Logistic regression models to predict solvent accessible residues using sequence- and homology-based qualitative and...

Area of Science:

• Computational Biology
• Structural Bioinformatics
• Machine Learning in Biology

Background:

• Predicting protein relative solvent accessibility (RSA) is crucial for understanding protein structure and function.
• Existing methods for RSA prediction vary in computational intensity and accuracy.
• Developing accurate and computationally efficient RSA prediction models remains an active research area.

Purpose of the Study:

• To develop and validate novel logistic regression models for predicting protein residue solvent accessibility.
• To evaluate the performance of models incorporating sequence-based and disorder propensity descriptors.
• To provide accessible tools for researchers to explore RSA prediction.

Main Methods:

• Utilized a domain-complete learning set of over 1300 proteins.
• Developed logistic regression models using qualitative (amino acid type) and quantitative (sequence entropy) descriptors.
• Incorporated homology descriptors from BLASTp alignments and residue disorder propensity.
• Trained models using dichotomous responses (buried/accessible) derived from RSA values from crystal structures.

Main Results:

• Achieved binary prediction accuracies comparable to existing methods using standard RSA thresholds (20% and 25%).
• Demonstrated incremental accuracy improvements by including the Lobanov-Galzitskaya residue disorder propensity descriptor.
• Reported 25% threshold accuracies of 76.12% and 74.79% on the Manesh-215 and CASP(8+9) test sets, respectively.

Conclusions:

• Logistic regression models with sequence and disorder propensity descriptors provide accurate protein solvent accessibility predictions.
• The developed models offer a computationally efficient and interpretable approach to RSA prediction.
• The provided software and datasets facilitate further research and educational exploration in RSA prediction.