Individual and Collective Contributions of Chaperoning and Degradation to Protein Homeostasis in E. coli

SUMMARY

The folding fate of a protein in vivo is determined by the interplay between a protein’s folding energy landscape and the actions of the proteostasis network, including molecular chaperones and degradation enzymes. The mechanisms of individual components of the E. coli proteostasis network have been studied extensively, but much less is known about how they function as a system. We used an integrated experimental and computational approach to quantitatively analyze the folding outcomes (native folding vs. aggregation vs. degradation) of three test proteins biosynthesized in E. coli under a variety of conditions. Overexpression of the entire proteostasis network benefited all three test proteins, but the effect of upregulating individual chaperones or the major degradation enzyme, Lon, varied for proteins with different biophysical properties. In sum, the impact of the E. coli proteostasis network is a consequence of concerted action by the Hsp70 system (DnaK/DnaJ/GrpE), the Hsp60 system (GroEL/GroES), and Lon.

INTRODUCTION

Protein homeostasis, or proteostasis, is achieved when an organism has enough natively folded proteins to carry out its essential functions but not enough misfolded and aggregated proteins to interfere with organismal fitness (Balch et al., 2008; Powers et al., 2009). In a simplified view of proteostasis, new proteins can have three fates: they can fold to their native states; they can misfold and/or aggregate; or they can be degraded (Figure 1). Proteins that experience the latter two fates are not functional. All organisms regulate the health of their proteomes via a collection of chaperones, folding enzymes, proteases and other components that together make up the proteostasis network, or PN (Kim et al., 2013; Powers and Balch, 2013).

Figure 1. Schematic of Kinetic Partitioning During Protein Folding In Vivo.

Protein folding in vivo begins with ribosomal synthesis (“Synthesis”), which yields unfolded protein molecules (“U”; note that we neglect the possibility of co-translational folding). Unfolded protein can fold to the native state (“N”), misfold to the misfolded state (“M”) or be degraded (“Degradation”). Misfolded protein can be degraded or self-associate to form aggregates (“A”; note that aggregation is reversible in principle, as shown here, but in many cases is irreversible in practice unless assisted by the proteostasis network). A protein’s folding energy landscape dictates its partitioning among the unfolded, native, misfolded, and aggregated states in vitro (red text). However, each folding process is modulated by components of the proteostasis network in vivo. In E. coli, DnaK, DnaJ, and GrpE (KJE; the Hsp70–Hsp40 system) opposes misfolding by binding to misfolded protein molecules and forcing them to resume the unfolded state. GroEL and GroES (GroELS; the chaperonin system) promote folding by encapsulating unfolded protein molecules and enabling them to fold in an isolated cavity. Degradation is carried out by proteases, in particular Lon. Finally, ClpB (Hsp104) collaborates with KJE to solubilize aggregates.

Since protein folding, misfolding, and aggregation equilibria are linked, PN components that modulate any of these processes indirectly affect the others. However, each PN component seems to have an “assigned responsibility”: a process that it affects most directly. Using the E. coli PN as an example, native folding is promoted most directly by GroEL and GroES, the E. coli chaperonin–co-chaperonin pair (Chapman et al., 2006; Horwich and Fenton, 2009). Misfolding and aggregation are opposed most directly by DnaK, DnaJ, and GrpE, the E. coli Hsp70–Hsp40–nucleotide exchange factor trio (Calloni et al., 2012; Mayer and Bukau, 2005; Sharma et al., 2010), and by the collaboration of this trio with the disaggregating chaperone ClpB (Doyle et al., 2013). Finally, many proteases degrade proteins in E. coli, but Lon appears to be the most important for degrading misfolded protein (Gottesman, 1996; Gur and Sauer, 2008).

While the main functions of individual PN components are fairly well understood, their contributions to the integrated, system-level function of the whole PN are not as clear (Bershtein et al., 2013; Dickson and Brooks, 2013; Hingorani and Gierasch, 2014; Kim et al., 2013; Powers et al., 2012; Wiseman et al., 2007). How do PN components complement each other? Do they perform multiple or redundant functions? To what extent does a protein’s folding energy landscape determine its route through the PN and its fate? We have used a combination of experiments and computational modeling to address these and related questions.

Here, we focus on how proteins with low stabilities behave when overexpressed in E. coli because expression of such proteins challenges proteostasis (Gidalevitz et al., 2006; Olzscha et al., 2011). We chose E. coli as a model organism because of the availability of FoldEco (Powers et al., 2012), a computational model of E. coli’s PN that is essential to answer the mechanistic questions posed above. In addition, E. coli is widely used as a microbial factory for producing heterologous proteins. Failures in proteostasis were quantified by measuring total expression levels and the amount of aggregated vs. soluble protein. These quantities report on the extent of degradation, aggregation, and native folding experienced by our test proteins, and therefore cover each of a protein’s potential fates. We interrogated the roles of the DnaK– DnaJ–GrpE pathway (KJE), the GroEL–GroES pathway (GroELS), and Lon by overexpressing these PN components individually or in combinations. The point at which, and the extent to which, proteostasis failed for the test proteins then informed us as to the limits of the E. coli PN.

The test proteins in this work are unstable variants of E. coli dihydrofolate reductase (EcDHFR), murine cellular retinoic acid-binding protein 1 (MmCRABP1), and a de novo designed retroaldolase enzyme (RA114.3). These proteins span a range of origins (endogenous E. coli vs. mammalian vs. de novo designed, respectively) and folds (αβα sandwich, β barrel, and α/β barrel, respectively; Figure 2) (Bjelic et al., 2014; Kleywegt et al., 1994; Liu et al., 2014; Sawaya and Kraut, 1997), and have no significant sequence similarity. By examining how each member of this diverse group partitions between being soluble, aggregating, and being degraded as a function of the composition of the E. coli PN, we hoped to extract general lessons about the attributes of the PN as a system in its interactions with as broad as possible a selection of proteins, as well as lessons about the dominant contributors to the PN’s various functions. It is important to note here that this undertaking requires the assumption that the PN of E. coli handles heterologous proteins and its own endogenous proteins similarly. While a PN component from one organism generally cannot complement the loss of the orthologous component in another organism, chaperones from one organism are generally capable of assisting the folding of proteins from another. For example, upregulation of E. coli chaperones improved the expression yields for most of a set of 64 heterologous proteins (de Marco et al., 2007). Thus, KJE, GroELS, and Lon from E. coli appear to be quite general in their selection and handling of substrates.

