A variational expectation-maximization framework for balanced multi-scale learning of protein and drug interactions

Area of Science:

• Computational biology
• Bioinformatics
• Drug discovery

Background:

• Protein interactions are crucial for biological processes and drug development.
• Current computational methods often analyze molecular networks or structures separately, lacking a unified approach.
• Existing multi-view learning methods struggle with imbalanced data and underfitting at different scales.

Purpose of the Study:

• To develop a novel multi-scale representation learning framework for integrating diverse biological data.
• To enhance the accuracy of molecular interaction prediction by effectively fusing information from atomic structure and molecular network scales.
• To address the limitations of current methods in handling imbalanced multi-scale learning.

Main Methods:

• Introduced MUSE, a multi-scale representation learning framework utilizing a variant expectation-maximization algorithm.
• Employed an alternating iterative procedure for optimizing different scales, enabling mutual supervision.
• Integrated atomic structure and molecular network information within a unified framework.

Main Results:

• MUSE demonstrated superior performance in predicting protein-protein, drug-protein, and drug-drug interactions compared to state-of-the-art models.
• The framework achieved high accuracy in atomic-level protein interface prediction.
• Successfully fused multi-scale information through iterative optimization and mutual supervision.

Conclusions:

• MUSE provides an effective multi-scale learning strategy for computational drug discovery.
• The framework offers a promising approach for integrating diverse biological data scales.
• The developed method has potential for broader applications in computational biology and drug discovery.