Trace the Cancer of Unknown Primary Origin and Molecular Subtype Via Machine Learning

Knowledge of a tumor’s primary site of origin and molecular subtype plays a critical role in the choice of treatment regimen and prognosis. We developed accurate machine learning based predictors of cancer primary site of origin and molecular subtype using gene expression data spanning 33 primary cancer types and molecular subtypes from 11 primary cancer types from The Cancer Genome Atlas (TCGA). Validation using external datasets shows that both primary site and molecular subtype predictors can robustly predict the primary site of origin and molecular subtype from microarray data, metastatic tumor data, and patient-derived xenograft (PDX) data. The predictors are available as open source software, freely available for academic non-commercial use.