A growing volume of microbiology research has been dedicated to characterizing drug-drug interactions in combination antibiotic treatment, because understanding how and why antibiotics interact with each other has the potential to help clinicians improve treatment efficacy, reduce dosage-related side effects and slow the emergence of antibiotic resistance (Tamma et al., 2012; Worthington and Melander, 2013; Mitosch and Bollenbach, 2014; Bollenbach, 2015; Wood, 2016). The use of high-throughput assays for measuring bacterial growth rate has accelerated the identification and characterization of antibiotic interactions (Yeh et al., 2006; Wood et al., 2012; Ocampo et al., 2014; Chevereau and Bollenbach, 2015; Zhou et al., 2015; Nguyen et al., 2016). Pairwise interactions between antibiotics can be classified as additive, synergistic or antagonistic (Figure 1A) based on whether their combined effects on bacterial growth inhibition are equal to, greater than or less than those of their individual effects, respectively (Loewe, 1928, 1953; Bliss, 1939; Yeh et al., 2006; Foucquier and Guedj, 2015). Suppression is a special type of antagonism in which the combined growth inhibitory effect is weaker than that of one antibiotic or both antibiotics alone (Yeh et al., 2006).

Our current understanding of the origins of antibiotic interactions still lacks comprehensive mechanistic underpinnings. There are several well-studied examples that explain why drugs that target the same pathways tend to interact synergistically. For example, trimethoprim and sulfamethoxazole inhibit two steps in the tetrahydrofolic acid synthesis pathway. Their synergistic interaction has been attributed to the drugs’ ability to potentiate each other’s inhibitory effects on key metabolite synthesis within this pathway (Minato et al., 2018). The synergism between two ribosome-targeting antibiotics of the macrolide family, lankacidin and lankamycin, can be explained by their ability to target neighboring sites in the ribosome (Auerbach et al., 2010). For antibiotics that target unrelated pathways, the ability of one drug to affect the uptake or efflux of another has been proposed to be a mechanism that results in drug interactions (Cokol et al., 2011; Brochado et al., 2018). For example, the synergism between streptomycin and penicillin was proposed to be due to the cell membrane-damaging effect of penicillin, which in turn facilitates bacterial cell penetration by streptomycin (Plotz and Davis, 1962). Conversely, antagonism can emerge when one drug limits the proton motive force-dependent uptake of another drug or increases its efflux through efflux pumps, resulting in decreased intracellular concentration of the latter (Brochado et al., 2018).

The suppressive effect of spiramycin, a ribosome-targeting antibiotic, on trimethoprim has been attributed to non-optimal regulation of ribosomal genes, which leads to an imbalance between cellular DNA and protein contents in the presence of trimethoprim-induced DNA stress (Bollenbach et al., 2009). This mechanism also underlies the antagonism between ciprofloxacin and tetracycline and has been proposed to hold more generally for antibiotic pairs comprising a DNA synthesis inhibitor and a protein synthesis inhibitor (Chait et al., 2007; Bollenbach et al., 2009). More broadly, polysaccharide and ATP synthesis in E. coli are key cellular processes that shape pairwise drug interactions between six antibiotics with different modes of action; supplementing growth medium with small molecule adjuvants that target these cellular functions alters drug-drug interactions in a predictable way (Chevereau and Bollenbach, 2015).

Erythromycin shares similarities with spiramycin with respect to their inhibitory effect on translation elongation (Kirst and Sides, 1989). Both are macrolide-type ribosome-targeting bacteriostatic antibiotics that bind to the 50S large ribosomal subunit, block nascent polypeptides (Hansen et al., 2002), and trigger the cold shock response in E. coli (van Bogelen and Neidhardt, 1990). Erythromycin is expected to suppress trimethoprim. However, the interaction between erythromycin and trimethoprim was experimentally characterized as additive (Ocampo et al., 2014) or predicted to be additive against E. coli (Wood et al., 2012). What causes this apparent lack of a suppressive interaction between erythromycin and trimethoprim? To address this question, we applied an RNA-Seq transcriptomics approach to capture the effects of single and combined antibiotic treatments on gene expression in E. coli. Gene regulation responses often show conflicts when antibiotics that generate different transcriptional responses are combined. Bacteria often resolve such conflicts by prioritizing gene expression to one of the two drugs or by averaging the gene expression due to both drugs (Bollenbach and Kishony, 2011; Young and Elowitz, 2011). We explored whether this phenomenon could help us identify genetic determinants that shape the interactions between erythromycin and trimethoprim. Notably, our study revealed that when the two antibiotics are combined, the expression of sulfate reduction genes escaped the stronger overall effect of trimethoprim on the transcriptome. This unexpected transcriptomic response underlies the lack of a suppressive interaction in a sulfate-limited growth environment. In addition, we showed that the presence of an intact sigma factor-binding protein gene crl also contributes to a suppressive interaction between the two antibiotics. Overall, our findings highlight the roles of gene expression prioritization patterns of an essential metabolic pathway and a key genetic background variation in shaping drug interactions.