Figure 2. Test Proteins Used in this Work.

(A) Structures of wild type test proteins: EcDHFR (PDB ID: 1RX2) (Sawaya and Kraut, 1997), MmCRABP1 (PDB ID: 2CBR) (Kleywegt et al., 1994), and RA114 (PDB ID: 4OU1) (Liu et al., 2014). The sites of the mutations in the destabilized test proteins are highlighted as pink spheres. (B) Bands corresponding to the test proteins in SDS-PAGE gels run on samples derived from E. coli overexpressing the wild type (“wt”) and mutant forms of the test proteins for 2 h at 30 °C. The lane labeled “total” is from pre-centrifugation cell lysates and the lanes labeled “soluble” and “aggregated” are from the supernatants and pellets, respectively, after cell lysates were centrifuged for 10 min at 13,500×g. The wild type variants of the test proteins form little or no aggregates when overexpressed in E. coli, but the mutants aggregate substantially. See also Figure S1.

RESULTS

The Test Proteins and their Overexpression in the E. coli Proteostasis Network in the Absence of Other Perturbations

The test proteins studied here are the M42T/H114R mutant of EcDHFR (m-EcDHFR), the R131Q/Y133S mutant of MmCRABP1 (m-MmCRABP1), and the E10K/D120V/N124S/L225P mutant of RA114.3 with a C-terminal His tag (m-RA114). These proteins are small- to medium-sized (DHFR: 159 amino acids; CRABP1: 137 amino acids; RA114: 258 amino acids) and monomeric (Figure 2). That each of these mutants is less stable than the corresponding wild type protein is demonstrated by their susceptibilities to urea denaturation for m-EcDHFR and m-RA114 (Figure S1A, B), and was reported previously for m-MmCRABP1 (Budyak et al., 2013).

Each of these test proteins was expressed by isopropyl β-D-1-thiogalactopyranoside (IPTG) induction of a plasmid under control of the lac promoter in E. coli K12 DE3 (HMS174) cells growing in LB media at 30 °C. After 2 h induction, the cells were lysed, and aggregated and soluble proteins were separated by centrifugation. Total, soluble, and aggregated (i.e., in the pellet) protein fractions were analyzed by SDS-PAGE and the absolute amount of test protein in each was determined by comparison to a calibration line constructed using purified recombinant protein (Figure S1C; Supplemental Experimental Procedures). The total protein concentrations present at the end of the 2 h expression period were 498 ± 58, 106 ± 5, and 385 ± 40 μM (mean ± SEM) for m-EcDHFR, m-MmCRABP1, and m-RA114, respectively. Substantial aggregation was observed for each test protein, with the aggregated fractions amounting to 46 ± 3%, 76 ± 1% and 86 ± 1% of total protein for m-EcDHFR, m-MmCRABP1, and m-RA114, respectively (Figure 2B, Figure 3). Thus, m-EcDHFR appears to be the best behaved of our test proteins, while m-MmCRABP1 and m-RA114 are more aggregation prone. Expression of the wild type versions of these proteins under the same conditions resulted in negligible (wt-EcDHFR and wt-MmCRABP1) to low (wt-RA114) levels of aggregates (Figure 2B).

Figure 3. Folding Fates of Test Proteins upon Overexpression in E. coli under Adapted-Basal Conditions and with Individual PN Components Upregulated.

(A) Experimental scheme for overexpression of test proteins. All experiments were performed at 30 °C. (B) Bar graph showing the results of overexpressing m-EcDHFR in E. coli under various conditions. Left axis: cytoplasmic concentration of m-EcDHFR as determined by quantitative analysis of gels like those shown in Figures 2B and S2E. Right axis: concentration of m-EcDHFR relative to the total concentration of m-EcDHFR (soluble + aggregated) produced under adapted-basal conditions. These concentrations are referred to in the text as [Sol]_rel,X and [Agg]_rel,X for soluble and aggregated forms of a given test protein “X”. White bars: total protein concentration under each condition. Blue bars: concentration of soluble protein under each condition (i.e., the concentration in the supernatant after lysis and centrifugation). Red bars: concentration of aggregated protein under each condition (i.e., the concentration in the pellet after lysis, centrifugation, and re-suspension). Error bars: standard errors of the mean. Numbers above bars: percentage of the total protein that is soluble or aggregated under each condition ± SEM. (C) As (B), but for m-MmCRABP1. (D) As (B), but for m-RA114. The “high” and “low” upregulation levels resulted in respective concentration increases of ~4- to 8-fold and ~2- to 3- fold for the major PN components (DnaK, GroEL, and Lon; Figure S2C,D). The “medium” upregulation level (for Lon only) resulted in a ~4-fold increase in the concentration of Lon (Figure S2C, D). See also Figure S2 and Table S1.

Heterologous protein expression can cause stress and lead to the upregulation of PN components (Gasser et al., 2008; Hoffmann and Rinas, 2004). We therefore measured the levels of GroEL, DnaK, and Lon before and 2 h after induction of m-EcDHFR and m-RA114. Compared to an empty vector control, DnaK and GroEL increased ~20% to 40% and Lon ~80% to 160% (Figure S2A). Similar increases in PN component levels were observed when wt-RA114, which is more stable than m-RA114 but still aggregates, was overexpressed. In contrast, overexpression of wt-EcDHFR, which is stable and well behaved, resulted in much smaller changes in PN component levels (Figure S2A). These results suggest that it is the overexpression of aggregation-prone proteins, and not the overexpression of proteins per se, that causes PN component levels to increase. The PN as it exists after being perturbed by expression of the test proteins will be referred to as the “adapted-basal” PN.

