Structural transition in antimicrobial resistance dynamics of Corynebacterium striatum

Abstract

Background

Antimicrobial resistance (AMR) is typically interpreted as a process of gradual accumulation of resistance traits. However, resistance systems may instead undergo structural transitions under changing selective pressures. We investigated whether long-term AMR dynamics in Corynebacterium striatum reflect accumulation or reorganization of a constrained phenotypic state space.

Methods

Longitudinal resistance data (2013–2025) were analyzed using segmented regression, multinomial modeling, entropy-based diversity metrics, and Markov transition analysis to characterize temporal dynamics, architecture distributions, and system-level behavior.

Results

A significant structural breakpoint was identified around 2020 (Davies’ test p < 0.001). The estimated breakpoint location showed limited uncertainty based on model-based confidence interval estimation. Prior to 2020, aminoglycoside-containing resistance architectures were prevalent, whereas after 2020 they collapsed and were replaced by aminoglycoside-negative, fluoroquinolone-associated configurations. Phenotypic diversity declined significantly over time, indicating contraction of the resistance state space. Markov analysis demonstrated convergence toward a stable stationary distribution dominated by two related architectures, with increased dynamical stability after the breakpoint.

Conclusions

AMR dynamics in C. striatum reflect a structural transition and restructuring within the simplified state space rather than progressive accumulation of resistance traits. The system converges toward a low-diversity phenotypic attractor, highlighting the importance of dynamical systems approaches in understanding resistance evolution.

Graphical abstract

Highlights

• •Structural transition in AMR dynamics identified around 2020.
• •Resistance evolution is consistent with restructuring within a simplified state space rather than simple accumulation.
• •Aminoglycoside-associated architectures collapse after 2020.
• •Phenotypic diversity declines, indicating contraction of state space.
• •System converges to a stable low-diversity resistance attractor.

1. Introduction

Antimicrobial resistance (AMR) is commonly interpreted as a cumulative process driven by progressive acquisition of resistance determinants (Holmes et al., 2016). Within this framework, temporal increases in resistance are viewed as gradual expansion of multidrug-resistant phenotypes (Davies & Davies, 2010). However, this paradigm assumes that resistance evolution follows largely monotonic and additive trajectories, which may not adequately capture the dynamics of complex clinical systems.

Alternative perspectives from evolutionary epidemiology and dynamical systems theory suggest that resistance may evolve through structural transitions within a constrained phenotypic state space (Blanquart, 2019; Cen et al., 2017). Under changing selective pressures, bacterial populations may reorganize toward a limited number of high-fitness configurations, resulting in abrupt shifts in resistance architecture rather than gradual accumulation (Blanquart, 2019). Such behavior can be interpreted as transitions between attractor-like configurations in a structured resistance landscape (Cen et al., 2017).

Despite these theoretical advances, most clinical AMR studies remain descriptive and focus on prevalence trends or individual resistance determinants, with limited attention to the structure and dynamics of phenotypic state spaces (van Duin & Paterson, 2016).

Corynebacterium striatum has emerged as an important multidrug-resistant opportunistic pathogen in healthcare settings (Lee et al., 2005; Orosz et al., 2022), yet its long-term resistance dynamics remain poorly characterized at the system level. This species has emerged as an opportunistic pathogen of increasing clinical relevance, particularly in immunocompromised patients and those with prolonged hospital exposure. It is commonly associated with device-related infections, respiratory tract infections, and bloodstream infections. A major concern is its capacity to develop multidrug resistance, which significantly limits therapeutic options (Orosz et al., 2022, 2023).

Resistance in C. striatum is mediated by both intrinsic and acquired mechanisms. Fluoroquinolone resistance is primarily associated with mutations in the quinolone resistance-determining regions of gyrA and parC genes. Rifampicin resistance typically arises from mutations in the rpoB gene. Aminoglycoside resistance is frequently mediated by aminoglycoside-modifying enzymes, while macrolide resistance is often linked to erm genes encoding ribosomal methylases. Tetracycline resistance may involve efflux pumps and ribosomal protection proteins (McMullen et al., 2017).

