E. coli transcription factors regulate promoter activity by a universal, homeostatic mechanism

Abstract

Transcription factors (TFs) may activate or repress gene expression through an interplay of different mechanisms, including RNA polymerase (RNAP) recruitment, exclusion, and initiation. However, TF function can vary depending on the identity of the regulated promoter, and the principles behind this context-dependence remain unclear. We demonstrate a relationship between the basal activity of a promoter and its regulation by a specific TF. Specifically, fold-change in expression scales inversely with basal activity: activation is weaker and repression is stronger on stronger promoters. This inverse scaling applies to both activators and repressors, suggesting a common underlying mechanism where TFs regulate expression by stabilizing RNAP binding at the promoter. The consequence of this relationship is that TFs buffer expression by affecting constant regulated expression levels across promoters of different basal activity, ensuring homeostatic control despite genetic or environmental changes.

Transcription factors (TFs) are crucial determinants of gene regulation, functioning through a myriad of regulatory mechanisms to ensure precise control of cellular processes. TFs can bind to specific DNA sequences, often located near the genes they regulate, and either promote or inhibit transcription. The regulatory mechanisms employed by TFs are diverse and complex, and regulation may alter the rate of one or more steps in the multi-step process between RNA polymerase (RNAP) binding and promoter clearance (1–4). The complexity of predicting the regulatory function of a specific TF in various genetic and physiological contexts arises from the propensity of TFs to regulate multiple steps of the transcription process, combined with the intricate interplay of the number and types of TFs, their binding strengths, binding site locations, and the promoter strength (5–8). In particular, TFs are usually classified based on their net regulatory function (activation or repression) rather than their mechanisms of regulation. Due to this classification scheme, it can seem surprising that same TF can have both activating and repressing interactions with different promoters, even in very similar contexts. Examples of such have been seen in both prokaryotic (9–11) and eukaryotic systems (12).

Here we measure the relationship between TFs and promoters in a controlled way in E. coli. We systematically alter constitutive expression levels of a promoter through several methods including perturbations to basal promoter sequence, as well as by perturbing physiological conditions such as growth media or availability of polymerase. Each of these perturbations alters the constitutive expression rate of the promoter while holding other TF-related features (such as TF identity, binding site position, and sequence) constant, thereby allowing us to confidently measure the relationship between TF function and promoter identity.

General gene regulation model to interpret TF-promoter relationship

To interpret the relationship between regulation and promoter strength, we use a simple thermodynamic model of gene regulation that we and others have proposed previously (13–18). In this model, regulation by a single TF is coarse-grained into activity on two steps of the transcription process (Fig. 1A). The first mechanism of regulation affects RNAP recruitment and stability at the promoter and is parameterized by β (which is confined to be a positive, real number); β > 1 corresponds to TFs with positive stabilizing or recruiting interactions with polymerase and β < 1 corresponds to negative interactions with polymerase through effects such as steric hindrance. The second mechanism alters transcription initiation and promoter clearance and is parameterized by α (again defined as a positive, real number). Once again, α > 1 corresponds to TFs that increase the rate of transcription initiation, while α < 1 corresponds to TFs that inhibit or slow the rate of this process. As such, parameter values > 1 represent activation of one step of the transcription process and values < 1 represent repression of that step. Historically, in vitro approaches studying TF function have tried to interpret which steps of the transcription process are regulated by a TF. A common model for the function of proximally-binding repressors is that of steric-hindrance where TF-binding occludes the binding of RNAP to the promoter (β < 1 in our model). Various in vitro studies have identified steric hindrance as one of the possible mechanisms of action for common repressors (19, 20), including LacI (21, 22). On the other hand, activation is often ascribed to positive “recruiting” interactions between specific domains of the TF and different subunits of the RNAP (so-called type I and II activation) (23).

Figure 1: Modeling promoter dependence of TF function.

(A) We use a general thermodynamic model of TF function that considers TF function through two distinct mechanisms: stabilization/destabilization and acceleration/deceleration. (B) The predicted expression from two conditions: constitutive (no TFs) and regulated (saturating TFs). (C) Fold-change at saturating TF concentration, FC, from this model, predicts a dependence on the constitutive expression level of the regulated promoter, C. (D) Predictions for FC vs constitutive expression level of the promoter, C for stabilizing TFs (β> 1) for a range of α. In all cases, FC decreases with promoter strength. (E) Predictions for destabilizing TFs (β < 1) show the opposite trend for any value of α.

We make no assumption a priori, about which step or steps the TF will regulate; instead, the general model introduced here allows interpretation of TF function in vivo through the two mechanisms introduced above: stabilization/destabilzation and acceleration/deceleration (12, 14, 16, 24). The full, general thermodynamic model is derived in SI section “Derivation of the general thermodynamic model” (Fig. S1). However, the model can be simplified by considering saturating TF concentrations, so that TF-unbound states in the regulated condition can be ignored, the thermodynamic model makes a simple prediction for both the constitutive expression levels (C) and regulated expression levels (R), Fig. 1B. The relationship between the fold-change in gene expression at a saturating TF concentration, FC, and the expression level of the constitutive promoter, C is then shown in Fig. 1C.

