Fig. 5: Model performs with >80% accuracy when trained on cells treated with 1 × EUCAST ciprofloxacin and tested on cells treated with 20 × EUCAST.

a Representative, correctly classified images of the ribosome phenotypes of strains R4 and S2 treated at 1 × EUCAST (0.5 mg/L) for 30 min. Scale bar, 2 μm. b The susceptible-resistant classifier trained on E. coli treated with 20 × EUCAST ciprofloxacin (black circles) is compared to the classifier trained on E. coli treated at 1 × EUCAST (blue triangles). The 20 × EUCAST dataset is composed of 6 clinical isolates (S1, S2, S3, R1, R2, R3) whereas the 1 × EUCAST dataset is composed of 2 clinical isolates (S2, R4). Each data point represents a biological replicate. The 20 × EUCAST model was tested on 28,448 holdout test images from 2 biological replicates of the 6 clinical E. coli isolates treated at 20 × EUCAST ciprofloxacin for 30 min. The 1 × EUCAST model was tested on the same 20 × EUCAST dataset. For every isolate, the 20 × EUCAST model is more likely to call resistant cells resistant and less likely to call susceptible cells resistant. However, the 1 × EUCAST model maintains an accuracy of 73.8 ± 5.3% on cells from susceptible strains and an accuracy of 89.6 ± 3.8% on cells from resistant strains, despite being trained on images of cells treated at a different concentration and classifying a previously unseen strain (R4).