Drug resistance is a largely unsolved problem in oncology. Despite the explanatory power of the genetic model of cancer initiation, most treatment resistance is unexplained by genetics alone. Even when known resistance mutations are present, they are often found in a small proportion of the cells in the tumour. So where is the cellular memory that leads to treatment failure? New evidence suggests resistance is multi-factorial, resulting from the contribution of heritable genetic and epigenetic changes, but also non-heritable phenotypic plasticity. However, cell plasticity has proven hard to study as it dynamically changes over time and needs to be distinguished from clonal evolution where cell phenotypes change because of Darwinian selective bottlenecks. Here we dissected the contribution of different evolutionary processes to drug resistance by perturbing patient-derived organoids with multiple drugs in sequence. We combined dense longitudinal tracking, single cell multi-omics, evolutionary modelling, and machine learning archetypal analysis. We found that different drugs select for distinct subclones, an essential requirement for the use of evolutionary therapy with sequential drug treatment. The data supports a model in which the cellular memory is encoded as a heritable configuration of the epigenome, which however produces multiple transcriptional programmes. Those emerge in different proportions depending on the environment, giving rise to cellular plasticity. Epigenetically encoded programmes include reactivation of developmental genes and cell regeneration. A one-to-many (epi)genotype-phenotype map explains how clonal expansions and non-heritable phenotypic plasticity manifest together, including drug tolerant states. This ensures the robustness of drug resistance subclones that can exhibit distinct phenotypes in changing environments while still preserving the cellular memory encoding their selective advantage.

Epigenetic heritability of cell plasticity drives cancer drug resistance through one-to-many genotype to phenotype mapping