Benchmarking co-folding methods to predict the structures of covalent protein-ligand complexes

Area of Science:

• Drug discovery and structural biology
• Computational chemistry and bioinformatics

Background:

• Targeted covalent inhibitors (TCIs) offer enhanced drug properties but their rational design is challenging.
• Accurate prediction of covalent protein-ligand complex structures is crucial but lacks robust benchmarks.
• Co-folding approaches show promise in biomolecular modeling but their performance in covalent complex prediction is underexplored.

Purpose of the Study:

• Introduce CoFD-Bench, a benchmark dataset for evaluating methods predicting covalent protein-ligand complex structures.
• Systematically assess classical docking and deep learning co-folding models on covalent complex prediction tasks.
• Provide insights into the strengths and limitations of current computational methods for TCI design.

Main Methods:

• Developed CoFD-Bench, a dataset of 218 covalent complexes.
• Evaluated classical docking tools (AutoDock-GPU, CovDock, GNINA) and co-folding models (AlphaFold3, Chai-1, Boltz-1x).
• Assessed ligand RMSD accuracy, protein-ligand interaction recovery, performance on novel pairs, and computational efficiency.

Main Results:

• Co-folding methods outperform classical docking in accuracy and interaction recovery.
• Co-folding performance degrades on novel pocket-ligand pairs, whereas classical methods show stable but modest results.
• Co-folding methods are computationally slower than classical approaches; AlphaFold3 shows potential for identifying covalent residues via noncovalent co-folding.

Conclusions:

• CoFD-Bench provides a rigorous evaluation framework for covalent complex prediction methods.
• Co-folding models offer higher accuracy for TCI design but face scalability challenges.
• Findings guide future development of co-folding-based TCI design strategies and model improvements.