Detecting repeated cancer evolution from multi-region tumour sequencing data


Session type:


Giulio Caravagna1,Ylenia Giarratano2,Daniele Ramazzotti3,Ian Tomlinson4,Trevor Graham5,Guido Sanguinetti2,Andrea Sottoriva1
1Institute of Cancer Research,2University of Edinburgh,3Stanford University,4University of Birmingham,5Barts Cancer Institute



Recurrent successions of genomic changes, both within and between patients, reflect repeated evolutionary processes that are valuable for anticipating cancer progression. Multi-region sequencing allows the temporal order of some genomic changes to be inferred within a tumour, but the robust identification of repeated evolution across patients remains an unmet challenge.


Here we present REVOLVER (Repeated Evolution in Cancer), a machine learning method based on Transfer Learning that overcomes the stochastic effects of cancer evolution and noise in the data, and identifies hidden evolutionary patterns in cancer cohorts. Our method is based on a joint analysis of a full cohort of cancer patients, compared to previous methods that analyze each patient independently. 


When applied to multi-region sequencing datasets from lung, breast, renal and colorectal cancer (768 samples from 178 patients), our method detected repeated evolutionary trajectories. These were used to stratify the cohorts in subgroups of patients whose tumours are similar, from an evolutionary point of view. The groups reproduced in large-scale cross-sectional single-sample cohorts (n=2,935), showing different survival outcome associated to repeated evolutionary trajectories.


Detecting repeated evolution in cancer is critical for the implementation of evolutionary approaches to disease management. Stratifying patients based on their recurrent evolutionary patterns helps to predict future steps of malignant progression, thus potentially informing optimal and personalised clinical decisions. Our Transfer Learning approach combines high-quality multi-region sequencing data of driver alterations and phylogenetic theory to detect the hidden signal of repeated evolution within multiple tumour types. The repeated evolutionary trajectories we identified were associated with subsets of patients with distinct prognosis, demonstrating the likely clinical value of stratifying patients based on how their tumours evolved.