Background: New treatments are urgently needed for advanced prostate cancer, with organoids established from patient-derived xenografts (PDX; human cancer samples grown in mice), emerging as a promising preclinical model system. While high-throughput drug screening using these 3D cultures can accelerate therapeutic discovery, accurately interpreting organoid responses remains challenging due to their complex morphology and diverse phenotypes. Furthermore, traditional endpoint viability assays provide limited insight into drug effects on organoid structure and behaviour. Machine learning approaches offer new possibilities for extracting and analysing multidimensional features from microscopic images of treated organoids.
Objective: Here, we aimed to develop an image-based high-content screening assay integrated with machine learning methods to enable more sophisticated analyses of drug responses in prostate cancer organoids.
Study design and results: We screened 882 compounds, selected from phase I-IV clinical trials or approved cancer therapeutics, on organoid cultures established from six PDXs of diverse phenotypes of prostate cancer, as well as two cell line-derived spheroid models. Using an automated high-throughput platform, we captured brightfield and Hoechst-stained fluorescent images of treated organoids at the end of the assay. Our machine learning pipeline extracted morphological and textural features to discriminate between viable and non-viable organoids. This approach successfully identified previously reported drug sensitivities, validating our screening platform, while also revealing novel therapeutic vulnerabilities specific to different prostate cancer phenotypes.
Conclusion: This approach demonstrates the power of combining automated high-content imaging with machine learning analyses in preclinical studies. Our platform presents a valuable advancement for drug screening and discovery, offering enhanced capability to analyse complex organoid phenotypes and drug responses.