automated quantitative toxicogenomic dose-response modeling

Automated Quantitative Toxicogenomic Dose-Response Modeling Burgoon, L.D.1,2, Boverhof, D.R.2, Zacharewski, T.R.1,2

1Toxicogenomic Informatics and Solutions, LLC, Lansing, Michigan 48909

2Department of Biochemistry & Molecular Biology, National Food Safety & Toxicology Center

and the Center for Integrative Toxicology, Michigan State University, East Lansing, MI 48824

AbstractToxicogenomics applied to dose-response studies yields hundreds of putative biomarkers of exposure and toxicity. Phenotypic anchoring, functional annotation, and pathway analysis may distinguish differential gene expression associated with toxicity from adaptive responses or exposure indicators thus facilitating the identification and development of mechanistically-based biomarkers of toxicity that can be used clinically or in population studies to monitor and assess risk. Typically linear or other low-dose extrapolation models are used to identify points of departure. Manually fitting these mathematical models is difficult and time-consuming, especially when many responses (e.g., differential gene expression) do not exhibit sigmoidal, linear, or exponential dose-response characteristics. ToxResponseModeler, a Java application, applies high throughput fitting 1) sigmoidal, 2) linear, 3) quadratic, 4) exponential, and 5) modified guassian function models to address this limitation. It identifies the most suitable model for each gene using the particle swarm optimization algorithm (PSO) to identify the optimal set of parameters. The best fitting models are then compared, and the optimal model is chosen for each gene based on the Euclidean distance between the predicted model and observed data. Doses can then be calculated at nth percent effective dose (EDn) including points of departure, using the optimal model. Vehicle-based points of departure are also calculated based on the intersections of the 95% or 99% confidence intervals for the vehicle and dose-response data. Collectively, this yields a point of departure with a known confidence interval based on measurement variance. The utility of ToxResponseModeler is demonstrated using published toxicogenomic dose-response hepatic data from TCDD treated C57Bl/6 mice. This work has been supported in part by NIEHS Superfund grant P42 ES04911.

Continuous Dose-Response Models

ToxResponseModeler Results by Gene Function

This work supported in part by the Superfund Basic Research Program: P42 ES04911

● E-mail: [email protected] ● http://www.txisllc.com ● http://dbzach.fst.msu.edu

ToxResponseModeler:Automated Dose-Response Modeling (ADRM) Application

Summary• The ToxResponseModeler utilizes an automated dose response modeling (ADRM) algorithm based on particle swarm optimization (PSO) to identify the best mathematical model for continuous dose response data

•Results from the ToxResponseModeler identified the distribution of ED50 and probabilistic point of departure (POD) values

•The ToxResponseModeler identified putative TCDD-sensitive functional gene categories

•Highly sensitive and novel putative BOEs were identified using the POD data, which will be further characterized in the future

Figure 1: ADRM Algorithm and Probabilistic Vehicle-based Point of Departure

The ADRM fits a mathematical model to the dose-response data from a feature (i.e., gene, metabolite, etc…). The parameters for the model are identified using the Particle Swarm Optimization (PSO) algorithm. PSO is an iterative algorithm that finds the best combination of parameter levels for the mathematical model resulting in a best fit model. At the start of the algorithm PSO assigns particles to cliques, and particles are only influenced by members of its clique. The goal of each particle is to “move” toward the goal, only using information from within the clique.

Using the optimal model, parameters such as the ED50 or EC50 can be identified. The model can also be used to calculate confidence intervals that can be used to identify probabilistic vehicle-based points of departure.

Model Name

Model Parameters

Exponential k = scaling parameterc = shift parameter

Linear m = slopeb = intercept

Normal β = shape parameter {Real}λ = ymin

γ = rate parameter > 0ε = scale parameter {Real}

Quadratic a = direction and scaleb = shapec = y-intercept

Sigmoidal y0 = minimum

response levela = max (Y) – y0

b = scaling; b > 0j = slope

ckaY x

cbxaxY 2

)(10 jxeb

ayY

bmxY

2

22exp

2xY

Table 1: Continuous Dose-Response ModelsFive classes of model that best represent the range of shapes seen in toxicogenomic data are simultaneously fit to the continuous dose-response data. The dose term for all of the models is denoted by the variable x.

TCDD Dose-Response Study Design

Figure 2: Dose-Response Design

Dose-response data were obtained from Boverhof, et al (2005, Tox Sci 85 (2): 1048-1063).

Animals were treated with 0.1mL of vehicle or 0.001, 0.01, 0.1, 1, 10, 100, or 300ug/kg TCDD and sacrificed 24hrs post exposure.

Summary of ToxResponseModeler ResultsTCDD (24hr)

Differentially Expressed Genes (Features) 238 (278)

Genes Exhibiting Sigmoidal Dose-Response 157 (78%)

ED50 Range (ug/kg; based on features) 0.01 – 177.70

ED50 0 – 2ug/kg 75 features

ED50 2 – 50ug/kg 129

ED50 2 – 10ug/kg 88

ED50 10 – 30ug/kg 34

ED50 30 – 50ug/kg 7

ED50 50+ug/kg 11

Probabilistic Point of Departure (POD) (ug/kg) 0.01 – 266.09

Table 2: Summary of Results

278 active features were identified from the study, corresponding to 238 active genes.