The qualitative nature of the relationship between fold-change and constitutive expression level of the unregulated promoter is determined by whether or not the TF is stabilizing (β > 1) or destabilizing (β < 1). If a TF is stabilizing (TF-RNAP interactions are favorable, β > 1), a TF will enact lower fold-change on stronger promoters (Fig 1D). That is to say that repressors will repress more (red line, Fig 1D), and activators will activate less on strong promoters (black line, Fig 1D) compared to weaker ones. The opposite relationship is predicted if the interactions are destabilizing (TF-RNAP interactions are unfavorable, β < 1), Fig 1E. A consequence of our model, which considers two mechanisms of regulation, a destabilizing TF can still activate, and a stabilizing TF can repress, contingent on the TFs role on the other step of transcription. Furthermore, the “role” (activation or repression) can be contingent on the strength of the promoter (see green line in Fig. 1D and E).

Some conserved properties of these fold-change curves can be seen by rescaling how we plot the theory. First, we normalize the fold-change, FC, by the fold-change in the weak promoter limit, FC(C → 0) = αβ. Fig. 2A shows the relationship between this normalized fold-change metric and constitutive expression; stabilizers decrease fold-change, while destabilizers increase fold-change with increasing constitutive promoter activity. Second, Fig. 2B shows that if we also rescale the constitutive promoter activity, C, to |β – 1|C/C_max, all data is expected to collapse to one of two possible behaviors differentiated entirely by the nature of the TFs interactions with RNAP; i.e. if the TF is a stabilizer or destabilizer. Within this figure, we can define three regimes: weak regulation, strong destabilization and strong stabilization. For weak regulation, where |β – 1|C/C_max ≪ 1, the fold-change is insensitive to constitutive expression level of the promoter and thus any differences in constitutive expression level are expected to persist through to the regulated expression level such that the fold-change remains constant (top panel, Fig. 2C). The strong stabilization regime occurs when (β – 1)C/C_max > 1 and is possible only for stabilizers (β > 1), this corresponds to a regime where fold-change scales with the inverse of constitutive expression level (C⁻¹). This produces a behavior, shown schematically in the middle panel of Fig. 2C, where stabilizing TFs will elicit the same regulated levels of expression independent of the constitutive expression of the promoter; if the constitutive expression of the promoter doubles, the regulated level will remain unchanged. Lastly, for strongly destabilizing TFs ((β – 1)C/C_max → ߝ 1), we expect TFs to drive divergent levels of regulated expression from promoters with relatively small differences in constitutive levels (bottom panel, Fig. 2C). This behavior is expected strictly for destabilizers. Although we have focused on saturating TF concentrations to this point in our analysis, even at sub-saturating concentrations we expect the same qualitative scaling behavior but with different scaling parameters that also depend on TF concentration and TF binding affinity (see SI section “Data collapse for non-saturating TF concentration” and Fig. S2).

Figure 2: Properties of the relationship between TF function and promoter strength.

(A) Fold-change, when normalized by the fold-change of the weakest promoter, (FC/FC(C → 0)) bifurcates into two behaviors based on the nature of TF-RNAP interactions (β > 1 or β < 1). (B) Renormalizing the constitutive promoter strength by (β – 1) collapses all possible curves into only two distinct curves. These curves show different behaviors in three distinct regimes (C) A schematic demonstrating the behavior of each regime in (b). The “strong stabilization regime” has a feature where differences in constitutive expression of the promoter are buffered by TF regulatory activity. (D) Example data for the TF CpxR on three natural promoters, regulated at similar relative binding positions on the promoter. The fold-change is qualitatively different depending on which promoter is regulated by the TF.

As an example of the complex dependence between TF regulatory function and promoter identity in endogenous promoters, consider the TF CpxR which regulates more than 40 different promoters in E. coli (25). Fig. 2D shows data for the fold-change as a function of CpxR concentration for three CpxR-regulated promoters (ldtCp, yccAp, and efeUp). These promoters have CpxR binding sites at approximately the same position relative to their transcription start site (TSS). However, two of these promoters are activated by CpxR (ldtCp and yccAp) while the other (efeUp) is repressed. Although some context-specific details (such as TSS, 5′ untranslated region, ribosome binding site, etc.) are different between these three promoters, one notable difference is that each core promoter gives rise to significantly different constitutive (unregulated) expression level. The inset to Fig. 2D shows the measured expression of each promoter in a CpxR knockout strain. CpxR activates the two weak promoters and represses the stronger promoter, which has 100–fold higher constitutive expression. Previous studies have suggested that CpxR acts primarily through positive, stabilizing interactions with RNAP (β > 1) (10), which is qualitatively consistent with the trend seen in our data in Fig. 2D and with our prediction that a TF can switch from activating to repressing gene expression dependent on the core promoter strength. We do not require that the intrinsic function of the TF changes or is “context specific”, it is a basic expectation of our model. However, there are many uncontrolled aspects of this measurement; the TF binds to slightly different positions on the promoter, the binding site sequences are different, and there may be differences in other regulatory factors involved at each endogenous promoter. Hence, we sought to measure the dependence of regulatory function on promoter strength in a system that controls for these contextual confounds.

Measuring the relationship between TF function and promoter identity on a synthetic promoter library

We chose eight different TFs identified from a previous study (16) to examine the relationship between their regulatory effect and the activity of the regulated promoter. These TFs were chosen due to clear evidence of regulatory interactions on a synthetic promoter cassette which contained only a single binding site for the TF (16) (details in Tables S1 and S2). An overview of the experimental method is outlined in Fig. 3A. To create a library of promoters with a spectrum of constitutive strengths, we mutated the −35 region of the promoter (153 possible combinations of single and double mutations in the −35 regions Fig. S15) and randomly sampled 96 different clones. This library typically showed a range of constitutive expression levels varying from 100 to 1000 fold relative to the weakest promoter. We measured the expression of each promoter variant both in our library of inducible TF-mCherry strain (24) at maximum induction conditions (TF⁺⁺), and in a strain where the TF has been deleted (TF⁻). Ideally, this will provide saturating TF concentrations (i.e. where increased TF concentration does not change fold-change), however in some cases TFs do not reach saturation (Fig. S4). We expect this to change our quantitative interpretation of the scaling relationship but without altering the overarching conclusions (see “Data collapse for non-saturating TF concentration in the thermodynamic model”); most important is that TF concentration is the same for all promoters. Unless otherwise stated, all TFs used in this study are fused with mCherry and expressed from aTC inducible promoter (we do not find the mCherry tag perturbs function, see SI section “Influence of mCherry tag on TF function”). We then plot the measured fold-change of each promoter against the constitutive expression level of that promoter (Fig. 3B).

