Benchmarking all-atom biomolecular structure prediction with FoldBench.

Area of Science:

• Computational Biology
• Structural Biology
• Biophysics

Background:

• Accurate prediction of biomolecular complex structures is crucial for understanding biological processes and designing therapeutics.
• Deep learning models have advanced biomolecular structure prediction, including proteins, nucleic acids, ligands, and ions.
• Existing benchmarks lack comprehensiveness, hindering rigorous assessment of model performance and generalizability.

Purpose of the Study:

• To introduce FoldBench, an extensive benchmark dataset for evaluating biomolecular complex structure prediction models.
• To assess the performance and generalizability of current deep learning models across diverse prediction tasks.
• To identify key challenges and limitations in predicting various biomolecular interactions.

Main Methods:

• Developed FoldBench, a dataset comprising 1522 biological assemblies across nine distinct prediction tasks.
• Evaluated multiple deep learning models on FoldBench, analyzing performance across different interaction types.
• Investigated performance dependencies on factors like training set similarity and molecular diversity.

Main Results:

• Ligand docking accuracy decreases with reduced ligand similarity to training data, mirroring protein-protein interaction trends.
• Antibody-antigen predictions remain highly challenging, with failure rates exceeding 50% for current methods.
• AlphaFold 3 demonstrated superior accuracy across most prediction tasks compared to other evaluated models.

Conclusions:

• FoldBench provides a crucial resource for rigorous evaluation of biomolecular complex prediction models.
• Significant progress has been made, but limitations persist, particularly in predicting novel ligand interactions and antibody-antigen complexes.
• Future model development should focus on improving generalizability and addressing specific challenging interaction types.