Individual Upregulation of KJE, GroELS, or Lon Moderately Decreases Test Protein Aggregation

To assess the effects of KJE, GroELS, and Lon on the test proteins, we introduced them into pBAD expression vectors (Figure S2B) so that their expression could be titratably induced by arabinose. Co-transformation with separate plasmids carrying the test protein and the PN components enabled us to express each independently. The expression of KJE, GroELS, or Lon was induced by adding arabinose 1 h before induction of the destabilized test protein (Figure 3A). The level of upregulation of the PN components was controlled by the concentration of arabinose added. The “high” and “low” upregulation levels resulted in respective concentration increases of ~4- to 8-fold and ~2- to 3-fold for the major PN components (DnaK, GroEL, and Lon; Figure S2C, D). The “medium” upregulation level, which was used only for Lon because the high-Lon conditions caused a drastic decrease in total protein levels (see below), increased Lon ~4-fold (Figure S2C, D).

Upregulation of KJE decreased, but did not eliminate, aggregation for each of the test proteins. The aggregated fraction decreased significantly from 46 ± 3% to 25 ± 4% of the total protein for m-EcDHFR (p = 0.001, one-tailed t-test); from 76 ± 1% to 46 ± 5% for m-MmCRABP1 (p < 0.001, one-tailed t-test); and from 86 ± 1% to 37 ± 3% for m-RA114 (p < 0.001, one-tailed t-test) under the high-KJE upregulation conditions (Figure 3B–D, Figure S2E, Table S1). In addition, the total test protein concentrations decreased by 30–40% under the high-KJE upregulation conditions (Figure 3B–D, Table S1). Low-KJE upregulation conditions also decreased total concentrations of the three proteins, but to a lesser extent. We cannot exclude the possibility that this result is due to direct delivery of substrates to proteases by DnaK or DnaJ (Sherman and Goldberg, 1992). However, based on modeling results with FoldEco that are presented in a later section, it appears more likely that the observed decreases in protein concentration are due to proteins that are rescued from misfolding and/or aggregation by KJE being degraded before they can fold or re-aggregate.

Upregulation of GroELS similarly decreased aggregation of m-EcDHFR and m-RA114. The aggregated fractions decreased from 46 ± 3% to 22 ± 9% for m-EcDHFR (p = 0.004, one-tailed t-test) and from 86 ± 1% to 39 ± 3% for m-RA114 (p < 0.001, one-tailed t-test) under the high-GroELS upregulation conditions (Figure 3B, D; Table S1). In contrast, GroELS upregulation did not significantly decrease the aggregation of m-MmCRABP1 (from 76 ± 1% to 66 ± 11%; p = 0.26, one-tailed t-test) (Figure 3C, Table S1). This observation suggests that m-MmCRABP1 is not a good substrate for GroELS; we examine this possibility in a later section.

Upregulation of Lon decreased the levels of both the soluble and aggregated forms of the test proteins, but, especially at the low and medium upregulation levels, the decrease was larger for the aggregates than for the soluble protein for two of the three (m-MmCRABP1 and m-RA114; compare Figure 3B to Figures 3C and D).

Upregulation of KJE, GroELS, or Lon in Pairs Further Decreases Test Protein Aggregation

To determine the effects of joint upregulation of KJE, GroELS, and Lon on our test proteins, we introduced pairs of these systems into the same expression vectors with one system under an arabinose promoter and the other under a tetracycline promoter (Figure S2B). We then repeated the test protein overexpression experiments described above (Figure 3A), except that the PN pathways were upregulated in pairs using inducer concentrations at the high upregulation level. In all cases, the levels of the PN components did not increase as much as when they were overexpressed on their own (~2- to 4-fold instead of 4- to 8-fold; Figure S3A).

GroELS+Lon and KJE+Lon were the most effective pairs of PN components for suppressing test protein aggregation (Figure 4, Table S1). The GroELS+KJE combination was less effective at suppressing aggregation (i.e., the fraction aggregated was higher under this condition) than either GroELS+Lon or KJE+Lon for m-MmCRABP1 and m-RA114 (for m-MmCRABP1: p = 0.0025 and 0.0016 for GroELS+KJE vs. KJE+Lon and GroELS+Lon; for m-RA114: p = 0.001 and 0.023 for GroELS+KJE vs. KJE+Lon and GroELS+Lon, one-tailed t-test). The same trends are apparent for m-EcDHFR, although the null hypothesis cannot be rejected with the same level of confidence (p = 0.089 and 0.176 for GroELS+KJE vs. KJE+Lon and GroELS+Lon, one-tailed t-test).

Figure 4. Folding Fates of Test Proteins upon Overexpression in E. coli under Adapted-Basal Conditions and with PN Components Upregulated in Pairs or after Overexpression of I54N σ³².

(A) As in Figure 3B, but with multiple PN components upregulated. (B) As in Figure 3C, but with multiple PN components upregulated. (C) As in Figure 3D, but with multiple PN components upregulated. The levels of DnaK, GroEL, and Lon increased by ~2- to 4-fold when they were upregulated in pairs or via overexpression of I54N σ³² (Figure S3A). See also Figure S3 and Table S1.

Simultaneous Upregulation of KJE, GroELS, or Lon via Expression of σ³² Virtually Eliminates Aggregation

As noted above, pair-wise upregulation of KJE, GroELS, and Lon resulted in lower levels of these PN components than upregulating them individually. Extrapolating this result suggests that it could be difficult to attain sufficiently high levels of upregulation if all three PN pathways were upregulated using arabinose- and tetracycline-induced expression systems. We therefore sought another way to simultaneously upregulate KJE, GroELS, and Lon.

