All strains were derived from E. coli NCM3722, which was obtained from the Coli Genetic Stock Center (Yale, United States). All strains, plasmids, and primers used are given in Supplementary Table S2 and the plasmid sequences are given in Supplementary Table S3. Phusion DNA polymerase, restriction enzymes, and T4 ligase used in plasmid construction were purchased from Thermo Fisher Scientific (Waltham, MA, United States). Primers were synthesized by Integrated DNA Technologies (Coralville, IA, United States).

Cell growth and nutrient shift experiments were performed in M9 minimal media with corresponding carbon source [M9 minimal salts supplemented with 75 mM MOPS at pH 7.4, 2 mM magnesium sulfate, 1 mg/L thiamine hydrochloride, 10 μM iron(II) sulfate, 100 μM calcium chloride, 3 μM ammonium heptamolybdate, 0.4 mM boric acid, 30 μM cobalt(II) chloride, 15 μM copper(II) chloride, 80 μM manganese(II) chloride, and 10 μM zinc sulfate]. Media was supplemented with sodium salts of each carbon source at a concentration of 72 mM carbon: 4 mM oleate, 36 mM acetate, 24 mM pyruvate, 24 mM glycerol, 18 mM (S)-malate, and 18 mM succinate. For experiments with co-utilization of glycerol and FA, the ratios of glycerol and oleate were adjusted to maintain a total of 72 mM carbon atoms in the media. For experiments with co-utilization of carbon and glyoxylate, 9 mM glyoxylate was used in all conditions, along with following concentrations of each carbon source such that 72 mM of carbon atoms was maintained: 3 mM oleate, 24 mM acetate, 18 mM pyruvate, 18 mM glycerol, 13.5 mM (S)-malate, and 13.5 mM succinate. All cultures were supplemented with appropriate antibiotic for selection (Ampicillin, 100 μg/ml; Kanamycin, 50 μg/ml).

Single colonies of the acyl-CoA biosensor strain were grown in 3 ml LB media for 2–3 h, then washed twice in M9 without a carbon source. Cultures were shifted into 3 ml M9 media supplemented with specified glycerol and oleate concentration and grown for an additional 3–5 h. Cells were again washed twice in M9 without a carbon source and were transferred M9 media supplemented with the specified glycerol/oleate ratios in a 96-well plate. Cells were diluted to a density of ~2 cells/μl so that the glycerol/oleate ratio does not significantly change during cell growth. Cells were grown for 9 h; then, growth was halted by addition of 100 μg/ml of rifampicin and incubated on ice for at least 15 min prior to measurement. Samples were analyzed with a Guava easyCyte HT5 flow cytometer (Luminex Corporation, Austin, TX, United States) with blue 488 nm excitation laser and the green 525 nm emission filter.

All nutrient shifts were performed following previous methods (Radzikowski et al., 2016). Specifically, cells were cultivated in a pre-shift medium at 37°C and were kept in exponential growth phase for 14–17 h. Cells were then collected and centrifuged in a pre-chilled centrifuge (4°C) at 4,500 rcf for 10 min. Supernatant was discarded and cells were resuspended in chilled M9 without carbon source. Three washes were performed. Finally, cell density (OD₆₀₀) was normalized to 0.5 in M9 without carbon, and then diluted 1:5 into pre-warmed oleate media so that the final cell density was 0.1, and the final oleate and ampicillin concentrations were 4 mM and 100 μg/ml, respectively, and with a final culture volume of 25 ml. At times indicated, 1 ml culture was transferred to centrifuge tubes pre-filled with 200 μl of phosphate buffered saline (PBS, pH 7.4). Cells are centrifuged at 4,500 rcf, 4°C, and washed four times in PBS. Finally, cell pellets were resuspended to a final volume of 1 ml, and serial dilutions were also performed in PBS. For each diluted culture, 10 μl was transferred into LB agar plates and incubated for 12 h at 37°C. Colonies were counted to determine the colony forming units per ml of culture (CFU/ml).

