Context-aware gene dependency modeling via graph attention networks for precision oncology

Understanding context-specific gene dependencies is essential for uncovering therapeutic vulnerabilities and advancing precision oncology. However, existing computational methods often fail to capture sample-specific effects or provide biological interpretability. We developed GATDep, a graph attention network-based framework that predicts sample-specific gene dependency scores by integrating transcriptomic data with curated gene-gene interaction networks. Each sample is represented as a context-aware graph instance, where node features are derived from gene set variation analysis, enabling pathway-level encoding of cellular states. GATDep frames dependency prediction as a node-level regression problem and outperforms state-of-the-art models across multiple benchmark datasets. The model provides interpretable attention weights and feature attribution maps via GNNExplainer, revealing biologically meaningful gene interactions. When applied to CRISPR and RNAi screening datasets, GATDep accurately identifies both known and novel dependencies and shows strong correlations with drug sensitivity. Furthermore, analysis of TCGA patient cohorts demonstrates significant associations between predicted dependencies and key clinical features, including microsatellite instability, chemotherapy response, and pathway-level perturbations. GATDep offers an extensible and interpretable framework for modeling gene essentiality at the sample level. By combining network topology with transcriptomic context, it enables the discovery of functionally and clinically relevant gene dependencies, providing a foundation for drug prioritization and precision cancer therapy.