Enhancing Evolutionary Couplings with Deep Convolutional Neural Networks.

Area of Science:

• Computational biology
• Structural bioinformatics
• Machine learning in genomics

Background:

• Protein function is dictated by its 3D structure, not just its gene sequence.
• Evolutionary information from multiple sequence alignments aids in predicting structural constraints.
• Limited related sequences hinder accurate prediction of residue-residue contacts for many proteins.

Purpose of the Study:

• To introduce DeepContact, a CNN-based method for accurate protein contact prediction.
• To leverage co-evolutionary motifs for improved contact probability inference, especially with sparse sequence data.
• To establish a consistent metric for assessing contact prediction accuracy across diverse proteins.

Main Methods:

• Developed DeepContact, a convolutional neural network (CNN) approach.
• Utilized co-evolutionary motifs discovered by the CNN to infer contact probabilities.
• Applied the method to the CASP12 blind contact prediction challenge.

Main Results:

• DeepContact significantly enhances the accuracy of residue-residue contact predictions.
• Achieved top performance in the CASP12 blind contact prediction task.
• Converted coupling scores into interpretable probabilities, enabling consistent metric assessment.
• Demonstrated substantial improvements in precision-recall for contact prediction.

Conclusions:

• DeepContact offers a powerful tool for accurate protein contact prediction, particularly when sequence data is scarce.
• The method facilitates a paradigm shift towards improved template-free protein structure modeling.
• The probabilistic output provides a standardized measure for evaluating contact prediction methods.