Figure 3: Measuring the relationship between TF function and promoter activity for synthetic promoter variants.

(A) Experimental approach to measure FC and constitutive expression levels of random promoter variants. (B) Each data point in our plots represents the constitutive and regulated level of expression for one promoter variant. (C) For 8 TFs measured here, the relationship between TF function and promoter strength conforms well to the predicted scaling of strong stabilizing interactions (black dashed lines). Subplots (i)-(iv) show TFs binding downstream of the promoter, while (v)-(viii) show TFs binding upstream of the promoter. Furthermore, panels (i)-(v) show TFs considered to be repressors and (vi)-(viii) show activators. (D) As expected from the theory for stabilizing TFs, the regulation by both CpxR and LacI cause a robust level of regulated expression from promoters with disparate constitutive levels of expression. Each color is a unique promoter variant for regulated (solid lines) and constitutive expression (dashed lines).

Fig. 3C shows the data for the eight TFs: five repressors (LacI, MngR, PdhR, AscG, and AcrR) and three activators (CpxR, MetR, and SoxS). The top row of regulatory interactions is for TFs with binding sites downstream of the promoter (centered at positions ranging between +11 to +16.5 from the TSS) and the bottom row features regulatory interactions from TFs with binding sites upstream of the promoter (centered at positions between −52.5 to −64 from the TSS); see Table S1 for details. The dashed line in each plot is a fit to the theory in Fig. 1C with αβ and β as fit parameters. The maximum promoter activity, C_max is set equal to the level of the promoter with the maximum expression in each dataset (range shown in Fig. S6A). By setting C_max this way, we establish a lower bound on the maximum possible expression. Intuitively, if C_max is increased, the value of β increases proportionally (see Fig. S6B,C). Each of these eight TFs shows the predicted scaling relationship for stabilizing TFs independent of function (activation or repression) or binding location on the promoter. However, for two activators, MetR and SoxS, the data for the weakest promoters fall into the regime where FC is constant with promoter strength consistent with weaker stabilizing interactions for those TFs as measured by lower values of β from the fit.

A common model of repression used by us and others in the past, is regulation purely through negative interactions with RNAP (steric hindrance) (6, 10, 26, 27). In our general model used here, this corresponds to setting β < 1 and α = 1. In this case, the model specifies that the fold-change at weak promoters will equal β and the fold-change of strong promoters will increase and eventually will be equal to 1 for the strongest promoters; our data shows that neither of these holds true. This goes against the steric hindrance model which fails to describe every repressive dataset in our study (Fig. 3C (i)–(v)). Furthermore, a purely decelerating mechanism (β = 1, α < 1) also fails to explain our data as we would expect a flat relationship between FC and C. In our model, β > 1 is a requirement to capture the observed inverse scaling of fold-change with promoter strength. Therefore, we conclude that there is a common positive interaction between TF binding and RNAP availability at the promoter in every regulatory interaction measured here and that the net regulatory effect is determined by the magnitude of impact on the second step, transcription initiation (α in our model). This positive, stabilizing interaction may arise through direct interactions with RNAP, indirect mechanisms such as changing the DNA’s local propensity to bind RNAP, or through transient rearrangements of TF binding state that are permissive to expression (28). In Fig. S9 we summarize many previous studies on LacI function specifically, and the implications those results have when interpreted through our model. In the SI (section “Data interpretation in the kinetic model framework for repressors” and Fig. S7 and Fig. S8), we outline implications of our data in the kinetic framework, although interpretation becomes more complex, our central findings are unchanged.

An important consequence of the stabilizing relationship seen for most TFs tested is that as constitutive expression levels change, the regulated levels remain constant for TFs with relatively high values for β. This is demonstrated in Fig. 3D for two TFs, LacI and CpxR. We plot the single-cell distribution of constitutive (dashed lines) and regulated (solid lines) expression for several promoters sampled from the a range of constitutive expression levels in the library where each color represents a unique promoter sequence. The constitutive expression levels vary roughly 200-fold over these promoters. However, the corresponding measurement of the regulated expression levels (solid lines) vary by only roughly 2-fold. Thus, due to this inverse scaling relationship between TF function and promoter activity, the TF acts to buffer changes to expression from induced mutations in the promoter for both an activator and a repressor.