KJE, GroELS, and Lon, as well as many other PN components, are in the regulon of the heat shock transcription factor, σ³² (Guisbert et al., 2008; Zhang et al., 2014). We therefore overexpressed the I54N mutant of σ³², which evades the post-translational regulation of σ³² (Guisbert et al., 2008; Yura et al., 2007; Zhang et al., 2014), to simultaneously increase the levels of KJE, GroELS, and Lon. Induction of I54N σ³² for 1 h prior to inducing the test proteins yielded 3- to 4-fold increases in the levels of DnaK, GroEL, and Lon (Figure S3A), as reported previously (Zhang et al., 2014).

Using I54N σ³² expression to upregulate the σ³² regulon nearly eliminated aggregation of the test proteins (Figure 4A–C). To determine the extent to which this result was due to KJE, GroELS, and Lon upregulation and not to other PN pathways that are part of the σ³² regulon, we examined the effect of upregulating ClpB (Doyle et al., 2013), or HtpG, the E. coli Hsp90 (Pearl and Prodromou, 2006), together with KJE, since both of these chaperones can cooperate with KJE (Doyle et al., 2013; Genest et al., 2011; Nakamoto et al., 2014). Upregulating ClpB+KJE or HtpG+KJE by expressing them simultaneously did not decrease the aggregated fractions of the test proteins beyond what was observed by upregulating KJE alone, even at the low level of upregulation (compare Figure S3B to the low-KJE upregulation results in Figure 3B–D). Simultaneous upregulation of IbpA and IbpB, the small heat shock proteins of E. coli (Kuczynska-Wisnik et al., 2002; Thomas and Baneyx, 1998), tended to increase the extent of aggregation of m-EcDHFR and m-MmCRABP1 relative to adapted-basal conditions, although this increase was only significant for m-EcDHFR (p = 0.003, one-tailed t-test). Upregulation of IbpA and IbpB had no effect on m-RA114 (Figure S3B).

Taken together, these results show that upregulating the PN using σ³², which evolved to counter the protein folding stress caused by heat shock, enables E. coli to suppress aggregation even for highly destabilized proteins at high expression levels, consistent with previous results (Zhang et al., 2014). Moreover, our results argue that KJE, GroELS, and Lon are primarily responsible for the effects observed with σ³² overexpression.

Analysis of Test Protein Folding Fates

Understanding how the E. coli PN manages proteostasis for our test proteins requires a quantitative analysis of our data. Thus, we first extract protein-specific trends, so far as they exist, by a phenomenological method. We then model how a test protein’s response to the PN reports on its energy landscape by using FoldEco, a mechanistic model for proteostasis in E. coli.

Phenomenological Models

To quantify how KJE, GroELS, and Lon affect our test proteins, we fit the overexpression data for each test protein to the phenomenological models below:

$${[\text{Agg}]}_{\text{rel},\mathrm{X}}=c_{\text{Agg},\mathrm{X}}+a_{\mathrm{K},\mathrm{X}}{[\text{DnaK}]}_{\text{rel}}+a_{\mathrm{G},\mathrm{X}}{[\text{GroEL}]}_{\text{rel}}+a_{\mathrm{L},\mathrm{X}}{[\text{Lon}]}_{\text{rel}}$$ (1)

$${[\text{Sol}]}_{\text{rel},\mathrm{X}}=c_{\text{Sol},\mathrm{X}}+s_{\mathrm{K},\mathrm{X}}{[\text{DnaK}]}_{\text{rel}}+s_{\mathrm{G},\mathrm{X}}{[\text{GroEL}]}_{\text{rel}}+s_{\mathrm{L},\mathrm{X}}{[\text{Lon}]}_{\text{rel}}$$ (2)

where [Agg]_rel,X and [Sol]_rel,X are the concentrations of the aggregated and soluble forms of test protein “X” normalized to the total concentration under adapted-basal conditions; [DnaK]_rel, [GroEL]_rel, and [Lon]_rel are the concentrations of the PN components relative to their adapted-basal concentrations; a_K,X, a_G,X, and a_L,X are the gradients of [Agg]_rel,X with respect to [DnaK]_rel, [GroEL]_rel, and [Lon]_rel; s_K,X, s_G,X, and s_L,X are the gradients of [Sol]_rel,X with respect to [DnaK]_rel, [GroEL]_rel, and [Lon]_rel; and c_Agg,X and c_Sol,X are the model intercepts. Values for [Agg]_rel,X and [Sol]_rel,X under the various PN conditions can be read off the right-hand axes in Figures 3B–D and 4A–C. Values for [DnaK]_rel, [GroEL]_rel, and [Lon]_rel under all conditions are shown in Figures S2D and S3A. The gradient parameters quantify the efficacies of the PN components for a given test protein. For example, a protein that benefits greatly from GroELS would have a large, positive value of s_G,X (indicating that [Sol]_rel,X increases sharply as GroELS is upregulated) and/or a large negative value of a_G,X (indicating that [Agg]_rel,X decreases sharply as GroELS is upregulated). Because of the differences in the expression levels of the three test proteins and the inherent non-linearity of aggregation kinetics with respect to protein concentration, one must be cautious when comparing the magnitudes of the gradient parameters of different proteins. However, differences in the signs of the gradient parameters, and whether or not they differ significantly from 0, are not subject to such concerns. Also, we have chosen to use relative concentrations of the PN components in our model rather than their absolute concentrations for ease of presentation. However, we also report the gradient parameters scaled to the absolute test protein and PN component concentrations in Table S2.

The qualities of the fits of Equation (1) to the normalized concentrations of aggregated protein are moderate to good for all three test proteins (adjusted R² = 0.73 for m-EcDHFR, 0.83 for m-RA114, and 0.68 for m-MmCRABP1; Figure 5A, red data points, see Table S1 for fit residuals). The parameters a_K, a_G, and a_L are negative for each test protein, indicating that all of the PN components decrease aggregation (Figure 5B, red bars). KJE decreases the extent of aggregation the most for a given fold-change in its concentration, followed by Lon and GroELS (a_K,X < a_L,X < a_G,X for all test proteins X). This result is consistent with KJE specifically antagonizing misfolding and aggregation, whereas GroELS and Lon affect these processes less directly.

