Docking-score landscapes shape active-learning performance across Vina, Glide, and SILCS.

The rapid expansion of large chemical libraries has created a need for virtual screening workflows that are both efficient and accurate. Active learning (AL) offers a scalable strategy by iteratively training surrogate models to prioritize promising compounds and reduce the number of required docking calculations. However, direct benchmarking of active-learning protocols across docking engines remains limited. In this study, we ask whether docking scores generated by different docking engines affect AL performance and what factors underlie these differences. To do so, we compare four active-learning virtual screening workflows, Vina-MolPAL, Glide-MolPAL, SILCS-MolPAL, and Schrödinger's active-learning Glide, across multiple protein targets and library sizes. Performance was assessed by recovery of top-scoring molecules according to each workflow's respective docking engine, score-prediction accuracy, chemical diversity, and computational cost. Vina-MolPAL achieved the highest top-1% recovery at a 1% batch size, whereas SILCS-MolPAL achieved comparable recovery at a larger batch size. Latent-embedding analyses suggest that the docking-score landscape strongly influences active-learning performance. In addition, combining active learning with SILCS provides a computationally efficient, membrane-aware approach for screening compounds at transmembrane binding sites.