ThreaderAI

ThreaderAI predicts protein tertiary structures by using deep residual neural networks (ResNets) to compute residue-residue aligning probability matrices for template-based modeling and by constructing template-query alignments via dynamic programming to improve template selection and alignment for proteins with distant homologs.

Key Features:

• Deep Learning Integration: Uses deep residual neural networks (ResNets) to predict residue-residue aligning probability matrices by integrating sequence profiles, predicted sequential structural features, and predicted residue-residue contacts, framing alignment as a pixel classification problem.
• Dynamic Programming: Constructs template-query alignments by applying dynamic programming to the predicted residue-residue aligning probability matrices.
• Performance Superiority: Shows improved accuracy relative to HHpred, CNFpred, and CEthreader, with reported increases of 56% in TM-score over HHpred, 13% over CNFpred, and 11% over CEthreader for fold-level similarity on SCOPe, and improvements of 16%, 9%, and 8% respectively on CASP13's TBM-hard dataset.

Scientific Applications:

• Structural biology: Provides tertiary-structure models to support elucidation of protein function when experimental structures are unavailable.
• Drug design: Supplies structural templates for structure-based drug design efforts.
• Enzyme engineering: Offers predicted structural models useful for enzyme redesign and engineering.
• Functional annotation: Facilitates annotation of novel proteins by providing structural hypotheses for proteins with distant homologs.

Methodology:

Predicts residue-residue aligning probability matrices using deep residual neural networks that integrate sequence profiles, predicted sequential structural features, and predicted residue-residue contacts, then applies dynamic programming to those matrices to construct optimal template-query alignments.