Advancing Ligand Binding Affinity Prediction with Cartesian Tensor-Based Deep Learning

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 structure pairs, PBCNet2.0 achieves zero-shot performance comparable to computationally intensive physics-based simulations. Our prioritization experiments show that PBCNet2.0 speeds up binding affinity optimization by 718% while reducing resource use by 41%. Through extensive retrospective experiments, we demonstrate that PBCNet2.0 intrinsically comprehends protein-ligand interactions, showing high sensitivity to intermolecular interactions and exceptional perception of spatial geometric information. Strikingly, PBCNet2.0 exhibits an emergent capability to predict affinity changes induced by binding residue variations, highlighting its potential for identifying resistance mutation. We prospectively validated these capabilities on two targets ENPP1 and ALDH1B1, where PBCNet2.0 successfully identified affinity shifts arising from subtle molecular interactions and conformational differences, and pinpointed critical binding residues with an 83% hit rate. This combination of computational efficiency, spatial geometric perception of binding site, and generalizable affinity prediction establishes PBCNet2.0 as a transformative tool for developing pharmacological probes for all human proteins.