Target-driven machine learning-enabled virtual screening (TAME-VS) enables prioritization of AKR1C3 inhibitors
Machine learning-guided virtual screening has the potential to accelerate early-stage drug discovery, but its practical utility is often limited by uncertain generalization to different chemical scaffolds and by insufficient prospective evaluation un...
Key Details
Machine learning-guided virtual screening has the potential to accelerate early-stage drug discovery, but its practical utility is often limited by uncertain generalization to different chemical scaffolds and by insufficient prospective evaluation under realistic experimental constraints. To address these challenges, we conduct a target-driven virtual screening study using TAME-VS against aldo-keto reductase 1C3 (AKR1C3), an enzyme implicated in cancer progression and chemoresistance. By integrating automated target expansion, bioactivity data acquisition, and supervised learning, we construct classification models that leverage homolog-derived information to address data sparsity. Retrospective validation using inhibitor series absent from the training data demonstrates generalization to chemically different scaffolds beyond similarity-based prioritization. Prospective validation on newly designed derivatives shows that model-guided prioritization enriches active inhibitors relative to rational medicinal chemistry design alone under constrained synthesis capacity. S34-1035 and S34-1040 were further identified as potent and selective AKR1C3 inhibitors, and were shown to restore doxorubicin sensitivity in resistant breast cancer cells, establishing functional relevance. Building on these findings, a publicly accessible online server with a graphical user interface was further developed for the validated workflow of TAME-VS, enabling reproducible application by the broader community.
Why This Matters
[Draft — editorial context needed]