Introduction

As a natural reservoir of antibiotic-resistance genes (ARGs), including those encountered in human pathogens1,2, soil plays a pivotal role in the emergence, evolution, and dissemination of antibiotic resistance across diverse ecosystems3,4. Listeria, a bacterial genus that includes two pathogenic members, L. monocytogenes and L. ivanovii, is commonly found in soils5. L. monocytogenes causes listeriosis in vulnerable human populations with a notable fatality rate of 20-30%6,7, while L. ivanovii rarely causes listeriosis in humans and is primarily a pathogen of ruminant animals8. Listeria can be broadly divided into two groups: sensu stricto and sensu lato, based on the relatedness of species to L. monocytogenes, with sensu stricto species being more closely related9. The standard treatment for listeriosis is a combination of penicillin and aminopenicillins (ampicillin or amoxicillin)9 or ampicillin and gentamicin10. While the incidence of resistance among clinical L. monocytogenes to these antibiotics remains low at present, intrinsic resistance to cephalosporins exists, and increased resistance to penicillin, trimethoprim, and rifampicin has been observed9,11,12,13,14. Furthermore, accelerated rates of antibiotic resistance in L. monocytogenes have been observed in food-associated environments15, possibly due to prolonged exposure to sublethal concentrations of antimicrobial agents in food processing and agriculture settings16,17, such as poultries15,18 and fresh produce factories19,20. Since Listeria can be transmitted from soils directly to humans21, or indirectly, via the food production chain22, they could be a key model for understanding how ARGs carried by soil microbes can be potentiated into human pathogens. Also, establishing a fundamental understanding of the ecological and evolutionary drivers of antibiotic resistance among soil-dwelling Listeria could help to better interpret current and future trends in antibiotic resistance patterns observed in food and clinical isolates. However, most studies of ARGs in Listeria have primarily focused on food-related and clinical isolates of L. monocytogenes18,23,24,25, resulting in an incomplete understanding of the dynamics of ARGs in Listeria in the environment.

Previous studies indicate that environmental factors, such as nutrient availability, temperature, pH, and exposure to natural or anthropogenic chemicals, can exert selective pressure favoring antibiotic resistance26,27. Genes essential for metabolism and behavior, including ARGs, have been observed to undergo positive selection (PS) to adapt to varying environments28,29. Apart from PS, environmental pressures can facilitate horizontal gene transfer (HGT), a pivotal pathway for the evolution of new resistant strains30. HGT typically occurs through three mechanisms, transformation (i.e., the uptake of free DNA from the environment), conjugation (i.e., the direct transfer of genetic material from one bacterium to another through physical contact typically encoded by plasmids or transposons), and transduction (i.e., transfer of genetic material via viruses)31. Existing evidence suggests that the acquisition of ARGs among L. monocytogenes is mediated by HGT. For example, the acquisition of tetracycline and trimethoprim resistances in L. monocytogenes has been experimentally linked to transposons like Tn916-Tn1545 and Tn619832,33. Despite the important role of environmental factors and HGT on the evolution of antibiotic resistance, the prevalence of ARGs, the extent of HGT and PS acting on them, and the influence of environmental factors on the distribution and evolution of ARGs in Listeria  in the soil environment remains largely unknown.

To bridge these knowledge gaps, we comprehensively examined the distribution of ARGs and associated HGT, PS, and environmental factors using a unique set of whole-genome sequencing data and paired environmental variables for 594 soil-dwelling Listeria isolates representing 19 species, including L. monocytogenes, that we collected across the United States (US). We identified five putatively functional ARGs, lin, mprF, sul, fosX, and norB. These ARGs were predominantly found among Listeria sensu stricto species and showed evidence of HGT within and/or across Listeria species, likely mediated by transformation, rather than conjugation and transduction. With ecological analysis and machine learning models, we also revealed evidence of environmental selection acting on the richness and genetic divergence of ARGs, likely mainly triggered by soil properties and surrounding land use, respectively. This study yields new insights into the dynamics of antibiotic resistance in soils and suggests that environmental disturbance may facilitate the emergence and spread of ARGs among bacterial species.

Results

Prevalence and spatial distribution of ARGs in soil-dwelling Listeria

In this study, a gene was defined as a “functional gene” if the sequence coverage exceeded 80% and no premature stop codon was detected, and as a “truncated gene” if the sequence coverage ranged between 30% and 80% or premature stop codon was detected34. A gene being either functional or truncated was referred to as a “present gene” (see Methods). We identified seven distinct ARGs in soil-dwelling Listeria: lin, mprF, sul, fosX, norB, dfrD, and mphB. Specifically, lin confers resistance to lincomycin, mprF to defensin, daptomycin, and gallidermin, sul to sulfamethoxazole, fosX to fosfomycin, norB to fluoroquinolones and nalidixic acid, dfrD to trimethoprim, and mphB to erythromycin, telithromycin, quinupristin, pristinamycin IA, and virginiamycin S35,36,37. Among these ARGs, the majority of lin, mprF, sul, fosX, and norB were functional, while dfrD and mphB were truncated and were each present in only 0.17% of Listeria genomes (Fig. 1a). Among the functional ARGs, lin was most prevalent among Listeria genomes (82.66%) followed by mprF (82.32%), sul (81.14%), fosX (60.77%), and norB (58.42%) (Fig. 1a).

