in silico modeling approaches towards …establishing robust design and control ... * undey et al.,...

PROCESS DEVELOPMENTCENK ÜNDEY, PhD

IN SILICO MODELING APPROACHES TOWARDS ESTABLISHING ROBUST DESIGN AND CONTROL STRATEGIES: BIO/PHARMA INDUSTRY PERSPECTIVE

4th PQRI/FDA Conference on Advancing Product Quality, Rockville, MD, Apr.9-11, 2019

2

• Process and product development paradigm– Prior knowledge, empirical, first principles, ML approaches

• Case Studies:– First principles modeling based UFDF for buffer composition

predictions– Drug product T/P filling process modeling – Process monitoring and predictive control examples– Incorporating PAT and models towards control strategy

OUTLINE

3


• Case Studies:– First principles modeling based UFDF for buffer composition predictions– Drug product T/P filling process modeling– Process monitoring and predictive control examples– Incorporating PAT and models towards control strategy

OUTLINE

4

IN SILICO MODELING HAS SIGNIFICANT POTENTIAL TO INCREASE SPEED AND EFFICIENCIES OF PROCESS AND PRODUCT DEVELOPMENT

Pre-PT FIH Development

Clinical Support

Commercial Process

DevelopmentCommercial

AdvancementCommercial

Support

First principles predictive modeling(reduced wet exp.)

Cell Line SelectionCell Line Development

Increased data flow, digitalization and in silico model use across the product lifecycle

Process and Product PerformanceManagement

Predictive modeling for QTPP

-40

-30

-20

-10

0

10

20

Dis

solv

ed O

xyge

n

Cul

ture

pH

Air

Flow

O2

Flow

CO

2 Fl

ow

Bio

reac

torL

evel

Agi

tatio

n

Ves

sel P

ress

ure

Tem

pera

ture

Con

tribu

tion

Variable

N-3_N-2_N-1_NDModXPS Contribution M5:N

(0010066284:0.0638889 - Average:0.0638889), Weight=RX[4]

SIMCA-Batch On-Line Client 3.3.0.2 - 2010-09-28 14:59:20 (UTC-5)

Robust design (6σ)Patient centric

Feedback loop

Design Qualification Continued Verification

5

RISK ASSESSMENT, DESIGN SPACE AND CONTROL STRATEGY ARE KEY IN SETTING UP PERFORMANCE-BASED CONTROL

* CMC Biotech WG, 2009, A-Mab: A Case Study in Bioprocess Development. V2.1

*

6

PRIOR KNOWLEDGE ASSESSMENT AND AGGREGATING HISTORICAL DATA FOR PLATFORM PROCESSES OFFER UNIQUE ADVANTAGES TOWARDS PROCESS DESIGN AND PROCESS PERFORMANCE QUALIFICATION (PPQ)

Courtesy of Dr. Roger Hart

7

• Much of the current modeling has its roots in engineering, biology, chemistry, physics and fluid mechanics (mass transfer, kinetics, etc.)

• In some cases, current scientific understanding and/or analytical resolution insufficient to utilize mechanistic modeling

ADVANCES IN COMPUTATIONAL POWER AND SCIENTIFIC UNDERSTANDING AREENABLING MORE MODELING OF BIOPHARMACEUTICALS PD & MFG

Q11 - Design and conduct studies (e.g., mechanistic and/or kinetic evaluations, multivariate design of experiments, simulations, modeling) to identify and confirm the links and relationships of material attributes and process parameters to

drug substance CQAs

• Convergence of disciplines: in silico modeling to enable Smart PDcomputational platforms

• in silico experiments for Process Characterization• Predictive modeling for RM selection

and variation control• in silico tooling and device design

8

WE ARE ADVANCING THE USE OF FIRST PRINCIPLES MODELING IN PROCESS, PRODUCT DEVELOPMENT AND IN MANUFACTURING

TODO: Chrom

Fluid VelocityAppliedForce

9

THE MODELING PROCESS (IDEALIZED)

4. SolutionAnalysis

3. NumericalApproximation

2. ProblemFormulation

1. DomainDefinition

10

Calibration and validation of first-principles model required a small number of targeted experiments

