Brian M. Larsen, Michelle Stein, Madhavi Kannan, Yilin Zhang, Veronica Sanchez-Freire, Andrea Cancino, Luka Karginov, Aly A. Khan, and Ameen A. Salahudeen
Patient derived Tumor Organoids (TOs) are emerging as patient-representative models that recapitulate clinical responses to candidate therapeutics. Yet standard methods of interpreting in vitro treatments of TOs have been developed from monoclonal, rapidly proliferating 2D cell lines that are not amenable for TOs which have limited biomass and intra-tumoral clonal heterogeneity. To address these challenges, we developed and optimized a drug screening platform more applicable for the unique characteristics of TOs. The platform couples high content fluorescent confocal imaging analysis with a robust statistical analytical approach to measure hundreds of discrete data points of TO viability from as few as 10^3 cells. We validated this approach through evaluating responses to hundreds of small molecule inhibitors as well as a panel of chemotherapeutic agents in TO models derived from different patients.
The platform was highly reproducible with minimal intra- and inter-assay variance (well:well variance = ns, plate:plate variance = ns, by ANOVA). Drug responses were both robust and reliable with a Z’ value of 0.8. QC of TOs was performed to remove outlier TOs by size and remaining TOs were normalized by mean vehicle proportion survival. To compare differential therapeutic toxicity between TOs from different patients, we developed a linear model to evaluate differences in proportion of surviving cells across equivalent therapeutic concentrations, identifying highly significant differences (P < 10-12). Intriguingly, the linear model not only uncovered heterogeneity of responses between TOs derived from different patients, but also identified organoid clonal populations derived from the same patient with differential drug response offering a window into uncovering functional intra-tumoral heterogeneity.
Lastly, we substantially increased throughput by applying a machine learning algorithm to predict therapeutic response via TO morphological changes from light microscopy. Employing this algorithm eliminated the need for fluorescent labeling leading to increased assay throughput by 3-4 fold amounting to 96 and 384 well plate acquisitions in as little as 5 and 15 minutes respectively. In summary, we describe a high-throughput platform able to measure TO therapeutic response with high statistical confidence and exquisite inter-assay reproducibility. This approach can be utilized in research settings to elucidate heterogeneity of therapeutic responses within and among patients, and may be utilized in the clinical laboratory to potentially guide precision oncology treatments.
VIEW THE PUBLICATION