Gaussian Processes for Predictive QSAR Modeling of Chromatographic Processes.

Area of Science:

• Biopharmaceutical Manufacturing
• Chemical Engineering
• Computational Chemistry

Background:

• Chromatography is essential for protein purification in biopharmaceutical production.
• Current process design and optimization are complex and time-consuming due to extensive design spaces.
• Market demands require faster, more efficient, and generalized process development frameworks.

Purpose of the Study:

• To introduce Gaussian processes (GPs) as a machine learning methodology for predictive modeling in chromatography.
• To evaluate the performance of GPs for resin and solvent condition selection in quantitative structure-activity relationship (QSAR) modeling.
• To demonstrate the utility of GPs for model-assisted optimization and interpretability in bioprocess development.

Main Methods:

• Application of Gaussian processes for predictive modeling in chromatography.
• Quantitative structure-activity relationship (QSAR) modeling for resin and solvent condition selection.
• Comparative analysis of GP predictive power against other machine learning algorithms.
• Derivation of feature importances from GP models.

Main Results:

• Gaussian processes demonstrate predictive power comparable to other leading machine learning algorithms.
• GPs provide crucial confidence estimates for predictions, enabling model-assisted optimization.
• Feature importances can be derived from GPs, offering interpretability similar to random forests.

Conclusions:

• Gaussian processes present a robust and efficient framework for biopharmaceutical chromatography process design and optimization.
• The interpretability and confidence estimation capabilities of GPs enhance their suitability for complex bioprocess development.
• This machine learning approach can accelerate the delivery of biopharmaceuticals to market.