BindMol splits the binding pocket into hotspots: higher novel hit rates than GEAM on five targets and more frequent recovery of type-I/II binding modes
Synopsis
Generative models for target-based drug design have long compressed protein–ligand interactions into a single docking score, hiding functional differences between binding modes on the same target. BindMol clusters the binding pocket into hotspots from reference complexes, assigns each hotspot its own reward dimension, and optimizes the conflicting objectives with preference-conditioned multi-objective reinforcement learning. On parp1, fa7, 5ht1b, braf and jak2 its novel hit ratios are 42.617, 24.580, 42.149, 30.767 and 44.864, above GEAM at 40.635, 17.903, 36.667, 27.887 and 42.025, and it recovers the crystallographic type-I and type-II binding modes more often on braf and jak2.
Interpretation
The generation objective changes from a single docking score to a reward vector decomposed by binding-pocket hotspots, making multiple binding modes on one target explicitly optimizable. Most prior generative methods reduced ligand–protein interactions to one scalar docking score, although compounds on the same protein can adopt different binding modes with distinct functional and selectivity profiles. On parp1, fa7, 5ht1b, braf and jak2, BindMol reaches novel hit ratios of 42.617, 24.580, 42.149, 30.767 and 44.864, above GEAM at 40.635, 17.903, 36.667, 27.887 and 42.025 and MOLLEO at 35.725, 20.651, 32.489, 25.112 and 39.120.
Molecules are assembled fragment by fragment, and the reward combines a global docking term with local interaction terms for each hotspot. The fragment vocabulary is scored and filtered by an energy model, so high-scoring fragments act as stabilizing anchors of the ligand–protein complex instead of coming from a default fragment library. Each completed molecule is docked once and PLIP reuses that pose to compute hotspot interactions, with every method sharing an evaluation budget of 3,000 completed molecules and docking calls.
Multi-objective trade-offs are handled by Envelope SAC, which learns a policy conditioned on preference vectors and approximates the convex hull of the Pareto front. Collapsing the reward vector into a scalar with fixed weights prevents the agent from dynamically exploring trade-offs between hotspots. In a comparison sharing the fragment action space, hotspot reward vector, training protocol, docking budget and three random seeds with GPI, C-MORL and PreCo, BindMol averages a 37.01% hit rate against 32.15%, 32.35% and 32.87%.
Perspective
The demonstrated scope is binding-affinity optimization on five protein targets and multi-property optimization on GuacaMol; training uses a fixed receptor, while the crystallographic-mode recovery evaluation re-docks generated hits into independent crystal structures that carry induced-fit, side-chain, DFG-loop and C-helix differences. Binding-mode statistics come from docking-predicted PLIP interaction fingerprints, and crystallographic-mode recovery is assessed only on braf and jak2 because only these targets provide directly comparable type-I/DFG-in and type-II/DFG-out paired labels.
The reward uses simplified interaction counts and static docking poses without weighting interaction types by energetic contribution; static docking may reward physically strained poses, so interaction modes are predicted rather than experimentally confirmed; IFP cluster counts may still depend on sample size and should be read alongside entropy and effective-mode statistics; crystallographic-mode recovery covers only braf and jak2.