Leveraging Uncertainty in Machine Learning Accelerates Biological Discovery and Design.

Area of Science:

• Computational biology
• Machine learning in drug discovery

Background:

• Machine learning (ML) models often fail when encountering data outside their training distribution.
• Quantifying prediction uncertainty is crucial for ML models to reliably handle novel biological phenomena.

Purpose of the Study:

• To demonstrate the utility of robust uncertainty prediction in biological discovery.
• To enable ML models to gracefully handle novel biological data and accelerate scientific breakthroughs.

Main Methods:

• Leveraging Gaussian process-based uncertainty prediction on pre-trained features.
• Training a model on a small dataset (72 compounds) for predictions on a larger library (10,833 compounds).

Main Results:

• Identification and experimental validation of compounds with nanomolar affinity for diverse kinases.
• Discovery of compounds exhibiting whole-cell growth inhibition of Mycobacterium tuberculosis.
• Demonstration of uncertainty prediction's generalizability across protein engineering and single-cell transcriptomics.

Conclusions:

• Robust uncertainty prediction is essential for reliable biological discovery using machine learning.
• Integrating uncertainty quantification into the ML experimental lifecycle enhances AI's role in scientific advancement.