A generalizable deep learning framework for structure-based protein-ligand affinity ranking.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Estimating protein-ligand binding affinities is vital for drug discovery.
• Current methods face a trade-off between accuracy and speed.
• Machine learning (ML) models struggle with generalizability, failing on novel proteins or chemical structures.

Purpose of the Study:

• To develop a generalizable machine learning model for predicting protein-ligand binding affinities.
• To overcome limitations of current ML models that fail on unseen data.
• To investigate the hypothesis that spurious correlations hinder ML model generalizability.

Main Methods:

• Introduced CORDIAL (COnvolutional Representation of Distance-dependent Interactions with Attention Learning), a deep learning framework.
• Designed CORDIAL with an inductive bias for learning distance-dependent physicochemical interaction signatures.
• Explicitly avoided direct parameterization of protein and ligand chemical structures, focusing on interactions.

Main Results:

• CORDIAL demonstrated maintained predictive performance and calibration through leave-superfamily-out validation.
• This approach effectively simulates prediction on novel protein families.
• Contemporary ML models showed degraded performance under similar validation conditions.

Conclusions:

• Encoding task-specific physicochemical principles into ML architectures enhances generalizability.
• CORDIAL offers a validated strategy for developing robust, generalizable models for structure-based drug discovery.
• The interaction-only approach in CORDIAL is effective for predicting binding affinities across diverse protein families.