To characterize the interactions between trimethoprim (TMP) and erythromycin (ERY) against an E. coli MG1655 wild-type strain, we set up a two-dimensional concentration gradient of the two antibiotics in LB medium and measured the optical density (600 nm) of bacterial cultures over time using kinetic assays. All experiments were carried out in LB medium at 37°C due to the availability of existing literature that focused on the inhibitory effects of sub-lethal TMP or ERY treatments under similar experimental conditions (Chevereau and Bollenbach, 2015; Giroux et al., 2017; Chen and Wang, 2021). The maximum growth rates of bacteria were quantified during the early exponential phase. The isoboles in the contour plots represent pairwise combinations of antibiotic concentrations that resulted in the same exponential growth rates relative to the no-antibiotic treatment controls. The overall shape of the contours suggests that the interaction between TMP and ERY is almost additive (Figure 1B). We then performed the same experiment for TMP and spiramycin (SPR) and found that SPR has a clearly antagonistic, and even mildly suppressive effect on TMP (Figure 1B), which is in agreement with previous results despite the differences in experimental conditions (Bollenbach et al., 2009). Our results confirmed that ERY and SPR show different pairwise interactions with TMP under the growth conditions in this study.

To identify genes and cellular pathways that shape the interaction between TMP and ERY, we obtained the RNA-Seq profiles of exponentially growing cultures treated with single antibiotics at their half-maximal inhibitory concentrations (IC₅₀), as well as with a combination of these TMP and ERY concentrations using paired-end Illumina HiSeq (Table 1). Principal component analysis (PCA) was used to dissect how each treatment contributes to the total variance in gene expression and visualize the distribution of sample groups on a PCA score plot (Figure 2A). TMP and ERY triggered distinct transcriptomic responses and formed separate clusters. The position of the TMP + ERY combination treatment group is highly similar to that of TMP on PC1, which explains most of the variance (78%). It is mostly on PC2, which captures 16% of the variance, where the TMP + ERY group differs from the TMP group. Differential gene expression analysis relative to the no-antibiotic control group was performed (Supplementary File 1). The Venn diagram in Supplementary Figure 1A shows that the number of differentially expressed genes (Wald test: P < 0.01 after Benjamini-Hochberg correction) was much higher for TMP (2,443 genes) compared to ERY (1,008 genes). Linear regression analysis (Supplementary Figure 1B) revealed a high degree of similarity between the log₂ fold-changes of differentially expressed genes that are common to the TMP and TMP + ERY groups (coefficient of determination, R² = 0.925). Overall, the changes in transcriptomic profiles of the TMP + ERY group are driven by the stronger transcriptomic responses associated with TMP.

Next, we performed a two-way hierarchical clustering analysis to contrast the gene expression profiles of the top 100 genes with the greatest variations across all sample groups. Almost all the genes in the TMP + ERY group displayed gene expression responses that are consistent with those in the TMP group (Figure 2B). The only clear exceptions are the cysH gene and cysIJ operon from the assimilatory sulfate reduction pathway, which belongs to the sulfur metabolism gene ontology (GO) category. The expression levels of these three genes in the TMP + ERY group resemble those in the ERY and the no-drug control groups. Differential gene expression analysis indicates that TMP led to significant upregulation of cysH and cysIJ relative to the no-antibiotic control group, while the TMP + ERY and ERY groups showed expression levels that are similar to the control group (Figure 2B inset). In other words, ERY in the TMP + ERY combination treatment lowered the expression of cysH and cysIJ relative to that in the TMP-treatment group.