Fig. 1: National ARG profiles of soil-dwelling Listeria.
figure 1

a Prevalence of both present (blue) and functional (red) ARGs across Listeria genomes. b Proportion of functional ARGs among different Listeria species, both sensu stricto (green) and sensu lato (orange). c Richness of both present (blue) and functional (red) ARGs in different Listeria species. Error bars, mean ± standard deviation. For Listeria sensu stricto (green) species, sample size N is as follows: L. monocytogenes (177), L. innocua (33), L. marthii (14), L. farberi (5), L. cossartiae (11), L. swaminathanii (1), L. seeligeri (98), L. welshimeri (141), L. ivanovii (2), and L. immobilis (9). For Listeria sensu lato (orange) species, N is as follows: L. rocourtiae (2), L. booriae (90) L. aquatica (1), L. fleischmannii (3), L. grandensis (3), L. grayi (1), L. weihenstephanensis (1), L. portnoyi (1), and L. newyorkensis (1). d Spearman’s correlation between the genetic similarity to L. monocytogenes and average richness of functional ARGs in Listeria species. Genetic similarity was calculated based on pairwise average nucleotide identity (ANI) between genomes for a given species and L. monocytogenes. ρ represents Spearman’s rank correlation coefficient. The line and shaded area depict the best-fit trendline and the 95% confidence interval (mean ± 1.96 standard error of the mean, SEM) for the linear regression. e Richness of functional ARGs among Listeria genomes across the US. The map was divided into eastern and western regions based on the longitudinal coordinate of the center of the US (−95°). Circles and crosses indicate genomes with and without functional ARGs, respectively, and are color-coded by species. Circle size is proportional to the richness of functional ARGs in each genome. f Richness of functional and present ARGs among Listeria genomes compared between the eastern and western US. A two-sided Mann-Whitney U test P < 0.05 was considered statistically significant. N = 417 and 177 for the eastern and western regions, respectively, for both present and functional ARG richness. Box plots show the interquartile range (IQR), with the line representing the median and whiskers extending to 1.5 times the IQR.

Full size image

Overall, the high richness of functional ARGs was consistently observed in all sensu stricto species, especially L. monocytogenes, L. innocua, L. marthii, L. farberi, L. cossartiae, and L. swaminathanii, but not in sensu lato species (Fig. 1b). While ARGs were present in sensu lato species, nearly all of them were truncated and the overall prevalence was lower than sensu stricto species (Fig. 1c). Specifically, lin, mprF, and sul were consistently present in all genomes of sensu stricto species (n = 491, Supplementary Fig. 1a), with each being functional in 100.00%, 99.59%, and 98.17% of these genomes, respectively (Fig. 1b). Functional fosX was found in all sensu stricto species, except for L. seeligeri, L. immobilis, and L. ivanovii. Functional norB was found in all sensu stricto species, except for L. welshimeri (Fig. 1b). Among sensu lato species, functional ARGs were only detected in fosX in one L. booriae genome and one L. rocourtiae genome (Fig. 1b). Notably, ARG richness in Listeria species was highly correlated with their genetic similarity to L. monocytogenes for both present and functional ARGs (Spearman's ρ = 0.88 and 0.88, P = 1.2e-06 and 1.3e-06, respectively; Supplementary Fig. 1b, Fig. 1d). This correlation indicates that species more closely related to L. monocytogenes tend to manifest a higher richness of ARGs.

The richness of both present and functional ARGs displayed spatial heterogeneity across the US (Supplementary Fig. 1c and Fig. 1e, respectively). Of note, eastern regions exhibited significantly higher present and functional ARG richness compared to western regions (Mann-Whitney U P = 1e-18 and 1.6e-12, respectively; Fig. 1f). The geographic signal of ARG richness appeared to be driven by the distribution of species, as Listeria sensu stricto species were more prevalent in the eastern regions, especially L. monocytogenes, which harbored high ARG richness (Fig. 1e, Supplementary Fig. 1c).

Evidence of HGT and PS acting on ARGs in soil-dwelling Listeria

To gain insights into the phylogenetic origins and HGT of the five functional ARGs (i.e., lin, mprF, sul, fosX, and norB), a gene tree was constructed for each ARG, depicted in Fig. 2a–e. If a gene is inherited vertically without any HGT, we anticipate the topology of a gene tree aligning with that of the phylogenetic tree, constructed based on more informative genetic variants, such as core single nucleotide polymorphisms (SNPs). Based on this notion, the comparison of gene trees and corresponding core SNP-based trees (Supplementary Figs. 2a-e) suggests that lin (Fig. 2a), fosX (Fig. 2b), and norB (Fig. 2c) likely undergo HGT between sensu stricto species. For example, for lin (Fig. 2a), we observed clades that included a mix of L. marthii and L. cossartiae isolates, as well as L. innocua and L. farberi isolates. Similarly, fosX (Fig. 2b) displayed a comparable pattern among sensu stricto species, with clades containing a mix of L. marthii and L. cossartiae isolates, as well as L. welshimeri and L. monocytogenes isolates. Notably, the apparent HGT of fosX between L. welshimeri and L. monocytogenes isolates serves as an example of gene transfer between pathogens (L. monocytogenes) and non-pathogens (L. welshimeri). Regarding norB (Fig. 2c), evidence of HGT was observed between L. innocua and L. farberi isolates, as well as among L. monocytogenes (pathogen), L. innocua (non-pathogen), and L. seeligeri (non-pathogen) isolates. While gene trees of mprF (Fig. 2d) and sul (Fig. 2e) did not have major clades of isolates from different species, they may have undergone HGT within a species, which is not easily observable in the tree comparison. Indeed, multiple statistical tests assessing the congruence between trees, including bootstrap proportion using resampling of estimated log-likelihoods (RELL)38, Kishino-Hasegawa (KH)39, Shimodaira-Hasegawa (SH)40, expected likelihood weight (ELW)41, and approximately unbiased (AU)42 tests, indicated a significant difference between the gene tree and the corresponding core SNP-based tree for all ARGs (P < 0.001 for all; Supplementary Table 1). These results suggest that ARGs in soil-dwelling Listeria have evolved differently from the overall species history due to HGT.

Fig. 2: Evidence of HGT of functional ARGs among Listeria isolates.
figure 2

a–e Maximum likelihood gene tree for a lin, b fosX, c norB, d mprF, and e sul. Trees were constructed using sequences for lin, fosX, norB, mprF, and sul detected in 491, 361, 347, 489, and 482 genomes, respectively, with 1,000 bootstraps. The evolutionary models used for constructing the trees were TN + F + I + R4, HKY + F + I + R3, TVM + F + I + R4, GTR + F + I + R5, and K3Pu+F + I + R3 for lin, fosX, norB, mprF, and sul, respectively. Trees were rooted by midpoint. Bootstrap values > 80% are indicated by grey circles. f Total number of homologous recombination events detected across Listeria species (orange) and within species (blue) for each ARG.

