High-content image analysis of colorectal cancer patient-derived tumroids identifies high intra- and inter-patient heterogeneity associated with drug activity and mode-of-action

Background: Ex vivo culture of colorectal cancer (CRC) patient-derived tumoroids reveals a rich heterogenous pool of growth morphologies, within and among models. Quantitative analysis of tumoroid morphologies can uncover new biology and predictive relationships. Studies have shown that compact vs. cystic morphology reflects drug sensitivity, with cystic tumoroids being sensitive to Wnt inhibitors and resistant to MEK inhibitors as compared to compact tumoroids. Methods: Individual tumoroids (n=4049) were annotated based on visual assessment of brightfield, nuclear and nuclear dead stain multi-channel images. ConvNeXt v2 classifiers were trained on the resulting dataset, using brightfield or all three channels as inputs. Results: We developed a phenotyper algorithm (iCANdy) stratifying CRC tumoroids into 10 classes: compact, budding, acinar, cystic (thick-walled), cystic (thin-walled), invasive, fibroblastic, monolayer, single/apoptotic and mixed/other. Performance was similar for brightfield and multi-channel inputs. Classifying all tumoroids according to the nearest morphology yielded an F1-score of 0.80 and balanced accuracy of 0.75, the most ambiguous class being mixed/other as expected. iCANdy was applied on brightfield images of 10.2 million tumoroids from a total of 22k drug treatment/control wells from 57 tumoroid models, using 37 drugs or drug combinations. For untreated controls, the dominant morphology in terms of total area varied across models with compact (39%), invasive (32%), single/apoptotic (12%), budding (9%), and cystic (thick-walled) (7%) being most common. We observed high intra-patient heterogeneity, the proportion of the dominant morphology ranging from 20% to 54% with a median of 36% (excluding single/apoptotic structures). Inter-patient heterogeneity was also associated with the activity of several drugs. While atorvastatin showed a similar single-oriented shift for both compact (82%) and invasive (100%) structures, other drugs such as 5-FU exhibited strong differences (68% and 18% respectively). Other compounds with limited impact on cell viability induced morphological phenotypes indicating mode-of-actions such as cytostasis or senescence. Interestingly, methotrexate caused an invasive-oriented shift for 80% of the cystic (thick-walled) structures but only 14% of the compact clusters. Conclusions: Quantitative phenotyping of tumoroids can identify novel predictive morphological relationships, drug mode-of-action and more accurately predict drug activity.