process modeling and control challenges in the ...€¦ · process modeling and control challenges...
TRANSCRIPT
Process Modeling and Control Challenges in the Pharmaceutical Industry
Phil Dell’OrcoTeam Leader, Process Safety and Design,
GSK Pharmaceuticals
Outline• Introduction: What do pharma engineers do?• Current Practice: what do we do for modeling and
control now?– Reaction engineering– Crystallization– Drying
• Better Practice: What are we working towards?• Some examples• Needs
– infrastructure– technical
Realities of Pharmaceutical Processes: High Attrition, Few Campaigns
• Numerous early stage candidates which will not survive from Phase I to Phase II/III
• Early objectives (first plant, large scale lab campaigns)
– enable chemistry– not much eng input
• Over 80% of batches in our pilot plants are one-off
• Perhaps only 5-6 total campaigns prior to filing
• Late stage objectives (Phase II supply and beyond)
– develop detailed models– determine critical parameters– understand equipment limitations
Early Toxicology, Phase I (30-50 candidates)
Phase III/NDA (1-3 candidates)
Number of Compounds
Introduction: Pharmaceutical Research and Development•We are a batch process industry
– essentially no continuous processes– why?
• Many products don’t make it to market• Batch reactors offer flexibility over continuous reactors in that
they are adaptable for many products and unit operations • Volumes continuously change through the lifetime of a product.
Again, batch reactors are adaptable• GMP
•We are an industry dominated by chemists– many plant campaigns are one-off; little engineering needed– sometimes difficult to get sufficient eng. Input to projects
Chemistry Dominated IndustryRatio of Chemical Engineers to Chemists Across Industry
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8 GSK (total)GSK (w/o plant)MerckMerck AGPharmaciaNovartisEli Lilly Astra ZenecaBMSDuPont MerckAventisSchering PloughAbbott
1:3 1.5 1.7 1:10 1:12
FTE ratio Chemical Engineers
to Chemists*
In general, chemists dictate the way processes are conceived and executed. Engineering input is often secondary
Tech
nolo
gy T
rans
fer
Process Control D
esign
Plant Configuration (1st make)
Plant Configuration (2nd make)
Plant Configuration (3rd make)
Plant Configuration (4th make)
Plant Configuration (5th make)
Final process
Rte optim
isation
Rte developm
ent
New
route
Rte selectionsupport
Process flowoptimisation
Unit operationoptimisation
Che
mic
al E
ngin
eerin
g A
ctiv
ity
Route Design Chemistry Optimisation Robustness StudiesSynthetic Chemistry Activity
Process Engineering - Input Level to Projects
Process designoptimisation- Heat/mass transfer-crystallization/particle eng.-Separations- Environmental, etc..
P.A.T.
Time
Key Unit Operations- Synthetic Chemicals
• Extractions: In general, every process stage has an average of greater than 1 extraction.
• Distillations (Batch): on average, 1 distillation per stage• Reactions: A given chemical stage has 1-4 chemical
transformations• Crystallizations: Critical for defining physical properties of final
drug product• Filtration: Both final product and intermediates • Drying: Every final product is dried. Probably most empirical
field.• Environmental: solvent recovery, biotreatment, incineration• Chromatography: infrequent currently, but becoming more
common
Key Unit Operations: Biopharmaceuticals
• Fermentations- most critical step. Determines complexity of separations chain and product yield
• Ultrafiltration- removes cellular material from fermentation broth
• Exclusion chromatography- purification by size separation
• Ion chromatography- purification by ionic charge• Dialysis- buffer exchanges
Manufacturing Limitations to Complex Control StrategiesWe are in general “Low Tech” with respect to Reaction Engineering
• Many of our vessels in pilot plant and manufacturing are jack of all trades, but master of none
– nearly 70% of existing GSK vessel capacity is glass lined, radial impeller (e.g., retreat curve, crow’s foot, cryobat)
– little flexibility for difficult mixing problems
• Roughly 50% of our manufacturing capacity relies on crude control
– “banded” control using following options: steam, hot water, coldwater, fridge, jacket hold, jacket drain
• We have little in the way of sophisticated process analytics which might provide opportunity for better control.
