Many pathogenic cellular populations, including solid tumours, are densely packed. However, little is known about how long-ranging spatial correlations - an inherent feature of these systems - reshape the evolution of therapy resistance. Modelling the emergent phenomena that couple the mechanical interactions of individual cells to evolutionary outcomes on the population level is inherently challenging. In my presentation, I will discuss an integrated top-down-bottom-up approach that combines concepts from active granular matter physics and stochastic numerical models with agent-based simulations and data from genetically tailored microbial experiments. Using this strategy, I will show how spatial population expansion and a clone-size-dependent selection conspire to create an “inflation-selection balance”. The ensuing stabilization of slower-growing resistant clones facilitates their evolutionary rescue of via de novo cost-compensatory mutations and, consequently, therapy failure. Collective cell dynamics also shift the evolutionary dynamics during therapy. I will discuss how understanding tumors in the context of active granular matter may help to harness the full potential of evolution-based therapies. Finally, I will discuss how these physical effects could be integrated with mathematical models to leverage the power of reinforcement learning.