Integrating unsupervised language model with multi-view multiple sequence alignments for high-accuracy inter-chain...

Area of Science:

• Computational Biology
• Structural Biology
• Bioinformatics

Background:

• Accurate identification of inter-chain contacts is crucial for determining protein complex 3D structures and biological functions.
• Existing methods for predicting inter-chain contacts from amino acid sequences have limitations.

Purpose of the Study:

• To develop a novel deep learning method, ICCPred, for accurate prediction of inter-chain contacts in protein complexes using only amino acid sequences.
• To improve the accuracy of inter-chain contact prediction compared to state-of-the-art approaches.

Main Methods:

• Developed ICCPred, a deep learning pipeline utilizing a deep residual network architecture.
• Integrated a pre-trained language model with three multiple sequence alignments (MSAs) from diverse biological perspectives.
• Trained and evaluated the model on 709 non-redundant protein complexes.

Main Results:

• ICCPred significantly improved the accuracy of inter-chain contact prediction.
• The method outperformed existing state-of-the-art approaches.
• Analysis revealed that pre-trained transformer language models effectively capture co-evolutionary diversity from MSAs, enhancing prediction accuracy.

Conclusions:

• ICCPred offers a highly accurate deep learning-based approach for predicting inter-chain contacts from protein sequences alone.
• The method provides a valuable tool for large-scale protein-protein interaction annotation and structural biology research.