Answers are grounded in this paper and cite their sources.
Motivation Predicting gene essentiality can reveal cancer vulnerabilities, but whether graph neural networks benefit from biologically meaningful connectivity rather than generic graph structure remains unclear. Results We evaluated 2,741 genes across three tissues using repeated gene-level cross-validation. GraphSAGE outperformed a capacity-matched multilayer perceptron and achieved comparable mean squared error to random forest, with higher Matthews correlation coefficients. The biological interaction network outperformed 25 degree-preserving rewired and 25 fully random networks, with no randomized realization matching its performance. Removing network-derived features minimally affected the biological network but substantially impaired randomized networks, suggesting that message passing exploits biological connectivity. Permuting stable gene-level or complete gene feature profiles also reduced performance, demonstrating the importance of biologically aligned features. However, per-gene correlations remained near zero, indicating limited prediction of context-specific variation. Adding global attention did not improve performance over local message passing. These findings support complementary contributions from biological network organization and aligned gene features to generalization across unseen genes. Availability and implementation Data are publicly available from the Cancer Dependency Map and STRING databases. Code is available at https://github.com/Whole-Cell-Signalling/gnn-sko-essentiality.
No takes yet. Share an insight, caveat, or question.
Rahimi et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: