Background Tumor organoids (TO) have emerged as compelling cancer models due in part to their conserved properties of tumor heterogeneity and 3D structure of the tumor-immune microenvironment. A principal challenge in high-throughput screening experiments is tracking and labeling the somatic states of TOs during clonal and subclonal selection and expansion in the presence of immune effector cells. In this work, we train and evaluate multiple computer-vision based models for the identification of driver mutations at single-organoid resolution.

Methods DNA from TO lines was sequenced using the Tempus xT assay. Lines were cultured in the presence of natural killer effector cells and regularly imaged via brightfield confocal microscopy. Individual TOs were segmented from the original image and processed with three feature extractors: an unsupervised pre-trained convolutional neural network (CNN), a self-supervised neural network, and CellProfiler,an industry-leading microscopy analysis software suite. Secondary predictor models were trained on each feature set to predict the mutational status of several cancer driver genes using soft-labels derived from sample-level sequencing. For each secondary predictor, splitting observations into training and testing datasets was stratified by sample, mutation status, and cancer type. Classification performance was evaluated across each of the three feature sets and combinations thereof for each of the intended gene targets. The overall workflow is shown in figure 1.

Results We find that models using human-interpretable features generated by CellProfiler generally outperform both pre-trained CNNs and self-supervised computer vision models in predicting mutation status in TOs, and that combining feature sets generally confers an additive benefit for predicting gene mutation (figure 2). Of the eight genes evaluated in this study, the most performant model was observed in predicting KRAS (AUROC=0.82) and PTEN (AUROC=0.79) and least effective for TP53 (AUROC=0.57). Finally, ranked feature importance scoring suggests that morphological features most useful in predicting mutation status were tumor size, perimeter, and metrics of spatial intensity distribution.

Conclusions We present a single-organoid featurization and classification strategy for single-organoid gene mutation prediction from brightfield morphological features, enabling a high-throughput, low-cost approach for inferring an individual organoids’ somatic state. This approach can be used for rapid assessment of heterogeneous clonality in progenitor cell populations in the context of high-throughput screenings.