Full size image

To provide more evidence of HGT of ARGs among Listeria species, we further detected homologous recombination, a biological process used in HGT by bacteria to exchange genetic material, using nine methods implemented in Recombination Detection Program v4 (RDP4)43, including RDP, GENECONV, BOOTSCAN, MAXCHI, CHIMAERA, SISCAN, PHYLPRO, LARD, and 3SEQ (see Methods). Recombination events were observed for all ARGs (Fig. 2f). A total of 20 recombination events were observed in lin, with 14 occurring within species and six across species (Fig. 2f). These included events among L. marthii, L. cossartiae, and L. swaminathanii isolates as well as between L. farberi and L. innocua isolates (Supplementary Data 1), which were also reflected in the tree comparison (Fig. 2a). Recombination was also detected within L. monocytogenes in lin (14 events; Supplementary Data 1). For fosX, three recombination events were observed, and they all occurred across species (Fig. 2f, Supplementary Data 2), including those among L. monocytogenes, L. welshimeri, and L. cossartiae isolates that were also detected in the tree comparison (Fig. 2b), and between L. innocua and L. farberi isolates. For norB, 18 recombination events were detected, with four occurring within species and 14 across species (Fig. 2f). For instance, recombination detected among L. monocytogenes, L. innocua, and L. seeligeri isolates supports the observation in the tree comparison that HGT may occur between pathogens and non-pathogens (Supplementary Data 3, Fig. 2c). For mprF, 14 recombination events (12 within species and two across species) were detected (Fig. 2f), with some involved between L. ivanovii (pathogen) and L. immobilis (non-pathogen) isolates as well as between L. marthii and L. cossartiae isolates (Supplementary Data 4). For mprF, two recombination events were detected across species (e.g., between L. ivanovii and L. immobilis isolates and between L. cossartiae and L. marthii isolates) and 12 events were detected within a species (e.g., L. innocua, L. cossartiae, L. welshimeri, L. monocytogenes, and L. seeligeri; Supplementary Data 4). For sul, four recombination events were detected, with two occurring between two closely related species (i.e., between L. marthii and L. swaminathanii isolates and between L. innocua and L. welshimeri isolates; Supplementary Data 5). For the other two recombination events, one was observed within L. innocua, and the other was within L. seeligeri (Supplementary Data 5). Taken together, these results compellingly suggest that HGT of ARGs occurs both within and across Listeria species.

To understand if maintaining ARGs may offer advantages in fitness to adapt to certain environmental stressors, we performed a gene-wide test for PS for each functional ARG. Results showed that mprF and sul exhibited a better fit to the unconstrained model that allows PS compared to the constrained model (LRT = 13.932 and 17.539, P = 4.72 e-04 and 7.77e-05, respectively; Fig. 2d, e), indicating the presence of PS acting on these two genes. For lin, fosX, and norB, there was no evidence supporting the presence of PS acting on them (P > 0.05 for all).

The probable mechanism of HGT of ARGs among Listeria species

To investigate the potential mechanisms underlying the HGT of ARGs among Listeria species, we employed a predictive approach focusing on mobile genetic elements (MGEs), including prophages, insertion sequences (IS), composite transposons, and plasmids. The presence of ARGs within these MGEs could indicate their role as carriers, informing transduction (via phages) and/or conjugation (via plasmids/transposons) as potential mechanisms of HGT of ARGs in Listeria. Among the 594 genomes examined, 1,338 prophages were identified (Fig. 3a). Out of these prophages, 14.7% (n = 197) were classified as ‘intact’. Of note, we found 12 ‘incomplete’ prophages in L. booriae, 11 of which presented lin and one of which presented norB (Supplementary Data 6). However, all lin presented on prophages were found to be truncated (Supplementary Data 7). The presence of remnants of lin in the prophages suggests a historical role of bacteriophages in transferring ARGs from other species to L. booriae. Apart from these observations, the only instance of recent HGT of a functional ARG potentially mediated by transduction was found in a norB gene located within a prophage, PHAGE_Bacill_Spbeta_NC_001884, of L. ivanovii isolate L7-1049 (Fig. 3b, Supplementary Data 6). The prophage region, delimited by the left (attL) and right (attR) attachment sites, comprised the norB gene and genes encoding other phage-related proteins and hypothetical proteins (Fig. 3b). The norB gene identified in this prophage was positioned in the first clade of L. ivanovii, L. seeligeri, and L. immobilis isolates in the gene tree (Fig. 2c). Homologous recombination was also detected among these species in this gene (Supplementary Data 3). These results suggest that overall HGT of functional ARGs mediated by transduction is rare in soil-dwelling Listeria, but may has occurred among L. ivanovii, L. seeligeri, and L. immobilis isolates in norB.

Fig. 3: Overview of the phylogeny and prevalence of ARGs, motility genes, competence genes, and MGEs in Listeria.
figure 3

a The maximum likelihood phylogenetic tree of the 594 Listeria genomes using core SNPs with 200 bootstraps. This tree was adapted from a previously published phylogenetic tree constructed using these genomes29. The tree was rooted by midpoint and features branches color-coded by Listeria species. The outer ring annotations include the presence/absence of ARGs, motility genes, and competence (Com) genes. A filled box indicates the presence of a functional gene; an empty box indicates the presence of a truncated gene; and white space indicates the absence of a given gene. MGEs are presented either as counts (for plasmids) or as proportions (for ISs and prophages) in the outer ring annotations. b Visualization of a prophage carrying a functional norB gene in L. ivanovii isolate L7-1049. norB, genes encoding phage-related proteins, and genes encoding hypothetical proteins are shown in red, yellow, and purple, respectively. The left and right attachment sites for the phage are referred to as attL and attR, respectively. c, d The number of c functional competence genes and d functional motility genes compared between genomes with ARGs involved and not involved in homologous recombination events. A two-sided Mann-Whitney U P < 0.05 was considered statistically significant. N = 385 and 209 for genomes with ARGs involved and not involved in homologous recombination events, respectively, for both functional competence and motility genes. Box plots show the IQR, with the line representing the median and whiskers extending to 1.5 times the IQR.

