ProtoBind-Diff: Protein-Conditioned Discrete Diffusion for Structure-Free Ligand Generation and Scoring

Structure-conditioned generative models are limited by sparse, biased experimental complexes: fewer than $30,000$ in PDBbind, concentrated on well-studied targets. We introduce ProtoBind-Diff, which removes that dependency: a structure-free masked diffusion model conditioning generation on primary sequence alone via frozen ESM-2 embeddings and cross-attention, trained on 1.17M active protein-ligand pairs from BindingDB. The same weights provide a zero-shot Pointwise Mutual Information (PMI) score ranking any molecule against any sequence, requiring no known actives. On held-out scaffolds, PMI separates a target's actives from its measured inactives (mean ROC-AUC $0.70$) better than AutoDock Vina ($0.54$) or Boltz-1 ipTM ($0.60$), and gives the best enrichment of any model tested under the more discriminative Boltz-1 score. Ranking $100,000$ purchasable compounds per target, we confirmed inhibitors of EGFR WT ($4/19$ below $10\,\mu$M; best IC50 = $0.44\,\mu$M) and TYK2 ($1/20$; IC50 = $0.75\,\mu$M), both sub-micromolar, from chemotypes not previously annotated against any kinase.