Protein threading using residue co-variation and deep learning.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Template-based modeling is crucial for protein 3D structure prediction.
• Challenges persist in alignment generation and template selection for proteins without close templates.

Purpose of the Study:

• To introduce DeepThreader, a novel method enhancing protein threading.
• To improve alignment generation and template selection using deep learning and co-variation information.

Main Methods:

• Deep learning predicts inter-residue distance distribution from co-variation and sequential data.
• An ADMM algorithm integrates predicted distance and sequential features for sequence-template alignment.
• The method leverages residue co-variation and deep learning (DL).

Main Results:

• Predicted inter-residue distance aids protein alignment and template selection, particularly for proteins without close templates.
• DeepThreader significantly outperforms popular methods like HHpred and CNFpred.
• DeepThreader shows superior performance compared to EigenTHREADER, a contact-assisted threading method.

Conclusions:

• Deep learning and co-variation information effectively improve protein threading.
• DeepThreader offers a significant advancement in protein 3D structure prediction for challenging cases.
• The method demonstrates superior performance over existing state-of-the-art techniques.