On the encoding of proteins for disordered regions prediction

Area of Science:

• Protein bioinformatics
• Computational biology
• Structural bioinformatics

Background:

• Intrinsically disordered proteins (IDPs) lack stable 3D structures but are vital for biological processes.
• Predicting IDPs is essential for inferring protein structure and function.
• Machine learning methods are widely used for IDP prediction, with feature encoding being a critical factor.

Purpose of the Study:

• To adapt a systematic methodology for evaluating feature encodings for predicting disordered protein regions.
• To assess the relevance of various feature encodings, including a novel one based on residue solvent accessibility.
• To benchmark the prediction performance against state-of-the-art methods using CASP and PDB datasets.

Main Methods:

• Utilized ensembles of extremely randomized trees for prediction.
• Adapted a systematic methodology for feature relevance assessment.
• Evaluated performance on proteins from the 10th CASP competition and a large PDB subset.

Main Results:

• A novel feature encoding residue proximity based on solvent accessibility was identified as the second most important predictor.
• Evolutionary information was found to be the most critical feature for prediction.
• The residue-independent approach achieved competitive accuracy compared to existing methods.

Conclusions:

• The proposed feature encoding based on solvent accessibility significantly contributes to disordered region prediction.
• The study underscores the importance of feature engineering in machine learning for protein structure prediction.
• The developed method provides accurate predictions and is accessible via a web application.