More Accurate Binding Affinity Prediction Using Protein Homology and Ligand-Based Transfer Learning

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of protein-ligand binding affinities is crucial for efficient drug discovery.
• Evaluating large chemical libraries and novel molecules from generative models demands rapid and precise methods.
• Existing methods may face challenges with scalability and accuracy for diverse datasets.

Purpose of the Study:

• To develop and validate UCBbind, a hybrid computational framework for efficient and accurate prediction of protein-ligand binding affinities.
• To integrate a similarity-based transfer learning module with a deep learning prediction module.
• To assess UCBbind's performance across diverse benchmark datasets relevant to drug discovery.

Main Methods:

• UCBind employs a hybrid approach, combining a similarity-based transfer module with a deep-learning prediction module.
• Experimental data from similar reference protein-ligand pairs is transferred when available.
• The deep learning module is utilized when no sufficiently similar reference data exists.
• Performance was benchmarked on CASF-2016, HiQBind (post-2020), and the COVID Moonshot database.

Main Results:

• UCBind achieved state-of-the-art predictive performance on multiple benchmark datasets.
• The framework demonstrated superior accuracy for test entries with high similarity to reference protein-ligand pairs.
• Performance gains were particularly notable for well-characterized reference proteins and ligands.

Conclusions:

• UCBind offers an efficient and accurate solution for predicting protein-ligand binding affinities in drug discovery.
• The hybrid framework effectively leverages existing data through transfer learning and deep learning.
• UCBind shows promise for supporting downstream computational tasks, including binding site prediction and classification.