DISULFIND

DISULFIND predicts the disulfide bonding state of cysteine residues and their connectivity within protein sequences to inform protein structural and folding analysis.

Key Features:

• Prediction Capabilities: Predicts the disulfide bonding state of cysteines and their pairwise connectivity from amino acid sequence data.
• Optional Known-State Input: Accepts known disulfide bonding states as input to refine connectivity predictions.
• Recursive Neural Networks: Employs recursive neural networks (RNNs) that operate on graph representations to model complex connectivity patterns.
• Graph Representation: Represents proteins as graphs with vertices labeled by real-valued vectors to capture residue-specific features.
• Evolutionary Profiles: Incorporates multiple alignment profiles to include evolutionary information in predictions.
• Visualization and Confidence: Outputs predicted bonding states with confidence levels and the most probable disulfide connectivity patterns.
• Performance Benchmarking: Demonstrates improved accuracy on the SWISS-PROT 39 dataset compared to weighted graph matching algorithms.
• Integration with Other Methods: Integrates predictions with MetalDetector and other methods via decision trees to enhance precision and recall for cysteine/histidine classification and bonding assignment.

Scientific Applications:

• Protein folding analysis: Provides residue-level disulfide connectivity information useful for studying protein folding constraints.
• Structural biology: Informs structural models by predicting disulfide bridges that influence protein conformation and stability.
• Protein engineering: Assists design and stability assessment by identifying potential disulfide bonds from sequence data.

Methodology:

Uses recursive neural networks applied to graph representations with vertices labeled by real-valued vectors and integrates multiple alignment profiles to incorporate evolutionary information; optionally uses known disulfide states as input; results were evaluated on the SWISS-PROT 39 dataset and compared to weighted graph matching algorithms, and predictions can be combined with MetalDetector outputs via decision trees.