Addressing Small Data Challenges in Biopharmaceutical Development and Manufacturing: A Mini Review of Multi-Fidelity...

Area of Science:

• Biopharmaceutical Development
• Computational Biology
• Machine Learning

Background:

• Biopharmaceutical development is lengthy and costly, often exceeding a decade.
• Machine learning (ML) can accelerate this process, but requires large, diverse datasets.
• Current ML approaches face challenges with data acquisition for high-fidelity (HF) training.

Purpose of the Study:

• To review multi-fidelity ML techniques for biopharmaceutical surrogate modeling.
• To provide recommendations for identifying low-fidelity (LF) and HF data.
• To highlight opportunities for applying these methods beyond current research areas.

Main Methods:

• Review of surrogate modeling techniques including Gaussian processes, neural networks, and physics-informed ML.
• Integration of abundant LF data with limited HF data for model training.
• Analysis of data requirements and identification strategies for LF and HF datasets.

Main Results:

• Multi-fidelity ML significantly reduces development costs and timelines by optimizing data acquisition.
• Gaussian processes and neural networks are currently dominant, but emerging models show promise.
• Opportunities exist to apply these methods to downstream processing, not just upstream.

Conclusions:

• Multi-fidelity ML is a powerful strategy to accelerate biopharmaceutical development.
• Careful selection of LF and HF data is crucial for successful model implementation.
• Future research should explore advanced ML architectures like Transformers and diffusion models.