Measuring the relationship between TF regulation and promoter strength in different physiological conditions

The effect of physiological perturbations on constitutive expression has been studied using simple models that demonstrate the scaling between growth rate and constitutive expression (29, 30). Despite strong coupling between growth rates and transcription regulation little is known about how different TF functions are influenced by physiological perturbations (30). We measured the relationship between TF function and promoter strength by inducing changes to constitutive expression levels through physiological perturbation of growth rates using an array of carbon sources. We chose two TFs to examine: LacI and CpxR. We measured each promoter mutant library in six different minimal media conditions supplemented with a range of carbon sources that yielded doubling times between 55 (glucose) and 230 minutes (acetate). We expect significant changes to several global regulators including the RNAP concentration between these media conditions that impact the basal strength of all promoters in the library. However, the TF concentration is also perturbed by these growth conditions; most notably the TF concentration in the cell increases significantly in the slowest conditions (pyruvate and acetate). To ensure the validity of our hypothesis testing, it is essential that fold-change is measured either at saturating TF concentrations, where fold-change will be insensitive to the changes in TF concentrations, or in the case of sub-saturating TF concentrations, the level should be held constant across each media condition. We discuss the nature of data collapse in sub-saturating TF concentrations in SI section “Data collapse for non-saturating TF concentration”. For all media conditions included here, we find that the changes in TF level due to media do not perturb LacI concentration below the expected saturating level for our binding site affinity (i.e. the differences in TF concentration will not alter fold-change, cyan data points in Fig. S4A). However, CpxR is at sub-saturating concentrations throughout our range and thus we expect regulation to be sensitive to TF concentration (brown points in Fig. S4A). Thus, for CpxR, we attempted to equalize the number of TFs in different growth conditions by altering inducer conditions such that each had comparable mCherry signal. We chose to exclude the two slowest growth media (pyruvate and acetate) where the TF concentration was significantly higher than for the other conditions. Including those data points (Fig. S5B) shows the same scaling relationship but with a higher y-intercept, as expected for increasing TF concentration in non-saturating conditions (see SI section “Data collapse for non-saturating TF concentration” and Figs. S2,S3).

Perturbing the growth rate changes the fold-change from the promoter library, with slower-growth rate systematically showing lower fold-changes, however, the scaling relationship is preserved across the entire dataset for both TFs (Fig. 4A). Fig. 4B shows the distribution of slopes from a straight-line fit of our theory for FC vs constitutive expression level for each promoter across the measured conditions for both CpxR and LacI. Most promoters have a scaling relationship with a slope of −1 (see example in Fig. 4B inset), consistent with what is expected from strong stabilization. This scaling is seen across individual promoters with different growth rates as well as across the collection of promoters at a single growth rate and this relationship is invariant between genetic or physiological changes for both activator and repressor. In Fig. 4C we show the single-cell expression distribution for a single promoter across various growth rates. Once again, although the constitutive expression level of the promoter depends strongly on growth rate, regulation is largely capable of buffering these changes where the stable level of expression is set by the identity of the TF regulating it; CpxR collapses the expression from the promoter into a higher expression state while regulation by LacI also collapses the data but in a repressed state. Once again, by global perturbation to the constitutive expression through physiological changes, we find the same fundamental mechanisms of regulation that “restore” the regulated state expression level.

Figure 4: Scaling of regulation is conserved across different methods of perturbations to constitutive expression.

(A) Constitutive expression level altered via changes in the growth rate. The plot shows fold-change against the constitutive level of expression for the promoter mutants regulated by LacI (circles) or CpxR (squares) when grown in media containing a variety of carbon sources (different colors). (B) Distribution of slope obtained from a linear fit of FC as a function of constitutive expression for any single promoter in different media. Inset shows the straight-line fit to the data for one of the promoters grown in 6 different carbon sources. Most promoters show the expected relationship with slope −1. (C) Representative single-cell distribution of the fluorescence of unregulated (dotted lines) and regulated expression (dashed lines) across different growth media. Although this changes the constitutive expression levels of the promoter library, the regulated expression levels both for activation (CpxR) and repression (LacI) are relatively unchanged. (D) Perturbing constitutive expression via perturbation to RNAP concentration (through changes to σ²⁸ concentration) or promoter sequence. Measurement of regulation by (i) LacI and (ii) CpxR for specific perturbations. Each color on the plot represents a different promoter and each data point represents a fixed concentration of σ²⁸. (E) Single-cell distribution of regulated and constitutive expression for changing σ²⁸ concentration (i, different colors represent different vanillic acid concentration), or promoter strength (ii, different color represent unique promoters as in (d) at 0.625μM vanillic acid). Perturbations in constitutive expression (dashed lines) produce relatively little change in the regulated expression level (solid lines).

Promoter-TF relationship is predictive for regulation of alternative sigma factor promoters

We next designed a system where we could systematically perturb only the constitutive expression levels. Ideally, we want to change expression levels of our promoter by altering the availability of RNAP, however, limiting any subunit of the core σ⁷⁰-RNAP holoenzyme induces extreme physiological changes (29, 31, 32), making it hard to isolate the regulation of just our one target gene. As such, we designed a system to control the availability of an alternative sigma factor protein, σ²⁸ by expressing it under an inducible system. The advantage of this system is that the σ²⁸-RNAP holoenzyme recognizes an orthogonal promoter sequence to the housekeeping σ⁷⁰, which is the primary sigma factor in E. coli (33, 34). The promoters recognized by σ²⁸ are involved in flagellar synthesis (35–37) and the mechanisms of control (anti-sigma factors) are known (38, 39). In this system, we have direct orthogonal control over both the physiological (via σ²⁸-RNAP concentration) and the genetic component (via promoter sequence) of the constitutive promoter activity.

