Fig. 4

Results on test set for pan-cancer sub-type predictions: a Comparison of MethylNet derived pan-cancer classification of test set (n = 1676) to UMAP+SVM method. 95% confidence intervals for each score were calculated using a 1000 sample non-parametric bootstrap; b Hierarchical clustering of average embedding cosine distance between all pairs of cancer subtypes. Cancer subtypes from both axes are colored by cancer superclasses, derived using the hierarchical clustering method. The clustering of similar MethylNet embeddings is concordant with known biology of tissue/cancer type difference. Skin and connective tissue cancers, and bile and liver cancers in Cluster 1. All kidney cancers in Cluster 2. Bladder, uterine and cervix cancers in Cluster 3. Pairing of colon and rectal cancers, both adrenal cancers in Cluster 4. A tie between lung adenocarcinoma and mesothelioma in Cluster 5, both of which may develop in similar locations. Pairings between stomach and esophagus cancer, and pancreas and prostate cancers in Cluster 6. Brain cancers in Cluster 7. Thymoma, Diffuse Large B-Cell lymphomas in Cluster 8. While the lung cancers were not paired together, they experienced a high degree of embedded similarity. The connectivity between the lung squamous cell cancer and its neighboring types prevented the two cancers from being grouped together