Pan–Pharmacological Drug–Target Interaction Prediction with 3D–Informed Protein Encoding at Scale

Accurate prediction of drug–target binding affinity across multiple pharmacological endpoints remains challenging, as most deep learning methodologies focus on a single metric and face a trade–off between incorporating structural information and computational throughput. Here we present OmniBind, a multitask framework that resolves both constraints by encoding protein tertiary structures as discrete token sequences and integrating them with amino acid sequence features through a gated fusion mechanism. Trained on over two million compound–protein pairs from BindingDB, OmniBind simultaneously predicts four pharmacological endpoints in a single forward pass, providing a pan–pharmacological profile across multiple affinity endpoints for each compound–target pair. Across adversarial and temporal benchmarks, OmniBind consistently outperforms state–of–the–art architectures, demonstrating that structural encoding captures physicochemical interaction principles rather than ligand–protein co–occurrence patterns. Attention analysis reveals biologically meaningful binding site recognition, with the model selectively weighting the ABL1 gatekeeper residue T315 and responding to its drug–resistance mutation T315I. As a demonstration of practical utility, proteome–wide screening of 20,421 human proteins recovered 85.7% of known clinical targets of clozapine within the top 200 predictions, correctly resolving its pharmacological profile despite structural similarity to clomipramine. OmniBind provides an accurate, structurally informed, and interpretable platform for multi–endpoint drug-target interaction prediction, with applications to lead optimization, off–target safety assessment, and drug repositioning.