Title: Algorithm Selection in Protein-Ligand Docking via Molecular Embeddings

Abstract:

Docking algorithm selection is inherently context-specific, as no single approach demonstrates consistent reliability across diverse structural, chemical, and procedural conditions. To address this, we introduce MolAS, a streamlined model designed to predict the performance of individual docking algorithms. MolAS leverages attentional pooling and a shallow residual decoder to analyze pretrained embeddings of proteins and ligands. Evaluated across five docking benchmarks, the model utilizes datasets ranging from hundreds to a few thousand labeled complexes. In these tests, MolAS delivered an absolute performance gain of up to 15 percentage points compared to the single-best solver (SBS). Furthermore, it successfully bridged 17% to 66% of the performance gap between the Virtual Best Solver (VBS) and the SBS.

Investigations into selection frequencies, margin-conditioned reliability, and the structure of benchmark-level oracles reveal that MolAS performs optimally in scenarios where the workflow-defined oracle landscape exhibits low winner entropy and a distinct, separable region for top-performing solvers. However, its efficacy diminishes when protocol mismatches occur, as these shifts alter solver rankings and modify the induced labels. These findings indicate that, within the studied domain, robustness is constrained more by the instability of solver hierarchies driven by workflow and protocol factors than by limitations in representational capacity. Consequently, MolAS is best positioned as an in-domain selector for fixed pipelines and as a diagnostic instrument for evaluating the suitability of docking algorithm selection tasks.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC