MPI Distributed Computing Algorithm to Develop Multilevel Parallel Expectation
Maximization (MPEM) Estimation Methods for Population Pharmacokinetic/Pharmacodynamic Analysis
Author: Chin Feng Ng1, Chee Meng Ng1
Affiliations: 1University of Kentucky at Lexington, USA
Objectives: In theory, two levels of parallelization can be used to accelerate the performance of the parametric
expectation-maximization (PEM) estimation methods in population pharmacokinetic/pharmacodynamic (PPKPD)
analysis1. However, all PEM algorithms in the commercial PPKPD software (NONMEM and Phoenix) only use a
single level of parallelization (L1P). The limitations of the L1P become apparent for complex PPKPD analysis as
solving a large system of ordinary differential equations (ODEs) with thousands of simulated random model
parameters sets (ISAMPLE) for each individual subject is extremely compute-intensive and become the rate-limiting
step in developing complex PPKPD model. In this study, we used a hybrid OpenMP/MPI distributed computing
algorithm to develop multilevel parallel parametric expectation-maximization (MPEM) that designed to break the
computational bottleneck of the existing PEM algorithms in the commercial software for PPKPD analysis.
Methods: A MPEM method based on Monte-Carlo PEM algorithm facilitated by Maximum a Posterior
(MCPEMMAP) with two levels of parallelization was developed and written in c. The first parallelization level (L1P)
used MPI to assign the E-step computation of individual subject to different computing nodes. The second
parallelization level (L2P) used OpenMP to breakdown and allocate the E-step computation within the same subjects
to separate computing nodes. Three different PPKPD dataset with 100 simulated subjects and intensive sampling
designs were used for the analysis: 1) two compartment linear PK model with analytical solution (PK2A), 2) two
compartment linear PK model expressed as a system of two ODEs (PK2N), and 3) PKPD model with targeted-
mediated drug disposition that consisted of three ODEs and 20 models parameters (TMDD). Runtimes of the
MEPM with a single CPU core (M-S), and L1P/L2P combinations (M-L12P) were recorded and compared.
Runtime of MCPEMMAP in NONMEM 7.3 (METHOD=IMPMAP) with a single CPU core (N-S) and L1P only (N-
L1P) was determine and used as an external reference for comparison. The performance of different methods were
assessed in the linux-based 20 computing-cores cluster with Intel Xeon E3-1240 v3 3.5GHz.
Results: The effects of additional L2P on model runtimes were depended on model complexity (Table 1). The best
model runtimes was achieved with L12P combination of 10x2 for PK2N and TMDD. The calculated time gain for
this L12P combination (10x2) from L1P only (20x1) was 8.30% and 13.1% for PK2N and TMDD, respectively.
Due to relatively inefficient of the OpenMP distributed computing algorithm, increased the workload of the
OpenMP compared to MPI from 10x2 to 5x4 increased the model runtimes.
Conclusions: The implementation of extra parallelization level L2P is able to overcome the computational
bottleneck of L1P in accelerating the performance of parametric EM in PPKPD analysis. The time gain of the L12P
compared to L1P only is increased with increased model complexity.
References: 1. Ng CM. AAPSJ 2013;15(4):1212-21
Table 1: Runtimes of Different Methods (min)
PK2A PK2N TMDD
NO
N M E M
Serial (N-S) 1 17.9 60.3 343
L1
P x
L2
P
(N-L
1P
)
20 x 1 2.00 6.20 50.3 M
PE
M
Serial (M-S) 1 7.48 34.2 101
L1
P x
L2
P
(M-L
12
P)
20 x 1
0.667
2.65
14.5
10 x 2
0.850 2.43 12.6
5 x 4 1.07 4.22 18.7
ISAMPLE=10,000 and Number of EM Iteration=100; The maximum computing cores for L1P and L12P in this
study was 20.
W-100
Impact and Implementation of a Quality Management System in a Pharmacometric Workflow
Brittany Walker1, Bela Patel1, Vikram Sinha1
1Merck & Co., Inc.
Objectives: The pharmacometrics community relies on various scientific platforms, processes and standards to
integrate knowledge, enable drug development decisions, and enhance regulatory submissions. As the impact of the
pharmacometric work continues to expand, the scrutiny of this work expands with it. Notably, the FDA is placing an
increased emphasis on quality, as demonstrated by an increased number of warning letters in recent years1,2.
Furthermore, costs incurred as a result of data quality issues are generally found to be higher than the costs of
ensuring high quality data3. To ensure high quality work and impact, a Quality Management System (QMS)
Framework is needed to implement cohesive solutions that will facilitate a consistent and reproducible
pharmacometrics workflow.
