reprof

Reprof predicts protein secondary structure and residue accessibility from amino-acid sequences and multiple sequence alignments using neural network models that incorporate evolutionary information.

Key Features:

• Two-Layered Feed-Forward Neural Network: Reprof employs a two-layered feed-forward neural network trained on a non-redundant database of 130 protein chains.
• Incorporation of Evolutionary Information: Uses multiple sequence alignments to incorporate evolutionary and protein family information, improving prediction accuracy by 6–8 percentage points over single-sequence methods.
• Three-State Prediction Accuracy: Reports overall three-state accuracies of 70.8% for globular proteins and 70.2% for membrane proteins.
• Balanced Secondary-Structure Predictions: Produces balanced predictions across alpha-helix, beta-strand, and loop, with approximately 65% correct prediction for strand residues.
• Position-Specific Reliability Index: Defines a position-specific reliability index with accuracy exceeding 82% for residues predicted with high reliability.
• Segment Length Prediction: Produces realistic predictions of secondary-structure segment lengths.

Scientific Applications:

• Comparison to Circular Dichroism: Prediction results are comparable to circular dichroism spectroscopy for assessing secondary-structure content.
• Validation and Benchmarking: Method performance was validated by sevenfold cross-validation and additional evaluation on 26 recently solved protein structures.
• Secondary-Structure Prediction for Diverse Proteins: Applied to predict secondary-structure elements in both globular and membrane proteins with reported accuracy metrics.

Methodology:

Predictions are generated by a two-layered feed-forward neural network trained on a non-redundant set of 130 protein chains that accepts either amino-acid sequences or multiple sequence alignments as input (alignments recommended to incorporate evolutionary information); the implementation is in Perl.