Deep learning for rapid, single-cell antimicrobial susceptibility testing

Thesis

The rise of antimicrobial resistance (AMR) is one of the most pressing global healthcare challenges, already causing an estimated 1.2 million preventable deaths annually and rising. Crucial to the management of AMR is rapid and specific diagnosis, allowing early and optimized intervention. Unfortunately, current gold-standard antimicrobial susceptibility tests are low-throughput and can take up to 48 hours to produce clinically relevant insights. In this thesis, we propose and evaluate a novel AST approach, based on the deep-learning of single-cell phenotypes directly associated with antimicrobial susceptibility. The phenotypes are revealed by widefield fluorescence microscopy and evaluated automatically by a deep-learning pipeline built on convolutional neural networks (CNNs). We demonstrate our Deep Antimicrobial Susceptibility Phenotyping (DASP) can robustly recognise susceptibility phenotypes associated with 4 representative antibiotics of major antibiotic families, in Escherichia coli, with over 80% single cell accuracy. We then deploy our models trained on susceptible lab strains, to clinical isolates of Escherichia coli treated with one of the antibiotics. Here, we demonstrate the distribution of single-cell phenotypic classification decisions is a reliable indicator of isolatesusceptibilityaroundafixedtreatmentpoint, revealingstatisticallysignificant (p<0.001) differences between untreated and treated cell populations in susceptible isolates, and no difference in resistant isolates. Further, we evaluate the limit of detection, and show this population-level output is indeed sensitive to the resistance status of single cells. Lastly, we investigate the relationship between treatment concentration, the minimum inhibitory concentration (MIC) of the isolate, and the DASP output, and compare this against the gold-standard growth assay. Here, we show that DASP has potential to produce equivalent information to the current gold-standard, but an order on magnitude faster. We conclude this thesis with an outlook on the developmental and mechanistic principles of the phenotypes by studying their time evolution.

Authors: Zagajewski, A

• DOI: 10.5287/ora-xkmo0jozv
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: antimicrobial susceptibility testing; deep learning; single cell
• Subjects: Drug resistance in microorganisms; Deep learning (Machine learning)