Molecular embedding-based algorithm selection in protein-ligand docking

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Molecular docking algorithm selection is challenging due to context-dependent performance.
• No single algorithm reliably performs across all structural, chemical, and protocol variations.
• Existing methods often lack adaptability and generalizability.

Purpose of the Study:

• To introduce MolAS, a lightweight algorithm-selection model for molecular docking.
• To predict per-algorithm performance using pretrained protein and ligand embeddings.
• To improve upon the single best solver (SBS) and approach the virtual best solver (VBS) performance.

Main Methods:

• Utilized pretrained protein and ligand embeddings.
• Employed attentional pooling and a shallow residual decoder for performance prediction.
• Evaluated MolAS across five molecular docking benchmarks with hundreds to thousands of labeled complexes.

Main Results:

• MolAS achieved up to a 15 percentage-point absolute improvement over the single best solver (SBS).
• The model closed 17-66% of the gap between the virtual best solver (VBS) and SBS.
• Performance was most effective with low oracle entropy and separable top-solver regions, but degraded under protocol mismatch.

Conclusions:

• MolAS offers a robust, embedding-based approach to docking algorithm selection.
• Its effectiveness is linked to oracle landscape characteristics and protocol stability.
• The model serves as an in-domain selector and a diagnostic tool for assessing selection well-posedness.