Fig. 4: Minimalist classifiers for FFPE primary and metastasis cases.
From: Minimalist approaches to cancer tissue-of-origin classification by DNA methylation
For FFPE primary cases (a, b; n = 339, 12 cancer types), the model trained on TCGA fresh primary data (a) was far less accurate (74.6% vs. 97.6%, left panels) and less confident (right panels) on independent validation of FFPE primary cancers than the model trained on combined TCGA and randomly selected FFPE primary cases (b). Since the two models used the same 55 probes, a, b suggest that probes identified via TCGA analyses are robust—although classifiers may need to be re-trained on new data. For FFPE brain metastasis (c, d; n = 45, including three CUPs, four cancer types): c) t-SNE plot shows that FFPE brain metastasis cases (Ys, n = 42) are well-separated, and cluster with their fresh primary counterparts (dots); CUPs are indicated by black open circles. d The heat map/confusion matrix for the classifier trained on FFPE brain metastases, validated on the 39 FFPE brain metastases of known origins (six unique probes, 95.2% accuracy); CUPs are not shown here. e A single probe (cg22280705) accurately separated all nine lymph node metastases and 98.5% of 32 melanoma and 235 breast carcinoma primaries on validation. BLCA bladder carcinoma, BRCA breast invasive carcinoma, CORE colorectal adenocarcinoma, DLBC diffuse large B-cell lymphoma, GBMLGG glioma (glioblastoma and low-grade glioma), GCT germ cell tumor (intracranial), HNSC head and neck squamous cell carcinoma, KIPAN pan-kidney cohort (clear cell, chromophobe, and papillary renal cell carcinoma), LUAD lung adenocarcinoma, LUSC lung squamous cell carcinoma, PRAD prostate adenocarcinoma, SKCM skin cutaneous melanoma, UCEC uterine corpus endometrial carcinoma.
