PiLSL

PiLSL predicts synthetic lethality between gene pairs in human cancers by learning pairwise interaction representations using a graph neural network integrated with multi-omics data.

Key Features:

• Pairwise interaction learning: Focuses on representing interactions between gene pairs rather than individual gene embeddings.
• Enclosing graph construction: Constructs an enclosing graph for each gene pair from a knowledge graph to capture potential interactions and dependencies.
• Attentive embedding propagation layer: Employs an attention-weighted embedding propagation layer within the GNN to assess edge importance and learn latent features specific to the pairwise interaction.
• Multi-omics integration: Fuses learned latent features with explicit features extracted from multi-omics data to produce robust gene-pair representations.
• Graph neural network framework: Uses a graph neural network architecture to learn from enclosing graphs and propagated embeddings.
• Mechanistic insight via attention-weighted paths: Highlights weighted paths within enclosing graphs through its attention mechanism to support interpretation of SL mechanisms.
• Performance and generalization: Demonstrates improved prediction performance over existing baselines across various scenarios and strong generalization.

Scientific Applications:

• Synthetic lethality prediction: Predicts SL interactions between gene pairs in human cancers.
• Therapeutic target prioritization: Supports broadening the spectrum of anti-cancer therapeutic targets by identifying candidate SL pairs.
• Mechanistic interpretation of SL mechanisms: Provides interpretable, attention-weighted paths in enclosing graphs to aid biological understanding of SL.

Methodology:

For each gene pair, an enclosing graph is constructed from a knowledge graph; an attentive embedding propagation layer within a graph neural network weights edges to learn latent pairwise features; the learned latent features are fused with explicit multi-omics features to form final gene-pair representations.