Detecting distant-homology protein structures by aligning deep neural-network based contact maps.

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein structure prediction

Background:

• Accurate protein structure prediction is crucial for understanding biological functions and drug design.
• Template-based modeling (TBM) is the most accurate method but struggles with distant homology.
• Existing TBM methods face challenges in identifying templates when evolutionary relationships are weak.

Purpose of the Study:

• To develop a novel method, CEthreader, for improving template-based protein structure prediction.
• To enhance the recognition of structural templates, particularly for proteins lacking clear homologous relationships.
• To overcome the limitations of current TBM approaches in identifying distant-homology templates.

Main Methods:

• CEthreader predicts residue-residue contacts using evolutionary precision matrices and deep residual convolutional neural networks.
• Predicted contact maps are integrated with sequence profile alignments for template recognition from the Protein Data Bank (PDB).
• The method was evaluated on two independent benchmark sets comprising 1,153 non-homologous protein targets.

Main Results:

• CEthreader identified 176% more correct templates (TM-score >0.5) for hard targets lacking homologous templates compared to state-of-the-art methods.
• The method showed significant improvements in detecting correct templates, even when excluding proteins from the same SCOPe Superfamily.
• CEthreader's advantage stems from effectively coupling contact maps with profile alignments for recognizing global protein folds.

Conclusions:

• CEthreader offers a significant advancement in template-based protein structure prediction, especially for proteins with weak homologous relationships.
• The integration of ab initio contact map prediction with profile alignments provides a powerful strategy to improve accuracy.
• This approach enhances the ability to annotate protein functions and design novel compounds by improving structural model reliability.