NEXUS: Spatially Informed Graph Attention Learning of Pathology Foundation-Model Embeddings for Cancer Tissue-of-Origin Prediction

Cancer of Unknown Primary (CUP) accounts for 3–5% of cancer diagnoses and refers to metastatic cancer in which the tumor origin remains ambiguous despite extensive diagnostic testing, complicating treatment decisions. Molecular assays can help, but they require additional tissue samples, introducing concerns regarding tissue stewardship and increasing costs. Previous architectures for predicting tumor origin have shown that H&E whole-slide images contain useful morphological information, while spatial pathology studies suggest that tissue organization may also matter. However, it is still unclear whether tissue topology provides additional tissue-of-origin information beyond the features captured by pathology foundation models. We developed NEXUS, a Kolmogorov-Arnold Graph Attention Network trained on UNI2-h pathology foundation-model embeddings from TCGA patients, with embeddings modeled according to their spatial organization using a fixed-radius proximity graph. Across site-stratified five-fold cross-validation, NEXUS achieved 93.46% Top-1 accuracy across 26 cancer origins. To test whether the model was truly benefiting from spatial organization, we disrupted the original spatial arrangement by randomly reassigning tile locations and rebuilding the proximity graph from the shuffled coordinates. Performance dropped by about 0.93%, suggesting that spatial organization provides useful information beyond morphology alone. NEXUS also achieved 93.74% accuracy on an independent CPTAC cohort, supporting generalization across patient populations. Overall, NEXUS reveals that tissue of origin may be reflected not only in morphologic patterns, but also in tissue topology, offering more robust spatially informed learning and grounded biological insight in digital pathology.