To quantify the length of time where the cells are transiently tolerant to ampicillin, we fit the colony counting data to a modified two-population Gompertz model, which has been widely used to describe kinetics of bacterial growth (Zwietering et al., 1990) and mortality (Kirkwood, 2015):

where p is the persister fraction, μ T T and μ P E R are the maximum killing rates of the transient tolerant and persister population, respectively, and τ T T and τ P E R are the transient tolerance times of the transient tolerant and persister fractions, respectively. Next, we assume that the tolerance time of the transient tolerant fraction and persister fraction are equal, i.e., τ T T = τ P E R = τ . Thus, only the maximum rate of killing distinguishes the transient tolerant and persister populations.

To fit the data, CFU/ml counts were first normalized to the CFU/ml from time point zero and converted to a log₁₀ scale. All data from replicates of each shifting experiment were fit to the two-population Gompertz model using the MATLAB R2020B curve fitting tool (cftool). An initial transient tolerance time was determined; the data were then re-normalized to the average CFU/ml counts prior to the tolerance time for each replicate. The re-normalized data were fit again, and the parameters along with their 95% CIs and SEM were determined.

Cells were grown in different pre-shift carbon sources at steady state as described for Colony Counting Assays. Cells were centrifuged and washed in M9 without carbon source for four times. Washed cells were then transferred to a Falcon 96-Well Imaging Microplate (Corning, NY, United States) and diluted to an initial OD₆₀₀ of 0.08 in oleate media with or without ampicillin, with a final volume of 150 μl. An Infinite F200PRO plate reader (TECAN, Männedorf, Switzerland) was used to make automated OD₆₀₀ and fluorescence measurements (Excitation: 514 nm, Emission: 552 nm) every 6 min with constant shaking and 37°C temperature control. Fluorescence measurements are normalized by OD₆₀₀ to give an estimate of FadD concentration.

Lag phase was determined by a similar protocol as described previously (Basan et al., 2020). The steady state growth rate after nutrient shift was first calculated by fitting a line to the natural log of the OD₆₀₀ for the first hour after the culture density increases 8-fold from its initial density. The time point where this line intersects the initial culture density is the end of the lag phase. The fluorescence measurement at the closest measured time point is used as YFP/OD₆₀₀ at the lag time. The FadD threshold was determined by taking the average YFP/OD level at the end of the lag phase across all nutrient shift conditions without glyoxylate. The accumulation time was determined as the first time point where the YFP/OD₆₀₀ was above this threshold and the YFP/OD₆₀₀ at all subsequent time points also exceeded the threshold.

For kinetic optical density measurements in the presence of ampicillin, the transient tolerance time was calculated as the last local maximum in the log of the OD for each time series. To determine this, a moving window slope of 11 time points was used twice to calculate the first and second derivatives of the log OD. The last time point where the first derivative was closest to zero and where the second derivative was negative was taken as the tolerance time measured by OD₆₀₀. Additionally, all fluorescence measurements after the measured tolerance time were normalized by the OD₆₀₀ at the tolerance time to give an estimate of the FadD concentration which is unbiased by the reduction in OD due to cell death.

Cells were cultivated to steady state growth to an OD₆₀₀ of 0.2 and 2 ml of cells were collected and stored in RNA/DNA Shield (Zymo Research, Irvine, CA, United States). Total RNA was extracted from cells using the Quick-RNA Miniprep Plus Kit (Zymo Research) following manufactures protocols. Contaminating genomic DNA was removed by DNase I treatment of RNA on collection column, following the Quick-RNA Miniprep Kit protocol. cDNA was synthesized using Revert Aid First strand cDNA Synthesis (Thermo Fischer Scientific) with random hexamer primers following the manufactures protocol. Negative control reactions without the use of reverse transcriptase were performed to evaluate the potential presence of contaminating genomic DNA. Two microliter of cDNA was amplified using Power SYBR green PCR Master Mix (Thermo Fischer Scientific) and gene specific primers (Supplementary Table S2). qPCR reactions for each biological replicate, gene, and growth condition were performed in triplicate. qPCR assays were performed on a QuantStudio3 (Thermo Fischer Scientific) following standard thermal cycling conditions recommended by the manufacturer. Expression levels of the condition invariant gene gyrA were used as a control for normalization between samples (Said-Salman et al., 2019). Fold changes of each gene of interest were calculated following the 2^−ΔΔCT method.

