Interpretable high-order knowledge graph neural network for predicting synthetic lethality in human cancers

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Synthetic lethality (SL) is a key strategy in cancer therapy, relying on identifying gene interactions.
• Current methods using knowledge graphs (KGs) and graph neural networks (GNNs) with attention mechanisms have limitations in explanation fidelity and capturing complex biological structures.

Purpose of the Study:

• To develop a novel KG-based GNN model, DGIB4SL, for more accurate and interpretable synthetic lethality prediction.
• To generate multiple, faithful explanations for SL gene pairs, overcoming the limitations of single-explanation methods.

Main Methods:

• Proposed DGIB4SL, a KG-based GNN incorporating a novel Diverse Graph Information Bottleneck (DGIB) objective.
• Integrated a determinant point process constraint into the information bottleneck objective.
• Utilized 13 motif-based adjacency matrices to encode high-order gene interaction structures.

Main Results:

• DGIB4SL demonstrated superior performance compared to state-of-the-art baseline methods in SL prediction.
• The model successfully generated multiple, diverse, and faithful explanations for SL gene pairs.
• The method effectively captured and encoded high-order biological structures.

Conclusions:

• DGIB4SL provides a more robust and interpretable approach to synthetic lethality prediction in cancer research.
• The ability to generate multiple explanations offers deeper insights into the diverse biological mechanisms underlying synthetic lethality.
• This advancement holds significant potential for developing novel cancer therapies.