We then examined the wider implications of the distinct expression patterns for cysH and cysIJ. Sulfate reduction genes are known to be differentially expressed when E. coli is treated with TMP (Sangurdekar et al., 2011; Mitosch et al., 2017). The assimilatory sulfate reduction pathway (Figure 3A) scavenges extracellular sulfate and reduces intracellular sulfate to sulfide via several intermediate steps (Sekowska et al., 2000). This pathway is activated by the transcriptional activator CysB and the N-acetyl-L-serine inducer (Ostrowski and Kredich, 1989; Neidhardt, 1996). The end-product of the pathway, sulfide, depletes the N-acetyl-L-serine inducer via the O-acetyl-L-serine precursor, and thus exerts negative feedback on the expression of cys genes (Neidhardt, 1996). CysB is also negatively regulated by itself and N-acetyl-L-serine (Jagura-Burdzy and Hulanicka, 1981; Neidhardt, 1996). We hypothesized that the distinct expression patterns of cysH and cysIJ in the TMP + ERY group extends to other sulfate reduction genes, because they share the same transcriptional regulatory mechanism.

We performed a separate two-way hierarchical clustering analysis for cysB and the sulfate reduction genes of Figure 3A using row Z-scores and average log₂ fold-change relative to the no-drug control. With the exception of cysB and the periplasmic sulfate-binding protein gene sbp, the expression profile of all the cys genes in the sulfate reduction pathway followed the same trend as cysH and cysIJ (Figures 3B,C). To exclude the possibility that this was a consequence of differences in growth rates across the four sample groups, we compared the TMP (IC₅₀) and ERY (IC₅₀) groups with an additional TMP + ERY sample group with approximately 50% reduction in exponential growth rate. This was achieved by reducing TMP and ERY concentrations for this combination treatment group. Two-way hierarchical clustering analysis confirmed that the TMP (0.2 × IC₅₀) + ERY (0.6 × IC₅₀) group clustered with ERY rather than with TMP (Supplementary Figure 3), which is consistent with the trend in Figure 3B. This suggests that the sulfate reduction genes are generally restrained from responding to TMP when ERY is also present.

Treating E. coli with TMP in LB medium causes oxidative stress (Giroux et al., 2017), which is expected to compromise sulfate reduction and promote auto-oxidation of sulfide (Dolla et al., 2006). LB medium is also considered to be a sulfate-poor growth medium and contains approximately 0.1–0.15 mM of sulfate (Kertesz, 2000). The upregulation of the sulfate reduction pathway in the presence of TMP is a transcriptomic signature of sulfur limitation and decrease in intracellular sulfide availability (Neidhardt, 1996; Gyaneshwar et al., 2005). Thus, it is consistent with the negative feedback transcription regulation for this pathway that counteracts sulfur limitation (Neidhardt, 1996). Further links between oxidative stress and upregulation of the sulfate reduction pathway had previously been established by transcriptomic studies in which E. coli cultures grown in LB medium were treated with oxidizing agents such as chlorine, hydrogen peroxide, paraquat and sodium salicylate (Pomposiello et al., 2001; Wang et al., 2009). These are consistent with the TMP-induced upregulation of assimilatory sulfate reduction genes we observed. However, it is conceivable that other TMP-induced effects on one-carbon metabolism, as well as methionine and S-Adenosyl methionine (SAM) synthesis could also have contributed to the observed upregulation of the assimilatory sulfate reduction pathway.

Using the TMP versus no-drug control and TMP + ERY versus TMP pairwise comparisons, we performed gene set analysis on the assimilatory sulfate reduction pathway and neighboring nodes of the sulfur metabolism network (KEGG annotation). ERY in the combination treatment suppressed the upregulation of cys sulfate reduction genes when TMP + ERY was compared against TMP (Supplementary Figure 4B). A volcano plot for the TMP + ERY versus TMP comparison also confirmed that cys genes were amongst some of the most downregulated genes (Supplementary Figure 5). An intriguing observation that emerged from the pathway analysis was the upregulation of tauD and ssuE/ssuD, whose gene products allow taurine and alkanesulfonates, respectively, to be utilized as alternative sulfur sources under sulfate starvation (van der Ploeg et al., 1996; Eichhorn et al., 1999). The lack of an expected upregulation of the sulfate reduction pathway in the TMP + ERY group implies a transcriptional response to sulfur limitation that results in lower fitness.

