Porter5

Porter5 predicts protein secondary structure from amino acid sequences, providing three-state and eight-state secondary structure assignments for computational studies.

Key Features:

• State-of-the-Art Architecture: Uses an ensemble of cascaded Bidirectional Recurrent Neural Networks (BiRNNs) and Convolutional Neural Networks (CNNs) to capture sequential dependencies and local patterns in protein sequences.
• Novel Input Encoding: Employs input encoding techniques that support both single-sequence and evolutionary profile-based inputs for neural network processing.
• Comprehensive Training Dataset: Trained on an extensive set of protein structures to improve robustness and generalizability across diverse datasets.
• High Accuracy Rates: Reports 84% accuracy (81% Sequence Order Verification) for three-state predictions and 73% accuracy (70% SOV) for eight-state predictions in independent tests.
• Homology-Free Training: Can be retrained on SCOPe-based datasets eliminating homology between training and testing samples while maintaining high accuracy.

Scientific Applications:

• Protein Structure Prediction: Predicts secondary structure from amino acid sequences to aid understanding of protein folding and structural modeling.
• Functional Annotation: Provides insights into structure and function for unknown proteins based on predicted secondary structure.
• Drug Design and Discovery: Enables modeling of target protein secondary structure to inform rational drug design.

Methodology:

Implements an ensemble of cascaded Bidirectional Recurrent Neural Networks (BiRNNs) and Convolutional Neural Networks (CNNs); BiRNNs capture sequential dependencies and CNNs identify local patterns. Models are trained on large protein-structure datasets and can be retrained on SCOPe-based homology-free datasets; inputs use encodings for single sequences and evolutionary profiles.