Deep Confident Steps to New Pockets: Strategies for Docking Generalization.

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Accurate blind docking is crucial for biological breakthroughs, but current methods struggle with generalizability across diverse protein targets.
• Existing benchmarks do not adequately evaluate the generalization capabilities of docking methods.
• Machine learning (ML)-based docking models exhibit limited ability to generalize to unseen proteins.

Approach:

• Developed DockGen, a novel benchmark utilizing protein ligand-binding domains to rigorously assess docking model generalizability.
• Analyzed scaling laws for ML-based docking, identifying data and model size as key factors for improvement.
• Introduced synthetic data strategies to enhance model generalization capacity.
• Proposed Confidence Bootstrapping, a new training paradigm leveraging diffusion and confidence models for improved multi-resolution generation.

Key Points:

• Existing ML docking models show weak generalization on the new DockGen benchmark.
• Scaling data, model size, and using synthetic data significantly improves generalization.
• Confidence Bootstrapping enhances docking performance on unseen protein classes.
• The study sets new state-of-the-art performance across benchmarks.

Conclusions:

• DockGen provides a rigorous assessment for generalizable blind docking.
• Optimized ML models and Confidence Bootstrapping advance the goal of accurate, proteome-wide blind docking.
• This work paves the way for more reliable computational drug discovery and biological insights.