Current Practice: Reaction Engineering
Collect kinetic data, often
semi-quantitative
Estimate parameters, or determine
boundary conditions
Determine crude optimum either
graphically, through DOE
Program DCS Sequence to obtain crude
optimum
• Do to large number of early phase compounds, expedient but not perfect methods required (80% engineering)
• We often use sub-optimal data (IR traces, reaction calorimetry, HPLC PAR’s) and model in these domains
• While parameter estimation is often numerically performed (not always!), optimization tends to be crude
• Tools: MatLab, HiQ, DynoChem
“Crude” Modeling: Using ReactIR data to improve a reaction• Rate constant k1 (formation of methylthiol product and rate constant of impurity
formation k2: k1 / k2 ≈ 75– Improve ratio of rate constants by changing reaction temperature.
Time (s)300000 6000 12000 18000 24000 30000
Conc. (mol/L)
1.2
0.0
0.3
0.6
0.9Reactant, predProduct, predImp, predReactant, actProduct, act
• For 100% conversion of starting thiol:
– 1.048 eq MeI– 94.4% Product– 5.6% Impurity
• More MeI only increases di-Methyl impurity %
Ar SH Ar SMe + KI + H2OKOH, MeI
H2O
Desired Practice: Reaction Engineering
Collect kinetic data,
quantitative
Estimate statistically significant
parameters, or determine boundary conditions
Use parameters to define an
optimal operating policy
Program DCS Sequence to
obtain optimum
Improve optimization
based on further experimentation
• Would like to improve level and quality of modeling• Will never be “perfect” due to time constraints, but better
optima, better extrapolation possible• Broaden use of optimization tools?
Reaction Engineering: Biopharmaceuticals• Biopharmaceutical fermentations tend to be very complex systems, with
many potentially important parameters– shear– dissolved gas levels (CO2, O2)– pH– fermentation by-products
• Many side reactions possible for proteins: similar reactive sites (e.g., a monoclonal antibody may have multiple histidine groups)
• Downstream purification operations are extensive: perhaps 10-20 unit operations to get fermentation product into a formulated product
• Single batches typically run for ~ 15 days. End value of batch can be 1 million US for 1000 L fermentor
Biopharmaceutical Modeling Problem: Deamidation of a Protein
][][
][][2][
][*2**][
2
21
10
deAmidAkdt
deAmidBd
deAmidAkmABkdt
deAmidAd
mABkkViabilityNdt
mABdcell
=
−=
−=Neat Protein/Buffer Protein/Buffer w/Cells
• Many unknowns in model: how do various parameters affect cell growth rate constant (different for almost any two experiments)
• How would optimum harvest time vary across natural variability of ko
Need for new process analytical technologies: mechanistic inferences, potentially kinetic information to drive model development. Example: electrospray mass spectrometry, real-time data
Typical Setup for Reaction Monitoring by ES/MS
PUMP #2
PUMP #3
PUMP #1 HPLC VIALS
PUMP #4
MassSpectrometer
100:1 FlowSplitter
TemperatureController
Proposed Reaction Mechanism Using Observed Intermediate Reaction Mechanism proposed to be that proposed by Knoevenagel, rather than that proposed by Hann and Lapworth
CH O
OO
OH
R
H
O
R'N
HC O
OO
OH
RN
+
H HN
H
O+
R'
HH
C O
OO
OH
R
N+
H
R' C O
OO
OH
RR
O
O
R
-H2O
-CO2
N
H
-
O
O
OO
H
RN
H
R'
[A][P]
[A-] [P+] [B] [BP+][A-]
[A-][L][M][N]
m/z 419
m/z 338
m/z 320
m/z 86
Current Practice: Crystallization
Data Collection: solubility, MSZW,
kinetic time scales, crude particle size
measurements
Preliminary Crsytallization
Design
Models: typically simple
Robustness studies around
mixing, seeding, addition rates, etc.
Final Crystallization
Design
• Semi-quantitatively driven– analytical issues: what is particle size?
