Machine learning enhanced process design in protein a chromatography

Area of Science:

• Biopharmaceutical Process Development
• Computational Chemistry
• Machine Learning in Drug Discovery

Background:

• Quality by Digital Design (QbDD) aims to accelerate biopharmaceutical development by reducing reliance on physical experiments.
• Identifying the design space is a critical bottleneck in QbDD, often requiring expensive, high-fidelity models.
• Current methods for design space identification are time-consuming and resource-intensive.

Purpose of the Study:

• To develop a machine learning-enhanced method for efficient design space identification in QbDD.
• To leverage synthetic data and transfer learning to overcome limitations of traditional approaches, especially under data scarcity.
• To demonstrate the applicability of this approach across different data availability scenarios (high, moderate, low).

Main Methods:

• Utilized a transfer learning framework combined with synthetic datasets generated from mechanistic models.
• Developed an artificial neural network (ANN) model for classifying feasible design regions.
• Evaluated the ANN model's performance under high, moderate, and low data availability (HDA, MDA, LDA) conditions.

Main Results:

• The data-driven method showed strong performance in HDA.
• Transfer learning significantly improved model accuracy in MDA and was crucial for performance in LDA.
• The machine learning approach effectively identified feasible design regions, demonstrating its utility.

Conclusions:

• Machine learning, particularly with transfer learning, offers a powerful and efficient solution for design space identification in QbDD.
• This approach can significantly reduce the cost and time associated with early-stage biopharmaceutical process design.
• The developed method holds potential for streamlining QbDD implementation and accelerating drug development timelines.