Methods: Building a QMS Framework based upon a strong understanding of the existing pharmacometric workflow
requires a thorough understanding of its gaps and inefficiencies. An independent gap analysis and/or mock audit are
effective methods to obtain this information. In the pharmacometrics space, workflow requirements and related gaps
typically fall into three categories: data integrity (quality control, maintaining blinding in blinded studies), daily
operations (SOP’s, guidances), and platforms (databases, systems, tools). A unified QMS team was formed, with
sub-teams each responsible for the continuous understanding and improvement of a specific aspect within each of
these three categories. These sub-teams facilitated focused yet cohesive solutions to frequently complex gaps within
the pharmacometric workflow.
Results: Implementation of a quality by design QMS structure encouraged development of a synergistic,
standardized, and consistent workflow (Figure). Advancements in data integrity will increase the quality, reliability
and traceability of analyses. Improvements in daily operations created a more efficient and reproducible
pharmacometric workflow, providing sufficient documentation and audit readiness. Enhancements to IT platforms
(particularly the creation of an integrated modeling workbench) enabled a continuous data flow with corresponding
audit trail, in addition to enabling new technologies and pharmacometric methodologies.
Conclusions: While QMS efforts require up front resources, the cost reduction associated with an increase in quality
and efficiency, and the reduction in re-work, remediations, and regulatory risk is expected to outweigh the initial
resourcing expenditure. The outcomes of this effort are expected to have long term impact on the quality and
reliability of the pharmacometric work for both internal decision making, external submissions and, publications.
The sustainability of these benefits relies on wide-spread adoption of the QMS Framework throughout the
pharmacometric community.
Figure: Proposed QMS framework for a typical pharmacometric workflow.
References: 1. FDA Data Dashboard. (2018, November 1). Retrieved from
https://datadashboard.fda.gov/ora/cd/inspections.htm 2. FDA Data Dashboard. (2018, November 1). Retrieved from
https://datadashboard.fda.gov/ora/cd/complianceactions.htm 3. Haug, Anders, Frederik Zachariassen, and Dennis
Van Liempd. "The costs of poor data quality." Journal of Industrial Engineering and Management 4.2 (2011): 168-
93.
W-101
Disease Progression Modelling for Duchenne Muscular Dystrophy Using North Star Ambulatory Assessment
and Forced Vital Capacity
Authors: Xiaoxing Wang1, Lutz O Harnisch1, Camille Vong1, Beth A Belluscio1, Doug Chapman1, Vivek S Purohit1
and the CINRG Investigators2
Affiliations: 1. Pfizer Inc. 2. Cooperative International Neuromuscular Research Group
Objectives: This is a longitudinal study using secondary data collection from an existing Duchenne Muscular
Dystrophy (DMD) natural history data base. The intent of this study is to quantitatively describe the disease
progression of DMD in boys at various ages and stages of their disease to inform future clinical development
programs.
Methods: Longitudinal natural history data were collected by the Cooperative International Neuromuscular
Research Group (CINRG) and later provided to Pfizer. Independent longitudinal models using age as the time
variable, were developed for the North Star Ambulatory Assessment (NSAA) total score for ambulatory boys
(n=156, age=4-20 years), and for the forced vital capacity percent predicted value (FVC%p) for non-ambulatory
boys (n=236, age=7-32 years), respectively.
NSAA total score was Rasch transformed into a linearized scale (0-100) [1]. Data was analyzed with nonlinear
modelling approach as implement in NONMEM. Data on the boundaries were considered censored and handled
using M3 method [2].
Covariates were assessed both graphically and statistically for their influence on the natural disease progression.
They included race, genotype, baseline body weight, baseline calculated height and baseline steroid use status.
Stable steroid use was defined as continuous use for ≥6 months at the first available record of the specific endpoint.
Results: NSAA vs. age was described adequately using a 2-phase linear model, including an early
growth/improvement phase and a later decline phase. The maximum NSAA which can be achieved by the end of the
growth phase is higher, on average, in patients with stable steroid treatment at baseline (77.1 vs. 66.7 points), and
who have a mutation amenable to exon 44 skipping (89.2 vs. 66.7 points).
FVC%p vs. age in non-ambulatory subjects was described adequately using a sigmoidal-shape model structure
adapted from the Hill equation. The estimated parameters included a baseline function at an age of 7 years, a time to
reach 50% baseline (ET50) and a Hill coefficient. Patients receiving stable steroid treatment at baseline have a larger
ET50 (10.99 vs. 9.07 years), suggesting a less rapid decline in pulmonary function.
The model can adequately describe the disease progression and variability within both the ambulatory and non-
ambulatory populations, based on the assessment of model diagnostics and VPC plots.
Conclusions: Using natural history data, longitudinal models were developed which adequately describe disease
progression for key endpoints in ambulatory and non-ambulatory DMD boys. Differences among subgroups based
on race, genotype and baseline steroid use were quantified. Steroid treatment was shown to have a significant effect
on disease progression in both ambulatory and non-ambulatory populations. Via simulations, the developed models
can be utilized to inform future clinical development programs in DMD.
References: 1. Mayhew AG, et al. Dev Med Child Neurol. 2013;55(11):1046-52. 2. Ito K, et al. J Pharmacokinet
Pharmacodyn 2012;39(6):601–618