Full size image

A total of 4,023 ISs were identified among the 594 genomes (Fig. 3a), with 18.3% (n = 735) being classified as ‘complete’. IS3 and IS1182 families constituted 66.4% (n = 488) and 15.5% (n = 114) of the complete ISs, respectively. Of note, we identified an IS-associated ARG that matched the functional fosX located on the negative strand in L. welshimeri isolate L7-1846, showing 100% identity and coverage. IS3 transposition involves a copy-out-paste-in process that requires at least two copies of IS344. With the copy of IS3 transposase being located downstream of fosX on the positive strand, we consider that the IS3 transposase under this configuration would not be sufficient for gene transfer. Also, no composite transposons carrying ARGs were detected. In addition, a total of 81 plasmids were identified (Supplementary Data 7). The most dominant plasmid incompatibility (Inc) family and group were Inc18 (98.8%, n = 80) and repUS25 (96.3%, n = 78), respectively. None of these plasmids, however, were found to carry any ARGs. These results suggest that HGT of functional ARGs mediated by conjugation is rare in soil-dwelling Listeria.

Given that both conjugation and transduction did not appear to be substantial contributors to the HGT of functional ARGs in soil-dwelling Listeria, we further investigated whether natural transformation might play a role. We predicted the presence of 12 competence genes, which signify a bacterium’s capacity to uptake foreign DNA, or extracellular DNA (eDNA), from its surroundings, and integrate it into its genome45. Results showed that Listeria sensu stricto species possessed more competence genes than sensu lato species (Fig. 3a). Among Listeria sensu stricto species, competence genes were uniformly present, with over 90% functionality observed for several key competence genes, including comK, coiA, cinA, comEC, and comEB (Fig. 3a). On the contrary, among Listeria sensu lato species, specific competence genes such as comGD, comG, and comGF were completely absent (Fig. 3a). The sole functional competence gene identified among Listeria sensu lato species was comEC, identified in three L. fleischmannii isolates (L7-1629, L7-1641, and L7-1645). A further comparison of the count of functional competence genes revealed that genomes with ARGs involved in recombination events had 11 functional competence genes on average, significantly higher than genomes with no recombination events detected (Mann-Whitney U P = 4.9e-82; Fig. 3c). This result suggests the importance of competence in homologous recombination in ARGs of Listeria.

As motility plays a crucial role in enabling bacteria to move, providing adaptive advantages in new environments46 and facilitating DNA uptake47, we further predicted motility genes among  Listeria genomes. Consistent with competence genes, Listeria sensu stricto species possessed more motility genes compared to sensu lato species (Fig. 3a). Specifically, all 31 motility genes identified were present in every Listeria sensu stricto species, with 90% of these genes being identified as functional in more than 99% of isolates, except for L. immobilis, the sole sensu stricto species identified thus far that lacks motility48. In contrast, almost none isolates of sensu lato species harbored any functional motility genes. The high motility in Listeria sensu stricto species may increase their chance of being exposed to diverse DNA pools in the environment and facilitate HGT. Indeed, we found that genomes with ARGs involved in homologous recombination have 30 functional motility genes on average, significantly higher than genomes without ARGs involved in recombination (Mann-Whitney U P = 2e-62; Fig. 3d). Collectively, the high prevalence of functional competence and motility genes in Listeria sensu stricto species, where functional ARGs predominate, and their associations with homologous recombination suggest an essential role of transformation in HGT of ARGs in Listeria.

Environmental factors associated with the richness and genetic divergence of ARGs in soil-dwelling Listeria

To assess the potential influence of the environment on ARG acquisition and evolution in Listeria, we performed a series of analyses, including Spearman’s partial correlation analysis, variation partitioning analysis (VPA), multidimensional scaling (MDS) analysis, Mann-Whitney U tests, and machine learning-based analysis. Considering the high correlation between ARG richness and the genetic similarity of genomes to L. monocytogenes (Spearman’s ρ = 0.88, P < 0.001; Fig. 1d, Supplementary Fig. 1b) and a geographic signal likely driven by species (Fig. 1e, Supplementary Fig. 1c), genetic similarity could potentially confound the correlations between environmental variables and ARG richness. Thus, Spearman’s partial correlation analysis, controlling for the genetic similarity of Listeria species to L. monocytogenes, was performed. Thirteen out of 34 environmental variables were significantly correlated with ARG richness (adjusted Spearman's P < 0.05 for all). Among these, seven variables, including aluminum, forest, zinc, manganese, iron, longitude, and developed areas with < 20% impervious surface, exhibited a positive correlation, while the remaining six: copper, wetland, molybdenum, magnesium, pH, and potassium, showed a negative correlation (Fig. 4a). To further quantify the contributions of different environmental variable categories to the observed variation in ARG richness, VPA was conducted. To exclude potential confounding effects, genetic similarity to L. monocytogenes was included in this analysis. Of note, the climate category was excluded because no climatic variables were found to be significant (Fig. 4a). As expected, genetic similarity alone accounted for the majority of the variation (adjusted R2 = 80.53%; Fig. 4b). Soil properties accumulatively explained 7.37% of the variation, similar to surrounding land use, which explained 6.74 % (Fig. 4b). Geolocation accounted for less than 1% of variation (Fig. 4b). MDS analysis further showed that isolates with and without functional ARGs formed significantly different clusters based on environmental conditions, which explained 26% of the variation of ARG presence (PERMANOVA P < 0.001, R² = 0.260, pseudo-F = 6.768; Fig. 4c). As functional ARGs were predominately detected in Listeria sensu stricto species (Fig. 1c), we hypothesized that environmental variables significantly differ between the habitats of Listeria sensu stricto and sensu lato species. Indeed, a total of 13 environmental variables (five soil property variables, including aluminum, one geolocation variable, three climatic variables, and four surrounding land use variables) were significantly different, and more than half of them (minimum and maximum temperatures, coverage of pasture, cropland, barren, and shrubland, and precipitation) had a fold change of greater than 1 or less than −1 (adjusted Mann-Whitney U P < 0.05 for all; Fig. 4d). These results collectively suggest that environmental conditions, particularly soil properties and surrounding land use, are important to the prevalence of ARGs in Listeria.

