PiLSL: pairwise interaction learning-based graph neural network for synthetic lethality prediction in human cancers.

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• Synthetic lethality (SL) offers novel anti-cancer therapeutic targets.
• Current methods often overlook gene pairwise interactions and lack mechanistic insights.
• Machine learning approaches typically focus on individual gene representations.

Purpose of the Study:

• To develop a novel graph neural network (GNN) model, PiLSL, for predicting synthetic lethality.
• To enhance SL prediction by learning pairwise gene interactions.
• To provide mechanistic explanations for SL interactions.

Main Methods:

• Constructing enclosing graphs for gene pairs from knowledge graphs.
• Employing an attentive embedding propagation layer within a GNN to learn weighted gene pair representations.
• Integrating latent features with multi-omics data for robust gene representation.

Main Results:

• PiLSL significantly outperforms existing baseline methods in SL prediction.
• The model demonstrates strong generalization capabilities across diverse scenarios.
• PiLSL offers interpretable insights into SL mechanisms through attention-weighted paths.

Conclusions:

• PiLSL effectively predicts synthetic lethality by focusing on pairwise gene interactions.
• The model's interpretability aids in understanding SL mechanisms.
• PiLSL advances the translation of synthetic lethality into cancer therapeutics.