TripletRes

TripletRes predicts protein inter-residue contact maps by integrating coevolutionary signals from multiple sequence alignments with deep residual convolutional neural networks to support protein structure prediction.

Key Features:

• Coevolutionary Feature Integration: Integrates a triplet of coevolutionary matrices—covariance matrix, precision matrix, and pseudolikelihood maximization—to capture evolutionary couplings and minimize information loss during model training.
• Deep Residual Neural Networks: Employs deep residual convolutional neural networks for end-to-end learning from discretized distance profiles to predict contact maps.
• Multiple Sequence Alignments from Genomic Databases: Derives deep MSAs from extensive whole-genome and metagenome databases to supply coevolutionary data.
• Benchmark Performance: Demonstrated improved accuracy on 245 non-homologous proteins from CASP and CAMEO, including ≥58.4% improvement in top-L long-range contact precision for CASP11&12, 44.4% for CAMEO, and 71.6% precision for top-L/5 long-range contacts on 31 FM targets from CASP13.

Scientific Applications:

• Ab initio protein folding: Provides contact-map predictions as constraints to guide ab initio folding simulations.
• Template-free structure modeling: Enables construction of three-dimensional structures for proteins lacking homologous templates in the Protein Data Bank (PDB).
• Medium- and long-range contact prediction: Predicts medium- and long-range contacts used to improve computational structural biology analyses.

Methodology:

Collecting deep MSAs from comprehensive whole-genome and metagenome databases; extracting and fusing covariance, precision, and pseudolikelihood maximization coevolutionary feature matrices; and training deep residual convolutional neural networks via end-to-end learning from discretized distance profiles to produce contact-map models.