Protein remote homology detection and structural alignment using deep learning.

Area of Science:

• Computational biology
• Bioinformatics
• Protein structure prediction

Background:

• Accurate protein alignment is crucial for understanding sequence-structure-function relationships in biotechnology.
• Existing methods struggle to identify homologous proteins with low sequence similarity.
• There is a need for advanced computational tools to bridge this gap.

Purpose of the Study:

• To develop novel deep learning methods for identifying and aligning proteins with low sequence similarity.
• To enhance the discovery of remotely homologous proteins for biotechnological applications.

Main Methods:

• Developed TM-Vec for predicting TM-scores (a measure of structural similarity) directly from protein sequence pairs.
• Developed DeepBLAST for structurally aligning proteins using only sequence information by identifying homologous regions.
• Trained and validated both methods on diverse protein datasets.

Main Results:

• TM-Vec accurately predicts structural similarity without requiring intermediate structure computations.
• DeepBLAST outperforms traditional sequence alignment methods and matches structure-based alignment performance.
• Both methods demonstrate superior identification of remotely homologous proteins compared to state-of-the-art approaches.

Conclusions:

• TM-Vec and DeepBLAST offer powerful deep learning solutions for protein sequence alignment, particularly for distantly related proteins.
• These methods advance the exploitation of sequence-structure-function relationships in biotechnology.
• The developed tools enhance protein annotation and discovery in large biological sequence databases.