Accurate prediction of solvent accessibility using neural networks-based regression

Area of Science:

• Computational biology
• Bioinformatics
• Structural bioinformatics

Background:

• Accurate prediction of relative solvent accessibility (RSA) is crucial for protein structure prediction and functional annotation.
• Existing machine learning methods often rely on classification with arbitrary boundaries, potentially limiting accuracy.

Purpose of the Study:

• To develop a novel, improved method for predicting protein residue RSAs.
• To overcome limitations of classification-based approaches by using nonlinear regression for continuous RSA approximation.

Main Methods:

• Developed a consensus predictor combining multiple feedforward and recurrent neural networks for nonlinear regression of real-value RSAs.
• Trained on 860 protein structures from PFAM and validated on 603 non-homologous Protein Data Bank structures.
• Compared regression approach with classification and semi-continuous (thermometer encoding) predictors, incorporating weighted approximation and evolutionary profiles.

Main Results:

• The regression-based method achieved mean absolute errors of 15.3-15.8% and correlation coefficients of 0.64-0.67 on control sets.
• Accuracy was higher for buried residues than exposed residues, aligning with RSA variability.
• Outperformed classification algorithms, achieving ~77% accuracy on two-class problems with a 25% RSA threshold.

Conclusions:

• The novel nonlinear regression method provides accurate and improved prediction of protein residue RSAs.
• This approach offers advantages over traditional classification methods for RSA prediction.
• A web server is available for RSA prediction and results visualization.