Learning the Signatures of Cancer (360G-Wellcome-203735_Z_16_A)

Cancer is a genetic disease that is the second leading cause of death worldwide. Developing effective personalised therapies requires characterisation of the genetic factors driving malignancy. This is challenging as cancer is highly complex, heterogeneous, and dependent on cellular context. Cancer stratification aims to group cancers that share similar features, and are therefore likely to respond similarly to treatment, however, current stratification methods ignore many important genetic and epigenetic markers that likely influence cancer pathology, which would result in sub-optimal treatment. We propose to use whole genome-and-epigenome profiling and machine learning to extract clinically meaningful features of the host and cancer genomes that can be used to improve patient stratification and reveal novel cancer subtypes. As a proof of principle, we will apply these methods to predict the site of origin in patients with metastatic cancer but unknown primary (CUP), which could help improve diagnosis and prognosis for patients with this complex disease. We envision the robust stratification of cancer patients using genome profiling could lead to direct prediction of optimal treatment decision for all cancer patients.