Accelerating molecular probe discovery and lead optimization requires accurate and efficient binding affinity prediction. Here we present PBCNet2.0, a Cartesian tensor-based Siamese neural network for protein-ligand relative binding affinity prediction. Trained on 8.6 million protein-ligand complex pairs, PBCNet2.0 achieves zero-shot accuracy similar to computationally intensive physics-based simulations while remaining highly efficient. Retrospective prioritization experiments show that PBCNet2.0 improves optimization efficiency by 7.18-fold and reduces resource use by 41%. Mechanistic analyses indicate that the model captures intermolecular interactions and encodes spatial geometric constraints, enabling sensitivity to subtle effects such as fluorine orthogonal multipolar interactions. Notably, although not trained on mutation data, PBCNet2.0 exhibits an emergent capability to predict affinity changes induced by binding pocket residue variations, supporting resistance analysis. We prospectively validated these capabilities on ENPP1 and ALDH1B1, accurately resolving affinity shifts from minor interaction and conformational differences and identifying critical binding residues with a hit rate of five out of six selected residues.
© 2026. The Author(s), under exclusive licence to Springer Nature America, Inc.