Fig. 4: Environmental conditions associated with the richness and genetic divergence of ARGs in Listeria.
figure 4

a Spearman’s partial correlation between the richness of ARGs in Listeria and environmental variables, controlled for genetic similarity to L. monocytogenes. Positive and negative correlation is indicated by purple and orange, respectively. b Venn diagram of VPA showing the variation of the functional ARG richness explained by environmental variable categories with significant correlations detected (i.e., geographic locations, soil properties, and surrounding land use) and genetic similarity to L. monocytogenes. Residuals indicate unexplained variation. c MDS analysis for genomes with (orange) and without (blue) ARGs based on environmental conditions. Each dot represents a genome. PERMANOVA P = 0.001 (R² = 0.260 and pseudo-F = 6.768). Ellipse size is determined by two times the standard deviation from the mean. d Volcano plot illustrating the significance (two-sided Mann-Whitney; y-axis) of the difference in environmental variables (fold change; x-axis) compared between Listeria sensu stricto and sensu lato species. Variables above the red dashed line have an adjusted P < 0.05. Dots are color-coded by environmental variable categories. e Receiver operating characteristic (ROC) curve for the prediction of ARG presence from environmental variables using a random forest model (auROC = 0.76). Each curve reflects one evaluation using holdout data, repeated 10 times, denoted by light blue lines. The dark blue line represents the mean performance across these repetitions. f Mantel tests between the sequence dissimilarity and environmental variables for a given ARG. Positive and negative correlation is indicated by blue and yellow, respectively. g Venn diagram of VPA showing the variation of the genetic divergence of norB explained by environmental variable categories. For a and f P-values were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate (BH-FDR) method. Significance levels are denoted by “*”, “**”, “***”, and “****” for P < 0.05, P < 0.01, P < 0.001, and P < 0.0001, respectively, and “ns” indicates not significant.

Full size image

Given the observed correlations between environmental variables and ARG richness, we hypothesize that the presence of ARGs in Listeria is predictable from environmental variables using machine learning models. To test this hypothesis, we compared different machine learning algorithms (see Methods). After hyperparameter tuning (Supplementary Data 8), a random forest model, which utilized logistic loss to assess the quality of splits, had a maximum depth of three, considered the logarithm base two of the total number of features to determine the best split, and comprised 100 trees in the forest, was identified as the best model for predicting the presence of ARGs with environmental variables. This model achieved a mean area under the receiver operating characteristic curve (auROC) of 0.76 (Fig. 4e), indicating acceptable discrimination ability49, and a mean area under the precision-recall curve (auPR) of 0.95 (Supplementary Fig. 3a). To interpret outputs from the best-performing model, we employed Shapley Additive exPlanations (SHAP)50 to evaluate the importance of each feature in the prediction. Among the top ten most predictive features, aluminum emerged as the most influential feature (Supplementary Fig. 3b), aligning with the results from the Spearman's partial correlation analysis (Fig. 4a). Except for magnesium and developed areas with < 20% impervious surface, the remaining most predictive features (i.e., latitude, organic matter, maximum and minimum temperature, wind speed, calcium, and surrounding shrubland coverage) were not significant in the correlation analysis (Supplementary Fig. 3b, Fig. 4a). This result suggests that machine learning models can capture complex relationships that are often undetectable through conventional statistical methods51. Despite that the SHAP values of features were overall not high (less than 0.030; Supplementary Fig. 3a), the model effectively used a combination of these features to make relatively accurate predictions, as shown in Fig. 4e. This is common in complex models where the predictive power comes from the collective contribution of many features rather than a few dominant ones50,52,53.

Lastly, to investigate the interplay between the genetic divergence of ARGs and environmental factors, Mantel tests were conducted to assess the correlations between the sequence dissimilarity of each ARG and the distance of each environmental variable. Seven variables—geographic distance, pH, potassium, precipitation, maximum and minimum temperatures, and surrounding forest coverage—exhibited a significant positive correlation with sequence dissimilarity universally in all five functional ARGs (Mantel P < 0.05 for all; Fig. 4f). Among these ARGs, mprF exhibited significant correlations with the greatest number of environmental variables (n = 19). To delineate the contribution of environmental variable categories to the variation of ARG sequence dissimilarity, we further conducted VPA. Among all five ARGs, norB sequence divergence was the most affected by environmental variables, collectively accounting for 16.56% of the explained variation (Fig. 4g). For fosX, mprF, lin, and sul, environmental factors collectively contributed to 12.99%, 7.93%, 6.79%, and 6.26% of the explained variation, respectively (Supplementary Figs. 4a-d). Despite the varying contributions of environmental variables to the sequence divergence in different ARGs, a consistent pattern emerged that surrounding land use was the most influential factor across all ARGs, independently (and collectively) explaining 2.02% (11.13%), 3.04% (6.48%), 1.35% (4.77%), 1.57% (4.42%), and 1.48% (3.45%) of the variation for norB, fosX, mprF, lin, and sul, respectively, followed by soil properties. These results suggest that environmental conditions, particularly surrounding land use and soil properties, are important to the genetic diversification of ARGs in Listeria.

Discussion

This study investigated the dynamics of ARGs in soil-dwelling Listeria, including sensu stricto species, which are more closely related to L. monocytogenes, and sensu lato species. We identified five functional ARGs, lin, mprF, sul, fosX, and norB, predominantly in Listeria sensu stricto species. Most of these five ARGs are still present in Listeria sensu lato species but were found to be truncated, suggesting that carrying ARGs may cause metabolic costs in these species54. In contrast, maintaining at least some of these ARGs in the genomes might increase fitness in Listeria sensu stricto species evidenced by the PS detected in two ARGs, mprF and sul. The large discrepancy in the prevalence of ARGs between Listeria sensu stricto and sensu lato species may be partly attributed to the different conditions in soil properties, climate, and surrounding land use these two bacterial groups encounter.

The five functional ARGs identified in this study were previously reported in L. monocytogenes and some other Listeria sensu stricto species23,35,55,56. The current treatment protocol for listeriosis involves a combination of penicillin and aminopenicillins (ampicillin or amoxicillin)9 or ampicillin and gentamicin10. ARGs conferring resistance to these antibiotics used in clinical treatment, including ampR, aacA4, and aadC, were not detected in this study, which suggests that antibiotics used in clinical settings had limited impact on soil-dwelling Listeria in nature. Consistent with our findings, ARG surveillance of L. monocytogenes in food-related and clinical settings in France23, Denmark57, and Spain58 report that acquired resistance is limited in L. monocytogenes and this pathogen remains susceptible to antibiotics over time. Therefore, while emerging resistance in L. monocytogenes is observed for certain clinical-use antibiotics, like penicillin11 and rifampicin14, resistance to the antibiotics used in patients with listeriosis (aminopenicillins and gentamicin) remains rare9,23.

