Attractome: From Theory to Generative Models for Cancer Dynamics and Control

Cancer progression couples intracellular gene-regulatory dynamics with eco-evolutionary selection, yet current approaches often model these axes separately. We introduce the Attractome, a unified framework that treats malignant phenotypes as attractors and their persistence via frequency-dependent selection. Effective barrier height between phenotypic basins links stochastic transition rates, mutational accessibility, and minimal control effort for reprogramming. We formalize a joint stability criterion: long-lived malignant phenotypes, including cancer stem cell states, must be both dynamically stable as attractors and evolutionarily stable under ecological feedback. We derive early-warning signatures of impending transitions, including critical slowing down, variance amplification, flickering, and spatial correlation, measurable in time-resolved single-cell and spatial genomic data. We outline a roadmap in which generative AI models learn quasi-potentials and barrier geometry from single-cell data for risk prediction and patient-adaptive intervention design.