The increasing prevalence of multidrug-resistant C. striatum strains in healthcare settings highlights its clinical importance and underscores the need for improved understanding of its resistance dynamics and epidemiological behavior. In this study, we test the hypothesis that AMR dynamics in C. striatum reflect structural reorganization of a constrained phenotypic state space rather than progressive accumulation of resistance traits. The terminology derived from dynamical systems theory is used as an interpretative framework for the observed statistical patterns, rather than as a formal demonstration of underlying dynamical mechanisms.

2. Materials and methods

2.1. Data collection

Clinical microbiological data were collected from a single tertiary-care center between 2013 and 2025. For incidence trend visualization (Fig. 1), extended surveillance data from 2002 were included. Isolates identified as C. striatum were included. Duplicate isolates were filtered using a first-isolate approach. To minimize sampling bias, only the first isolate per patient per year was included in the primary analyses. Sensitivity analyses using all isolates yielded comparable results.

Fig. 1.

Epidemiological trends of Corynebacterium striatum. A) Annual number of clinical C. striatum isolates (grey bars) and proportion of multidrug-resistant (MDR) isolates (black line). (B) Proportion of C. striatum among all clinical bacterial isolates over time.

Antimicrobial susceptibility testing was performed according to EUCAST guidelines in effect during the study period. Susceptibility results were interpreted as binary resistant/susceptible categories for downstream modeling.

2.2. Definition of resistance architectures

Resistance profiles were grouped into phenotypic architectures based on three antibiotic classes: fluoroquinolones (FQ), rifampicin (RIF), and aminoglycosides (AG). Binary resistance states were combined into composite architectures (e.g., FQ⁺RIF⁻AG^-).

This representation defines a finite discrete phenotypic state space of size 2³, consisting of all possible combinations of class-level resistance. Rare architectures (n < 10) were collapsed into an “Other” category in analyses sensitive to sparsity. This reduced representation allows modeling of resistance evolution as transitions within a finite and interpretable phenotypic state space.

2.3. Segmented regression

Temporal trends in multidrug resistance (MDR) were analyzed using segmented logistic regression, with calendar year as a continuous predictor. MDR was defined as resistance to at least two of the three antibiotic classes considered.

Structural breakpoints were evaluated using Davies’ test. The location of the breakpoint was estimated using segmented regression, and uncertainty was assessed using model-derived confidence intervals. Breakpoint estimation was confirmed by model fit comparison. Model parameters were estimated separately for pre- and post-breakpoint segments, and interaction terms were used to assess differences in temporal slopes.

2.4. Multinomial modeling

Temporal changes in resistance architecture distribution were assessed using multinomial logistic regression, with resistance architecture as the outcome and calendar year as a continuous predictor. Odds ratios were calculated to quantify the annual change in probability of each architecture relative to a reference category. Model predictions were used to generate time-resolved estimates of architecture probabilities.

2.5. Diversity analysis

Phenotypic diversity was quantified using Shannon entropy:

$$H=-{\sum }_i{p}_i\log p_i$$

where $p_i$ denotes the relative frequency of each resistance architecture.

Temporal trends in diversity were evaluated using linear regression to capture the overall direction of change over time. This approach focuses on long-term trends and does not explicitly model short-term fluctuations or potential nonlinear dynamics.

2.6. Markov modeling

Patient-level resistance transitions were modeled using empirical first-order Markov transition matrices constructed from temporally ordered isolates within individual patients. Transition probabilities were estimated as row-normalized frequencies of observed state transitions. To assess robustness, transition patterns were evaluated at the level of aggregated transition frequencies, and dominant transitions were consistently observed across the dataset. Stationary distributions were derived from empirical transition matrices and interpreted in the context of dominant, high-frequency resistance architectures.