We designed a system where the expression of σ²⁸ from the FliA gene can be induced with vanillic acid, and the endogenous copy of the gene is knocked out. We then swapped our σ⁷⁰ specific promoter used previously with a promoter specific for σ²⁸. We see that increasing the concentration of vanillic acid results in higher expression from σ²⁸ promoters due to the elevated levels of active σ²⁸-RNAP holoenzyme. The dashed lines in Fig. 4E(i) show the achievable range in our inducible system for one promoter with different vanillic acid concentrations. To pair with this control of σ²⁸, we designed five promoters that span our measurable range of expression levels. The dashed lines in Fig. 4E(ii) show the distribution of expression for single cells from each promoter at intermediate vanillic acid concentrations. As a final means of expanding our range of expression, we make these measurements in a strain with the endogenous anti-σ²⁸ factor FlgM (which competitively binds to free σ²⁸ and thus lowers available σ²⁸-RNAP holoenzyme availability (40)) as well as in a strain with FlgM knocked out.

Fig. 4D shows these measurements for regulation by a repressor, LacI(i), and an activator, CpxR(ii). Each promoter (represented by different colored points) is measured at eight different σ²⁸ induction levels both with and without the endogenous expression of the anti-sigma factor, FlgM. The result is a constitutive expression range over more than 3 orders of magnitude. The fold-change scales over this range as expected for a single set of α and β parameters for each regardless of whether genetic or physiological perturbations changed the constitutive expression level. In Fig. 4E(i–ii), we show once again how this data collapses when regulated by LacI for both physiological (top panel) and genetic (bottom panel) perturbations. Different CpxR promoters do not align perfectly with theoretical predictions; however, within each promoter type, the inverse scaling relationship is preserved. The absence of a complete data collapse for CpxR may occur due to sub-saturating TF concentration or limitation to the σ²⁸ availability.

Measuring the relationship between TF function and promoter identity in natural promoters

To this point, we have observed a specific scaling relationship between TF function and constitutive promoter strength that is pervasive in simple synthetic promoters designed to be regulated only by a specific TF. Here we extend this concept to determine the applicability of this relationship in naturally occurring promoters with complex regulatory architectures such as regulation by DNA looping, multiple binding sites for the same TF, or interfering binding sites by other TFs.

In Fig. 5A, we show measurements of several endogenous promoters regulated by the TF, SoxS (MarA and Rob, the other isoforms of SoxS are deleted from the strain to avoid any cross-talk): poxBp (yellow), decRp (blue), and fldAp (red) (41). These promoters were chosen because their endogenous binding sites are located at a similar position relative to the promoter (Fig. S10A). We created 96 random mutants of each promoter using the same strategy as described above for the synthetic promoters. Unlike the synthetic promoters that have minimal changes between different TFs, the natural promoters differ in every aspect including the 5′ UTR region, ribosomal binding site (RBS) affinity, and transcriptional start site. These differences in translation efficiency and mRNA stability make comparing constitutive expression levels across different promoters challenging. To normalize for these differences, we measured the relative mRNA and protein expression levels between each of the 3 promoters and corrected the constitutive expression by normalizing C by the ratio of mRNA to protein expression levels of each promoter (see Fig. S10B). Correcting for these features helps reduce the non-transcriptional differences across the three promoters to compare their transcriptional activity directly. The effect of this multiplicative correction factor amounts to translation of the data horizontally along the x-axis and does not influence the scaling of any one dataset.

Figure 5: Promoters with complex regulatory architecture show stabilizing relationship with mutant promoter library.

(A) The top panel shows three native promoters of E. coli regulated by SoxS. The relative position of the binding site is shown as a complete green square with the center of the binding site marked underneath. The bottom panel shows the plot of fold-change against the corrected constitutive expression level for the three native promoter mutant libraries. (B) The top panel shows a schematic representation of the DNA looping mediated regulation by LacI. The bottom panel shows the massively parallel reporter assay data from (44) plotted using un-induced/induced as a proxy for the fold-change and induced expression as the constitutive expression. Data corresponding to the strong proximal binding sites (LacO1(i) and LacOsym(ii)) and all 10 different distal binding sites are plotted. (C) Measure of regulation for a promoter architecture with two binding sites for the TF of interest (dark blue squares) around the promoter (red square). (D) Measure of regulation for promoters with binding sites for other TFs (black squares) in addition to the binding site for the TF of interest.

The collection of promoter mutants shows regulation of up to 70-fold activation down to 3-fold repression between the three promoters. Each promoter shows a good agreement with the predicted scaling relationship of stabilizing interactions. The relationship is nearly universal between the three promoters across the entire range of fold-change, however, the fold-change values for the fldAp promoter are systematically 4–fold higher. This may be due to slightly stronger regulatory interactions from SoxS at −61.5 compared to the other two positions (16) or perhaps due to slight differences in binding site affinity of SoxS across the three unique binding sites. While measuring binding affinity is beyond the scope of this paper, we assessed the proximity of the annotated binding sites on decRp, poxBp, and fldAp to the SoxS consensus binding site using the TF PSSM browser in RegulonDB (41). This tool conducts a MEME analysis and generates p-values for the annotated binding sequences. Lower p-values signify sequences that closely resemble the consensus, whereas higher p-values indicate a greater similarity to a random sequence. We expect the SoxS binding site of decRp and poxBp to be of similar strength (p-values of 1.14 × 10⁻³ and 2.74 × 10⁻³ respectively), while the site in fldAp is significantly closer to the consensus (p-value 1.76 × 10⁻⁴) (42). Converting distance from consensus sequence to affinity is not straightforward, but this qualitatively supports the cause of the shift in fldAp data. Furthermore, we highlight that SoxS can switch roles from activation to repression just by changing the strength of the core promoter. It is noteworthy that SoxS has been previously thought to be involved in stabilizing RNAP by acting as a co-sigma factor (43), consistent with our data. However, the ability to repress strong promoters implies that it must also repress the initiation or a further downstream step of transcription.

