A generic method for assignment of reliability scores applied to solvent accessibility predictions.

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biochemistry

Background:

• Estimating the reliability of real-value predictions in biology is challenging.
• Existing methods for discrete predictions are not suitable for continuous biological features.
• A reliable prediction requires an associated confidence score.

Purpose of the Study:

• To develop a method that predicts relative surface accessibility of amino acids.
• To simultaneously predict the reliability of each accessibility prediction.
• To implement reliability scores as Z-scores for enhanced interpretability.

Main Methods:

• An ensemble of artificial neural networks was trained on experimentally solved protein structures.
• The model predicts relative amino acid exposure and assigns reliability scores during training.
• Reliability is quantified using a Z-score, integrated directly into the prediction process.

Main Results:

• The method achieved a Pearson's correlation coefficient of 0.72 on the CB513 benchmark set.
• Performance is comparable to the leading public method, Real-SPINE.
• The Z-score reliability measure proved more informative for distinguishing accurate predictions across all exposure levels.

Conclusions:

• The developed method accurately predicts amino acid surface accessibility and provides reliable confidence scores.
• The Z-score reliability metric enhances the ability to trust predictions, outperforming existing approaches.
• This approach offers a more robust way to assess the reliability of continuous biological feature predictions.