Similar to many other nutrients, FA uptake in E. coli is regulated by a positive feedback loop (Figure 1A). In this loop, the fatty acyl-CoA ligase FadD controls the uptake and activation of extracellular free FAs to acyl-CoAs, which are then catabolized by a series of FA degradation (Fad) enzymes to acetyl-CoA for use in central metabolism. Expression of fadD and the fad regulon is repressed by the transcriptional regulator FadR, whose DNA-binding activity is further inhibited by acyl-CoAs, forming a simple positive feedback loop. Modeling of this FA uptake loop shows that bistability in FA uptake can appear under some parameters and the parameters for E. coli may be near the bistable regime (Mannan and Bates, 2021). In the bistable regime, a sub-population of cells would have low FadD expression due to FadR repression, thus low FA uptake rate to activate FadR and to support rapid cell growth. Another sub-population would maintain a high FadD expression level, thus can keep high intracellular acyl-CoA level to both antagonizes FadR’s DNA-binding activity and to support cell growth. If bistability in FA uptake does occur, the rapid growing sub-population will be killed by ampicillin, whereas slow-growing cells may display tolerance to ampicillin which only targets growing cells.

To test the possible presence of bistability in FA uptake, we transformed E. coli cells with an engineered FadR-based acyl-CoA biosensor (Zhang et al., 2012) that reports intracellular acyl-CoA level (Xiao et al., 2016; Liu and Zhang, 2018) and indirectly reports expression of fad genes. To characterize the acyl-CoA biosensor activity under different conditions, the strain was grown to steady state in defined media containing mixtures of glycerol and oleate at different ratios. Flow cytometry showed that the mean activity of the acyl-CoA biosensor increased as the ratio of FA-to-glycerol increased (Figure 1B), consistent with the biosensor’s behavior in glucose/FA media (Liu and Zhang, 2018). At single-cell level, the biosensor activity showed unimodal distribution under all conditions (Figure 1C), without a distinctive sub-population above the detection limit of the method (~1% of the population). These results suggest that although positive FA uptake loop can support bimodality/bistability under some parameters (Mannan and Bates, 2021), the actual parameters for this strain of E. coli do not support a large bimodal population. Even so, a much smaller bimodal population may exist below the detection limit of flow cytometry.

Thus, to more sensitively determine whether two distinct sub-populations exist with different killing kinetics, we next conducted nutrient shifting experiments (Radzikowski et al., 2016) from different ratios of glycerol/FA mixtures to pure FA accompanied by ampicillin treatment (Figure 1D). When pure FA was used in the pre-culture as a negative control without nutrient switching (FA-to-FA), a typical bi-phasic killing kinetics was observed (Figure 1E): nearly immediate and rapid killing followed by a small population (0.008%) of antibiotic tolerant cells, close to previously reported levels of persistent cells in active cultures (Balaban et al., 2019). Although we observe a short period without killing (less than 1 h), which may be caused by starvation and cold stress from the washing procedures to provide temporary tolerance to ampicillin (Heinemann et al., 2020), this tolerance is not maintained for a large fraction of the population. Overall, this result proves that the switching procedure used in this study did not generate an elevated level of persisters. In contrast, when switching from glycerol to FA, the killing curve displayed much longer initial tolerant period, where nearly all cells survived ampicillin treatment during the first 8 h. This transient tolerance period is followed by a rapid killing where 98% of the transiently tolerant cells were killed. Finally, the rate of killing reduced with ~2% population having elevated antibiotic tolerance. When mixtures of glycerol and FA were used in the pre-culture, the killing curves also displayed an initial transient tolerant period followed by rapid killing. The length of this tolerance period increased as the amount of FA in the pre-culture decreased (Figure 1E). These results further confirm that the long transient tolerance to ampicillin was caused by specifically by nutrient shifting rather than the washing procedure used in this study and suggests a connection between the prior acyl-CoA activity with the tolerance period.

To examine whether the transient tolerant exists when switching from other gluconeogenic carbons, we performed additional switching experiment by replacing glycerol with four other carbon sources: acetate, pyruvate, malate, and succinate. These carbon sources represent different entry points into the central metabolism (Figure 2A), thus would allow us to compare how different metabolic state of cells in the pre-culture affect transient tolerance. All transitions displayed similar tri-phasic killing kinetics with an initial transient tolerance followed by rapid killing of susceptible cells and a slow killing of persistent cells. The period of transient tolerance varied significantly with different carbon sources (Figure 2B). While transitions from acetate had a relatively short transient tolerance time (3.3 ± 0.4 h, 95% CI), transitions from pyruvate displayed extremely long tolerance (44 ± 3 h, 95% CI). However, despite great variability in the timing, all conditions showed that majority of cells was eventually killed by ampicillin, indicating that the initial antibiotic tolerance is only temporary after gluconeogenic-to-FA nutrient shifts.