Evidence of HGT was observed for all functional ARGs within and/or between Listeria sensu stricto species, but not between Listeria sensu stricto and sensu lato species. This observation supports the notion that HGT tends to display a bias toward individuals and species that are more closely related59. HGT of ARGs has been observed among both clinical and food isolates from various L. monocytogenes clonal complexes, with tetracycline resistance identified as the most prevalent acquired resistance phenotype60. This has been primarily attributed to the presence of composite transposons like Tn916-Tn1545 carrying tetracycline resistance (Tn916-carrying tetM genes) in L. monocytogenes23,32,33. However, we did not identify any tetracycline resistance genes (tetM and tetS) in soil-dwelling Listeria in this study. This is likely attributed to the widespread use of tetracycline in clinical and food-related environments61,62,63, whereas the baseline tetracycline concentrations in less disturbed environments like soils might be low64.

Given the limited instances of acquired resistance observed from transposons, ISs, plasmids, or prophages in this study, we propose that conjugation and transduction may not be the primary mechanisms for the HGT of ARGs in soil-dwelling Listeria. Instead, we found that natural transformation is the most likely mechanism. Natural transformation relies on the recipient bacterium that expresses the competence machinery65 and largely depends on the uptake and incorporation of exogenous naked DNA from the environment into the genomes of competent recipient organisms66. Despite the presence of genes associated with competence machinery, L. monocytogenes has not been recognized as naturally transformable in lab settings67. The absence of competence in L. monocytogenes has been attributed to the truncation of the comK gene, which cleaved into two parts by a 42-kb region containing several open reading frames (ORFs) encoding phage-related products67. The regulation of the competence system relies on the formation of a functional comK gene via prophage excision68. In this study, we detected a complete set of competence genes, including comK not truncated by prophages, in most genomes of Listeria sensu stricto species, where HGT events of ARGs were observed and associated with. Despite unsuccessful attempts to transform L. monocytogenes with an intact comK gene under lab conditions in a previous study67, our results suggest that Listeria may require unusual conditions (beyond competence minimal medium, at 37 °C, and selection on Brain Heart Infusion agar supplemented with chloramphenicol) for competence67, which complex soil environments may uniquely be able to provide, facilitating HGT of ARGs in this bacterium via natural transformation.

Understanding the associations between environmental factors and ARGs is crucial for unraveling the dynamics and evolution of antibiotic resistance under environmental disturbance. In this study, multiple layers of analysis from different perspectives all suggest that environmental selection, likely triggered by soil properties and surrounding land use, plays a role in the evolution of ARGs in Listeria. Soil properties (e.g., pH and nutrients) have been widely reported to influence ARGs in soils69,70,71. In this study, aluminum and magnesium were found to be the most influential soil properties for ARG richness in Listeria, supported by both correlation and machine learning-based analyses. A previous study suggests that nanoalumina can enhance the uptake of ARGs by facilitating the transfer of plasmid-mediated ARGs72. Additionally, the persistence of aluminum in soils may exert prolonged selective pressure on bacteria to maintain ARGs through co-selection mechanisms73. These mechanisms can all lead to a positive correlation between ARG richness and aluminum observed in Listeria in this study. In contrast, the correlation between magnesium and ARG richness in Listeria was negative. This may be due to the competitive dynamics in magnesium-limited environments, where fungal-bacterial competition for magnesium can increase the fitness and survival of bacteria possessing ARGs, as these genes provide a competitive advantage under stress conditions74. Furthermore, magnesium-modified biochar has been found to reduce the spread and abundance of ARGs by influencing the bacterial community structure and inhibiting HGT75. Given the complexity of natural environments and the observed associations are correlational not causal, further experimental investigations are needed to better understand the influence of soil properties on ARG richness.

Besides soil properties, surrounding land use, particularly forest coverage, was found to be important for the richness of ARGs in soil-dwelling Listeria as well. This is consistent with the finding that forests, irrespective of their location and type (boreal, cold, temperate, or tropical), exhibited the highest richness of ARGs in their soils in a previous global study4. Wildlife in forests (e.g., deer76, bird77, reptiles78, and rodents79) can serve as carriers of ARGs. These genes can be subsequently excreted into soils through fecal matter as wildlife moves around, contributing to ARG richness in bacteria in adjacent environments. In addition, surrounding land use was identified as the most important factor for the genetic diversification of ARGs in Listeria. Prior research has highlighted the contribution of land use to the evolution of antibiotic resistance80. For example, in cropland areas, the extensive use of antibiotics to enhance crop productivity selects specific ARGs81. A plausible explanation for the potential influence of surrounding land use on the genetic diversification of ARGs is that land use could influence soil properties in adjacent natural environments, which indirectly imposes selective pressure on ARGs, leading to genetic diversification. For instance, surrounding cropland and pasture coverage, in which we previously detected a positive correlation with soil magnesium29, were found to be associated with the sequence dissimilarity of mprF and norB, with the former showing evidence of PS in this study. Overall, the importance of surrounding land use reflects potential anthropogenic effects on the dynamics of antibiotic resistance in the natural environment.

In summary, by leveraging a national reconnaissance of Listeria genomes, we showed the genetic and ecological context of their ARGs in the soil environment. ARGs are predominately found in Listeria sensu stricto species. Considering the limited occurrence of prophages and plasmids carrying ARGs and the presence of a full set of putatively functional competence genes in Listeria sensu stricto species that were correlated with homologous recombination involvement in ARGs, we propose that natural transformation may be the more plausible route for the HGT of ARGs in soil-dwelling Listeria. In contrast, HGT of ARGs appears to be often achieved via conjugation in food and clinical isolates, suggesting that Listeria isolates from different environments may employ distinct HGT mechanisms for ARG acquisition. We also identified evidence of environmental selection most likely triggered by soil properties and surrounding land use in the acquisition and diversification of ARGs in Listeria, highlighting the importance of monitoring the impact of environmental disturbance on the dynamics of antibiotic resistance. Overall, this study provides a baseline knowledge of evolutionary and ecological processes governing the distribution and diversity of ARGs in the soil environment and demonstrates the usage of Listeria as a model organism for understanding the impact of environmental changes on ARG mobilization and evolution.

