Eva Wehrhahn 1, Pernilla Lagerbäck 2, Thomas Tängdén 2, Elisabet I Nielsen 1
1 Department of Pharmacy, Uppsala University (Uppsala, Sweden), 2 Department of Medical Sciences, Uppsala University (Uppsala, Sweden)
Introduction/Objectives: Carbapenem-resistant Gram-negative bacteria are becoming increasingly prevalent worldwide [1]. Improved understanding of the mechanisms of bacterial killing could potentially improve the use of these last-resort antibiotics. Carbapenems can cause bacterial morphology changes, such as bulges or elongated filaments [2]. Pharmacokinetic-pharmacodynamic (PKPD) models based on time-lapse microscopy images have previously been developed to find associations between antibiotic concentration and changes in bacterial counts and morphology [3]. Literature indicates that even shape changes within bacterial morphology classes may affect their division and survival. It has been shown previously that longer cells, i.e. filamented bacteria, can divide more rapidly [4], and that bulge stability is related to turgor pressure expressed through various shape metrics, including radius [5]. This study aimed to explore the incorporation of time-varying shape covariates into existing PKPD models to improve mechanistic understanding and model fit.
Methods: The development of the existing PKPD model has been previously described [3]. Briefly, an image analysis pipeline was used to generate bacterial morphology class counts from time-lapse microscopy experiments [6] with Escherichia coli exposed to ertapenem and Pseudomonas aeruginosa exposed to meropenem at 0-16x minimum inhibitory concentration for up to 24 hours. Morphology class change, killing, and growth in relation to drug concentration were then described using PKPD models built in NONMEM [7]. Bacterial morphology classes included healthy rod, filament, bulged filament, and bulged bacteria. In P. aeruginosa, a mixture model described variability in the presence of resistant filaments. In the current project, bacterial size and shape measurements that were expected to inform on cell division and death based on existing literature were extracted from the image analysis pipeline. Median image filament, bulged filament, and bulge class area and maximum radius were utilized as proxies for cell length and cell/bulge radius, respectively. Correlations between these time-varying covariates were evaluated, and covariates with an R2 of at least 0.5 were not included in the same model. Covariate relationships were tested with linear, exponential, power, and log-linear functions [8] in NONMEM. Covariates were added individually to parameters in PKPD models, and then sequentially based on reductions in objective function value (OFV). A minimum decrease of 6.63 (p=0.01) was required for covariate inclusion. After forward addition of covariates, covariates with low OFV reduction, high parameter uncertainty, minimal impact on parameter value for 90th and 10th percentiles of the covariates, and little or negative impact on goodness-of-fit plots were removed. Final models were evaluated using Visual Predictive Checks built using Xpose4 and compared using Akaike information criterion (AIC) [9].
Results: Forward addition of covariates in the E. coli + ertapenem model resulted in the addition of five covariate relationships , however two covariates were removed due to high parameter uncertainty and minimal impact on the parameters or model fit. Two covariates were added in the P. aeruginosa + meropenem model. Addition of covariates to both models resulted in reduced AIC (-461 dAIC E. coli, -122 dAIC P. aeruginosa) and reduced bulge residual error (-0.031 E. coli and -0.034 P. aeruginosa additive error on log10 scale). In E. coli + ertapenem, the remaining three covariate relationships suggest that increased bulged filament area reduces bulged filament dormancy rate (0.033-0.001 h-¹ at 10/90th covariate percentile), and increased bulge area increases bulge death (0.144-0.639 h-1 at 10/90th covariate percentile) and dormancy rates (0.004-0.031 h-¹ at 10/90th covariate percentile), resulting in improved model fit for bulge and bulged filament observations. The covariates in the P. aeruginosa + meropenem model described an increase in bulged filament death rate as bulged filament area increases (0.43-1.30 h-¹ at 10/90th covariate percentile), and an increase in filament growth rate as filament area increases (0.287-1.742 h-¹ at 10/90th covariate percentile). After adding covariates, removing the mixture model was tested, which found that covariates (predominantly filament area on filament growth rate) were able to describe the regrowth trends described using the mixture model, albeit with a significant increase in AIC (+67) compared to when the mixture model was included.
Conclusions: The inclusion of time-varying covariates improved model fit and reduced residual error. The included covariates have mechanistic implications, which may help explain antibiotic effects and build better models that may be of use for improving antibiotic treatment in the future.
References:
[1] Global antibiotic resistance surveillance report 2025: summary. https://www.who.int/publications/i/item/B09585
[2] Cushnie TPT, O’Driscoll NH, Lamb AJ. Cell Mol Life Sci CMLS. 2016 Dec;73(23):4471–92.
[3] Wehrhahn E et al. (2025). Abstr 11745 [www.page-meeting.org/?abstract=11745]
[4] Wehrens M et al. Curr Biol CB. 2018 Mar 19;28(6):972-979.e5.
[5] Wong F et al. Front Microbiol. 2021;12:712007.
[6] oCelloScope – Automated Microscopy – Live Cell Imaging – BioSense Solution. BioSense Solutions. https://biosensesolutions.dk/overview/
[7] Bauer RJ. CPT Pharmacomet Syst Pharmacol. 2019 May; 8(8): 525-537.
[8] Lindbom L et al. Comput Methods Programs Biomed. 2004 Aug;75(2):85–94.
[9] Keizer RJ et al. CPT Pharmacomet Syst Pharmacol. 2013 Jun 26;2(6):e50.
Reference: PAGE 34 (2026) Abstr 11868 [www.page-meeting.org/?abstract=11868]
Poster: Drug/Disease Modelling - Infection