This paper introduces AAnet, a novel neural network method for Archetypal Analysis (AA) designed to resolve the continuous, spatially localized cell states that characterize intratumoral heterogeneity, especially in cancer. The authors developed AAnet to overcome the limitations of traditional methods like clustering and linear AA when dealing with the highly nonlinear geometry of single-cell RNA sequencing (scRNA-seq) data, allowing it to learn extreme cellular states ("archetypes" or ATs) and map cells as a convex combination of these ATs in a latent space. Applying AAnet to a triple-negative breast cancer (TNBC) model, the study identified five distinct ATs in primary tumors and metastases—including proliferative, oxidative, hypoxic, cell damage/death, and immunostimulatory states—and further used spatial transcriptomics to reveal that these ATs possess unique and structured organizations within the tumor that correlate with specific microenvironment (ME) ATs and paracrine signaling. Finally, the research highlights the metabolic vulnerability of the hypoxic AT, showing that targeting the glucose transporter GLUT3 significantly inhibits tumor growth and metastasis, and validates the functional ATs by finding similar metabolic and immune signatures in human breast cancer samples.

References:

• Venkat A, Youlten S E, San Juan B P, et al. AAnet resolves a continuum of spatially-localized cell states to unveil intratumoral heterogeneity[J]. Cancer Discovery, 2025.