PSSpred

PSSpred predicts protein secondary structure using an ensemble of neural networks trained on profile-based sequence information to improve secondary-structure assignment accuracy.

Key Features:

• Neural Network Training Algorithm: Employs neural networks trained using the Rumelhart error backpropagation method for secondary structure prediction.
• Multiple Sequence Alignments (PSI-BLAST): Uses PSI-BLAST to collect multiple sequence alignments and generate profile inputs for the predictors.
• Amino-Acid Frequency and Log-Odds with Henikoff Weights: Incorporates amino-acid frequency and log-odds statistics computed with Henikoff weights to construct sequence profiles.
• Integration of Predictors: Combines predictions from seven distinct neural network predictors, each trained on different profile data and parameter settings, to produce the final prediction.

Scientific Applications:

• Protein secondary structure prediction: Provides residue-level secondary structure assignments for structural bioinformatics analyses.
• Template-based structure prediction and function annotation: Supplies secondary structure constraints to support template-based modeling and functional inference.
• Fold-recognition benchmarking: Supports evaluation and comparison of profile-profile versus sequence-profile and sequence-sequence alignment approaches (see PMID:24018415).

Methodology:

PSI-BLAST is used to collect multiple sequence alignments and generate profiles; profiles are encoded using amino-acid frequency and log-odds with Henikoff weights and used to train neural networks via Rumelhart error backpropagation; final predictions are obtained by combining outputs from seven neural network models trained on varied profiles and parameters.