Interaction-Based Inductive Bias in Graph Neural Networks: Enhancing Protein-Ligand Binding Affinity Predictions From...

Area of Science:

• Computational Biology
• Machine Learning

Background:

• Machine learning (ML) models for protein-ligand binding affinity (PLA) prediction vary in inductive bias, affecting generalization and interpretability.
• Existing ML methods often lack alignment with biological binding mechanisms, limiting predictive power and mechanistic understanding.

Purpose of the Study:

• To propose an interaction-based inductive bias for ML models to improve PLA prediction.
• To develop an explainable heterogeneous interaction graph neural network (EHIGN) embodying this bias.
• To ensure predictions are grounded in biologically relevant atomic interactions.

Main Methods:

• Representing protein-ligand complexes as heterogeneous graphs with covalent and non-covalent interactions.
• Assuming PLA is the sum of pairwise atom-atom affinities from non-covalent interactions.
• Implementing EHIGN to model pairwise atom-atom interactions from 3D structures.

Main Results:

• EHIGN demonstrated superior generalization capability compared to state-of-the-art ML baselines in PLA prediction and virtual screening.
• Analyses confirmed the interaction-based inductive bias guides learning of physically realistic atomic interactions.
• EHIGN accurately predicted Nirmatrelvir efficacy against SARS-CoV-2 variants, providing meaningful explanations.

Conclusions:

• The proposed interaction-based inductive bias enhances ML models for PLA prediction.
• EHIGN offers improved generalization and interpretability by focusing on physically relevant atomic interactions.
• This method shows practical utility in drug efficacy prediction and understanding variant-specific interactions.