Multiclass cancer classification and gene selection using mutation information

A recent study showed that somatic mutations of well-known cancer driver genes, found in circulating tumour DNA, can be detected in the bloodstream of patients with different cancer types. This makes blood screening a potential approach for non-invasive early cancer detection. The few genes tested in that study, however, are by themselves insuf-ficient to identify the cancer type. Here, we investigated the extent to which mutations in general can distinguish between different cancer types. Using data from the cBioPortal for Cancer Genomics, we examined the somatic mutations in 5929 tumour samples spanning 22 cancer types. We cast the data into a binary matrix, using “1 ” and “0 ” to indicate the presence or absence of somatic mutations in a gene. We then em-ployed conventional machine learning techniques—namely logistic re-gression, support vector machines (SVM) and random forest—to classify the tumour samples. We found linear classifiers to work better than non