Fig. 1: GM workflow pipeline.
From: Optimizing drug design by merging generative AI with a physics-based active learning framework
The GM workflow involves two nested iterative processes: a lower-level iteration (inner AL cycle) and a higher-level iteration (outer AL cycle). During inner AL cycles, generated molecules are evaluated based on Quantitative Estimate of Drug Likeliness (QED), SA score, and Tanimoto similarity. Outer AL cycles use the Glide gscore for evaluation. The specific training set initially consists of molecules with known or predicted affinity to the target (initial-specific training set). Inner AL cycles enrich this set with molecules meeting QED, SA and similarity thresholds (temporal-specific set). In the outer AL cycles, accumulated molecules that meet the Glide gscore thresholds are retained in the specific training set (permanent-specific set), while those that do not are discarded. After candidate selection, generated molecules are validated using absolute binding free energy simulations (ABFE) and bioassays (see “Methods”).
