Multi-resBind

Multi-resBind predicts in vivo RNA binding sites and visualizes binding preferences using a residual network-based multi-label classifier to analyze CLIP-based experimental data.

Key Features:

• Residual multi-label deep-learning architecture: Employs a residual network-based multi-task/multi-label classifier to model binding across multiple RNA-binding proteins (RBPs).
• Comparison to prior models: Demonstrates enhanced prediction power relative to previous models such as DeepRiPe through multi-label learning.
• Improved prediction accuracy: Shows substantial improvements in area under the receiver operating characteristic curve (AUROC) and average precision when evaluated on PAR-CLIP datasets.
• Bias mitigation: Addresses biases inherent in CLIP-based experimental protocols to provide more reliable binding site predictions.
• Comprehensive evaluation: Includes extensive experiments assessing the impact of different input data types and loss functions on prediction performance.
• Biological insight generation: Uses a modified integrated gradients method to produce attribution maps that disentangle context-specific contributions to protein–RNA interactions.

Scientific Applications:

• RNA binding site prediction: Identification of in vivo RNA binding sites for multiple RBPs from CLIP-based datasets.
• Binding preference visualization: Visualization of binding preferences and patterns across sequences to aid interpretation of protein–RNA interactions.
• Mechanistic insight extraction: Generation of attribution maps to reveal context-dependent contributions relevant to gene expression regulation and interaction mechanisms.

Methodology:

Uses a residual network-based multi-label deep-learning framework trained and evaluated on CLIP-based datasets including PAR-CLIP, with experiments varying input data types and loss functions, and employs a modified integrated gradients method to produce attribution maps.