LeadFlow learns activity-aligned molecular generation from chemical series for structure-based molecular optimization

Structure-based molecular optimization requires supervision that captures scaffold modifications associated with experimental activity. Here we introduce LeadFlow, combining maximum common substructure supervision from assay-consistent chemical series, Bayesian Flow Networks for joint generation of three-dimensional coordinates and molecular graphs, and experimental activity-based direct preference optimization (DPO). On MolGenBench, LeadFlow recovered reference active scaffolds in approximately 400 of 600 series and exact active molecules in approximately 120 series. Controlled scaffold-definition ablations supported broader active-chemotype recovery and fused-ring elaboration. Activity-based DPO increased activity-weighted recovery relative to pretraining and docking-score-based preferences. In a prospective ClpP campaign, generation followed by computational filtering and medicinal chemistry selection yielded five synthesized compounds with lower mean EC₅₀ values than the starting compound A1. DC-ClpP-03 reached 38.7 nM, an approximately 156-fold potency improvement. These results support chemical-series supervision and experimental preferences for scaffold-conditioned molecular optimization.