EPSILON-CP: using deep learning to combine information from multiple sources for protein contact prediction

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Accurate protein contact prediction is crucial for determining 3D protein structures.
• Identifying correct residue-residue contacts is challenging due to the vast solution space.
• Existing methods combine evolutionary, sequence-based, or physicochemical information.

Purpose of the Study:

• To develop a novel protein contact predictor, EPSILON-CP.
• To integrate evolutionary, physicochemical, and sequence-based information for enhanced prediction accuracy.
• To address challenges of increased dimensionality and complexity through feature analysis.

Main Methods:

• Developed EPSILON-CP, a deep neural network-based contact predictor.
• Combined evolutionary, physicochemical, and sequence-based data sources.
• Performed critical feature analysis to reduce dimensionality and complexity.

Main Results:

• EPSILON-CP achieved 35.7% mean precision for top-L/10 long-range contacts on CASP11 targets, outperforming MetaPSICOV by 11%.
• Demonstrated a 17% improvement for 1.5L predictions.
• Found amino acid composition ineffective in meta-approaches and reduced the feature set by 75%.

Conclusions:

• Integrating diverse information sources is vital for accurate protein contact prediction.
• Critical feature analysis enhances performance in multi-source information methods.
• EPSILON-CP offers improved accuracy and is available as a webservice.