Eva Sörensson (1), Chenyan Zhao (1), Pikkei Wistrand-Yuen (2), Pernilla Lagerbäck (2), Thomas Tängden (2), Elisabet I Nielsen (1)
(1) Department of Pharmacy, Uppsala University, Uppsala, Sweden. (2) Department of Medical Sciences, Uppsala University, Uppsala, Sweden.
Introduction: Pseudomonas aeruginosa are Gram-negative, rod-shaped bacilli that can cause a variety of infections, ranging from relatively non-severe urinary tract infections to life-threatening lung and bloodstream infections [1-3]. Although previous studies have shown that P. aeruginosa exhibits morphological changes when treated with certain antibiotics, there is a lack of quantitative models characterizing the time and concentration dependency of such changes [4-8]. Time-lapse microscopy has emerged as an efficient method for evaluating antibacterial activities of antibiotics that enables real-time imaging of living cells under conditions similar to the conventional time-kill experiments [8]. This technique offers a great potential to obtain longitudinal, morphological insights into the bacteria-drug interactions.
Objectives: This study aimed to develop an image analysis method to process time-lapse microscopy images and to investigate the effect of various concentrations of meropenem and colistin on morphological changes in P. aeruginosa by developing a pharmacokinetic-pharmacodynamic (PKPD) model.
Methods: Data from previously performed time-lapse microscopy (oCelloscope) experiments were used, where P. aeruginosa ATCC 27853 were exposed to meropenem or colistin (0.5 – 16 xMICs) and images were taken every 15 min during a total period of 24 hours [8-9]. The first step of the study was to develop an image analysis methodology able to adequately identify and count bacteria in the time-lapse microscopy images, and to classify them into morphological groups. The images were evaluated using a CellProfiler pipeline [10-11] to count and obtain shape and corrected pixel intensity metrics for individual bacteria. Unsupervised learning was applied to identify the morphological clusters according to the bacterial shape and intensity. These clusters were then used in a RandomForest machine learning model to classify individual bacteria [12], which was iteratively improved using feature engineering, recursive feature elimination and hyperparameter tuning. A manual count of bacteria was performed on a selection of the oCelloscope images to act as a comparison. As the second step, a PKPD model was developed, to simultaneously describe the rates of change between morphological compartments and the change of bacterial counts.
Results: The morphological states observed when P. aeruginosa was treated with meropenem were healthy rods, filamentous, bulging, and faded/dying bacteria. Colistin-exposed bacteria appeared as healthy, bulging, aggregates, and faded/dying. Images with bacterial counts greater than ~4000 were too crowded to be adequately evaluated, which prevented low-exposure experiments from being counted past ~4.5 hours. The machine learning morphology classifiers obtained 77% (colistin) and 83% (meropenem) accuracy when compared to manual classification. The compiled bacterial morphology data set indicated that at low concentrations (<1xMIC), bacteria predominantly remained as healthy rods. For meropenem, at middle concentrations (1-2xMIC), healthy cells became filaments, but at higher concentrations (>=4xMIC) healthy rods mainly form bulges, with filaments being a potential transient state between the two. The higher concentrations the fewer the filaments. For colistin, some bacteria at 1-2xMIC formed aggregates, but at 1-4xMIC most cells become bulged or faded/dying. PKPD models were developed for the two drugs separately, to account for the different morphological states. In the developed models, the healthy bacteria were allowed to transfer to filaments, aggregates, and bulges. Filaments could also transfer to bulges. The rate constants of these changes were antibiotic concentration dependent. Bulges could not change into other morphologies other than faded/dying cells, and this change was concentration independent. Antibiotic adsorption and degradation were accounted for, and a pre-existing subpopulation model was used to describe observed bacterial regrowth.
Conclusions: To the best of our knowledge, this is the first study combining time-lapse microscopy, image analysis and PKPD modeling to quantify changes in antibiotic-induced bacterial count and morphology over time. This quantitative morphological information may increase the knowledge of the antibiotic-microbe interaction and may in the future contribute to a more in-depth understanding of the dynamics of antibacterial effects of antibiotics.
References:
[1] Foxman B, Brown P. Epidemiology of urinary tract infections: transmission and risk factors, incidence, and costs. Infect Dis Clin North Am. 2003 Jun;17(2):227–41.
[2] Høiby N, Ciofu O, Bjarnsholt T. Pseudomonas aeruginosa biofilms in cystic fibrosis. Future Microbiol. 2010 Nov;5(11):1663–74.
[3] Micek, S. T., Lloyd, A. E., Ritchie, D. J., Reichley, R. M., Fraser, V. J., & Kollef, M. H. . Pseudomonas aeruginosa bloodstream infection: importance of appropriate initial antimicrobial treatment. Antimicrobial agents and chemotherapy. 2005 ;49(4): 1306–1311. https://doi.org/10.1128/AAC.49.4.1306-1311.2005
[4] Sumita Y, Fukasawa M, Okuda T. Comparison of two carbapenems, SM-7338 and imipenem: affinities for penicillin-binding proteins and morphological changes. J Antibiot (Tokyo). 1990 Mar;43(3):314–20.
[5] Horii T, Kobayashi M, Sato K, Ichiyama S, Ohta M. An in-vitro study of carbapenem-induced morphological changes and endotoxin release in clinical isolates of gram-negative bacilli. J Antimicrob Chemother. 1998 Apr;41(4):435-42. doi: 10.1093/jac/41.4.435. PMID: 9598774.
[6] Choi J, Yoo J, Lee M, Kim E-G, Lee JS, Lee S, et al. A rapid antimicrobial susceptibility test based on single-cell morphological analysis. Sci Transl Med. 2014 Dec 17;6(267):267ra174.
[7] O’Driscoll NH, Cushnie TPT, Matthews KH, Lamb AJ. Colistin causes profound morphological alteration but minimal cytoplasmic membrane perforation in populations of Escherichia coli and Pseudomonas aeruginosa. Arch Microbiol. 2018 Jul;200(5):793–802.
[8] Ungphakorn W, Malmberg C, Lagerbäck P, Cars O, Nielsen EI, Tängdén T. Evaluation of automated time-lapse microscopy for assessment of in vitro activity of antibiotics. J Microbiol Methods. 2017 Jan;132:69–75.
[9] Fredborg M, Andersen KR, Jørgensen E, Droce A, Olesen T, Jensen BB, Rosenvinge FS, Sondergaard TE. Real-time optical antimicrobial susceptibility testing. J Clin Microbiol. 2013 Jul;51(7):2047-53. doi: 10.1128/JCM.00440-13. Epub 2013 Apr 17. PMID: 23596243; PMCID: PMC3697729.
[10] Stirling DR, Swain-Bowden MJ, Lucas AM, Carpenter AE, Cimini BA, Goodman A. CellProfiler 4: improvements in speed, utility and usability. BMC Bioinformatics. 2021 Sep 10;22(1):433. doi: 10.1186/s12859-021-04344-9. PMID: 34507520; PMCID: PMC8431850.
[11] Stirling DR, Carpenter AE,Cimini BA, CellProfiler Analyst 3.0: accessible data exploration and machine learning for image analysis, Bioinformatics, 2021, Nov; 37(21) 3992–3994,
[12] Pedregosa F, Varoquaux, Ga”el, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine learning in Python. Journal of machine learning research. 2011;12(Oct):2825–30.
Reference: PAGE 30 (2022) Abstr 10206 [www.page-meeting.org/?abstract=10206]
Poster: Drug/Disease Modelling - Infection