Graph contextualized attention network for predicting synthetic lethality in human cancers.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Synthetic lethality (SL) is crucial for targeted anticancer therapeutics, offering selective cancer cell killing.
• In silico prediction of SL interactions provides a cost-effective alternative to expensive wet-lab experiments.
• Existing methods struggle with capturing neighbor dependencies and predicting interactions for new genes.

Purpose of the Study:

• To propose a novel graph contextualized attention network (GCATSL) for improved synthetic lethality prediction.
• To develop a method that effectively captures local and global gene dependencies.
• To enhance the prediction accuracy for new genes without known SL partners.

Main Methods:

• Constructing multiple gene feature graphs from diverse data sources.
• Employing node-level attention to learn local and global gene representations.
• Utilizing feature-level attention to integrate gene representations across different feature graphs.

Main Results:

• GCATSL significantly outperforms 14 state-of-the-art methods across three datasets.
• The model demonstrates consistent performance improvements in synthetic lethality prediction.
• Case studies confirm GCATSL's effectiveness in identifying novel synthetic lethal pairs.

Conclusions:

• GCATSL offers a powerful and accurate approach for in silico prediction of synthetic lethality.
• The method advances the identification of potential targets for anticancer drug development.
• GCATSL provides a valuable tool for guiding experimental screening of synthetic lethal interactions.