Methods

Listeria genomes and environmental data

The genomic dataset of 594 Listeria isolates collected by us from minimally disturbed natural environments across the contiguous US (described in a previous study examining the mechanism underlying bacterial pangenome evolution29) was further analyzed to consider their carriage of ARGs in this study. The genome assemblies met the following quality control criteria: fewer than 300 contigs, N50 greater than 50,000, average coverage exceeding 30X, consistent presence of sigB allelic types based on both whole-genome sequencing extraction and PCR-based assays, and no detected contamination using kraken2–2.0.829. Also, all of these genome assemblies had a high completeness with 99.99% ± 0.03% (mean ± SD) (Supplementary Data 9) assessed by CheckM282. The genomes represent 19 Listeria species, predominantly L. monocytogenes (n = 177), followed by L. welshimeri (n = 141), L. seeligeri (n = 98), and L. booriae (n = 90). Other Listeria genomes in our dataset included L. innocua (n = 33), L. marthii (n = 14), L. cossartiae (n = 11), L. immobilis (n = 9), L. farberi (n = 5), L. fleischmannii (n = 3), L. grandensis (n = 3), L. ivanovii (n = 2), and L. rocourtiae (n = 2). Single genomes were available for L. swaminathanii, L. grayi, L. aquatica, L. weihenstephanensis, L. portnoyi, and L. newyorkensis. Among the species included in this study, L. monocytogenes, L. seeligeri, L. marthii, L. ivanovii, L. welshimeri, L. innocua, L. cossartiae, L. farberi, L. immobilis, and L. swaminathanii are classified as Listeria sensu stricto species (491 genomes total), while others are sensu lato species (103 genomes total)83.

Previously reported environmental data encompassing 34 variables29 paired with this genomic dataset were also examined further in this study. The environmental data include three geolocation (latitude, longitude, and elevation), 17 soil properties (moisture, total nitrogen, total carbon, pH, organic matter, aluminum, calcium, copper, iron, potassium, magnesium, manganese, molybdenum, sodium, phosphorus, sulfur, and zinc), four climatic (precipitation, wind speed, maximum and minimum temperatures), and 10 surrounding land use (open water, barren, forest, shrubland, grassland, cropland, pasture, wetland, and developed open space categorized as >20% and <20% impervious cover) variables29.

Detection of ARGs, competence genes, motility genes, and MGEs

ARGs, competence genes, and motility genes were identified through BLASTN searches84, using an E-value of 0.01 and without restrictions on percent identity, against a reference database sourced from the BIGSdb-Lm platform35. BIGSdb-Lm is a curated bacterial genome sequence database specializing in L. monocytogenes35. A total of 25 ARGs, 12 competence genes, and 31 motility genes were extracted from the platform (Supplementary Data 10)35. Gene with the highest bit-score was chosen for further analysis. Subsequently, the presence of premature stop codons (TGA, TAG, TAA), up to but not including the last stop codon85, and sequence coverage (%) were assessed for each detected gene using Python 3.6.8. Genes were categorized as putatively functional if their sequence coverage exceeded 80% and no premature stop codon was detected; truncated if its sequence coverage ranged between 30% − 80% or premature stop codon was detected; and absent if sequence coverage was less than 30% or no hits were observed in the BLASTN searches34. Based on this categorization, we further simplified the classification as present (including either truncated or functional) and functional34.

To assess possible bias in ARG detection using BIGSdb-Lm, we first applied the Comprehensive Antibiotic Resistance Database (CARD)37, which includes a more diverse set of reference ARGs than BIGSdb-Lm, to our dataset. We found no significant differences in the overall prevalences for present and functional ARGs between the predictions from BIGSdb-Lm and CARD (paired two-sided t-test P = 0.81 and 0.11, respectively; Supplementary Fig. 5a). Also, a consistent trend was observed where Listeria sensu stricto species harbor more functional ARGs than senso lato species (Supplementary Fig. 5b) and the functional ARG richness was strongly positively correlated with the genetic similarity to L. monocytogenes (Spearman’s ρ = 0.88 and 0.91, P = 1.7e-07 and 1.3e-06; Fig. 1d and Supplementary Fig. 5c for BIGSdb-Lm and CARD, respectively). Of note, using CARD, we did not identify any sul, an ARG commonly found in L. monocytogenes23,35, but it predicted fosXCC, an ARG with a similar function to fosX, but is commonly identified in Campylobacter coli rather than Listeria species86 (Supplementary Figs. 5a-b). Next, we obtained the reference genome of L. rocourtiae FSL F6-920 (accession number: AODK01) from the NCBI database and predicted its ARGs using both BIGSdb-Lm and CARD. This isolate was selected because it belongs to Listeria sensu lato species, and genotypic and phenotypic antibiotic resistance data for this isolate were published in literature87. Using BIGSdb-Lm, we identified three ARGs, sul, mprF, and fosX. fosX was classified as functional according to our criteria, aligning with its known phenotypic fosfomycin resistance in this isolate87, while sul and mprF were truncated due to the presence of a premature stop codon (Supplementary Data 11). Since this isolate was found to be phenotypically resistant to sulfamethoxazole (conferred by sul)87, the functionality of sul with a truncated feature might be attributed to its high coverage (94.6%) and the position of stop codons in a later part of the gene (Supplementary Fig. 6). Using CARD, however, none ARGs detected in this genome were functional (Supplementary Data 12). Although it predicted a fosX gene, this gene was truncated, and it did not identify any sul genes. Taken together, we conclude that our prediction of ARGs in Listeria genomes using BIGSdb-Lm is not biased, and it is robust and sensitive even for Listeria sensu lato species.

To predict MGEs, including plasmids, ISs, transposons, and prophages, we employed specialized programs, including PlasmidFinder288 (0.6 cutoff) to identify plasmids; ISEScan89 and ISAbR_finder90 to detect ISs and ISs associated with ARGs; TnFinder90 to identify composite transposons; and PHASTER91 for prophage prediction. All programs were employed with default settings if not specified. Subsequently, a comparison was made to determine if the functional ARGs were present in (for plasmids and prophages) or near (for IS, within 2000 bp90) the predicted MGEs, based on their genomic coordinates. Positive findings were visualized using Gene Graphics92.

