Robust Many-Objective Molecular Design with Preference-Gated GFlowNets

Molecular design rarely involves optimizing a single property: a useful candidate must combine strong target binding with suitable physicochemical properties, permeability, bioavailability, and toxicity, even when these objectives compete. We introduce PrefGateGFN, a plug-in for reaction-based GFlowNets that uses a user-defined preference vector to modulate molecular representations and action selection during synthesis-aware generation. We evaluate PrefGateGFN on five protein targets using tasks with 5, 9, and 14 objectives, against a multiplicative RxnFlow baseline and an RxnFlow implementation of MOGFN-PC. Our central result is robustness as the objective panel grows. From 5 to 14 objectives, PrefGateGFN retains 92% of its mean joint-success rate and 40% of its acceptable-region hypervolume; MOGFN-PC retains 37% and 26%, while the multiplicative baseline falls to zero on both measures. In the replicated nine-objective experiment, PrefGateGFN identifies 2,234 satisfactory molecules among 15,000 training-pool candidates, compared with 1,377 and 1,019 for the baselines. In an equally sized fresh-generation evaluation, the corresponding yields are 2,826, 1,274, and 2,078; among them, 1,795, 1,073, and 1,567 also achieve a predicted docking score of −8 kcal mol⁻¹ or lower. Controlled preference sweeps further show that increasing an objective’s preference improves its corresponding outcome for 86% of objectives in the 14-objective task. Together, these results show that preference gating provides a practical and controllable approach to many-objective molecular generation.