DeepLPI

DeepLPI predicts interactions between long non-coding RNAs (lncRNAs) and protein isoforms using multimodal deep learning and multiple instance learning to account for isoform-specific binding.

Key Features:

• Isoform-aware prediction: Predicts interactions at the protein isoform level, accounting for multiple isoforms encoded by a single gene.
• Sequence and structure integration: Integrates sequence and structure data to extract intrinsic features of lncRNAs and proteins.
• Expression-derived topological features: Uses expression data to derive topological features and incorporate co-expression information.
• Multimodal deep learning: Employs a multimodal deep learning neural network to combine heterogeneous feature types.
• Conditional random field integration: Merges the multimodal network output with a conditional random field model for joint prediction.
• Multiple instance learning (MIL): Applies MIL to address the scarcity of known lncRNA–protein isoform interactions.
• Validation and performance metrics: Validated on human interactions from NPInter v3.0, showing a 4.7% improvement in AUC and a 5.9% increase in AUPRC over state-of-the-art methods, with supporting human and mouse case studies.

Scientific Applications:

• lncRNA–protein isoform interaction prediction: Predicts potential interactions between lncRNAs and specific protein isoforms.
• Isoform-specific interaction identification: Identifies interactions that differ across protein isoforms encoded by the same gene.
• Novel interaction prioritization: Ranks and prioritizes candidate novel lncRNA–protein isoform interactions for follow-up.
• Co-expression-informed prediction: Refines interaction predictions using co-expression and expression-derived topological features.
• Cross-species analysis: Applied to both human and mouse lncRNA–protein interaction datasets.
• Functional role exploration: Supports investigation of potential functional roles of lncRNAs and protein isoforms in biological contexts.

Methodology:

Integrates sequence and structure data to extract intrinsic features, uses expression data to derive topological features, combines a multimodal deep learning neural network with a conditional random field, and employs multiple instance learning (MIL).