Next, we explore the common natural regulatory architecture of DNA looping, where a TF binds to two sites on the promoter simultaneously to repress expression. Yu et al. (44) examined regulation by LacI with any one of ten different LacI binding sequences at both the distal and proximal binding location using a collection of promoter sequences. In Fig. 5B we show the data with the two strongest sites (O1 and Osym) at the proximal position and any of the 10 different sequences at the distal position. The spacing between these two LacI binding sites is set to mimic the natural distance between O1 and O3 in the lac operon. Since the choice of the distal site will influence the level of regulation and thus the parameters of the model, we fit each data set to the theory and plot all of the data according to the rescaled axes suggested in Fig. 1C. As can be seen, the data from every examined looping architecture collapses to a single curve and obeys the expected scaling relationship observed for stabilizing interactions. This same relationship holds even for the weaker proximal sites, although the regulation becomes very weak (Fig. S11). Interestingly, one trend that can be seen in this data as predicted by our theory, the stronger the proximal site, the more the data skew to the “strong stabilization” part of the universal plot that scales with the inverse of the constitutive expression (Fig. S11).

We further examined other TFs acting on endogenous promoters with more complex regulatory features where we have no insight into the mechanisms of regulation. Fig. 5C shows two situations where regulation of the endogenous promoter occurs through two binding sites: (i) mngRp regulated by MngR and (ii) agaRp regulated by AgaR. Although naturally these promoters are wired for feedback, since we have deleted the endogenous TF gene, our measurements isolate the regulatory role for fixed, saturating TF concentration. Furthermore, Fig. 5D shows two other natural promoters regulated by TFs we have previously studied (UlaR, GntR (16)). These promoters have two binding sites for the controlled TF (UlaR or GntR) as well as additional TF binding sites; ulaGp features IHF regulation at an unknown location and gntTp features a CRP binding site. Again, in all cases, we find the same scaling relationship between fold-change and constitutive promoter strength. Interestingly, these promoters which are regulated by other TFs, show decreased response to promoter perturbations compared to synthetic promoters designed to be regulated by only a single TF. This is in-line with the idea that TFs are capable of stabilizing such genetic perturbations.

All data collapses to the same scaling law

To demonstrate the universal nature of this TF-promoter relationship in our measurements, in Fig. 6A we collapse all the data from this study onto the same plot by rescaling the constitutive expression of each data set by (β – 1)/C_max and the fold-change by 1/(αβ) akin to our theory plot in Fig. 2B. The data shows a strong, universal collapse to the stabilizing TF theory predictions. Furthermore, we add previously published data sets: from the activator AraC (45), from LacI looping (44), from the activator CRP (46), and from in vitro measurements of the TF, CarD from the bacterium Mycobacterium tuberculosis (47) (raw data shown in Figs. S11, S12, and S13). Strikingly, across nearly 6 orders of magnitude in the fold-change split between activation and repression (see inset to Fig. 6A), the relationship between TF function and constitutive promoter strength collapses to a single functional form. This includes data from genetic perturbations, physiological perturbations, and in vitro measurements. In all cases, the data conforms to a picture of TF regulatory function that has a conserved regulatory mechanism.

Figure 6: Universal collapse for all regulation data.

(A) All data from Figs. 3–5 are renormalized using the α and β obtained by fit to the theory in Fig. 1c. We also include data from refs. (44,45,47), which we fit the same way. The black line represents the zero-parameter theory line: f(x) = 1/(1+x). All data collapses to a single theory curve suggesting a conserved universal mechanism of action between all measured regulation. The inset shows a histogram of FC values for all points included in the figure. (B) Scatter plot demonstrating the “buffering” of expression for each of our data sets. The dynamic range of expression in the library, defined as the ratio of expression level between the highest and lowest expression member of the library, for regulated expression against constitutive expression. Blue points represent the range across all promoters, and red points correspond to the range of promoters in the “strong stabilizing” regime ((β – 1)C/C_max > 1. Dashed lines indicate buffering levels: black, no buffering; red, 10– fold; green, 100– fold; blue, 1000– fold.

This common relationship across all TFs studied here give rise to a general “buffering” of regulated expression levels from promoters of varying activities. To visualize this buffering behavior across the many different data sets from this study, Fig. 6B shows the magnitude of expression range in the promoter library (defined as the ratio of the highest expression level to the lowest expression level) when regulated against the expression range in the promoter library when expressed constitutively. The range of expression has a minimum value of 1, corresponding to all promoters with exactly the same expression level. Each data point here represents one TF acting on the promoter library; the blue points correspond to including the entire library, whereas the red points correspond to only those promoters in the “strong stabilization regime” ((β – 1)C/C_max > 1. The magnitude of buffering is demonstrated by a data point’s perpendicular distance below the black dashed 1 : 1 line. In nearly all cases, regulation narrows the expression range of the full promoter library (blue points) and in all cases it is narrowed for the strong promoters (red points). The scale of buffering is typical of order 10 – 100 fold (between the red and green dashed lines). This behavior is most pronounced if we include only strong promoters, where we rarely see a range greater than 10–fold. However, significant buffering is typical in the full promoter libraries. Each data point is labeled with a number that corresponds to the labels in Fig. S14, which contains information about the TF name and condition of the experiment and also show boxplots for all promoters in these data sets to give better intuition for the statistics of promoter expression range for each dataset.