The Markov assumption implies that the probability of transitioning to a given resistance state depends only on the current state. Stationary distributions were obtained as the dominant eigenvector of the transition matrix. The spectral gap was defined as 1 − λ₂, where λ₂ is the second largest eigenvalue (in magnitude). It determines the rate at which the Markov chain converges to its stationary distribution, with smaller values corresponding to slower mixing. No assumption of reversibility was imposed on the transition matrix; therefore, the spectral gap is interpreted as a general indicator of mixing dynamics rather than a strict measure of convergence rate under detailed balance conditions. The stationary distribution was estimated from pooled transition data and is interpreted as a time-averaged representation of the system, without assuming strict stationarity of the underlying process. A total of 504 sequential isolate pairs were used for Markov transition analysis.

2.7. Pre/post analysis

Separate models were constructed for pre-2020 and post-2020 periods to evaluate structural changes in system behavior.

Comparisons between periods included MDR temporal trends, resistance architecture distributions, phenotypic diversity, and Markov transition dynamics. This stratified analysis was used to assess whether the identified breakpoint corresponded to a qualitative change in system dynamics.

2.8. Statistical environment

All analyses were performed in R (version 4.5.2). Segmented regression was implemented using the segmented package, multinomial models using nnet, and Markov analyses using custom scripts. Statistical significance was assessed at a two-sided α level of 0.05. For visualization purposes, smoothing techniques were applied to highlight dominant temporal patterns. Detailed annual dataset characteristics, including isolate counts, patient counts, antimicrobial testing frequencies, resistance counts and proportions, and the number of transitions used for Markov analysis, are provided in Supplementary Table S1. The code used for the main analyses is provided in the Supplementary Material (Supplementary_Code.R).

3. Results

3.1. Incidence trends

The final dataset comprised isolates collected between 2002 and 2025, with annual sample sizes ranging from 1 to 312 isolates per year. The number of unique patients per year ranged from 1 to 172. Key descriptive statistics of the dataset are summarized in Supplementary Table S1. The number of Corynebacterium striatum isolates increased markedly over the study period, with a pronounced rise after 2019 (Fig. 1). By 2025, the organism represented a substantial proportion of clinical isolates, indicating increasing epidemiological relevance.

3.2. Resistance class dynamics

Resistance patterns differed across antibiotic classes (Fig. 2). Fluoroquinolone resistance remained consistently high throughout the study period, whereas rifampicin resistance showed temporal variability. In contrast, aminoglycoside resistance declined sharply after 2020 and became rare thereafter. These class-specific trajectories indicate that multidrug resistance dynamics were primarily driven by the presence or absence of aminoglycoside resistance.

Fig. 2.

Class-specific resistance dynamics of Corynebacterium striatum. (A) Temporal trends in fluoroquinolone (FQ) and rifampicin (RIF) resistance. (B) Temporal trends in aminoglycoside (AG) and macrolide (MAC) resistance. (C) Temporal trend in tetracycline (TET) resistance. Smoothed curves are shown to illustrate overall temporal patterns.

3.3. Structural shift in multidrug resistance dynamics

Temporal analysis of multidrug resistance (MDR) revealed a significant structural breakpoint in 2020 (Davies’ test p < 0.001). Prior to 2020, MDR prevalence showed no significant trend, whereas after 2020 a marked decline was observed (OR per year 0.63, 95% CI 0.46–0.87).

This change in slope indicates a transition in system dynamics rather than gradual temporal variation. Importantly, the redistribution of resistance architectures begins prior to the identified breakpoint (Fig. 5), suggesting that the 2020 breakpoint reflects a change in the rate of an ongoing restructuring process rather than its onset. The decline in MDR prevalence was consistent across specimen categories, suggesting a global shift rather than compartment-specific effects (Fig. 3, Fig. 4).

Fig. 5.

Temporal distribution of resistance states. Stacked area plot showing annual proportions of resistance states defined by combinations of fluoroquinolone (FQ), rifampicin (RIF), and aminoglycoside (AG) resistance. Proportions are calculated from observed isolate counts. The dashed vertical line indicates the year 2020.

Fig. 3.

