A Drug–Target Specificity Foundation Model for Off-target Prediction, Repurposing, and Generative Design

Molecular recognition - which small molecule binds which protein, and with what selectivity - governs the efficacy, safety, and discovery of every therapeutic, yet binding specificity is still determined by experimental screening or by computational methods that first predict three-dimensional structure. Transformer softmax attention is mathematically isomorphic to the Boltzmann distribution governing molecular binding at thermal equilibrium ¹ , an identity that prescribes a single sequence-native architecture: the Specificity Foundation Model (SFM), which computes molecular binding compatibility as a thermodynamic quantity directly from sequence ² . The framework was recently realized as prototype encoders across six molecular-recognition domains ³ . Here we report the small molecule drug-target protein SFM (dtSFM) as the first instance to pair a full-scale encoder with a generative decoder, trained on publicly available data consisting of 714,747 measured drug–protein interactions spanning 522,776 compounds and 22,964 proteins. Throughout, we verify binding predictions with AlphaFold 3 ⁴ as an orthogonal structural verifier that shares no architecture, training data, or representational basis with dtSFM. From this single dtSFM model we demonstrate the three sequence-native applications of drug discovery: off-target prediction, repurposing, and generative design. The dtSFM encoder retrieves a drug’s target, and a target’s drug, at 95% and 89% recall-at-10 in distribution, respectively. In the drug→target direction it screens off-targets at proteome scale, ranking the documented off-targets of clinical kinase inhibitors at a median of 30th out of 4,910 genes - the top 0.6% of the screen - when validated against a chemoproteomic panel ⁵ . In the target→drug direction it ranks the full 522,776-compound library against three immunology targets, identifying 46 novel candidates that pass AlphaFold-3 structural gating. The dtSFM cross-attentive decoder generates novel molecules for 16 targets, 850 of 1,200 (71%) designed candidates match the AlphaFold 3 structural confidence of the approved drug (iPTM ≥ 0.9 and interface PAE ≤ 1.67 Å), with the best candidates reaching iPTM 0.95–0.99 and interface PAE 0.79–1.37 Å. dtSFM brings computational thermodynamics to every stage where molecular recognition shapes drug discovery; experimental wet-lab validation is the immediate next step.