Integrating hybrid modeling and high throughput screening: A modular process development platform for flowthrough...

Area of Science:

• Biopharmaceutical Manufacturing
• Process Development
• Chromatography

Background:

• Biopharmaceutical industry demands efficient processes for complex therapeutics.
• Flowthrough and frontal loading chromatography offer high capacity and simplicity.
• Current high throughput screening (HTS) with mechanistic modeling faces challenges in model calibration and capturing complex effects.

Purpose of the Study:

• Develop machine learning (ML) models for predicting protein loading and yield in chromatography.
• Utilize HTS plate-based data as sole input for ML model training.
• Create a platform for seamless integration of HTS and hybrid modeling in process development.

Main Methods:

• Developed neural network and symbolic regression models using a database of general rate model (GRM) simulations.
• Created novel analytical expressions for ideal chromatography conditions, treating transport limitations as perturbations.
• Validated ML models against GRM simulations and experimental data, considering experimental uncertainty.

Main Results:

• ML models accurately predict protein loading and yield using only accessible HTS inputs.
• Neural networks showed highest accuracy; symbolic regression offered interpretability and simplicity.
• ML model predictive errors were comparable to or smaller than mechanistic simulations under typical experimental uncertainty.

Conclusions:

• Proposed platform integrates HTS and hybrid modeling for efficient bioprocess development.
• ML models translate HTS data into practical guidance for resin and condition selection.
• Predicted purities, yields, and productivities align well with experimental column performance.