ED50 ranges are reported as number of features as different features may probe different gene regions, and result in different ED50 and probabilistic point of departure (POD)values

Gene Name SymbolPOD (ug/kg)

ED50 (ug/kg) Class

Cell Cycle

cell division cycle 37 homolog (S. cerevisiae) Cdc37 0.64 0.57 Highly Sensitive

retinoblastoma-like 2 Rbl2 0.85 0.63 Highly Sensitive

cyclin D3 Ccnd3 1.08 0.84 Highly Sensitive

Jun proto-oncogene related gene d1 Jund1 0.75 0.90 Highly Sensitive

cyclin D3 Ccnd3 0.93 1.05 Highly Sensitive

amyloid beta precursor protein binding protein 1 Appbp1 13.07 5.89

Moderately Sensitive

RIKEN cDNA 4921532D01 gene

4921532D01Rik 7.68 7.02


thioredoxin-like 4 Txnl4 27.99 11.37Moderately Sensitive

cyclin-dependent kinase inhibitor 1A (P21) Cdkn1a 9.13 17.35


SMC4 structural maintenance of chromosomes 4-like 1 (yeast) Smc4l1 31.11 18.89


Oxidative Stress

glutathione S-transferase, alpha 4 Gsta4 0.83 0.79 Highly Sensitive

NAD(P)H dehydrogenase, quinone 1 Nqo1 0.25 0.99 Highly Sensitive

glutathione synthetase Gss 11.78 10.25Moderately Sensitive

glutathione reductase 1 Gsr 91.63 91.34 Resistant

Ubiquitin Pathway

huntingtin interacting protein 2 Hip2 8.86 4.81Moderately Sensitive

ubiquitin-conjugating enzyme E2E 2 (UBC4/5 homolog, yeast) Ube2e2 8.00 5.09


amyloid beta precursor protein binding protein 1 Appbp1 13.07 5.89


autophagy-related 12 (yeast) Atg12 5.02 8.61Moderately Sensitive

ubiquitin-conjugating enzyme E2H Ube2h 14.63 9.65


ubiquitin-like 5 Ubl5 13.33 9.90Moderately Sensitive

Gene Name SymbolPOD (ug/kg)

ED50 (ug/kg) Class

Apoptosis

huntingtin interacting protein 1 Hip1 0.90 0.74 Highly Sensitive

hypoxia inducible factor 1, alpha subunit Hif1a 0.69 0.84 Highly Sensitive

amyloid beta precursor protein binding protein 1 Appbp1 13.07 5.89 Moderately Sensitive

caspase 6 Casp6 10.16 6.18 Moderately Sensitive

Bcl-2-related ovarian killer protein Bok 266.09 44.26Moderately Responsive

Oxidoreductase

cytochrome P450, family 1, subfamily a, polypeptide 1 Cyp1a1 0.06 0.07 Highly Sensitive



xanthine dehydrogenase Xdh 0.72 0.59 Highly Sensitive

dehydrogenase/reductase (SDR family) member 3 Dhrs3 0.59 0.72 Highly Sensitive

UDP-glucose dehydrogenase Ugdh 0.44 0.91 Highly Sensitive

aldehyde dehydrogenase 16 family, member A1

Aldh16a1 0.87 0.95 Highly Sensitive

NAD(P)H dehydrogenase, quinone 1 Nqo1 0.25 0.99 Highly Sensitive

Lipid and Fatty Acid Metabolism

lipoprotein lipase Lpl 0.79 0.72 Highly Sensitive

very low density lipoprotein receptor Vldlr 4.63 6.08 Moderately Sensitive

acyl-CoA thioesterase 7 Acot7 6.78 8.19 Moderately Sensitive

very low density lipoprotein receptor Vldlr 8.15 8.33 Moderately Sensitive

abhydrolase domain containing 5 Abhd5 8.43 8.41 Moderately Sensitive

L-3-hydroxyacyl-Coenzyme A dehydrogenase, short chain Hadhsc 71.43 34.54

Moderately Responsive

Table 3: Dose-Response Modeling Assigns TCDD-sensitivities to Functional Pathways

Cell cycle, oxidative stress, apoptosis and lipid and fatty acid metabolism pathways exhibit genes which are highly sensitive, moderately sensitive, and resistant to TCDD exposure. The oxidoreductase genes tend to be highly sensitive to TCDD, and include members of the AhR gene battery. Genes within the ubiquitin pathway tend to be moderately sensitive. All of the genes listed here demonstrated a sigmoidal dose-response relationship.

Most Sensitive Biomarkers of Exposure

Gene NameGene

SymbolPOD

(ug/kg)ED50

(ug/kg)

a disintegrin and metallopeptidase domain 2 Adam2 0.01 0.01

cytochrome P450, family 1, subfamily a, polypeptide 1 Cyp1a1 0.05 0.10

whey acidic protein Wap 0.07 0.11

dermatan sulphate proteoglycan 3 Dspg3 0.08 0.11

cutA divalent cation tolerance homolog (E. coli) CutA 0.10 0.07

PAK1 interacting protein 1 Pak1ip1 0.12 0.56

Table 4: ToxResponseModeler Identifies Biomarkers of Exposure

Highly sensitive biomarkers of exposure (BOEs) represent genes whose expression at low exposure levels can be reliably differentiated from untreated and vehicle populations.

automated quantitative toxicogenomic dose-response modeling

Documents