A global machine learning based scoring function for protein structure prediction.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Biophysics

Background:

• Protein structure prediction is crucial for understanding biological function.
• Existing knowledge-based potentials for scoring protein models have limitations.
• Accurate scoring of protein decoys remains a significant challenge in structural biology.

Purpose of the Study:

• To develop a novel knowledge-based function for scoring protein decoys based on native structure similarity.
• To introduce a new metric, the 'fitness score,' to quantify model-native structure similarity.
• To overcome limitations of traditional knowledge-based potentials in protein structure assessment.

Main Methods:

• Constructed a feature set describing protein chain structure and sequence, including residue-residue and residue-solvent distances, and pairwise/four-body potentials.
• Established a relationship between features and underlying electromagnetic interactions.
• Trained a back-propagating neural network on native and decoy structures to predict the fitness score.

Main Results:

• The novel scoring function demonstrated superior performance compared to input knowledge-based potentials.
• The method achieved third place overall and second place for hard targets in the CASP10 experiment among approximately 200 participating groups.
• The approach successfully captured subtle structural details missed by other protein structure prediction methods.

Conclusions:

• The developed scoring function provides a robust and accurate method for evaluating protein models.
• This novel approach offers significant improvements in protein structure prediction accuracy.
• The freely available source code and executables facilitate further research and application.