• Data collection techniques: probably underdeveloped relative to other fields
Solubility Measurement: Example of basic collected data using IR methods. Automated, rapid, accurate, and routine.
0
0.02
0.04
0.06
0.08
0.1
0.12
0.14
0.16
0.18
0.2
200 300 400 500 600 700 800
Time (min)
Solu
te C
onc.
(g/
g so
lven
t)
0
10
20
30
40
50
60
Tem
pera
ture
(C)
Solute Concentration, TolueneSolute Conc., 1:1 EtAc/CyclohexaneSolute Conc., n-butanolTemperature
System calibrated in 0.5 days for 4 solvents; PLS fit to calibration standards. ~ 6 g material required, ~ 4 g recovered. In ~24 hours, solubility mapped out in 4 solvents.
Kinetic Information: Qualitative Desupersaturation Monitoring
• For a typical crystallization design, crystallization monitoring occurs after measuring solubility, MSZW
• Time scales for crystal growth under different seeding, cooling conditions are inferred from turbidity,
Lasentec, and IR signals• From this basis, boundary
conditions are determined• Qualitative Design 0
5000
10000
15000
20000
25000
30000
0 2000 4000 6000 8000
Time (seconds)
Lase
ntec
Cou
nts
(1-5
00 m
icro
ns)
0
20
40
60
80
100
120
Turb
idity
Out
put/T
emp
(C)
Lasentec Counts
Turbidity Output
Temp (C)
Desired Practice: CrystallizationData Collection:
solubility, MSZW, kinetic data, better particle
size, desupersaturation
measurements
Preliminary Crsytallization
Design
Use models for senstivity
analysis on systems
Additional kinetic data, paremeter
estimation
Validate model, improve
parameter estimation
Perform robustness experiments
based on sensitivity
Implementation of control on plant (as opposed to DCS sequence) is possible
due to high value of product
• Need for improved analytical methods for data collection• While we may never be able to model accurately in the domain of FDA
accepted particle size methods, we can use models to indicate sensitivity of process to changes. This would improve robustness of process.
• Due to high value of product at this stage, active control strategies on plant would attract investment
• Some good academic activity in this area (Rawlings, Braatz)
Drying: The forgotten field
• Drying influences product sizes perhaps more strongly than crystallization unit operations, yet there is very little research in drying modeling
• How do we explain/correlate particle breakage in filter dryers (f(mechanical properties, geometry))
• How do we explain agglomeration phenomena?
200
220
240
260
280
300
320
340
360
slurry
ex re
actor
after
drying
after
scree
ning
x90
0.13 m^2 filter dryer, non-sonicatedTumble Dryer, non-sonicated
020406080
100120140160180200
slurry
ex re
actor
after
drying
after
scree
ning
x90
0.03 m^2 filter dryer, non-sonicated0.13 m^2 filter dryer, non-sonicatedTumble Dryer, non-sonicated
Industry Needs: Infrastructure
•More hiring of graduates from control and modelingprograms
– ChE hires tend to be from crystallization, chemistry oriented programs---> tend to have weaker math/modeling skills
• Integration of best practice optimization algorithms, approaches into commercial software (e.g. MatLab)
• Training and implementation sessions from modeling and control community
Industry Needs: Technical (1)• Integration of detailed mixing effects into modeling• Consideration of pharma used process analytical tools
(Raman, IR, Lasentec?, NIR) in development of control schemes: how would we take complex data, reduce it quickly, and allow its feedback?
• Some academic effort into modeling drying processes• Close work with PAT community to improve model
development and control strategies for crystallization and reaction engineering
• New PAT’s to enable better models: electrospray MS? Ultrasound? Etc...
Industry Needs: Technical (2)
• Additional resource into crystallization modelingand control
• Detailed consideration of biopharmaceutical reaction engineering: how to best model cellular systems
Summary• As an industry, we are in general not state of the art in
modeling and control• Pharma is chemistry rather than engineering driven• Significant scale-up constraints due to plant configurations• Significant infrastructure and technical needs that can be
addressed by modeling and control community• Advancement in modeling is relatively straightforward;
advancement in control is longer term