Detecting repeated cancer evolution in human tumours from multi-region sequencing data

Carcinogenesis is an evolutionary process driven by the accumulation of genomic aberrations. Recurrent sequences of genomic changes, both between and within patients, reflect repeated evolution that is valuable for anticipating cancer progression. Multi-region sequencing and phylogenetic analysis allow inference of the partial temporal order of genomic changes within a patients tumour. However, the inherent stochasticity of the evolutionary process makes phylogenetic trees from different patients appear very distinct, preventing the robust identification of recurrent evolutionary trajectories. Here we present a novel quantitative method based on a machine learning approach called Transfer Learning (TL) that allows overcoming the stochastic effects of cancer evolution and highlighting hidden recurrences in cancer patient cohorts. When applied to multi-region sequencing datasets from lung, breast and renal cancer (708 samples from 160 patients), our method detected repeated evolutionary trajectories that determine novel patient subgroups, which reproduce in large singlesample cohorts (n=2,641) and have prognostic value. Our method provides a novel patient classification measure that is grounded in the cancer evolution paradigm, and which reveals repeated evolution during tumorigenesis, with implications for our ability to anticipate malignant evolution.