Abstract
Intratumoral heterogeneity, originating from genetic, epigenetic, and phenotypic cellular diversity, is pervasive in cancer. As these heterogeneous states employ diverse mechanisms to promote tumor progression, metastasis, and therapy resistance, emergence of cancer heterogeneity is one of the most significant barriers to curative treatment. Here we leverage deep learning approaches to develop a high-throughput image-analysis paradigm with subcellular resolution that quantifies and predicts colorectal cancer (CRC) patient outcome based on tissue architecture and nuclear morphology. We further combine this approach with spatial transcriptomics, multiplex immunohistochemistry, and patient-derived organoids to uncover a dynamic, co-evolutionary relationship between tumor architecture and cell states. We identify clinically relevant architectural interfaces in CRC tissue that diversify cellular identities by favoring distinct cancer stem cell states and thus promote evolution of tumor heterogeneity. Specifically, tissue fragmentation and associated compressive forces promote loss of classic stem cell signature and acquisition of fetal/regenerative stem cell states, which initially emerges as a hybrid state with features of epithelial-to-mesenchyme (EMT) transition. Additional tumor stroma communication then diversifies these states into distinct stem cell and EMT states at the invasive margins of tumors. Reciprocally, Wnt signaling state of tumor cells tunes their responsiveness to tissue architecture. Collectively, this work uncovers a feedback loop between cell states and tissue architecture that drives cancer heterogeneity and cell state diversification. Machine learning-based analyses harness this co-dependency, independently of mutational status or cancer stage, to predict patient outcome.