Predicting protein residue-residue contacts using random forests and deep networks

Area of Science:

• Computational biology
• Bioinformatics
• Protein structure prediction

Background:

• Predicting residue-residue contacts in proteins is crucial for understanding protein structure and function.
• Limited experimentally determined protein structures necessitate computational prediction methods.
• Contact prediction aids in reducing computational complexity for protein structure prediction and identifying functional regions.

Purpose of the Study:

• To develop and benchmark machine learning methods for residue-residue contact prediction.
• To evaluate the performance of various algorithms including random forests, direct-coupling analysis, support vector machines, and deep networks.
• To provide accessible tools for protein contact prediction.

Main Methods:

• Implemented and evaluated random forests, direct-coupling analysis, support vector machines, and deep networks (stacked denoising autoencoders).
• Trained models using amino acid sequences to predict contacting residue pairs.
• Blindly tested models on the CASP11 dataset for performance evaluation.

Main Results:

• Random forest predictors achieved top 10 prediction accuracy scores of 85.13% (short range), 74.49% (medium range), and 54.49% (long range).
• Ensemble deep network models (stacked denoising autoencoders with support vector machines) achieved scores of 75.51% (short range), 60.26% (medium range), and 43.85% (long range).
• Model performance was comparable to existing contact predictors from the CASP11 dataset.

Conclusions:

• Benchmarking diverse prediction methods is beneficial for the challenging task of contact prediction.
• Developed user-friendly tools for contact prediction with simple interfaces.
• Released a standalone C++ implementation of direct-coupling analysis and the RFcon web server for public access.