Richness, diversity, and spatial distribution of ARGs

The richness and Shannon-Wiener diversity index of present and functional ARGs were computed for each genome using the skbio library in Python 3.6.8. Since we observed a strong positive correlation between ARG richness and diversity (Spearman’s ρ = 1, P < 10e-30; Supplementary Fig. 7), and ARG richness is easier to interpret and calculate and has been widely used in other studies23,93,94, subsequent analyses were focused only on ARG richness. To evaluate the association between the ARG richness of Listeria species and their genetic similarity to L. monocytogenes, we averaged the previously reported pairwise average nucleotide identity (ANI)29 for each genome based on their respective species compared with one L. monocytogenes genome and correlated it with the average richness of ARGs (both present and functional) using Spearman’s rank correlation analysis. The spatial distribution of ARGs was visualized using Mercator Projection and the Basemap Matplotlib Toolkit v.1.2.1 in Python v.3.6.8. We then compared the ARG richness between the eastern and western regions determined by the longitude of the center of the US (−95°) using two-sided Mann-Whitney U test, with P < 0.05 indicating a significant difference.

Gene and phylogenetic tree construction, tree topology congruence tests, and detection of homologous recombination and PS

We annotated a previously published core-SNP-based phylogenetic tree of 594 Listeria genomes29 with details about species, ARGs, plasmids, competence genes, motility genes, ISs, transposons, and prophages using the Interactive Tree of Life (iTOL) webserver95.

Gene tree of each ARG was constructed using a maximum likelihood method with 1,000 bootstraps in IQ-TREE96 based on nucleotide sequences aligned using MUSCLE v.3.8.3197. IQ-TREE implements ModelFinder to select the best evolutionary model for the phylogenetic estimates98. The best-fit models, determined by the Bayesian information criterion (BIC), were TN + F + I + R4, GTR + F + I + R5, K3Pu+F + I + R3, HKY + F + I + R3, and TVM + F + I + R4 for lin, mprF, sul, fosX, and norB, respectively. A phylogenetic tree for genomes harboring each ARG was constructed based on core SNPs using RAxML v899 with 200 bootstraps and the GTR  +  G  +  I substitution model with ascertainment bias correction applied. Core SNPs of genomes included in the construction of each phylogenetic tree were determined using a custom Python script published previously29. All trees were visualized using the iTOL webserver95. To test the congruence in the topologies between the core SNP-based tree and gene tree for each ARG, five statistical methods, RELL38, KH39, SH40, ELW41, and AU42, were employed. A P value < 0.05 indicates significantly different tree topologies.

Nine detection methods (RDP, GENECONV, BOOTSCAN, MAXCHI, CHIMAERA, SISCAN, PHYLPRO, LARD, and 3SEQ) implemented in RDP443 were used to detect homologous recombination events in ARGs. Events detected by at least two methods were considered true positive. To assess whether PS occurs across the entire gene in each functional ARG, we utilized the BUSTED (branch-site unrestricted statistical test for episodic diversification) model in HyPhy (Hypothesis Testing Using Phylogenies)100. The aligned nucleotide sequences and the ARG trees constructed in the previous steps were used as input files. A likelihood ratio test (LRT) was performed to compare the unconstrained model, which allows PS, and the constrained model, which disallows PS. Statistical significance was determined by approximating the test statistic to a χ2 distribution. ARGs with a P < 0.05 were considered evidence of PS, at least at one specific site in a given ARG.

Assessment of the relationships between environmental variables and the richness and diversification of ARGs

A Spearman’s partial correlation analysis, controlling for the genetic similarity to L. monocytogenes, was performed to evaluate the associations between ARG richness and each environmental variable followed by a BH-FDR adjustment to account for multiple testing. This approach addresses potential confounding effects caused by genetic similarity to L. monocytogenes. Environmental variables with an FDR-adjusted P < 0.05 were considered significant. Following this, a VPA was conducted, controlling for genetic similarity to L. monocytogenes, to evaluate the relative contribution of each environmental category with significant variables identified (i.e., geolocation, soil property, and land use) to ARG richness. Both genetic similarity and environmental data were structured in matrix format, enabling the use of adjusted R2 in redundancy analysis (RDA) ordination to partition the variation. The respective adjusted R2 value in VPA was visualized as a Venn diagram using the ‘varpart’ function in the vegan package v.2.6-4 in R. Furthermore, MDS based on the Euclidean distance of environmental variables along with a permutational multivariate analysis of variance (PERMANOVA) test was used to compare overall environmental conditions for genomes with and without functional ARGs. PERMANOVA P < 0.05 indicates that there are significant differences in the environmental conditions between genomes with and without functional ARGs. Two-sided Mann–Whitney U tests were employed to identify significant difference for each of the environmental variables between samples positive for Listeria sensu stricto and sensu lato species followed by FDR correction.

Mantel tests were conducted to assess the relationships between the distance matrices of environmental variables and ARG sequences followed by FDR correction. Briefly, genetic dissimilarity of a given ARG between genomes was quantified using the Levenshtein distance101, while dissimilarity of a given environmental variable between genomes (excluding longitude and latitude)29 was calculated using Euclidean distance. Geographic distance was computed based on longitude and latitude using the haversine formula. VPA was then performed using the calculated distance matrices to assess the relative contribution of each environmental category with significant variables identified to the sequence dissimilarity of a given ARG.

Machine learning models to predict the presence of ARGs from environmental variables

To predict the presence of ARGs from environmental variables, we developed an end-to-end machine learning-based framework that embodies a series of individual software programs (e.g., scikit-learn and SHAP) written in Python 3.6.8 for data preparation, hyperparameter tuning, model training, and testing, model evaluation through cross-validation, and visualization. Samples were first cleaned and split into the training set (80%) and the testing set (i.e., the holdout set; 20%) in a stratified fashion. The training set was further split into five stratified folds for cross-validation, in which a collection of predefined models (i.e., classifiers based on decision trees, random forest, multilayer perceptron, support vector machines, and gradient boosting) were trained and tested with a random set of hyperparameters (Supplementary Data 8). The average auROC score was used to evaluate the performance of the models across the five rounds of cross-validation. To account for stochasticity introduced by the random splitting of samples and division of training data into five folds, we repeated these steps 10 times. We selected the best model and its hyperparameter set with the highest interquartile mean of the auROC scores out of the 10 repetitions among the predefined models. The interquartile means of the auROC and auPR scores of the best model that was exclusively trained on the training set were reported based on a single evaluation of the holdout data from each of the 10 repetitions. The importance of the features was quantified using SHAP50.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.