Figure 5. Dependences of Test Protein Folding Fates on Different PN Components Based on Phenomenological Fits of Overexpression Data.

(A) Plots of the experimental values of [Agg]_rel,X and [Sol]_rel,X of m-EcDHFR (left), m-MmCRABP1 (middle), and m-RA114 (right) from Figures 3 and 4 vs. the corresponding model-derived values from the fits of Equations (1) or (2). Red data points: data points for [Agg]_rel,X, fit with Equation (1). Blue data points: data points for [Sol]_rel,X, fit with Equation (2). Dashed line: Line through the origin with a slope of 1. The extent to which the data points fall on the dashed line indicates the goodness-of-fit of the model. The circled data points have the largest residuals in the fit of Equation (2) to the [Sol]_rel,RA114 data. (B) Bar graph showing the gradient parameters and their standard errors from the fits of Equations (1) (red bars) and (2) (blue bars). Negative values indicate that increasing the concentration of a PN component decreases the concentration of the aggregated (red bars) or soluble (blue bars) form of a test protein. Positive values indicate the opposite. The blue bars for m-RA114 are bordered by dashed lines because these parameter values were obtained from a poor quality fit. The p-values for the gradient parameters are indicated as follows: ***, p < 0.0001; **, p < 0.001; *, p < 0.01; and n.s., p > 0.05. Tests were not performed for s_K,RA114, s_G,RA114, or s_L,RA114 since the p-value for the fit as a whole was > 0.01. See also Tables S1 and S2.

The fits of Equation (2) to the normalized concentrations of soluble protein are good for m-EcDHFR and m-MmCRABP1 (adjusted R² = 0.82 for m-EcDHFR and 0.71 for m-MmCRABP1), but poor for m-RA114 (adjusted R² = 0.10) (Figure 5A, blue data points, see Table S1 for fit residuals). The poor fit for m-RA114 is likely due to the small range of [Sol]_rel,RA114 combined with its surprisingly low value when KJE+GroELS are jointly upregulated and its surprisingly high value when GroELS+Lon are jointly upregulated (highlighted data points in Figure 5A, right-hand panel). This observation suggests that the KJE and GroELS pathways may interfere with each other whereas the GroELS and Lon pathways cooperate to handle m-RA114.

The parameter s_L,X is negative for both m-EcDHFR and m-MmCRABP1, but s_L,DHFR is roughly the same as a_L,DHFR whereas s_L,CRABP1 is much smaller in magnitude than a_L,CRABP1 (Figure 5B). This observation indicates that upregulating Lon preferentially depletes aggregates for m-MmCRABP1, but not for m-EcDHFR. In addition, s_K,DHFR is close to 0, indicating that while KJE is very effective for diminishing aggregation for m-EcDHFR, it does not increase the concentration of soluble protein (Figure 5B). In contrast, s_K,CRABP1 is substantial and positive for m-MmCRABP1 (Figure 5B). The situation is reversed for the GroELS system: s_G,DHFR is substantial and positive but s_G,CRABP1 is much smaller (Figure 5B). These results suggest that m-MmCRABP1 may be a poor substrate for GroELS. While one should be cautious when interpreting this observation because of the expression level differences of m-EcDHFR and m-MmCRABP1, an analysis of the data using FoldEco (see below), which explicitly accounts for these expression level differences, corroborates this notion.

While fitting Equations (1) and (2) to our data has enabled us to quantify the effects of KJE, GroELS, and Lon on our test proteins, these equations are phenomenological and cannot inform us about the causes of a given protein’s behavior. Based on previous results with FoldEco, the extent to which a protein benefits from different chaperoning mechanisms should be a function of that protein’s folding energetics (Dickson and Brooks, 2013; Powers et al., 2012). Thus, the values of the best-fit parameters for Equations (1) and (2) should reflect the folding energetics of our test proteins. To explore this possibility, we used FoldEco to fit our data by using the folding energetics as adjustable parameters.

Analysis of Test Protein Folding Fates Using FoldEco: General

FoldEco comprises a system of ordinary differential equations that represent the kinetics of the processes undergone by proteins in vivo: synthesis, folding, misfolding and aggregation, interaction with KJE and GroELS, and degradation by Lon (Figure 1) (Powers et al., 2012). To fit FoldEco to our data, we have to vary the parameters in FoldEco until the model optimally matches the experimental data. FoldEco has many dozens of parameters, but many of these are likely to be independent of, or weakly dependent on, the nature of the protein (Powers et al., 2012). For example, chaperone–co-chaperone interaction parameters should not be differentially affected by bound substrates (that is, the effect of one bound substrate on chaperone–co-chaperone interaction parameters should be similar to the effect of another). Furthermore, chaperones bind to substrates promiscuously (Aoki et al., 2000; Landry and Gierasch, 1991; Rudiger et al., 1997; Wang et al., 1999)—GroEL can even chaperone a substrate composed entirely of D-amino acids (Weinstock et al., 2014)—so chaperone-substrate interaction parameters are likely to be similar for most substrates. Values for such parameters can be obtained from the literature, and the available data are broadly (though not perfectly) consistent with the assertions in the preceding sentences (see (Powers et al., 2012), and Figures S1–S5 therein). Thus, to a first approximation, the only adjustable parameters needed to fit FoldEco to our data are those for the folding energy landscapes and synthesis rates of the test proteins. These parameters include the folding rate and equilibrium constants (k_f and K_f); the misfolding rate and equilibrium constants (k_m and K_m); the aggregation rate and equilibrium constants (k_a and K_a, where K_a is the equilibrium constant for adding a misfolded monomer to an aggregate); and the steady-state protein synthesis rate (σ) (Figure 1). Importantly, FoldEco can account for how the different expression levels of the test proteins affect their overall behavior and in particular their tendency to aggregate via its built-in nucleated polymerization model for protein aggregation (Powers et al., 2012).