To validate our key observation that ERY in the TMP + ERY combination treatment suppresses the expression of the sulfate reduction pathway, we performed quantitative real-time PCR (qRT-PCR) on E. coli MG1655 that had been treated with TMP and TMP + ERY. We focused on the PAPS reductase gene cysH, which showed one of the highest variations in expression levels across sample groups (Figure 3B). cysH expression was normalized to the average expression of two stably expressed essential genes (according to our RNA-Seq based differential gene expression analysis), rpoB (encoding the RNA polymerase β-subunit) and secA (encoding the protein translocase subunit SecA) as internal reference genes. As expected, combining ERY with TMP resulted in an approximately twofold reduction in cysH expression level relative to the TMP-treated reference group (Figure 4A). Under sulfur limitation, supplementing growth media with sulfate is expected to stimulate the expression of the sulfate reduction pathway (Pasternak et al., 1965; Wheldrake and Pasternak, 1965). To verify this, we added 2 mM sodium sulfate to the LB medium that contained TMP + ERY treated MG1655 and observed an increase in average cysH relative expression compared to that without exogenous sulfate. This increase, however, did not fully offset the suppressing effect of ERY on cysH expression.

We hypothesized that a transcriptional response to sulfur limitation that is detrimental to fitness in the TMP + ERY group shapes the drug interaction between TMP and ERY. To test this hypothesis, we supplemented LB growth medium with 2 mM sodium sulfate and performed the same growth rate assays as before in TMP + ERY two-dimensional concentration space. Notably, ERY exerts a suppressive effect on TMP in the presence of additional exogenous sulfate, especially at higher TMP concentrations (Figure 4B). The disproportionately large beneficial effect of sulfate on normalized growth rate in the presence of both drugs altered the shape of the isoboles in the 2D drug concentration space (Supplementary Figure 6B).

To expand our characterization of TMP + ERY interactions to another key reference strain of E. coli, we performed additional growth rate assays using the BW25113 strain (Grenier et al., 2014). Although both BW25113 and MG1655 were derived from the E. coli K12 ancestral strain, we observed that ERY exerted a pronounced suppressive effect on TMP in the two-dimensional contour plot for BW25113 (Figure 5A). A major difference between BW25113 and the crl-deficient MG1655 strain used in this study is that the former carries an intact crl gene. Crl is a sigma factor-binding protein and transcription activator that exerts global regulatory effects on the transcriptome by increasing the affinity of RNA polymerase for the principal sigma factor of the stationary phase σ^S (Typas et al., 2007; Cavaliere and Norel, 2016; Cartagena et al., 2019; Xu et al., 2019). The 86 genes that are controlled by Crl in E. coli K-12 form 40% of the RpoS regulon (Santos-Zavaleta et al., 2019). The loss of crl can be detected in E. coli strains under laboratory settings that favored the inactivation of the general stress response (Faure et al., 2004), as well as in naturally-occurring pathogenic strains of E. coli (Provence and Curtiss, 1992; Asadi et al., 2018). We repeated the growth rate experiment using the BW25113 Δcrl knockout mutant from the Keio collection to mimic the effects of crl-inactivation in the MG1655 strain we used. As expected, deleting crl abolished the suppressive drug interaction previously observed (Figure 5B), yielding an antagonistic interaction. Deleting crl substantially altered the susceptibility of BW25113 to TMP. The IC₅₀ for TMP approximately doubled from 0.10 μg/mL to 0.21 μg/mL, while IC₅₀ for ERY only decreased marginally from 48 μg/mL to 42 μg/mL. On the vertical axes of the contour plots, it is evident that the change in shapes of isoboles was mainly driven by the increase in tolerance to TMP in the absence of crl (Figure 5). The remaining difference between TMP + ERY interactions in the MG1655 (additive; Figure 1B) and BW25113 Δcrl strains (antagonistic; Figure 5B) is most likely to be due to other variations in their genomes.

Applying a transcriptomic approach to identify genetic determinants that affect the interaction between two antibiotics with distinct modes of action is challenging, because the transcriptomic signatures of individual antibiotic treatments are superimposed. We used the interactions between TMP and ERY as a model system, because ERY and SPR show different interactions with TMP (Figure 1B) despite their similarities as ribosome-targeting antibiotics. TMP inhibits folic acid metabolism, depletes key metabolites and induces DNA damage (Sangurdekar et al., 2011; Giroux et al., 2017). Activating global stress responses such as the SOS response and the ppGpp-mediated stringent response has wide-ranging effects on the transcriptome (Courcelle et al., 2001; Khil and Camerini-Otero, 2002; Durfee et al., 2008; Traxler et al., 2008). Despite this, the transcriptomic response of the sulfate reduction pathway in the TMP + ERY group prioritized to that of ERY rather than to TMP (Figure 3 and Supplementary Figure 3). By supplementing the culture medium with exogenous sulfate and perturbing the affected pathway (Figure 4A), we successfully changed the additive interaction between TMP and ERY to a suppressive interaction (Figure 4B). These results demonstrate that analyzing prioritization patterns of conflicting transcriptomic responses can be a useful tool for dissecting drug interaction mechanisms.

