Enhanced compound-protein binding affinity prediction by representing protein multimodal information via a...

Area of Science:

• Computational biology
• Drug discovery
• Machine learning

Background:

• Predicting compound-protein binding affinity (CPA) is crucial for drug discovery but limited by challenges in representing multimodal protein information.
• Current machine learning methods for CPA prediction suffer from low accuracy due to inefficient representation of protein structure and sequence data.

Purpose of the Study:

• To develop a novel end-to-end architecture, FeatNN, for accurate and generalizable CPA prediction.
• To address the limitations in representing multimodal protein information by introducing a coevolutionary strategy.

Main Methods:

• Developed FeatNN, an end-to-end architecture employing a coevolutionary strategy to jointly represent protein structure and sequence features.
• Implemented a data-driven approach to utilize both high- and low-quality databases for optimizing model accuracy and generalization.
• Visually interpreted feature interactions between sequence and structure within the FeatNN architecture.

Main Results:

• FeatNN significantly outperforms state-of-the-art baselines in virtual drug evaluation tasks.
• The coevolutionary strategy effectively represents multimodal protein information, leading to improved CPA prediction.
• The rational data utilization method enhances both accuracy and generalization ability of the model.

Conclusions:

• FeatNN offers a superior method for CPA prediction by efficiently integrating multimodal protein information.
• The approach demonstrates feasibility for practical applications in virtual drug evaluation.
• FeatNN provides enhanced accuracy and generalization for predicting compound-protein binding affinity.