Evaluating generalization in protein-ligand cofolding methods.

Area of Science:

• Computational Biology
• Drug Discovery
• Structural Bioinformatics

Background:

• Deep learning has advanced protein structure prediction.
• Predicting protein-ligand interactions is crucial for drug discovery.
• Existing cofolding methods lack comprehensive performance evaluation due to insufficient benchmarking datasets.

Purpose of the Study:

• To comprehensively evaluate leading all-atom cofolding methods.
• To introduce a new, high-resolution benchmark dataset, Runs N' Poses, for protein-ligand systems.
• To assess the capabilities of current deep learning methods for predicting protein-ligand interactions.

Main Methods:

• Developed and utilized the Runs N' Poses benchmark dataset, containing 2,600 protein-ligand systems.
• Evaluated four state-of-the-art all-atom cofolding methods.
• Assessed method performance against the new benchmark, focusing on generalization beyond training data.

Main Results:

• Current cofolding methods demonstrate a tendency to memorize ligand poses from their training data.
• Performance on the Runs N' Poses dataset indicates limited generalization capabilities.
• The benchmark dataset highlights the shortcomings of existing approaches for novel predictions.

Conclusions:

• Existing cofolding methods are not yet suitable for de novo drug design due to memorization.
• The Runs N' Poses dataset provides a realistic benchmark for future method development.
• Further research is needed to develop cofolding methods that generalize effectively for accurate protein-ligand interaction prediction.