Discussion

The relationship between TFs and the promoters they regulate is an important determinant of gene expression levels. Any given TF may regulate a battery of promoters with distinct features and these promoters will be regulated in various physiological conditions. Therefore, an understanding of the principles that determine how these features govern regulated gene expression is crucial for determining how gene expression changes upon mutation or environmental conditions and how these affect physiology. Using in vivo measurements in E. coli combined with a model of gene regulation, we discover a universal principle that TFs function to stabilize RNAP at promoters and the fold-change in expression enacted by strongly-acting TF scales as the reciprocal of the constitutive activity of the promoter. This function applies to all TFs tested, both activators and repressors, acting on both synthetic and naturally occurring promoters. The relationship holds if we perturb the “promoter strength” through (1) random genetic mutations to the promoter (2) physiological perturbations to the growth rate that change the propensity of transcription, (3) through direct perturbation of RNAP availability, or (4) any combination of these perturbations.

It is particularly surprising that every repressor measured here shows this same “stabilizing” relationship with promoter strength. This is contrary to what is often considered the common model of repressor action: steric occlusion prevents RNAP from binding to the promoter in the presence of a repressor. This is sometimes termed the “competitive model” of TF function. This idea has been supported in some in vitro studies specifically for the function of LacI (21, 48). This model is intuitive since LacI binds very close to the lac promoter, with the 5′ end of the binding site at +1 on the promoter. However, there are conflicting in vitro studies which have suggested regulation by LacI occurs at the step of transcription initiation (49–51). Straney and Crothers specifically found that LacI increased RNAP binding by more than 100–fold in their in vitro assays (49), although these contradictory results have sometimes been dismissed as an artifact of low ionic concentrations in the in vitro experiments (21, 52). Our in vivo findings support both findings by Straney and Crothers: (1) LacI negatively regulates at the step of initiation and (2) LacI positively affects stability of RNAP at the promoter. In fact, we find this to be a general regulatory feature of TFs across every interaction measured in our study. Surprisingly, our study does not find any evidence for regulation by steric occlusion, as inferred from our model. However, for some repressors we see slightly higher than expected fold-changes for the strongest promoters. We suspect this might relate to competitive re-binding of the promoter after DNA replication. A feature that would be most relevant for very strong promoters.

An important feature of stabilization interactions is that regulation is “restorative” by nature. Perturbations to the base level of gene expression, either through mutations to the promoter region or changes to the physiological state which may up- or down-regulate global expression rates, will be compensated when the TF-promoter has a stabilizing relationship. On the other hand, destabilizing TFs, which we do not find to be common, would exacerbate perturbations to base expression levels. A speculative explanation for the prevalence of stabilization relationships observed could be that the robust relationship between TF and promoters is evolutionarily favored; A TF whose function compensates for perturbations will help make expression levels robust and help maintain homeostasis of the cell.

This observation of a conserved mode of regulation underscores the limitations of the labels “activator” or “repressor” as a characterization of TF function and highlights the importance of understanding the mechanisms of action that affect regulation. If it is the case that most TFs operate through stabilization, then many TFs activators of weak promoters or repressors of strong promoters would be “incoherent regulators” (10, 12, 14–16, 53) and their identity as a TF depends entirely on the basal strength of the regulated promoter.

Materials and Methods

Bacterial Strains

All strains used in this study are based on the collection of TF library strains as in (24) and synthetic reporters are from (16). Unless otherwise stated, all TFs used in this study are tagged with mCherry at the C-terminal and controlled by adding aTC. Reporters with native promoters are constructed by cloning the promoter sequence (starting from the transcription start site and moving upstream with all the regulatory regions included) upstream of the YFP reporter. For sigma factor titration, the regulator vanR and the promoter, vanCp are PCR amplified from the marionette strains (54) and combined by splice overlap extension PCR to a spectinomycin cassette and inserted by lambda red recombinase into the native locus of fliA (the gene for σ²⁸ sigma factor) replacing the native fliAp promoter with vanCp. The regulator vanR, is inserted bidirectionally to fliA gene. The λ prophages carrying Spec^R::vanCp::fliA is transduced into the control and library strains of CpxR and LacI and selected for spectinomycin resistance. To introduce the anti-sigma factor knockout, the λ prophage carrying the flgM knockout from the Keio collection is first transduced into the corresponding control strains, selected for kanamycin resistance, and then the kanamycin cassette is cured by the use of pCP20 plasmid. The λ prophages carrying the TF construct (library strain) and the Spec^R::vanCp::fliA constructs are then sequentially transduced into the control strains with the flgM knockout, via two tandem P1 transduction.

Creation of mutant library

Oligo pools homologous to the −35 region of the promoter are designed by replacing two of the wild-type sequences with random nucleotide ”N” as depicted in Fig. S15B. 15 oligos are designed for a −35 sequence resulting in 154 unique combinations of the −35 sequences. For DL5p, the −35 region is selected from the literature. For native promoters, the −35 sequence is either based on the predictions in regulonDB (41) or based on the Salis promoter calculator (55). For synthetic DL5p constructs with the binding site located downstream of the promoter or for the native promoters, the reporter plasmids are PCR amplified with primers that target regions immediately upstream and downstream of the −35 region of the promoter and the single-stranded oligo pool with the mutations is used as a bridge to complete the plasmid Fig. S15A. The amplified PCR products are digested with DpnI enzyme, purified, assembled using NEB Hifi DNA assembly reaction, and transformed to corresponding control strains where the TF of interest is knocked out. For synthetic DL5p with binding sites located upstream, the forward primer amplifying the plasmid carries mutations for the −35 region at the 5′ end and the reverse primer is immediately upstream of the −35 region. The plasmid is then blunt-end ligated using NEB’s KDL reaction mix and transformed to corresponding control strains where the TF of interest is knocked out. 96 different clones per mutant library are sequenced to exclude misassembled constructs and the rest of the constructs are transformed to the corresponding library strain for further measurement.

