a Schematic diagram of organoid cultures and screening workflow. Plates are generated by first depositing a buffer layer of 60% Matrigel to prevent cells from adhering to the bottom. Cell-containing 50% Matrigel is then added followed by media. Drugs are transferred using a 100 nl pin tool 2 days after cell seeding and serial bright-field imaging is performed in the following days. b Image analysis pipeline. 3D image stacks are first background corrected and projected into a single z-plane. High-level image features are then derived from the flattening layer of the RESNET-18 pre-trained on the ImageNet dataset. Principal component analysis is then used to explore response patterns (the percent of variance that is explained by each principal component is labeled to the axis) and k-nearest neighbor model is constructed for subsequent classification tasks. Assay quality is monitored using a 50/50 train/test split for each control category, which is shown in the confusion matrix. Negative (Neg), DMSO-treated wells; Positive (Pos), miR-200c induced wells; Media, cell-free wells. Act/Pred, Actual vs Predicted. c Principal component analysis of RESNET-18 features for experimental samples with control regions annotated (blue, negative region; dark green, positive region; red, media region). Increasing drug concentrations are shown from low (black) to high (yellow).