CDR-Informed Graph Neural Network for Plasma Metabolomics in Lung Cancer and Pulmonary Neuroendocrine Neoplasms.

Plasma metabolomics offers a promising avenue for the non-invasive classification of lung cancer. However, most classification frameworks treat lung cancer as a single entity, conflating non-small-cell lung cancer (NSCLC) with pulmonary neuroendocrine neoplasms (NENs), despite their distinct biology and treatment pathways. We apply a heterogeneous graph neural network to a three-class problem discriminating Control, NSCLC, and NEN from targeted plasma metabolomic profiling in 800 participants (466 NSCLC, 120 NEN, 214 Controls). Matched metabolites were annotated by Cell Danger Response (CDR) relevance class and expected direction of change, with these annotations encoded as metabolite features and direction-aware patient-metabolite edge weights. Across ten random seeds, the three-class model achieved a macro-F1 of 0.898±0.026 and macro-AUROC of 0.962±0.016. Collapsing the three-class posteriors to a cancer-versus-control score yielded an AUROC of 0.951±0.017, a sensitivity of 0.960±0.011, and an F1 score of 0.949±0.010. A separate cancer-only classifier distinguished NEN from NSCLC with an accuracy of 0.960±0.026 and an AUROC of 0.985±0.020. The gradient-times-input attribution identified one-carbon, glycine-serine, proline, and ornithine-arginine metabolic programs, with proline, C5DC, uric acid, and fumaric acid among the leading annotated metabolites. Removal of all 11 cohort-derived Class N annotations left NSCLC-versus-NEN AUROC essentially unchanged, and broader comparator analyses showed no measurable predictive advantage of the CDR prior over the otherwise matched concentration-weighted GNN. These findings indicate that predictive discrimination is driven primarily by the measured metabolomic features, whereas the CDR component provides a biologically structured framework for directional metabolite and pathway interpretation. Independent external and prospective validation is required to establish generalizability and clinical utility.