Learning Binding Affinities via Fine-Tuning of Protein and Ligand Language Models

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Accurate in silico prediction of protein-ligand binding affinity is crucial for drug discovery.
• Traditional methods like docking are often ineffective, while free energy methods are computationally expensive.
• Existing deep learning models lack generalization capabilities and current evaluation methods are insufficient.

Purpose of the Study:

• To introduce BALM, a novel deep learning framework for predicting binding affinity.
• To propose improved evaluation strategies for assessing model performance on new targets.
• To enhance the efficiency and adaptability of computational tools in early-stage drug discovery.

Main Methods:

• Utilized pretrained protein and ligand language models within the BALM framework.
• Developed improved evaluation strategies using diverse datasets and metrics.
• Curated BindingDB for model training and validation.

Main Results:

• BALM demonstrated generalization to unseen drugs, scaffolds, and targets.
• Outperformed traditional machine learning and docking methods in few-shot learning scenarios for targets like USP7 and Mpro.
• Showcased computational efficiency and high adaptability.

Conclusions:

• BALM provides an accurate, efficient, and adaptable solution for binding affinity prediction.
• Proposed evaluation methods enable more stringent assessment of machine learning scoring tools.
• BALM is practical for large-scale screening in early-stage drug discovery.