Temporal dynamics of multidrug resistance in Corynebacterium striatum. (A) Annual multidrug resistance (MDR) rates with 95% confidence intervals (grey shading). The red line represents the fitted segmented regression model, and the dashed vertical line indicates the estimated breakpoint (2020). (B) Estimated odds ratios (per year, log scale) for MDR prevalence before and after 2020. Error bars represent 95% confidence intervals; the dashed horizontal line indicates no change (OR = 1). (C) Comparison of MDR prevalence calculated using all isolates versus a first-isolate-per-patient-per-year approach.

Fig. 4.

Model-based predicted multidrug resistance probability across specimen categories. Model-derived predicted probabilities of multidrug resistance (MDR) over time stratified by specimen category. Smoothed curves are shown to highlight overall temporal trends.

3.4. Resistance architecture restructuring

Analysis of resistance architectures demonstrated a pronounced redistribution of phenotypic states. Aminoglycoside-containing architectures show a gradual decline beginning around 2014–2015, while aminoglycoside-negative configurations progressively increase over time. After 2020, this redistribution becomes more pronounced, with rapid decline of aminoglycoside-positive states and stabilization of a reduced set of dominant architectures.

Empirical distributions and multinomial model estimates were consistent, showing a directional shift toward a restricted subset of resistance states (Fig. 5). These findings indicate restructuring of the phenotypic state space rather than progressive accumulation of resistance traits.

3.5. Decline in phenotypic diversity

Shannon entropy analysis revealed an overall decreasing trend over time (β = −0.090 per year, p < 0.001). While short-term fluctuations are observed, particularly around 2014 and near 2020 (Fig. 6), the linear model captures the dominant long-term trend of increasing concentration in the phenotypic distribution. The decline becomes more pronounced after the 2020 breakpoint, consistent with the emergence of a more concentrated phenotypic configuration dominated by a limited number of resistance architectures.

Fig. 6.

Temporal trend of phenotypic diversity of Corynebacterium striatum resistance architectures. Shannon diversity index calculated annually for resistance architectures. Points represent observed values, the solid blue line indicates the fitted linear regression, and grey shading denotes the 95% confidence interval. The dashed vertical line marks the year 2020.

3.6. Department-level distribution

Department-level analysis showed that multidrug resistance was distributed across multiple clinical units without consistent clustering (Fig. 7). This pattern supports institution-wide dissemination rather than localized outbreaks.

Fig. 7.

Department-level distribution of multidrug resistance in C. striatum. Heatmap of multidrug resistance (MDR) rates across clinical departments over time. Each cell represents the annual MDR proportion within a given department, with darker shading indicating higher resistance rates.

3.7. Markov dynamics and system convergence

Markov transition analysis demonstrated strong persistence of resistance architectures, with a pronounced diagonal structure in the transition matrix (Fig. 8). The observed transition structure is dominated by high-probability self-transitions and a limited number of recurrent transitions between common resistance architectures, suggesting that the estimated transition probabilities are driven by stable and well-supported patterns rather than sparse or noise-driven events. The system converged toward a stationary distribution dominated by two aminoglycoside-negative architectures. These distributions were consistent with the empirical frequency of dominant resistance states, supporting the robustness of the estimated stationary configuration. Comparison of pre- and post-2020 models revealed a shift toward increased dynamical stability. The spectral gap decreased from 0.284 to 0.171, indicating slower mixing dynamics of the system. In this context, slower mixing is consistent with increased persistence of dominant resistance configurations and reduced transition between alternative states. (Fig. 9).

Fig. 8.

Empirical Markov transition matrix of simplified resistance states. Heatmap representation of row-normalized transition probabilities between resistance states derived from sequential isolates of individual patients. Columns represent the origin state (“From”), and rows represent the subsequent state (“To”). Cell shading indicates transition probability (0–1).

Fig. 9.

Stationary distributions of resistance states before and after 2020. Comparison of stationary probability distributions derived from Markov models fitted to pre-2020 and post-2020 periods. Bars represent stationary probabilities for each resistance state.

