Atomic-level protein-ligand recognition with PBCNet2.0 for probe discovery.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Accurate binding affinity prediction is crucial for accelerating molecular probe discovery and lead optimization in drug development.
• Current methods often face trade-offs between accuracy and computational efficiency.

Purpose of the Study:

• To introduce PBCNet2.0, a novel Cartesian tensor-based Siamese neural network for efficient and accurate protein-ligand relative binding affinity prediction.
• To evaluate the performance of PBCNet2.0 in retrospective and prospective drug discovery scenarios.

Main Methods:

• Development of PBCNet2.0, a deep learning model utilizing Cartesian tensors and a Siamese network architecture.
• Training on a large dataset of 8.6 million protein-ligand complex pairs.
• Validation through retrospective prioritization experiments and prospective studies on ENPP1 and ALDH1B1.

Main Results:

• PBCNet2.0 achieved zero-shot accuracy comparable to computationally intensive physics-based simulations.
• Demonstrated a 7.18-fold improvement in optimization efficiency and a 41% reduction in resource use.
• Showcased emergent ability to predict affinity changes from binding pocket mutations and identified critical residues in prospective validation.

Conclusions:

• PBCNet2.0 offers a highly efficient and accurate solution for protein-ligand binding affinity prediction.
• The model's ability to capture complex interactions and predict mutation effects supports its application in lead optimization and resistance analysis.
• Prospective validation confirmed PBCNet2.0's utility in resolving subtle affinity shifts and identifying key binding residues.