Image-based profiling and deep learning reveal morphological heterogeneity of colorectal cancer organoids

Kai Huang,Mingyue Li,Qiwei Li,Zaozao Chen,Ying Zhang,Zhongze Gu

DOI: https://doi.org/10.1016/j.compbiomed.2024.108322

Abstract:Patient-derived organoids have proven to be a highly relevant model for evaluating of disease mechanisms and drug efficacies, as they closely recapitulate in vivo physiology. Colorectal cancer organoids, specifically, exhibit a diverse range of morphologies, which have been analyzed with image-based profiling. However, the relationship between morphological subtypes and functional parameters of the organoids remains underexplored. Here, we identified two distinct morphological subtypes ("cystic" and "solid") across 31360 bright field images using image-based profiling, which correlated differently with viability and apoptosis level of colorectal cancer organoids. Leveraging object detection neural networks, we were able to categorize single organoids achieving higher viability scores as "cystic" than "solid" subtype. Furthermore, a deep generative model was proposed to predict apoptosis intensity based on a apoptosis-featured dataset encompassing over 17000 bright field and matched fluorescent images. Notably, a significant correlation of 0.91 between the predicted value and ground truth was achived, underscoring the feasibility of this generative model as a potential means for assessing organoid functional parameters. The underlying cellular heterogeneity of the organoids, i.e., conserved colonic cell types and rare immune components, was also verified with scRNA sequencing, implying a compromised tumor microenvironment. Additionally, the "cystic" subtype was identified as a relapse phenotype featuring intestinal stem cell signatures, suggesting that this visually discernible relapse phenotype shows potential as a novel biomarker for colorectal cancer diagnosis and prognosis. In summary, our findings demonstrate that the morphological heterogeneity of colorectal cancer organoids explicitly recapitulate the association of phenotypic features and exogenous perturbations through the image-based profiling, providing new insights into disease mechanisms.