Machine Learning for Predicting Cancer Severity

Cancer is an extremely heterogeneous disease, and this property becomes increasingly exacerbated as the disease progresses. Its heterogeneity can be reflected by protein signature, called its proteome, which is essentially a dataset indicating which proteins are highly expressed or lowly expressed in a tumour. Whilst no two cancer proteomes are the same, various patterns in the proteome can help distinguish certain cancers from others. There are prominent proteomic patterns that a Machine Learning (ML) technique can pick from different cancer types. However, identifying severity patterns across cancer types is challenging due to major proteomic differences interfering with ML performance. Accordingly, proteomic analyses are rarely performed on datasets consisting of multiple cancer types unless aiming to distinguish between two types. In this study, we tested three ML algorithms in classifying the TCGA (The Cancer Genome Atlas) PanCancer dataset, consisting of 32 different cancer types, into various clinically relevant metrics, such as stage, grade, and treatment response of tumour. On average, we achieved the best accuracies when employing a Support Vector Machine (SVM) classifier with a Radial Basis Function (RBF). The highest accuracies were accomplished when classifying based on pathological stage, suggesting a possible future application as a diagnostic tool, where cancers can be staged based on a quick ML classification rather than a lengthy evaluation by a pathologist.