NeBcon: protein contact map prediction using neural network training coupled with naïve Bayes classifiers

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Protein structure prediction is crucial for understanding protein function.
• Co-evolution based contact predictions show promise but require abundant sequence homologs.
• A need exists for methods that generate reliable contact maps for diverse protein targets, especially those with limited homologous sequences.

Purpose of the Study:

• To develop an efficient and reliable method for protein contact map prediction.
• To improve the accuracy of ab initio protein structure prediction, particularly for challenging targets.
• To integrate diverse contact prediction strategies into a robust pipeline.

Main Methods:

• Developed NeBcon, a pipeline utilizing the naïve Bayes classifier (NBC) theorem.
• Combined eight state-of-the-art contact prediction methods based on co-evolution and machine learning.
• Trained NBC posterior probabilities with intrinsic structural features using neural network learning.

Main Results:

• NeBcon improved the accuracy of the best co-evolution based meta-server predictor by 22% on 98 non-redundant proteins.
• Accuracy improvement reached 45% for difficult targets lacking sequence and structural homologs.
• Optimized NBC combination of complementary co-evolution and machine learning data was key to improvement.
• Neural network training enhanced the coupling of NBC probabilities and structural features, crucial for proteins with limited homologs.

Conclusions:

• NeBcon offers a significant advancement in protein contact map prediction.
• The pipeline demonstrates superior performance, especially for proteins with sparse homologous sequences.
• NeBcon enhances the success rate of ab initio protein structure prediction.