Predicting GD2 expression across cancer types by the integration of pathway topology and transcriptome data.

Background: The disialoganglioside GD2 is a key cancer therapy target due to its overexpression in several cancers and limited presence in normal tissues. However, experimental assessment is technically challenging and not routinely available. We developed a computational framework that integrates reaction activity derived from transcriptomic data with the glycosphingolipid biosynthesis pathway to predict GD2 expression. Methods: We computed Reaction Activity Scores from transcriptomic data and weighted the reactions of a glycosphingolipid metabolic network, refining edge weights with topology-based transition probabilities to account for enzyme promiscuity. Cumulative activities of GD2-promoting and -mitigating reactions served as features in a Support Vector Machine (SVM) to model GD2-associated differences between neuroblastoma and normal tissue. SVM decision values were used as a continuous proxy for GD2 expression. We validated the predicted GD2 scores across independent datasets by comparing them with literature-reported values and flow-cytometric confirmation of a model-predicted high-GD2 tumor. Copy-number alteration (CNA) data were integrated to identify candidate genomic biomarkers of GD2-positive samples. Results: Our SVM-based GD2 score achieved balanced accuracy of 0.80 with a linear kernel, selected due to reduced overfitting risk and interpretability, while matching the accuracy of more complex kernels. The model transferred reliably across six independent RNA-seq datasets and reproduced known GD2 expression patterns, outperforming a two-gene signature in capturing subtype-specific heterogeneity and avoiding overestimation in normal brain tissue. Pan-cancer analyses revealed heterogeneous GD2 expression in several cancer subtypes. Notably, we experimentally confirmed high GD2 expression in clear cell sarcoma of the kidney, consistent with model predictions. CNA analysis implicated B4GALNT1 amplification as a GD2-promoting factor in dedifferentiated liposarcoma. To facilitate adoption of our approach, we developed GD2Viz, an R package with an interactive Shiny application for score computation, visualization, and analysis of user data. Conclusion: Our computational framework provides a robust, interpretable, biologically grounded predictor of GD2 expression, offering greater consistency and clinical interpretability over existing gene-based signatures. Importantly, with over 20 GD2-directed trials ongoing, our approach may help prioritize tumor entities with high GD2 levels, delineate candidate patient subgroups, and generate testable hypotheses in underexplored cancers, thereby supporting patient stratification and eligibility screening for clinical trials.