The use of artificial intelligence (AI) is increasingly integral to the drug-discovery process, and among AI-driven methodologies, deep generative models stand out as one of the most promising approaches for hit identification and optimization. Here, we report a retrospective benchmarking analysis of a series of tubulin inhibitors, 3-aroyl-1,4-diarylpyrroles (ARDAP), using the deep-generative algorithm Molecule Optimization by Reinforcement Learning and Docking (MORLD) in combination with five docking software (QuickVina 2, AutoDock-GPU, PLANTS, GOLD, and Glide). Our results indicate that the performance of the MORLD/docking workflow is highly dependent on the availability of initial structural information; only the incorporation of a core constraint in Glide yields satisfactory predictions. To address this limitation, we developed a docking-free variant of MORLD that exploits receptor-derived shape similarity and pharmacophore alignment. Kernel-density estimation, convergence analysis, and SMARTS-based success-rate metrics confirmed that this Shape-Pharmacophore implementation autonomously generates chemically valid, SAR-consistent analogues of the reference compounds. Collectively, this work demonstrates a practical, structure-only driven paradigm for reinforcement-learning-based compound optimization, thereby extending the reach of AI-enabled drug design beyond traditional docking workflows.

Department of Drug Chemistry and Technologies Laboratory Affiliated with the Institute Pasteur Italy Cenci Bolognetti Foundation Sapienza University of Rome 00185 Rome Italy

Department of Organic Chemistry Faculty of Chemical Technology University of Chemistry and Technology Prague 16628 Prague 6 Czech Republic

Citace poskytuje Crossref.org