Protein-ligand binding affinity prediction using multi-instance learning with docking structures.

Area of Science:

• Computational Chemistry
• Structural Biology
• Machine Learning

Background:

• Deep learning models show promise in predicting protein-ligand binding affinity for drug discovery.
• Current methods often rely on co-crystal structures, which are not always available.
• Inaccurate predicted structures from molecular docking can reduce machine learning model accuracy.

Purpose of the Study:

• To develop a novel structure-based inference method for predicting binding affinity.
• To overcome the limitation of requiring co-crystal structures in binding affinity prediction.
• To leverage multiple molecular docking poses for improved prediction accuracy.

Main Methods:

• A novel structure-based inference method was developed.
• The method utilizes multiple molecular docking poses for each complex.
• Multi-instance learning with an attention network was employed to predict binding affinity.

Main Results:

• The proposed method was validated using PDBbind and SARS-CoV-2 main protease datasets.
• The binding affinity prediction performance was competitive with state-of-the-art methods.
• The method demonstrated effectiveness using docking poses instead of co-crystal structures.

Conclusions:

• The developed method enables binding affinity prediction without the need for co-crystal structures.
• This approach significantly broadens the applicability of binding affinity prediction models.
• The method offers a valuable tool for virtual high-throughput screening in drug discovery.