Real value prediction of solvent accessibility in proteins using multiple sequence alignment and secondary structure

Area of Science:

• Computational biology
• Bioinformatics
• Protein structure prediction

Background:

• Solvent accessibility is crucial for understanding protein function and interactions.
• Predicting solvent accessibility from amino acid sequences is a key challenge in bioinformatics.

Purpose of the Study:

• To develop a novel neural network-based method for accurate prediction of solvent accessibility.
• To leverage evolutionary information and secondary structure for improved prediction accuracy.

Main Methods:

• Utilized two feed-forward neural networks trained with back-propagation.
• Employed multiple sequence alignments from PSI-BLAST and secondary structure predictions from PSIPRED as input features.
• Developed a two-step prediction process: sequence-to-structure and structure-to-structure networks.

Main Results:

• Incorporating multiple sequence alignment improved prediction accuracy (Pearson's correlation coefficient from 0.53 to 0.63).
• Further enhancement was achieved by including secondary structure information (correlation coefficient increased to 0.67).
• The final method achieved a mean absolute error of 15.2% on independent datasets and 15.9% on CASP6 proteins.

Conclusions:

• The developed neural network method effectively predicts solvent accessibility using evolutionary and secondary structure information.
• The SARpred server provides a valuable tool for predicting residue solvent accessibility from protein sequences.
• The study demonstrates the utility of integrating diverse sequence-derived features for enhanced protein structure prediction.