It is important to note that the parameters derived from FoldEco fits of our data are “effective parameters”, since we are applying FoldEco’s generic folding mechanism to the test proteins. For example, EcDHFR folds through a multi-step mechanism (Frieden, 1990). Thus, the single folding rate constant derived from FoldEco fits to the m-EcDHFR expression data (k_f) subsumes the rate constants for the individual folding steps. In addition, the equilibrium denaturation of m-RA114 reveals at least one intermediate (Figure S1B). Thus, the best-fit folding equilibrium constant for m-RA114 encompasses the energetics for all of the states in the native conformational ensemble. Nevertheless, we expect that the effective parameters derived from fitting FoldEco will faithfully capture the essences of the true folding and misfolding processes.

The qualities of the FoldEco fits range from moderate to good, with the best-fit values of [Agg]_rel,X and [Sol]_rel,X deviating from their experimental values on average by 0.05 for m-EcDHFR, 0.12 for m-MmCRABP1, and 0.11 for m-RA114 (Figure 6; see Table S1 for fit residuals). Unfortunately, only the fit for m-EcDHFR permitted parameters to be estimated with acceptable precision (Table S3). These are discussed in the next section. The FoldEco fits for m-MmCRABP1 and m-RA114, despite not providing definite parameter estimates, nevertheless define some relationships among the parameters and thereby enable us to discern some general features of the in vivo folding energy landscapes of these two proteins.

Figure 6. FoldEco-Derived Fits of Experimental Data from the Overexpression of Test Proteins.

As in Figure 5A, except that the model-derived relative concentrations are determined from the FoldEco fits to the data. The circled points are those for m-MmCRABP1 under the low- and high-upregulation GroELS conditions, which have particularly large residuals. See also Figure S4, Figure S5, Table S1, and Table S3.

Analysis of the Folding Fate of m-EcDHFR Using FoldEco

The effective folding parameters for m-EcDHFR are: k_m = 0.06 s⁻¹; K_m = 0.3 (corresponding to ΔG_m = +0.7 kcal mol⁻¹); k_f = 0.3 s⁻¹; K_f = 130 (corresponding to ΔG_f = −2.9 kcal mol⁻¹); and k_a = 4×10⁹ M⁻¹ s⁻¹ (see Table S3 for the errors in these parameters). Only a lower limit could be determined for K_a, which was 1.4×10⁹ M⁻¹ (corresponding to ΔG_a = −12.7 kcal mol⁻¹). These parameters indicate that m-EcDHFR folds about five times faster than it misfolds, but its native state is not very stable, and once it misfolds it aggregates rapidly to form stable aggregates (Figure 7A). In fact, the aggregation rate constant is around the diffusion-controlled limit, possibly indicating that m-EcDHFR aggregates directly from the unfolded state, rather than through a monomeric misfolded intermediate as assumed in FoldEco.

Figure 7. Partitioning of the Three Test Proteins under Adapted-Basal Conditions at t = 2 h Based on FoldEco Simulations.

(A) Summary diagram for m-EcDHFR, laid out as in Figure 1. The sizes of the colored circles indicate the concentrations of the unfolded (U), native (N), misfolded (M), and aggregated (A) states. The radii of the circles are proportional to the cube roots of the concentrations. Cube roots are used to enable the lowest and highest concentrations to be shown on the same diagram. The circles for “Synthesis” and “Degradation” represent the total concentration of protein synthesized and degraded over the 2 h time course of the simulation. The numerical values of each concentration are written below the circles. Blue text: best-fit biophysical parameters from the fit of FoldEco to the data. Red numbers: percentages of misfolded protein molecules that aggregate, engage the KJE recovery pathway, or engage the Lon degradation pathway. Black italic numbers: percentages of degraded protein taken from the unfolded or misfolded states. (B) As in (A), but for m-MmCRABP1. (C) As in (A) but for m-RA114. Note that qualitative descriptors are used for the biophysical processes in (B) and (C). See also Table S3.

It is of interest to compare the parameters obtained from our FoldEco fits to the analogous parameters measured in vitro. However, because parameters for processes that are not on the folding pathway, like the rate and equilibrium constants for misfolding and aggregation, can be difficult to determine, we limit our comparisons to the folding equilibrium and rate constants, which can be determined using well-established methods. The in vitro free energy and rate constant for folding of m-EcDHFR were found to be ΔG_f = −2.6 kcal mol⁻¹ and k_f = 1.1 s⁻¹ by equilibrium denaturation (Figure S1A) and fluorescence-monitored refolding kinetics (Figure S4; k_f here refers to the rate of tertiary structure acquisition, not subsequent steps required for NADP binding that were identified in past mechanistic studies of wt-EcDHFR folding (Frieden, 1990)). These values were determined at [m-EcDHFR] = 3.5 μM and 25 °C; aggregation was not detected under these conditions. The values determined for these parameters from our FoldEco fit were ΔG_f = −2.9 kcal mol⁻¹ and k_f = 0.3 s⁻¹ (Table S3, Figure 7A). The good correspondence between these parameter estimates and their experimental values supports our use of FoldEco to understand the in vivo folding of our test proteins.

Some aspects of the sensitivities of m-EcDHFR proteostasis to the PN components (quantified in Figure 5B) can be explained by the FoldEco fit to the m-EcDHFR expression data. For example, Lon affects the levels of aggregated and soluble m-EcDHFR about equally (a_L,DHFR = −0.080 vs. s_L,DHFR = −0.076; Figure 5B). Since the relatively slow misfolding of m-EcDHFR is followed by fast formation of stable aggregates (Figure 7A), the misfolded state of m-EcDHFR has a short lifetime and a low concentration, and Lon cannot intercept and degrade it before it aggregates. In contrast, Lon appears to preferentially diminish the aggregated state for m-MmCRABP1 (a_L,CRABP1 = −0.129 vs. s_L,CRABP1 = −0.014; Figure 5B). The same is qualitatively true for m-RA114 (Figure 3D), although a quantitative comparison cannot be made for m-RA114 because of the poor fit of Equation (2).

