Deep Learning and Single Cell Phenotyping for Rapid Antimicrobial Susceptibility Testing

Abstract

The rise of antimicrobial resistance (AMR) is one of the greatest public health challenges, already causing up to 1.2 million deaths annually and rising. Current gold-standard antimicrobial susceptibility tests (ASTs) are low-throughput and can take up to 48 hours, with implications for patient care. We present advances towards a novel, rapid AST, based on the deep-learning of single-cell specific phenotypes directly associated with antimicrobial susceptibility in Escherichia coli. Our models can reliably (80% single-cell accuracy) classify untreated and treated susceptible cells, across a range of antibiotics and phenotypes - including phenotypes not visually distinct to a trained, human observer. Applying models trained on lab-reference susceptible strains to clinical isolates of E. coli treated with ciprofloxacin, we demonstrate our models reveal significant (p<0.001) differences between resistant and susceptible populations, around a fixed treatment level. Conversely, deploying on cells treated with a range of ciprofloxacin concentrations, we show single-cell phenotyping has the potential to provide equivalent information to a 24-hour growth AST assay, but in as little as 30 minutes.

Infection Inspection: Using the power of citizen science to help with image-based prediction of antibiotic resistance inEscherichia coli

“…In susceptible bacteria this can result in the compaction of the DNA and the inability to separate to dividing cells. Whilst our previously reported computer model could achieve a classification accuracy as high as 80% 16 there remains a degree of confusion with respect to certain images, especially near the minimum inhibitory concentration of the strain (Fig. 2b).…”

“…The accuracy for classifying resistant cells was similar, standing at 67.3% (Fig.2a). We also employed the same images to test a deep-learning model 16 . Compared to the volunteers, the model was less accurate in classifying resistant cells (62.5%; Fig.…”

“…The project dataset was made up of 49,074 individual images of ciprofloxacin-treated E. coli cells generated for previous work 16 . All bacteria had been chemically fixed and stained using 4′,6-diamidino-2-phenylindole (DAPI) as the nucleic acid stain and Nile Red as the membrane stain.…”

“…All bacteria had been chemically fixed and stained using 4′,6-diamidino-2-phenylindole (DAPI) as the nucleic acid stain and Nile Red as the membrane stain. The initial dataset was composed of 11,074 256x256 Red-Green-Blue (RGB) images of E. coli cells from 5 clinical strains (EC1, EC2, EC3, EC5, and EC6, reported previously 16 ) treated at 10 mg/L for 30 minutes, with clinical strains defined as ciprofloxacin-resistant (minimum inhibitory concentration [MIC] >0.5mg/L) or ciprofloxacin sensitive (MIC ≤0.25 mg/L) using European Union Council on Antimicrobial Susceptibility Testing (EUCAST) breakpoints 17 . A second dataset of 38,000 images included the same E. coli strains treated at 9 concentrations ranging from 16 mg/L to 0.001 mg/L ciprofloxacin for 30 minutes.…”

“…We are developing a diagnostic method that relies on a microfluidic device for the direct capture and identification of bacteria and associated antibiotic resistance from clinical samples using microscopy. We recently developed a deep-learning model which can classify individual E. coli cells as ciprofloxacin-sensitive or resistant with 80% accuracy (which results in high-confidence classifications of populations of bacteria) based on morphological changes to the sub-cellular structure 16 . The continued development of these single-cell, imaging-based classification methods requires robustness to bacterial heterogeneity, and an understanding of why certain cells within a sample are misclassified is essential.…”