DeepDISOBind

DeepDISOBind predicts intrinsically disordered regions (IDRs) in proteins that bind DNA, RNA, and other proteins to identify IDR-mediated binding interactions.

Key Features:

• Deep Multi-Task Learning Architecture: Employs a deep multi-task learning neural network that processes information-rich sequence profiles of protein sequences to predict binding IDRs for multiple partner types.
• Hierarchical Prediction Layers: Uses a common input layer that branches into specialized layers distinguishing nucleic acid-binding (DNA/RNA) versus protein-binding IDRs and further separates DNA- and RNA-specific prediction layers.
• Statistical Superiority: Demonstrates statistically significant improvements in predictive accuracy across DNA, RNA, and protein partners compared to single-task models and methods that combine disorder- and structure-trained tools.
• Protein-Level Propensities: Aggregates residue-level predictions into protein-level propensities that predict DNA- and RNA-binding proteins and identify protein hubs, validated on the human proteome.

Scientific Applications:

• Protein Function Annotation: Identifying potential binding partners mediated by IDRs to inform protein function annotations.
• Network Biology: Mapping cellular interaction networks by predicting protein hubs and IDR interactions with nucleic acids and proteins.
• Drug Discovery: Prioritizing candidate therapeutic targets by elucidating roles of IDR-mediated interactions in disease-related pathways.

Methodology:

Processes protein sequences through a deep neural network that leverages multi-task learning on information-rich sequence profiles; a hierarchical network structure branches and refines predictions for nucleic acid- versus protein-binding IDRs and for DNA- versus RNA-specific interactions.