This point can be further illustrated using the FoldEco simulation of m-EcDHFR under adapted-basal conditions. In this simulation, only 0.1% of the protein molecules that sample the misfolded state are degraded by Lon (Figure 7A, red numbers), and furthermore, of the almost 188 μM of m-EcDHFR that is degraded over the 2 h time course of the expression simulation, only 0.2% is taken from the misfolded state (Figure 7A, black numbers). The remaining 99.8% of the m-EcDHFR that is degraded comes from the unfolded state. Since the unfolded state is the progenitor of both natively folded and aggregated protein, Lon affects the levels of each similarly for m-EcDHFR.

The short lifetime of the misfolded state also explains why KJE decreases levels of aggregated m-EcDHFR without increasing soluble m-EcDHFR. KJE affects the partitioning between aggregated and soluble protein only when it directly reverses the process of misfolding by converting misfolded protein to unfolded protein. Because misfolded m-EcDHFR exists so briefly and at such a low concentration, DnaJ and DnaK, like Lon, cannot bind to it before it aggregates. In our FoldEco simulation of m-EcDHFR under adapted-basal conditions, only 1% of the m-EcDHFR molecules that sample the misfolded state are engaged by KJE (Figure 7A, red numbers). Although KJE, especially with the assistance of ClpB, can recover protein from the aggregated state and return it to the unfolded state, this process happens after the “product determining step” of the protein folding pathway—that is, after unfolded protein is partitioned between folding and misfolding/aggregation pathways—and therefore cannot influence the folding/misfolding decision. The protein that the ClpB+KJE pathway recovers from aggregates is, however, susceptible to degradation, which at least partly explains why upregulating KJE leads to lower total protein levels for m-EcDHFR as well as m-MmCRABP1 and m-RA114.

Analysis of Folding Fates of m-MmCRABP1 and m-RA114 Using FoldEco

Although the fits of FoldEco to the data for m-MmCRABP1 and m-RA114 did not permit precise estimation of the biophysical parameters for these two proteins, the fits did define some relationships between the parameters that illuminate some general features of the folding energy landscapes of m-MmCRABP1 and m-RA114. For example, the folding rate constant (k_f) of m-MmCRABP1 can be increased from its optimal value with very little decrease in the quality of the FoldEco fit provided that, first, the misfolding equilibrium constant (K_m) also increases, and second, the misfolding rate constant (k_m) is much greater than k_f so that misfolding is nearly at equilibrium with respect to folding. Then, increases in k_f and K_m have offsetting effects on the unfolded-to-native-state flux. Given that this flux is equal to the product k_f × [U] (where [U] is the concentration of unfolded protein), increasing k_f directly increases this product, while increasing K_m decreases this product by shifting the misfolding equilibrium away from the unfolded state, thereby decreasing [U]. The overall effect of these combined changes on the behavior of m-MmCRABP1 is therefore minimal.

Despite the uncertainties in the parameters from the FoldEco fits for m-MmCRABP1 and m-RA114, certain aspects of their folding energy landscapes are nevertheless clear. Both m-MmCRABP1 and m-RA114 misfold faster than they fold (k_m > k_f) but do not aggregate as fast as m-EcDHFR, resulting in appreciable accumulation of misfolded protein under adapted-basal PN conditions (Table S3; Figure 7B, C). This behavior can explain the response of m-MmCRABP1 and m-RA114 to KJE and Lon overexpression. The higher concentration of misfolded protein increases the utilization of the KJE recovery and Lon degradation pathways by almost 10-fold compared to m-EcDHFR (compare Figure 7A to Figure 7B and C, red numbers). Thus, Lon preferentially reduces the concentration of aggregates of m-MmCRABP1 and m-RA114 by degrading mainly misfolded protein. Similarly, KJE can directly convert misfolded m-MmCRABP1 and m-RA114 to the unfolded state, simultaneously increasing the levels of soluble protein and decreasing levels of aggregated protein.

The FoldEco fits, however, are unable to account for the poor chaperoning of m-MmCRABP1 by GroELS. According to the FoldEco fits, all of the test proteins should benefit similarly from GroELS upregulation. In reality, m-MmCRABP1 benefits much less from GroELS than do m-EcDHFR or m-RA114 (Figure 3B–D), as can be seen in the residuals of the FoldEco fit for [Agg]_rel,CRABP1 and [Sol]_rel,CRABP1 under the GroELS-low and -high upregulation conditions (Figure 6, middle panel, highlighted data points; Table S1). These residuals average 0.21 for m- MmCRABP1, compared to 0.05 for m-EcDHFR and 0.12 for m-RA114, suggesting that m-MmCRABP1 may interact differently with GroEL than the other test proteins do.

To explore this notion, we fit FoldEco to the expression data for m-MmCRABP1 while varying the equilibrium association constant between m-MmCRABP1 and GroEL (K_Gro–CRABP1). The best fit was obtained when K_Gro-CRABP1 was 1000-fold lower than its default value (10³ M⁻¹ vs. 10⁶ M⁻¹; Figure S5A), in which case the overall FoldEco fit improved by about 20% (mean residuals = 0.10 vs. 0.12), and the residuals for the GroELS upregulation conditions decreased from 0.21 to 0.15 (Figure S5B). The improved fit still did not yield precise estimates of the biophysical parameters of m-MmCRABP1, but it is clear that the fit improves because of the large decrease in K_Gro–CRABP1, suggesting that m-MmCRABP1 binds weakly to GroEL. This observation based solely on fitting the folding fate of this protein under different PN conditions is consistent with experimental findings in vitro that GroELS has no effect on MmCRABP1 folding and that there is no detectable binding between GroEL and incompletely folded MmCRABP1 variants (I. Budyak, H.-P. Feng, and L. M. Gierasch, unpublished results).