Shifts from glycolytic to gluconeogenic carbon sources have been shown to produce lag phase during which cell halts growth transiently (Basan et al., 2020). Since ampicillin is only effective on growing cells, cessation of growth after a nutrient shift offers a simple mechanism for transient tolerance. To test this, we measured growth kinetics of cultures switching from each gluconeogenic carbon source to FA without antibiotic (Figure 2C). The lag time of each shift was found to correlate well (R² = 0.82) with the transient tolerance time (Figure 2D). These results suggest that non-growing cells in the lag phase after a nutrient shift cause transient tolerance to ampicillin.

Since the lag time and transient tolerance period after the nutrient shift are well correlated, we sought to understand the molecular mechanisms which govern both phenomena. Growth rate in pre-culture conditions was shown to predict lag time for glycolytic-to-gluconeogenic switches (Basan et al., 2020) and therefore may also predict tolerance time on nutrient shifts from gluconeogenic carbons to FA. We measured pre-culture growth rates and found a weak negative correlation (R² = 0.35) between pre-culture growth rate and tolerance time (Figure 3A). Although the direction of correlation is consistent with previous studies (Basan et al., 2020), nutrient shifts from pyruvate to FA present a clear outlier, suggesting the presence of other mechanisms that control tolerance time more directly.

Previous studies illustrated that β-oxidation enzymes are essential for growth when FA is the sole carbon source (DiRusso et al., 1999), suggesting a possible role for β-oxidation enzymes in tolerance time during nutrient shifts to FA. Because FadD is a key component of the positive feedback loop controlling transport and activation of FAs for β-oxidation, we genetically fused a YFP to FadD to monitor activity of FA degradation during the nutrient shift. The FadD-YFP strain was grown in different gluconeogenic carbon sources and switched to FA in the absence of ampicillin. Cell growth and YFP fluorescence were simultaneously monitored after the nutrient shift. After shifting from different gluconeogenic carbon sources, the initial FadD levels were different and all lower than that without carbon shift (i.e., from FA to FA, Figure 3B), suggesting that the β-oxidation enzymes were not highly expressed in gluconeogenic carbons. The FadD concentration gradually increased over time during the lag phase. When cells resume steady state growth, FadD concentration from all switches reached the same threshold level (Figure 3C). Without carbon shift (from FA to FA), the FadD level remained above this threshold level, and the cells continued to grow without a lag phase. We measured the time needed to reach this threshold FadD level and found that it has a good correlation with the tolerance time (R² = 0.82; Figure 3D). Since the timing of accumulating FadD correlates well with transient tolerance, we wondered whether this timing of each phenomenon is simply determined by the amount of FadD initially present immediately after the nutrient shift. Although FadD accumulation time appears to decrease with increasing initial FadD concentrations (R² = 0.57; Figure 3E), transitions from pyruvate again provide a strong outlier with cells taking much longer to accumulate FadD to a threshold.

To verify whether these FadD dynamics are representative of the nutrient shifts in the presence of antibiotics, we performed similar experiments using the FadD-YFP fusion strain in the presence of ampicillin. Across all switching conditions, OD₆₀₀ initially decreased slightly followed by a leveling off period (Supplementary Figure S1A). Because this slight decrease is also observed in the absence of antibiotic, this decrease more likely represents reductive division of cells, which has been previously reported for cells under starvation conditions (Nyström, 2004), rather than killing by antibiotics. After this relatively level period (transient tolerance), the OD decreases substantially, corresponding to rapid killing of susceptible cells. In all switching conditions, the FadD level continued to increase during the transient tolerance period, indicating active transcription and translation. The time needed to reach a threshold FadD level still correlated well with the tolerance time (R² = 0.79, Supplementary Figure S1B). Additionally, the timing of FadD accumulation in the absence and presence of ampicillin correlated extremely well (R² = 0.997; Supplementary Figure S1C), indicating that the FadD dynamics during the nutrient shift are nearly the same prior to reaching the FadD threshold, regardless of the presence of ampicillin. Overall, these results suggest that FA metabolic activity was low during lag phase, resulting in slow cell growth and transient tolerance to ampicillin. A key level of FadD is needed to resume cell growth and active metabolism in FA medium.

