Logo: to the web site of Uppsala University

Abstract

Introduction

Pseudomonas aeruginosa are Gram-negative, rod-shaped bacilli that cause a variety of infections, including serious lung, bloodstream, and skin infections as well as minor urinary tract infections (1–4).  Studies have shown that antibiotics can cause morphological changes in P. aeruginosa, including filamentation, aggregation, and sphere/bulge formation (5–8). These changes have not been quantitatively described in a pharmacokinetic-pharmacodynamic (PKPD) model. PKPD models describe the interactions of pharmacokinetics, the amount of drug present over time, and the pharmacodynamics, or drug effect (9). This study uses time-lapse microscopy bacterial images, which captures live bacteria morphological changes bacteria over time (10) and aimed to evaluate the effect of the concentration of two antibiotics, meropenem and colistin, on P. aeruginosa bacterial shape in a PKPD model. 

Methods

oCelloScope time-lapse microscopy images of meropenem and colistin treated P. aeruginosa ATCC 27853 taken every 15 min for 24 hours were evaluated (10). Doses were relative to the minimum inhibitory concentration (MIC), ranged from 0.5-16xMIC (10). Images were evaluated using a CellProfiler pipeline to obtain bacterial shape and intensity metrics. Unsupervised learning was applied to establish morphological clusters suggested by the metrics output from the pipeline. A RandomForest machine learning model was then applied to classify bacteria into classes (11). This classification was used to generate morphology count data per image. The count data was used to create preliminary PKPD models for P. aeruginosa and colistin (PA+COL) and P. aeruginosa and meropenem (PA+MERO), respectively. The PKPD models described the bacteria growth and shape changes over time at different concentrations.      

Results 

In both antibiotic groups, healthy rod-shaped bacteria predominated at low concentrations (<1xMIC). Larger variability in morphology was witnessed at 1xMIC. For meropenem, healthy bacilli mainly formed filaments at 0.5-2xMIC, or bulges at >=4xMIC. Most filaments became bulges >=4xMIC. For colistin, healthy cells formed bulges (1-4xMIC), and aggregates showed little trend but were slightly more common at 0.5xMIC. Faded bacteria represented most likely dead bacteria. The machine learning classifier moderately well classified (77-78% test accuracy) bacterial morphology. Automated and manual count agreement was lower (68-85%) but comparable to manual intra- (84-93%) and inter-observer (87-92%) agreement. Post-processing of bacterial counts was used to reduce misclassification error. Changes in bacterial morphology class counts over time were described by a PKPD model. The preliminary model depicted a concentration-dependent drug effect delay. Regrowth was modelled with a pre-existing resistant subpopulation. Healthy bacteria transferred to bulge or filament compartments in a concentration-dependent sigmoid Emax models. The transfer of filaments to bulges was captured with a concentration-dependent Emax model. Bulges were only permitted to die, and loss was concentration independent. Preexisting models for drug adsorption and degradation in the labware during the experiments were included. 

Conclusion

In this study, we developed a pipeline to process time-lapse microscopic bacterial images and to quantify bacterial different morphological class counts. Preliminary PKPD models quantified the time-course of these antibiotic-induced morphology changes. Further work is planned to improve the pipeline and the PKPD model. This quantitative morphological information may help inform on the antibiotic-microbe interaction and eventually add to the understanding of the antibacterial effects of antibiotics.