Given the disproportionately large beneficial effect of sulfate addition on relative exponential growth rates when the two drugs were combined, we argue that the suppressing effect of cys upregulation by ERY is a transcriptional response that does not maximize fitness. The sulfate reduction pathway is a subset of sulfur metabolism, which is subject to strong negative feedback control, because L-cysteine is one of the most toxic amino acids at high concentrations (Datta, 1967; Avcilar-Kucukgoze et al., 2016). In addition to the ability of sulfide to deplete O-acetyl-L-serine (precursor to the N-acetyl-L-serine inducer), negative feedback is also mediated by the inhibitory effect of L-cysteine on the conversion of serine to O-acetyl-L-serine (Neidhardt, 1996). TMP upregulated cys genes of the sulfate reduction pathway and slightly downregulated cysB (Figures 3B,C), and these changes are consistent with the transcriptional regulatory mechanism of this pathway in response to sulfur limitation. The sulfate binding protein gene sbp is the only sulfate reduction gene that showed similar expression levels in both the TMP and TMP + ERY groups (Figure 4B), but sbp is known to be activated more specifically by sulfate starvation (van der Ploeg et al., 1996). Given how tightly sulfur metabolism is regulated, it is surprising that the cys genes displayed an unusual transcriptional response in TMP + ERY (Figures 3B, 4A). As a ribosome-targeting antibiotic, ERY is not expected to directly affect the intra- and extracellular availability of metabolites for sulfur assimilation. More metabolomic insights into the regulation of the cys genes in response to combination treatment may reveal why TMP + ERY treated cells displayed a transcriptional response that is unlikely to be effective in counteracting the effects of TMP in terms of oxidative stress and sulfate limitation.

We also demonstrated that the sigma factor-binding protein gene crl plays a significant role in determining the interaction between TMP and ERY against E. coli. This finding lends support to the growing recognition that interactions between the same combinations of antibiotics can sometimes differ between bacterial species, as well as between different strains of the same species (Brochado et al., 2018). For example, the interaction between nalidixic acid and tetracycline against several multidrug resistant clones of Acinetobacter baumannii and E. coli is synergistic, while their interactions in susceptible clones of the same species are not (Gaurav et al., 2021). A previous study showed that drug interactions between antibiotics representing six main modes of action are surprisingly robust against most single gene deletions in E. coli, and that only a very small minority of single gene knockout strains produced qualitative changes in drug interactions (Chevereau and Bollenbach, 2015). Therefore, crl represents a rare example of single gene deletions that caused such shifts in drug interactions. While it is not clear if there are any interactions between the Crl-RpoS regulon and the assimilatory sulfate reduction pathway, our findings hint at the possibility that the effects of specific gene expression patterns on drug interactions may be contingent on the presence or absence of other genes such as transcriptional regulators. Additional future efforts could also be directed toward systematic discovery of specific genetic contexts in which alterations in drug interactions arise due to variations between strains, including the gain of antimicrobial resistance genes and loss of transcriptional regulators.

In conclusion, drug interaction is a complex phenotype that is determined by antibiotic modes of action (Cokol et al., 2011; Ocampo et al., 2014; Chevereau and Bollenbach, 2015), gene expression patterns (Bollenbach et al., 2009; Bollenbach and Kishony, 2011), genetic variations across species and strains (Chevereau and Bollenbach, 2015; Brochado et al., 2018; Gaurav et al., 2021). Our study exemplifies that metabolite abundance in the growth environment and gene prioritization patterns can also affect drug-drug interactions. It is hoped that our work will pave the way for future studies that employ a combination of transcriptomic and complementary techniques to dissect drug-drug interactions.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found at: https://doi.org/10.5061/dryad.bk3j9kdcn.

QQ and TB conceived the study. SAA and QQ designed, performed, and analyzed the experiments. QQ performed the bioinformatic analyses. TB supervised and managed the project. All authors contributed to the original draft and edited subsequent versions of the manuscript.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.