SARpred

SARpred predicts the real value of solvent accessibility (SA) in protein sequences using neural network models that integrate position-specific scoring matrices (PSSMs) from PSI-BLAST and secondary-structure predictions from PSIPRED to leverage evolutionary information.

Key Features:

• Neural Network Architecture: Two feed-forward neural networks, each with a single hidden layer, constitute the predictive model.
• Learning Algorithm: Networks are trained using standard back-propagation.
• Input Data: Primary input consists of position-specific scoring matrices (PSSMs) generated by PSI-BLAST from multiple sequence alignments (MSAs).
• Two-stage Prediction: A sequence-to-structure network is trained on PSSMs and a structure-to-structure network refines predictions by incorporating secondary-structure information.
• Secondary Structure Integration: Secondary-structure predictions from PSIPRED are used as input to the second network to improve accuracy.
• Performance and Metrics: Use of PSI-BLAST PSSMs increased Pearson correlation from 0.53 to 0.63 and reduced MAE from 18.2% to 16%; adding PSIPRED secondary-structure information raised correlation to 0.67 and reduced MAE to 15.2% on nonhomologous nonredundant datasets, and evaluation on CASP6 yielded a correlation of 0.68 and MAE of 15.9%.

Scientific Applications:

• Structural Bioinformatics: Providing per-residue real-value solvent accessibility predictions to inform structure-function analyses, dynamics, stability, and biomolecular interactions.
• Protein Structure Modeling: Supplying SA estimates that can be used as inputs or restraints in protein modeling and refinement workflows.
• Functional Site Analysis: Assisting identification of solvent-exposed or buried residues for locating functional or interaction sites.

Methodology:

Generate PSSMs via PSI-BLAST from MSAs; train a feed-forward sequence-to-structure network on those PSSMs using back-propagation; and refine outputs with a second feed-forward structure-to-structure network that incorporates PSIPRED secondary-structure predictions.