• a fraction of conventional DOE-based process characterization approaches

More operating parameters and wider ranges explored via fast and inexpensive “in-silico” experiments

• investigating parameters difficult to study experimentally (e.g., bed compaction)

• replacing costly or time-demanding experiments(e.g., column size)

Reducing costly and time consuming experimentation

MODEL-BASED PROCESS DEVELOPMENT

11

Automated Cytopathic Effect (CPE) Detection via Deep Learning

WE HAVE BEEN ADVANCING DEEP LEARNING APPLICATIONS TO CREATE SIGNIFICANT EFFICIENCIES TOWARDS ACCELERATING PD

Deep CNN

ANN

https://www.tutorialspoint.com/python_deep_learning/python_deep_learning_deep_neural_networks.htm

12



OUTLINE

13

GOAL: ACCURATELY PREDICT pH AND EXCIPIENT CONCENTRATIONS IN PROTEIN DRUG SUBSTANCE

Challenge: High concentrations of charged mAb during UF/DF commonly result in offsets in pH/excipients between UF/DF pool and formulation buffer• Molecular basis for this offset is the selective retention or rejection of ions due to:

– Charge interactions between ions and the protein: ions attracted or repelled– Volume exclusion: solutes excluded from volume occupied by high concentration of mAb

14

FIRST PRINCIPLES UNDERSTANDING OF UF/DF AND BUFFER EXCIPIENTS WAS USED VIA IN SILICO MODELING TO REDUCE WET EXPERIMENTS DURING PC

Model Verification

Model Prediction

15

IN SILICO DOE: MECHANISTIC MODEL OF EXCIPIENT EXCHANGE DURING PROTEIN ULTRAFILTRATION & DIAFILTRATION (MUD)

• (Right→) In-silico DoE performed in MUD to explore the formulation buffer design space for targeted experimentation

• (Below↓) Worst-case buffer conditions were tested experimentally and results agreed with MUD predictions

Buffer Conditions

Result Source

Acetate (mM)

Phe(mM)

Sorbitol (%)

Osmolality (mOsm) pH

Buffer Formulation #1

| Exp. – Pred. |2.0 2.1 0 3 0.02

Buffer Formulation #2 0.3 0.4 0.31 7 0.02

Process robustness:Model accurately predicts UF/DF pool pH

16



OUTLINE

17

TIPCOM RECOMMENDATIONSFOLLOWED, SUBSTITUTINGLOGISTICALLY PREFERRED TUBING

EXTENDED DOSING TIME TOACCOMMODATE TUBING(TDOSE > 0.6 S)

CPK: 5.0-9.0

BENCHTOP FILLER

TIPCOM USE IN DP FILLING PROCESS DEVELOPMENT

• TARGET FILL VOLUME• VISCOSITY• DENSITY• VIAL GEOMETRY

SKU CHARACTERISTICS

• ORIFICE GEOMETRY• NEEDLE DIMENSIONS• TUBING DIMENSIONS• PINCH VALVE GEOMETRY

HARDWARE CHARACTERISTICS

Formalisms

INPUT ODEPDE(FEA,CFD)

FILLCHARACTERISTICS

MOTORDISPLACEMENT

DOSINGCONTROL

PRODUCTTEMPERATURE,

PRESSURE

FLOW INFILLING SET

FLOW INCOMPRESSED

TUBE

RECIPE

ALGEBRAIC

TIPCOM

DESIGN SPACE IN-SILICO PROCESS CHARACTERIZATION DESIGN CANDIDATES

LSL USLTarget

10

20

30

40

2.1 2.15 2.2 2.25

OFFLINE CONFIRMATION

<1 HOUR TO COMPLETE IN-SILICOPROCESS CHARACTERIZATION

32 FILLING SETS, 6720 RECIPE SETTINGS EXPLORED• 6 USABLE FILLING SETS IDENTIFIED• 2 FILLING SETS W/ RECIPES RECOMMENDED

TIPCOM-RECOMMENDED RECIPES

TIPCOM FILLING PROCESS CHARACTERIZATION PRIOR TO B6 BENCHTOPCHARACTERIZATION ENABLED:

33% TIME SAVINGS IN WET EXPERIMENTS

ELIMINATION OF EXPERIMENTS AT COMMERCIAL LINE

ASSURANCE OF OPTIMAL RECIPE SELECTION

HIGHLY CAPABLE FILLING PROCESS (CPK: 5.0-9.0)

AN EVOLUTION IN DEPLOYMENT

• IMPROVED USER INTERFACE

• CUSTOMIZED EQUIPMENT-BASED APPS

• SITE- AND LINE-SPECIFIC WORKFLOWS

• AUTO-GENERATED PDF REPORTS

• CROSS-PLATFORM

• NO INSTALLATION REQUIRED

• VALIDATED (TECH REPORT)

TIPCOM 3.0

18



OUTLINE

19

BIOPHARMACEUTICAL PROCESSES, COMPRISED OF CONSECUTIVE UNITS, GENERATE ABUNDANT DATA: PERFORMANCE-BASED CONTROL IS CRITICAL

20

• Per process characterization, culture duration and seeding density are positively correlated to PQA

• A reliable real-time prediction model for PQAs enables improved control

PERFORMANCE PREDICTION MODEL CAN BE USED TO ADJUST CULTURE DURATION FOR OVER SEEDED BATCHES

. . .

Seed

ing

Den

sity

Batches

Cul

ture

Dur

atio

n

Batches

Batches

PQA

Over-seeded batch

Batch kept at target

duration

High PQA

Batch kept at lower

duration

Seed

ing

Den

sity

Culture Duration

Acceptable PQA

Production Bioreactor BDS

21

MULTIVARIATE PREDICTIVE MODEL DEVELOPED AND DEPLOYED FOR PERFORMANCE-BASED CONTROL FOR SEED BIOREACTORS

Open-loop control

22

KPI’S AND PQA’S ARE PREDICTED TO CONTROL BIOREACTOR CULTURE HARVEST*

* Undey et al., 2015, Predictive Monitoring and Control Approaches in Biopharmaceutical Manufacturing, European Pharmaceutical Review, 20(4), pp. 63-69 .

Open-loop control

23

• Hierarchical modeling helps identifying variability within and across unit operations between process and product performance variables in real time

• Batch fingerprints are generated to compare batch behavior to historical batches

Unit Op 1 . . .

MV

Cha

rts

Key

/Crit

ical

OPs

MV

Cha

rts

for

CPP

/KPP

s

MV

Cha

rts

for

PQAs

. . .

. . .

Batch Fingerprint-UOP1

. . .

Unit Op 2 Unit Op NBDS

Phas

e N

Phas

e 1 . . .

Phas

e 2

ENHANCED CONTINUED PROCESS VERIFICATION (eCPV) CAN ENABLE HOLISTIC PERFORMANCE-BASED MONITORING AND CONTROL

Batch Fingerprint-UOP NBatch Fingerprint-UOP2

24

eCPV HELPS IDENTIFYING VARIABILITY WITHIN AND ACROSS UNIT OPERATIONS BETWEEN PROCESS AND PRODUCT PERFORMANCE VARIABLES

Unit Op 1 Unit Op 2

Phas

e 1

Phas

e 2

Phas

e N

Phas

e 1

Phas

e 2

. . . BDSPQA 1

PQA 2

PQA N

M54 M29

-2

-1.5

-1

-0.5

0

0.5

1

1.5

Unit Op 1 Unit Op 2

-10

0

10

20

30

40

50

Variable XM45 M46 M47 M48 M49 M50 M51 M52 M53

-10

-5

0

5

10

15

20

25

Phase N

Model suggested an increase in PQA2 in BDS pool was correlated to a Phase N in Unit Operation 2

10276302

10276304

10276307

10276318

1027632110276322

10276324

10287475

1028848110288482

10297587

10297588

1029772010297721

1029772210297723

10297724

10297725 102977261029772810297729

10297730

1030046410300465 10300467

10300468

10300469

10300470

10300471

10300472

10300473

1024677610246777 10276299

10276300

1027630510276306

10276317

10276319

10276320

10287474

10297585

1029758610297719

10297727

-200

-150

-100

-50

0

50

100

150

-150 -100 -50 0 50 100t[1]

Phas

e N. . .. . .

25

MACHINE LEARNING-BASED PREDICTIVE METHODS ARE PROVEN PROMISING: COMBINING DATA FROM DIFFERENT SCALES IMPROVED THE CQA PREDICTIONS

MFG Scale Only Combined

Training MFG Site A (8) MFG Site A (8) + Bench-scale (29)

Testing MFG Site B (6) MFG Site B (6)

Convolutional Neural Networks

Input Data(Predictors)

CQ

A

26



OUTLINE

27

CONTINUOUS MANUFACTURING BIOREACTOR OPERATION: CONTROL AND LOT STRATEGY

Key considerations for the Cell Culture process

• Supporting high cell densities for extended durations

• Cell separation at high cell densities

• Perfusion rates, media formulation, liquid handling

• Lot strategy

• Detect and segregate NC material

00 Time

Biom

ass

Initiate material collection

Cell Expansion

Lot #1

Steady-State Production

Lot #2 Lot #3 Lot #N

28

TARGETED HIGH MANNOSE CONTROL IS ACHIEVED USING MPC METHODSlide courtesy of Jack Huang

Automation – Closing the loop for PQA Control

29

OVERVIEW OF GLUCOSE FEEDBACK CONTROL STRATEGY

Glucose Control Logic:

1. The control equation is based on mass balance of the system

2. The control equation is checked at a set frequency to determine if glucose addition is required

Actual vs. Predicted Glucose

Residual Plot

MVA Model

Raman Spectra

Raman AnalyzerBioreactor

Real time Glucose

ValueDeltaV

Automation System

Logic:

If Glucose value is less than Glucose

target

Turn pump

on

Real-Time Glucose Feedback Control

Glucose MVA Model

Development

• Other typical cell culture measurements (e.g., lactate, glutamine, VCD, viability, pCO2, osmolality, etc.) can also be correlated to Raman spectra

Glucose Pump

(Source: Watson-Marlow)

Automation – Closing the loop with Process Analytical Chemistry Tools

30

MODEL PREDICTIVE CONTROL FRAMEWORK ENABLES PERFORMANCE-BASED CONTROL

31

EXPERIMENTS WERE USED TO DETERMINE THE APPROPRIATE LEVERS TO CONTROL PQA; FEED A AND FEED B ACT IN OPPOSITE DIRECTIONS AS IDEAL LEVERS FOR MPC


32

WE WERE ABLE TO MEET PQA TARGET WITHIN +/- 2.5% OF ITS DESIRED VALUE USING MPC IN PRODUCTION BIOREACTOR

• Two control points were available in this run• MPC used to add Feed A and/or Feed B at control points

Control Points


33

First-principlesmodels

PAT Method 1: MALS for Controlling %HMW species

PAT Method 2: First principles model for controlling %HMW species

* Rathore et al., 2010, Large Scale Demonstration of a Process Analytical Technology Application in Bioprocessing, Biotechnology Progress, 26(2), 448.

ROBUSTNESS AND CONTROL OF OPERATING SPACE CAN BE IMPROVED USING PAT AND FIRST PRINCIPLES MODELS

34

Every patient, every time

CONCLUSIONS

In silico first principles modeling:

• is an instrumental tool for rapid process design and knowledge-based manufacturing

• enables an efficient use of experimental efforts• enables a richer characterization of the robust design space• can enable next-generation process monitoring and control

applications

Augmented with PAT, sensors and in silico models offer advanced process performance management capabilities

Artificial Intelligence and Machine Learning-based in silico models are emerging, offering unique opportunities in bio/pharma

35

• Tony Wang• Seyma Bayrak• Sinem Oruklu• Pablo Rolandi• Xiaoxiang Zhu• Fabrice Schlegel• Richard Wu• Jack Huang

ACKNOWLEDGEMENTS

• Will Johnson• Chris Garvin• Myra Coufal• Gerd Kleemann• Deirdre Piedmonte• Stephen Brych• Justin Ladwig

• Roger Hart• Karin Westerberg

in silico modeling approaches towards …establishing robust design and control ... * undey et al.,...

Documents