4. Discussion

The present study demonstrates that long-term antimicrobial resistance dynamics in Corynebacterium striatum are not adequately explained by progressive accumulation of resistance traits but instead reflect a directional restructuring of a finite phenotypic state space. This distinction challenges the conventional linear paradigm of AMR evolution and supports a systems-level interpretation of resistance dynamics (Davies & Davies, 2010; Holmes et al., 2016).

The identification of a statistically robust structural breakpoint in 2020 is consistent with a transition in system behavior rather than gradual temporal drift. Such non-linear shifts are increasingly recognized in complex infectious disease systems, where ecological and evolutionary constraints can produce changes in population structure (Lipsitch et al., 2000). The observed transition was accompanied by a redistribution of resistance architectures, characterized by a decline of aminoglycoside-containing states and the emergence of two dominant aminoglycoside-negative configurations. The identified breakpoint coincides with the onset of the COVID-19 pandemic, which introduced substantial changes in hospital dynamics, including altered patient populations, increased ICU utilization, and shifts in antimicrobial usage patterns. Therefore, the observed transition cannot be interpreted as purely endogenous system restructuring. Instead, it likely reflects the combined effects of intrinsic resistance dynamics and external ecological perturbations.

Importantly, the restructuring of resistance architectures appears to begin several years before the identified breakpoint, as indicated by the gradual decline of aminoglycoside-containing configurations from approximately 2014–2015. The 2020 breakpoint therefore represents an inflection point in system dynamics, marking the acceleration and consolidation of an ongoing process rather than its initiation.

This restructuring was consistently observed across multiple analytical layers. Multinomial modeling demonstrated directional shifts in architecture probabilities, entropy analysis revealed increasing concentration of the phenotypic distribution, and Markov modeling indicated a tendency toward a stationary distribution dominated by a limited number of resistance states. Importantly, the observed decline in Shannon entropy reflects increased concentration of the observed distribution among dominant phenotypes and does not directly imply contraction of the underlying phenotypic state space. The reduced state space based on three antibiotic classes was introduced as a simplified representation to enable tractable analysis and visualization. Accordingly, the observed patterns should be interpreted within this reduced framework rather than as a complete characterization of the full resistance phenotype. Importantly, the stationary distribution in this context does not imply a time-invariant underlying process, but rather represents an empirical summary of transition patterns over the observed period. Given the presence of temporal heterogeneity and structural changes, it should be interpreted as an effective, time-averaged characterization of system behavior rather than a true equilibrium state. The reduction in spectral gap reflects slower mixing dynamics, consistent with increased persistence of dominant configurations (Levin et al., 1997). Thus, the observed concentration of states reflects redistribution within the defined simplified state space and does not necessarily imply contraction of the full phenotypic resistance space. Entropy-based measures should therefore be interpreted as describing distributional properties rather than the dimensionality of the full resistance phenotype space. Together, these findings are consistent with the interpretation that the observed dynamics may reflect a marked shift in the distribution of resistance states within the simplified representation. Rather than incremental acquisition of resistance determinants, the system appears to have reorganized toward a new equilibrium characterized by reduced diversity and dominance of specific phenotypic configurations. Comparable regime shifts have been described in microbial populations under strong selection pressure, where adaptive landscapes constrain accessible evolutionary trajectories (Mustonen & Lässig, 2009).

A plausible mechanistic explanation for this restructuring is clonal replacement driven by selection pressure. The disappearance of aminoglycoside resistance suggests either changes in antimicrobial usage or fitness costs associated with maintaining resistance determinants. Experimental and epidemiological studies have shown that resistance-associated fitness costs can shape population structure and drive replacement dynamics when selective pressures shift (Andersson & Hughes, 2010; Melnyk et al., 2015). Under such conditions, lineages lacking aminoglycoside resistance but retaining fluoroquinolone resistance may gain a selective advantage, leading to competitive exclusion of previously prevalent multidrug-resistant phenotypes.

This interpretation aligns with evolutionary epidemiology models, which predict that changes in transmission dynamics and antimicrobial exposure can induce bifurcations and transitions between alternative stable states (Blanquart, 2019; Cen et al., 2017). Within this framework, the observed increased concentration within the simplified state space reflects movement toward a low-entropy attractor dominated by a limited number of high-fitness resistance configurations.

