Tumor Origins Through Genomic Profiles
Menée à partir de données génomiques portant sur 7 791 tumeurs de 22 types différents puis validée sur 11 644 autres tumeurs, cette étude évalue la performance d'un algorithme d'apprentissage automatique, intégrant des profils mutationnels tumoraux (nombre de copies de gènes, présence de variants à simple nucléotide, de délétions, d'insertions et de réarrangements structurels), pour identifier avec précision le type et l'origine des tumeurs
One way to look at genome panels in cancer is as a collection of hundreds of individual genetic diagnostic tests, such as, EGFR mutation, EML4-ALK translocation, that can each be used to extract useful clinical information to guide therapy. However, the behavior of the collection of mutations can also act as a clinical parameter of value. For example, the tumor mutational burden (TMB), which scores the total mutational load within a tumor, is used to measure the proclivity of a tumor to respond to immuno-oncologic agents. In this issue of JAMA Oncology, Penson et al from Memorial Sloan Kettering Cancer Center advanced this concept further in describing an approach that uses artificial intelligence to assess higher meaning of the mutational profile from a 468-gene cancer panel, the Memorial Sloan Kettering–Integrated Mutation Profiling of Actionable Cancer Targets (MSK-IMPACT). From a training cohort of tumors from 7791 patients with a variety of cancers, they used single-nucleotide variations, indels, copy number changes, and structural rearrangements to build classifiers that could distinguish the tissue of origin of each tumor. They then validated the classifier in an independent test cohort of 11 644 patient tumors. Their results showed an accuracy of between 73.8% and 74.1% in predicting the correct tissue of origin with greater successes in some tumor types than others. The best predictor was for uveal melanomas, gliomas, and colorectal cancers, whereas, the poorest was for esophagogastric, ovarian, and head and neck cancer, cancers with greatest genomic mutational heterogeneity. A unique aspect of their predictor is that a probability score was assigned to each result that allowed the clinician to have an estimate of the certainty of the tissue assignment. Thus, even in those problematic tumors, misdiagnosis could be avoided by censoring the ambiguous cases.
JAMA Oncology , éditorial, 2018