Accurate contact predictions using covariation techniques and machine learning.

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein structure prediction

Background:

• Accurate prediction of residue-residue contacts is crucial for determining protein 3D structures.
• The Critical Assessment of protein Structure Prediction (CASP) competition benchmarks prediction methods.

Purpose of the Study:

• To present the performance of the CONSIP2 server, based on the MetaPSICOV method, in the CASP11 competition.
• To analyze the strengths and weaknesses of the MetaPSICOV approach for contact prediction.

Main Methods:

• Utilized the CONSIP2 server, employing the MetaPSICOV contact prediction method.
• Combined classical contact prediction features with three distinct covariation methods.
• Employed a two-stage neural network predictor.
• Integrated jackHMMer and HHblits for generating deep multiple-sequence alignments.
• Tuned feature contributions based on alignment depth.

Main Results:

• Achieved an average top-L/5 long-range contact precision of 27% on 40 target domains in CASP11.
• The MetaPSICOV method combines classical and covariation features effectively.
• Identified reliance on fixed parameters for initial sequence alignments and absence of domain splitting as factors limiting performance in some cases.

Conclusions:

• The MetaPSICOV method, implemented in CONSIP2, shows competitive performance in residue-residue contact prediction.
• Future improvements could involve adaptive parameter selection and domain splitting for enhanced accuracy.