The absence of strong sender-level clustering supports the hypothesis of institution-wide dissemination of dominant lineages rather than localized outbreaks. This pattern is consistent with previous observations that hospital-adapted opportunistic pathogens can spread diffusely across departments through patient movement and shared healthcare environments (D'Agata et al., 2009). Similarly, the consistency of trends across specimen categories indicates that the restructuring process reflects a system-wide ecological shift rather than niche-specific dynamics.

From a clinical perspective, the observed decline in MDR prevalence requires cautious interpretation. Although the proportion of isolates meeting MDR criteria decreased after 2020, this change coincided with the emergence of a highly homogeneous population dominated by fluoroquinolone-resistant phenotypes. Similar phenomena have been reported in other pathogens, where reductions in MDR metrics masked the stabilization of dominant resistant clones with limited therapeutic options (Baquero et al., 2009). Accordingly, reductions in MDR prevalence should not be interpreted as unequivocal improvement, but rather as potential indicators of structural reorganization within the resistance landscape.

Several limitations should be considered. The study is based on single-center observational data, which may limit generalizability. In addition, genomic data were not available to directly confirm clonal relationships or transmission pathways. Integration of genomic epidemiology would be essential to validate the proposed clonal replacement mechanism and to link phenotypic restructuring to underlying evolutionary processes (Didelot et al., 2012). Furthermore, antibiotic consumption data were not incorporated, limiting direct inference regarding selective drivers. The use of a linear model for entropy trends represents a simplification, as short-term fluctuations are present. However, the linear approach was chosen to capture the dominant long-term direction of change, and more complex nonlinear modeling was beyond the scope of the present study. The Markov modeling approach is based on empirical transition matrices and does not explicitly account for uncertainty in transition probabilities. However, the dominance of high-frequency transitions and the consistency between transition patterns and observed architecture distributions support the robustness of the inferred stationary behavior. In addition, the analysis did not explicitly adjust for external drivers such as antibiotic consumption patterns or pandemic-related changes in hospital organization. Therefore, the identified structural breakpoint should be interpreted as an empirical feature of the data rather than a purely intrinsic dynamical transition.

Despite these limitations, this study demonstrates that integrating segmented regression, multinomial modeling, entropy-based metrics, and Markov analysis provides a coherent framework for detecting structural transitions in AMR systems. Such approaches align with emerging calls to move beyond descriptive surveillance toward mechanistic and systems-level modeling of antimicrobial resistance dynamics (Colijn et al., 2010).

These findings suggest that structural transitions may represent a general mechanism shaping AMR dynamics in hospital pathogens.

5. Conclusions

Long-term surveillance of C. striatum reveals that antimicrobial resistance dynamics are driven by structural reorganization of a constrained phenotypic state space rather than progressive accumulation of resistance traits. A marked transition around 2020 reflects the acceleration and consolidation of an ongoing restructuring process that begins several years earlier.

These findings indicate that apparent reductions in multidrug resistance may reflect underlying system stabilization rather than true reversal of resistance burden. Interpreting AMR trends therefore requires a systems-level perspective that accounts for structural changes in resistance architecture, as such transitions may lead to the persistence of dominant and clinically challenging phenotypes despite declining MDR rates.

CRediT authorship contribution statement

Anna Lorencz: Writing – original draft, Methodology, Investigation, Formal analysis, Conceptualization. Bence Sajerli: Project administration, Methodology, Investigation, Formal analysis, Data curation. Katalin Burián: Writing – review & editing, Validation, Supervision, Funding acquisition. László Orosz: Writing – original draft, Visualization, Software, Resources, Methodology, Investigation, Funding acquisition, Formal analysis, Data curation, Conceptualization.

Funding statement

The authors received open access publication support from the University of Szeged, Hungary (Klebelsberg Library, Open Access Fund; application ID: 8641). No external research funding was received for this work.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Handling Editor: Dr Yiming Shao

Appendix A. Supplementary data

The following are the Supplementary data to this article: