How modeling and simulation are streamlining biopharmaceuticals webinar part 1

Concise summary

The webinar, presented by Dr. Edward Close (Siemens Digital Industry Software), explains how science-based mechanistic and hybrid bioprocess digital twins are used end-to-end across development, engineering, and manufacturing to speed up, de-risk, and optimize biopharmaceutical development and production.

Key scientific concepts and phenomena

• Bioprocess digital twin: a comprehensive virtual representation of a bioprocess used to understand, predict, optimize, monitor, and support decision-making for its physical counterpart.
• Degrees of digital integration:
  • Digital model: manual data transfer (offline).
  • Digital shadow: automatic physical → digital data flow; digital → physical still manual.
  • Digital twin: fully automated bi-directional data flow.
• Mechanistic models: differential-equation models derived from physics/chemistry/biology whose parameters have physical meaning; typically large (many thousands of equations) and simulate dynamics (critical quality attributes from critical process parameters).
• Hybrid models: mechanistic cores augmented by data-driven elements (e.g., neural networks, PLS) to fill gaps in first-principles knowledge while preserving mechanistic structure.
• Uses across the lifecycle:
  • R&D and engineering: often digital models or shadows.
  • Manufacturing: shadows or full twins.
  • Potential for automated process development: closed-loop experiments guided by a digital twin.
• Benefits: reduce physical experiments, shorten timelines, lower material/staff/lab costs, improve robustness, yield, and productivity, enable soft-sensing and real-time optimization, and support regulatory science priorities.

Case studies / technical examples

Chromatography gradient optimization (peptide & oligonucleotide polishing)

• Problem: product-related impurity (n−1) elutes close to the target.
• Method: rate-model chromatography calibrated with data from three linear-gradient experiments (samples every column volume; species concentrations measured by HPLC); adsorption isotherm calibrated.
• Outcome: model identified an optimal gradient in about one week using far less material; matched and was confirmed by experiment; replaced a months-long DOE.

Monoclonal antibody bioreactor feed optimization

• System: fed-batch production with three daily nutrient feeds and a temperature shift.
• Mechanistic model: segregated cell-population model (viable cycling, viable resting, dead, lysed) plus phenomenological kinetics for growth, death, lysis, nutrient consumption, metabolite secretion, and product formation.
• Data: 22 experiments (7 for calibration, 15 for external validation; samples typically taken daily).
• Challenge: poor prediction at high seeding densities and temperature-affected kinetics.
• Solution: hybrid model embedding an artificial neural network (ANN) as a correction factor applied in parallel with the kinetic model.
• Outcome: improved prediction for difficult runs; used with input variability to identify a more robust biocapacitance-based feed strategy versus fixed feeding; experimentally confirmed.

Single-pass tangential flow filtration (TFF)

• Method: mechanistic model calibrated to experimental runs; used to predict behavior during manufacturing-expected flow rate changes.
• Outcome: close match to experiments; used design-space/contour plots to demonstrate process robustness for regulatory support.

Model-building and deployment methodology (typical workflow)

1. Select mechanistic model(s) representing relevant unit operations (use established libraries where available).
2. Collect targeted calibration experiments (design experiments suitable for parameter estimation).
3. Calibrate model parameters (prefer physically meaningful parameters).
4. Validate externally with blind/holdout runs.
5. Where mechanistic gaps exist, introduce hybrid components (ANN, PLS) to model unexplained behavior; calibrate these components with additional data.
6. Use the validated model to:
   • Explore design space and robustness (contour plots, variability analysis).
   • Optimize control strategies (e.g., feeding, gradients).
   • Replace or augment physical experiments in development.
   • Deploy online for soft-sensing, monitoring, and real-time optimization (digital shadow/twin).
7. Iterate as more data and understanding become available.

Practical and operational points

• Mechanistic and hybrid models reduce experimental burden by encoding prior scientific knowledge.
• Hybrid approaches have accelerated industrial adoption by addressing gaps in first-principles knowledge.
• Models typically require fewer but better-targeted experiments; experimental design for calibration matters.
• Modeling and simulation are recognized as a regulatory science priority.

“In the last decade, Modeling and Simulation has become firmly established as a regulatory science priority.” — U.S. Food and Drug Administration (quoted in the webinar)

• Siemens provides an integrated library/platform to build, calibrate, and deploy end-to-end mechanistic/hybrid bioprocess twins.

Researchers and sources featured

• Dr. Edward (Ed) Close — Head, Bioprocess Modeling Competency Centre, Siemens Digital Industry Software.
• Angela — webinar host (named in the transcript).
• Siemens Digital Industry Software — provider of the described software/services.
• University College London — where Dr. Close obtained his EngD (in collaboration with Pfizer).
• Pfizer — collaborator on Dr. Close’s EngD.
• U.S. Food and Drug Administration (FDA).
• Industry publication by leaders from 14 major biopharmaceutical manufacturers and suppliers (referenced; no individual authors named).
• Leading academics and Siemens partners/customers (mentioned collectively; no individual names given).