Multi-task bioassay pre-training for protein-ligand binding affinity prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Protein-ligand binding affinity (PLBA) prediction is crucial for drug discovery.
• Current deep learning models for PLBA prediction are limited by data scarcity and generalization issues.
• Existing large-scale affinity datasets like ChEMBL have inconsistent labels and experimental conditions.

Purpose of the Study:

• To propose a novel pre-training framework, Multi-task Bioassay Pre-training (MBP), for structure-based PLBA prediction.
• To construct a comprehensive pre-training dataset, ChEMBL-Dock, containing over 300k affinity labels and 2.8M docked 3D structures.
• To enhance the generalization ability of PLBA prediction models by learning robust structural knowledge from varied and noisy data.

Main Methods:

• Developed the Multi-task Bioassay Pre-training (MBP) framework.
• Created the ChEMBL-Dock dataset integrating diverse affinity labels and 3D structures.
• Employed multi-task learning to predict different affinity labels and relative rankings within bioassays.

Main Results:

• MBP demonstrates significant capability in structure-based PLBA prediction.
• The pre-training approach effectively learns transferable structural knowledge.
• The developed ChEMBL-Dock dataset addresses limitations of existing affinity data.

Conclusions:

• MBP is the first affinity pre-training model for PLBA prediction.
• The framework shows great potential for advancing drug discovery.
• MBP offers a robust solution for handling noisy and diverse bioassay data.