Deep learning extends de novo protein modelling coverage of genomes using iteratively predicted structural...

Area of Science:

• Computational biology
• Structural bioinformatics
• Deep learning applications

Background:

• Traditional amino acid covariation methods struggle with small protein families, limiting genome-wide structural annotation.
• Recent advances in deep learning show potential for accurate residue-residue contact prediction from shallow sequence alignments.

Purpose of the Study:

• To introduce DMPfold, a novel deep learning method for protein structure prediction.
• To demonstrate DMPfold's capability in predicting inter-atomic distance bounds, hydrogen bond networks, and torsion angles for iterative model building.

Main Methods:

• Utilized deep learning to predict structural features including inter-atomic distance bounds, main chain hydrogen bond network, and torsion angles.
• Employed an iterative approach for protein model construction based on predicted features.
• Benchmarked DMPfold against existing methods using CASP12 domains and transmembrane proteins.

Main Results:

• DMPfold achieved higher accuracy than two popular methods on CASP12 test domains.
• The method demonstrated comparable performance for transmembrane proteins.
• Confident structural models were generated for 25% of Pfam 'dark families' (those without known structures) within a week on a modest cluster.
• 16% of human proteome UniProt entries lacking structures were successfully modeled.

Conclusions:

• DMPfold overcomes limitations of traditional methods, enabling structural prediction for previously intractable protein families.
• The deep learning approach facilitates efficient and accurate protein structure modeling, even with limited sequence data.
• DMPfold significantly expands the scope of structural annotation for genomic and proteomic datasets.