Rahimi, S.; Bonner, S.; Afzal, A.; Milo, M.; Morrissey, E.; Petsalaki, E.

Predicting gene essentiality across cellular contexts is a central challenge in computational biology, with implications for identifying cancer vulnerabilities. Graph neural networks (GNNs) integrate molecular interaction networks with gene-level features, but it remains unclear whether their performance gains arise from biologically meaningful connectivity or generic graph structure. Here, we systematically evaluate the role of network information in gene essentiality prediction using 2,741 genes across three tissues. We compare GNNs to feature-only baselines, including multilayer perceptron (MLP) and random forest (RF) methods, under a strict gene-level 5-fold cross-validation scheme to prevent information leakage. To isolate the role of network information, we assess models on the STRING protein-protein interaction network, a degree-preserving shuffled network, and a fully randomized network, with and without network-derived features. GNNs outperform feature-only models, reducing mean squared error and improving Matthews correlation coefficient across all tissues. However, these gains depend critically on biologically structured connectivity: performance degrades substantially under randomized topology and is not preserved by degree-constrained rewiring. Network features are largely redundant when using biologically meaningful graphs, as their information is recovered through message passing, but become important when topology is uninformative. Per-gene analyses reveal uniformly low correlations across models, highlighting intrinsic limits imposed by data variability. Graph Transformer models incorporating global attention do not outperform standard GNNs, indicating that predictive signals are predominantly local. Together, these results show that predictive gains arise from biologically structured connectivity rather than generic graph topology.