Automated culture and AI-enabled image analysis of compound responses in patient-derived colorectal cancer organoids

Organoids transformed biomedical research by providing physiologically relevant models for cancer studies, essential for investigating disease mechanisms and drug responses. However, manual organoid culture processes are labor-intensive and prone to variability, limiting widespread adoption. Additionally, extracting information from complex biological systems remains a major challenge in organoid research. Here, we present the development of methods for automated culture, expansion, and end-point assays using patient-derived organoids, combined with machine-learning approaches for image-based analysis of compound responses. Patient-derived colorectal cancer (CRC) organoids were cultured in Matrigel domes using the CellXpress.ai automated cell culture system, which enables automated seeding, media exchanges, imaging, and passaging. Imaging and media exchanges were set periodically, while passaging was either triggered by users or automatically, based on image analysis and phenotypes of organoids. Using this system, we successfully maintained organoid cultures for over a month and expanded them to quantities sufficient for multiple 96-well assay plates. Cultured organoids were treated with a panel of anti-cancer drugs representing diverse mechanisms of action. Dose-dependent effects on organoid morphology and cell viability were evaluated using automated confocal high-content imaging. Following drug treatment, organoids were either stained live with viability dyes or fixed and stained with a panel of markers for nuclei, cytoskeleton, mitochondria, and RNA content. Images were analyzed using a deep learning model in IN Carta analysis software. Machine learning tools offer an unbiased and comprehensive approach to analysis by automatically identifying organoids and scoring them as intact or damaged. The analysis protocol extracted a panel of quantitative features per organoid. These included morphological descriptors, intensity metrics across three fluorescent channels, textural features, and spatial distribution patterns. Features were extracted enabling multidimensional phenotypic profiling of organoid populations. Classification was performed using a combination of unsupervised and supervised machine learning. Unsupervised clustering grouped organoids into phenotypic clusters based on feature similarity, which were then curated into biologically relevant categories using user-defined criteria. The classification allowed evaluation of compound effects in a concentration-dependent manner. Unlike single-parameter readouts, this approach incorporates a broad morphological context. In summary, we demonstrated a fully automated workflow for evaluation of compound effects using 3D human organoids. The described approach enables AI-driven phenotypic profiling for assessing drug-induced effects in organoid models.