Fig. 3: Ribosome phenotype recognition is robust across biological replicates.

a Four biological replicates of E. coli MG1655 were tested for each antibiotic and for the untreated condition. To test phenotype robustness and repeatability, a holdout cross-validation was performed in which each model was trained and validated on images from three of the biological replicates and tested on images from the fourth. The training images received random data augmentations before being passed to the model, whereas the holdout dataset was passed directly to the model for testing. b The balanced accuracy of the ribosome phenotype classifier is shown for each antibiotic. Each point represents a biological replicate. The mean balanced accuracy is shown on each column and the error bars indicate the 95% confidence interval of the mean on the four biological replicates. c Confusion matrices for the ciprofloxacin (Cip) ribosome phenotype classifier. The total number of cells is a sum of the results from four experiments, each with a model trained on three biological replicates and tested on a fourth holdout replicate. The number of images in each class is shown, along with the percentage of cells for each treatment condition. The treatment condition is shown on the columns and the model’s predicted classification is shown on the rows. Right column: positive predictive value (PPV) and negative predictive value (NPV) of the model’s predictions are shown. Bottom row: accuracy of the model on antibiotic-treated cells (Sensitivity), accuracy of the model on untreated cells (Specificity), and the Balanced Accuracy (Accuracy) are shown. See Accuracy Metrics for details. d As in (c), for the gentamicin (Gent) ribosome phenotype classifier. e As in (c), for the chloramphenicol (Cam) ribosome phenotype classifier. f As in (c), for the carbenicillin (Carb) ribosome phenotype classifier.