Publications

Artificial intelligence is revolutionizing scientific discovery in medicine, with reinforcement learning (RL) emerging as a promising tool for optimizing therapeutic strategies. Yet applying RL to complex scenarios such as therapy dynamics in solid tumors is constrained by the challenge of constructing training environments that are both computationally efficient and mechanistically interpretable. Here, we introduce Reinforcement Failing, an AI-guided, human-in-the-loop discovery framework that shifts the focus from agent policy optimization to the refinement of the training environment itself. By combining multi-fidelity RL with group-relative performance evaluation across agent cohorts, Reinforcement Failing systematically reveals emergent mechanisms that first-principles models overlook. We apply this framework to adaptive therapy in solid tumors, which seeks to delay resistance-mediated treatment failure. In this setting, Reinforcement Failing uncovered a coupling between the mechanically driven collective motion of cells and spatially-heterogeneous proliferation that strongly influences therapy outcomes. We integrated these emergent physical mechanisms with existing domain knowledge into an augmented training environment, resulting in improved cross-environment therapeutic performance and exposure of potential pitfalls during translation. More broadly, these findings position Reinforcement Failing as a powerful artificial scientific discovery framework, capable of deciphering high-complexity processes at the interface of physics, machine learning, and medicine.

The evolution of therapy resistance in structured populations such as biofilms and solid tumours is shaped by emergent spatial organization, with profound consequences for evolution-based therapies. However, how treatment reshapes these patterns remains poorly understood. Here we show that intermittent treatment pulses transiently reconfigure the resource landscape, reorganize spatial growth zones and can enable resistant mutants to escape spatial confinement and drive therapy failure. We introduce a spatial evolution assay in which populations expand from single, genetically tailored yeast cells, enabling quantitative tracking of the full spatiotemporal trajectories of continually emerging resistant mutants under intermittent treatment. By integrating these observations with a mechanistically interpretable computational model in a real-to-sim-to-real loop, we identify a dynamic phase-transition-like boundary in schedule space that defines a candidate optimal balance between population control and sustained resistance confinement, which we test experimentally. Together, our results establish resource-mediated spatial confinement as a central organizing principle of resistance evolution and provide a mechanistic foundation for spatially informed, evolution-based therapies.

The mechanical properties of cells are dynamic, allowing them to adjust to different needs in different biological contexts. In recent years, advanced biophysical techniques have enabled the rapid, high-throughput assessment of single-cell mechanics, providing new insights into the regulation of the mechanical cell phenotype. However, the molecular mechanisms by which cells maintain and regulate their mechanical properties remain poorly understood. Here, we present a genome-scale RNA interference (RNAi) screen investigating the roles of kinase and phosphatase genes in regulating single-cell mechanics using Real-Time Fluorescence and Deformability Cytometry (RT-FDC). Our screen identified 82 known and novel mechanical regulators across diverse cellular functions from 214 targeted genes, leveraging RT-FDC’s unique capabilities for comprehensive, high-throughput mechanical phenotyping with single-cell and cell cycle resolution. These findings refine our understanding of how signaling pathways coordinate structural determinants of cell mechanical phenotypes and provide a starting point for uncovering new molecular targets involved in biomechanical regulation across diverse biological systems.