Transformation to the library strains

96 independent colonies per mutant library are inoculated in 1 mL LB media in deep-well plates and grown overnight. Cells are pelleted by centrifugation at 4000g for 30 minutes and suspended in P1, P2 and P3 buffers (from Zymo) in a 1 : 1 : 2 ratio. Plates are centrifuged at 4000g for 30 minutes and the supernatant is transferred carefully into 3 times the volume of absolute isopropanol to precipitate the plasmid. Plates are again centrifuged at 4°C for 1 hour at 1800g and isopropanol is decanted and the plates air-dried at room temperature for 3 – 4 hours. 50μL of nuclease-free is added to each well and the plates are incubated in 42°C water for 10 minutes to dissolve the pelleted plasmid. The plates are then incubated on ice with the chemically competent cells of the corresponding TF Library strains for 30 minutes, heat-shocked, and recovered in SOC for 1 hour. Plates are centrifuged again and the pellet is resuspended in 30μL SOC and patterned column-wise in LB- kanamycin plate using an 8-channel multi-channel pipette, plates are incubated overnight and individual colonies are obtained.

Growth and FACS measurement

The control and library strains of a given TF carrying independent promoter variants are grown overnight in 300μL LB supplemented with antibiotics, kanamycin, and carbenicillin (for library strains, chloramphenicol is also added). The strains are diluted in a ratio of 1 : 5000 in 300μL M9-minimal media supplemented with glucose in a 2 mL deep-well plates and grown at 37°C to an OD of 0.2 – 0.5. The theoretical prediction in Fig. 1C is valid for saturated TF concentrations, effectively removing TF concentration as a variable. However, achieving true saturation may be challenging, particularly for weak binding sites. In this study, we approximate fully induced TF concentration (TF⁺⁺) as equivalent to saturated levels. This assumption does not significantly impact our conclusions, provided that the TF concentration stays constant and relatively high. Except for SoxS, other TFs are induced at 25 ng/mL anhydrotetracycline (aTC). SoxS is induced with 6 ng/mL aTC as inducing SoxS to higher concentration reduced the growth of cells significantly. For MetR, the comparison is made between uninduced library strain and 25 ng/mL aTC as the MetR knockout cannot grow in M9-minimal media. For physiological perturbations, M9-minimal media is supplemented with different carbon sources (glycerol, galactose, L-arabinose, sodium pyruvate, or sodium acetate) to achieve different growth rates. Measurements for sigma factor variation are performed similarly to that for the synthetic promoters in M9-minimal media with glucose and 25 ng/mL aTC. Vanillic acid is serially diluted two-fold starting from 10μM concentration to achieve different levels of σ²⁸ concentration.

The strains were diluted 1 : 80 in M9-minimal media with no carbon or nitrogen source to arrest cells in a steady state and incubated on ice until measurements. The plates are measured for mCherry and YFP fluorescence using BD LSRFortessa with an HTS (X –20– Model : 656385) and using the default settings for high-throughput sample readout. The mCherry is measured using a PE-CF594 laser at 600V. For the YFP signal, it is important to choose a FITC voltage to accommodate the range of expression that is expected of the different mutants. We measured auto-fluorescence and the weakest promoters at different voltages of FITC to find the minimal voltage that could be used to differentiate the auto-fluorescence signal from the weakest signal (Fig. S16A,B). We measured all YFP fluorescence with the FITC at 300V. The CST calibration is performed every day.

Data analysis and fitting to the model

The raw data is extracted from the ‘fcs’ files with a custom-built Matlab code and unsupervised gates are applied as described in (27). Mean and standard error is calculated for YFP fluorescence after subtracting the signal from the autofluorescence sample. Any sample with less than 5000 events is excluded from the analysis (except for acetate the events less than 1000 events are excluded). Fold-change is calculated by taking the mean of the ratio of fluorescence from the library strain to the control strain in 2 independent experiments. The data is then fit to the model with α and β and β as fit parameters and C_max set by the promoter mutant with the maximum fluorescence for a given TF. Cells with negative fluorescence are excluded only for visual purposes when plotting the histograms of single-cell distribution of fluorescence.

Measuring translation differences for native promoters of SoxS

The highest constitutive mutant is selected from the promoter library for decRp, fldAp, and poxBp. The strains are grown as described above in multiple 300μL wells, pooled at steady state (to a total of 1mL), OD quantified, and cell pellets were frozen overnight. Total RNA is isolated using the monarch RNA purification kit from NEB following the manufacturer’s protocol. The resulting RNA is quantified using nanodrop and 2 ng/μL of total RNA is used in a single-step RT-qPCR reaction with NEB’s Luna RT-reagent. Separate reactions were set for the yfp and the kanamycin resistance gene. Both genes are expressed from the same reporter plasmid and the kanamycin resistance gene readout is used as a housekeeping gene control for normalization. Reactions were set up with two biological and two technical replicates each. Standard curves for both kanamycin gene and yfp are prepared using the 6 dilutions of the plasmid. The concentration of yfp mRNA level is first normalized with the concentration of kanamycin mRNA level and then by the total RNA yield per OD of the cell. This factor is then used to correct all the constitutive expressions of the corresponding promoters.

Data and materials availability:

All data is available in the supplementary data file.