Finally, we evaluated the single-cell behavior of FA metabolism using the acyl-CoA biosensor during nutrient transitions from pyruvate to FA, which has the longest transient tolerance time. We characterized the acyl-CoA biosensor activity under mixtures of pyruvate and FA and found that the population has a monomodal distribution of acyl-CoA biosensor activity under all conditions (Supplementary Figure S2A), consistent with the biosensor activity on glycerol/FA mixtures (Figure 1C). We then conducted nutrient switches from pyruvate to FA with ampicillin and measured the acyl-CoA biosensor activity at time points during the transient tolerance phase. Again, we observed monomodal populations at all-time points, with the biosensor activity of the whole population increasing sharply by 4 h after the shift, and then more gradually after 21 h (Supplementary Figure S2B). Thus, these results suggest that there is a single, large majority population before the shift and this whole population gradually transitions toward active FA metabolism after the shift, leading to growth and sudden killing by ampicillin for a large population at the end of the transient tolerance.

Given that the time for FadD to reach its threshold concentration correlates well with transient tolerance, we wondered what controls FadD accumulation. Particularly, shifts from pyruvate to FA produced the longest transient tolerance times from the above experiment. We searched the literature and did not find direct regulatory effects of pyruvate on fadD expression in E. coli. However, pyruvate activates the IclR transcription factor (Lorca et al., 2007) which represses expression the glyoxylate bypass operon aceBAK (Figure 4A). The glyoxylate bypass is necessary for FA utilization because β-oxidation produces only two-carbon metabolite precursors in the form of acetyl-CoA (Dolan and Welch, 2018). To examine whether pyruvate impacts the glyoxylate bypass or β-oxidation gene expression, we evaluated the expression level of the aceB and fadD genes via RT-qPCR for cells growing in glycerol, pyruvate, and FA (Supplementary Figure S3). In pyruvate, aceB expression is 1.7-fold lower compared to that in glycerol; however, fadD expression is not significantly different between glycerol and pyruvate conditions, indicating that the glyoxylate bypass rather than the FA transport genes is more blocked in pyruvate. During growth on FA, both aceB and fadD expression are significantly increased (13.7-fold and 5.0-fold increase, respectively), which indicates that the expression of both aceB and fadD genes needs to be increased for cells to grow on FA, in agreement with previous literature (Clark and Cronan, 2005).

In addition to pyruvate regulation, IclR activity is antagonized by glyoxylate (Lorca et al., 2007), an intermediate in the glyoxylate bypass (Figure 4A), thus forming a positive feedback loop which also governs FA utilization. Therefore, when cells are growing in pyruvate, accumulation of glyoxylate is slow, and the glyoxylate bypass enzymes remain at low levels. We hypothesize that switching from pyruvate to FA requires a gradual increase in the metabolic flux through the glyoxylate bypass to produce central metabolites (e.g., amino acids) for synthesizing β-oxidation enzymes. To test this hypothesis, we added glyoxylate to the pre-shift medium to activate the glyoxylate bypass enzymes, thus alleviating the bottleneck in the glyoxylate bypass enzymes. New nutrient shift experiments with glyoxylate displayed decreased lag phase as well as FadD accumulation times for pyruvate, succinate, and malate (Figures 4B,C). Among them, pyruvate exhibited the most drastically change, its FadD accumulation time decreased from 11.4 ± 1.3 h to 3.4 ± 0.5 h (95% CI; Figure 4C). Additionally, we performed similar growth kinetic experiments during nutrient shift to FA with ampicillin and observed similar reductions in the transient tolerance time measured by OD₆₀₀ (Supplementary Figure S4). The transient tolerant time shifting from pyruvate with glyoxylate supplement to FA decreased from >16 h without glyoxylate to 3.1 ± 0.5 h (95% CI). With the presence of glyoxylate in the pre-shift medium, the highest correlation between the transient tolerance period and FadD accumulation time was obtained (R² = 0.98; Figure 4D), demonstrating that both FadD accumulation and transient tolerance are accelerated by the removal of the glyoxylate bottleneck.