Generative modeling reveals the connection between nuclear morphology and gene expression

Abstract

How transcriptional programs shape nuclear morphology, and how morphological features in turn reflect and influence cell identity and function, remains poorly understood. Here, we introduce COSMIC, a bidirectional generative framework that enables quantitative decomposition of transcriptional variance reflected in nuclear morphology and morphological variance explained by gene expression. Pretrained on over 21 million segmented nuclei, COSMIC models cross-modal relationships by using a multimodal dataset acquired with IRIS, a technology that captures high-resolution images and transcriptomes from the same single cells. COSMIC learns relationships between gene expression and nuclear morphology, enabling the identification of genes whose expression covaries with nuclear morphological variation. In prostate cancer cells, COSMIC distinguishes chemotherapy-responsive and -resistant cell states through coordinated morphological and transcriptional changes, revealing morphology-associated genes linked to tumour state. Applied to spatial transcriptomics of zebrafish embryogenesis, COSMIC reveals how local cellular neighborhoods influence morphology-expression relationships. These results demonstrate that generative modeling powered by paired single-cell measurements can quantify the information shared between nuclear morphology and gene expression, opening new avenues for mechanistic discovery in both basic and translational cell biology.