Balanced-rank policy refinement of discrete molecular graph diffusion for multi-objective molecular generation

Generating drug-like molecules while balancing chemical validity, synthetic accessibility, novelty, and distributional fidelity is a multi-objective optimization problem. We present HybridMolGen as a categorical two-dimensional molecular graph diffusion backbone refined with proximal policy optimization (PPO), a balanced within-batch marginal-rank reward, and adaptive objective weights. The evaluated implementation neither generates Cartesian coordinates nor includes an independently evaluated SE(3)-equivariant coordinate generator or cross-attention property-conditioning module. In an eight-seed paired confirmation with 256 samples per condition, refinement increased synthetic-accessibility utility by 0.00610 (95% bootstrap CI 0.00106 to 0.01061), novelty by 0.00928 (0.00342 to 0.01465), and composite utility by 0.00632 (0.00153 to 0.01068). Hypervolume changed by − 0.00021 and satisfied the predefined non-inferiority gate. A frozen checkpoint was subsequently evaluated with the official MOSES implementation on 25,000 requested samples. Relative to the matched backbone, FCD/Test improved from 0.57250 to 0.55633 and scaffold similarity from 0.90221 to 0.91031, whereas validity changed from 0.87756 to 0.87488 and novelty from 0.94076 to 0.93936. The results support a controlled policy-refinement trade-off rather than universal superiority. Three-dimensional equivariant generation and explicit cross-attention conditioning remain future extensions requiring independent implementation and evaluation.