BetatPred2

BetatPred2 predicts beta-turns in proteins using neural network models that integrate evolutionary (PSSMs from PSI-BLAST) and secondary-structure (PSIPRED) information to inform protein structural analysis.

Key Features:

• Neural Network Architecture: Employs two feed-forward back-propagation neural networks each with a single hidden layer, where the second network refines predictions using outputs from the first and secondary-structure data.
• Use of PSI-BLAST and PSIPRED: Trains the first network on position-specific scoring matrices (PSSMs) generated by PSI-BLAST and uses PSIPRED secondary-structure predictions as input to the second network.
• Dataset and Validation: Reported overall prediction accuracy of 75.5% evaluated on 427 nonhomologous protein chains using sevenfold cross-validation.
• Performance Metrics: Q(pred) 49.8%, Q(obs) 72.3%, and Matthews correlation coefficient 0.43.
• Comparative Performance: Reported metrics indicate BetatPred2 outperforms existing beta-turn prediction methods.

Scientific Applications:

• Structural biology: Prediction of beta-turns to support analysis of protein folding and conformational changes.
• Protein stability and function studies: Informing investigations into protein stability, functional sites, and intermolecular interactions.
• Structure annotation: Assisting annotation of protein structures with beta-turn locations.
• Enzyme mechanism studies: Supporting studies that relate local turns to enzymatic activity and mechanism.
• Drug design: Contributing structural information relevant to ligand binding and drug-target interactions.
• Biomaterials development: Providing structural insights useful for engineering novel biomaterials.

Methodology:

Generates PSSMs with PSI-BLAST, obtains secondary-structure predictions with PSIPRED, and applies a two-tiered system of feed-forward back-propagation neural networks (single hidden layer) with validation by sevenfold cross-validation on 427 nonhomologous protein chains.