Atomic-level protein-ligand recognition with PBCNet2.0 for probe discovery - PubMed
3 hours ago
- #Binding affinity prediction
- #Protein-ligand recognition
- #Machine learning
- PBCNet2.0 is a Cartesian tensor-based Siamese neural network for protein-ligand relative binding affinity prediction.
- It was trained on 8.6 million protein-ligand complex pairs, achieving zero-shot accuracy comparable to physics-based simulations while being highly efficient.
- The model improves optimization efficiency by 7.18-fold and reduces resource use by 41% in retrospective prioritization experiments.
- PBCNet2.0 captures intermolecular interactions, spatial geometric constraints, and subtle effects like fluorine orthogonal multipolar interactions.
- It exhibits emergent capability to predict affinity changes from binding pocket residue variations, despite not being trained on mutation data.
- Prospective validation on ENPP1 and ALDH1B1 accurately resolved affinity shifts from minor differences and identified critical binding residues with a high hit rate.