Structure–function conditioned molecular generation with PocketGPT enables prospective discovery of PARP1 inhibitors

Structure-based molecular generation has become increasingly prominent in drug discovery; however, many generated binders still fail to satisfy the physicochemical and functional constraints required for pharmacological viability. Here, we present PocketGPT, a pocket-conditioned generative framework that demonstrates structure-aware molecular generation can be effectively enabled without explicit three-dimensional ligand construction, by jointly conditioning protein pocket context and medicinal-chemistry constraints during generation. By co-conditioning physicochemical property targets, fragment priors, and protein–ligand interaction fingerprints, PocketGPT enables structure–function co-generative control, allowing directional modulation of drug-like properties while partially retaining binding-relevant interaction patterns across diverse targets. In a prospective design–make–test campaign targeting poly(ADP-ribose) polymerase 1 (PARP1), PocketGPT identified a novel inhibitor with nanomolar binding affinity (K_d = 4.88 nM) and favorable tolerability in mice. Together, these results demonstrate that joint structural and pharmacological conditioning, embedded directly in the generative process, constitutes a viable and controllable representation for structure-aware molecular generation, offering a complementary pathway to geometry-first approaches for producing experimentally actionable chemical matter.