We note that adding more adjustable parameters does not dramatically improve the FoldEco fits for m-EcDHFR or m-RA114, nor does varying parameters beyond K_Gro–CRABP1 yield further substantive improvements to the FoldEco fit for m-MmCRABP1. Thus, the default, literature-derived parameters used in FoldEco appear to be sufficiently accurate in most cases, and when they are not, a lack of fit can indicate an unusual interaction between a PN component and a substrate.

DISCUSSION

Our results show that increasing the levels of PN components enables cells to better handle situations that involve protein misfolding, consistent with previous studies of the effect of chaperone upregulation on, for example, the yields of overexpressed heterologous proteins (de Marco, 2007; de Marco et al., 2007; Makino et al., 2011; Zhang et al., 2014), protein evolution (Bershtein et al., 2013; Bogumil and Dagan, 2012; Queitsch et al., 2002; Tokuriki and Tawfik, 2009), and stress tolerance (Feder et al., 1996; Welte et al., 1993). The “pro-folding” (GroELS), “anti-misfolding” (KJE), and “concentration control” (Lon) arms of the E. coli PN form an efficacious triad for maintaining proteostasis for a broad range of proteins, in that the concentration of soluble protein is generally maintained or increased by chaperone upregulation even when the total amount of protein decreases (see Figures 3 and 4). These arms are especially effective when upregulated in their native ratios as part of the σ³² transcriptional program (Figure 4). The two quantitative models of E. coli proteostasis, the phenomenological Equations (1) and (2) and FoldEco, provide insight into how these systems contribute to proteostasis, and how their contributions depend on their substrates’ folding energy landscapes.

The two models, which generally explain 60–80% of the variation in our data (Figures 5A and 6), share two important features: first, the only components of the PN that they include are KJE, GroELS, and Lon; and second, they do not account for any direct mechanisms by which these components can collaborate. The success of the models in fitting the data in light of the first feature suggests that KJE, GroELS, and Lon are the primary contributors to the proteostasis of our test proteins. This does not mean that other PN components, like ClpB, HtpG, and IbpA/IbpB are not important, but rather that their importance lies outside generic proteostasis in the absence of environmental stress. ClpB is vital for recovery from heat shock (Doyle et al., 2013; Squires et al., 1991; Weibezahn et al., 2004), but it is not necessary for maintaining normal growth rates under non-stress conditions (Squires et al., 1991). Overexpressing ClpB in our system has little effect on the folding fates of our test proteins, probably because protein disaggregation by ClpB+KJE is simply not as fast as protein synthesis. Similarly, HtpG may be important for the folding of specific substrates (Yosef et al., 2011), but it is not necessary for either normal growth or thermotolerance (Bardwell and Craig, 1988).

The ability of the models to fit our data without invoking collaboration among KJE, GroELS, and Lon suggests that these PN components act independently under our experimental conditions and that their effects are therefore largely, though not perfectly, additive. This assertion is corroborated by the residuals of the fits of our models to the test protein data, ~70% of which are less than ±0.1 and ~90% of which are less than ±0.2 (Table S1). However, we emphasize that many examples of collaboration between chaperone systems have been reported. In eukaryotes, the Hsp70 and Hsp90 systems work together in the proteostasis of, for example, steroid hormone receptors (Pratt and Toft, 2003; Wegele et al., 2006). PN components may also “hand off” substrates from one to another—for example, DnaK to GroEL (Langer et al., 1992)—and this sequential activity is important for in vivo protein folding. However, the large amounts of unfolded protein that are being produced in our experiments may lead to unoccupied chaperones binding to newly synthesized protein before they can accept the transfer of a substrate from another chaperone, thereby suppressing chaperone collaboration.

While the major PN systems in E. coli appear to act independently under high protein folding loads, there may still be occasional exceptions to this rule. Such an exception may have caused the poor fit of Equation (2) to the [Sol]_rel,RA114 data. While we can only speculate that direct interactions between PN components is the reason that Equation (2) does not fit the [Sol]_rel,RA114 data well, it remains an intriguing possibility that some client proteins can access chaperone collaboration pathways that others cannot.

FoldEco simulations allow us to go beyond the phenomenological treatment of Equations (1) and (2) because they enable correlations between folding energetics and chaperone mechanisms. For example, the response of m-EcDHFR to KJE and Lon was due to a combination of slow misfolding and fast aggregation. Even the cases where a good fit of the model to the data was elusive could be a clue to underlying biophysics. The poor fit of FoldEco to the data for m-MmCRABP1 under GroELS upregulation suggested weak binding of m-MmCRABP1 to GroEL. It should be noted that some aspects of proteostasis that could contribute to protein misfolding and aggregation are not yet modeled in FoldEco, like the effect of translation rates on proteostasis. Given that the m-MmCRABP1 and m-RA114 genes were not codon optimized for E. coli, such effects could contribute to the high aggregation propensities observed for these proteins. However, previous FoldEco simulations of luciferase expression found that ignoring the effect of co-translational folding did not greatly diminish the performance of FoldEco (Powers et al., 2012).

FoldEco has enabled us, in a sense, to invert the usual reductionist process for studying in vivo protein folding. Rather than analyzing our test proteins’ folding in vitro and then using this information to rationalize their behavior in vivo, we have used FoldEco to translate the behavior of our test proteins in vivo into information about their folding energy landscapes. In addition, since FoldEco embodies our best understanding of chaperone mechanisms based on the wealth of biochemical literature that exists on this subject, using FoldEco to fit our data constitutes a test of this understanding. We believe that the results of these fits were good enough to indicate that this understanding is fundamentally sound, albeit incomplete. We expect that as FoldEco is refined and expanded, its ability to explain the nuances of how proteostasis is managed for proteins with different folding energy landscapes will improve, providing a foundation for understanding processes like organismal stress responses and protein evolution that are intimately linked to protein folding energetics.

EXPERIMENTAL PROCEDURES

E. coli Strains and Plasmids

E. coli K12 strain HMS174 (DE3) was used for the overexpression experiments. Vectors under pBAD and/or pTet promoters on low copy number plasmids were used to express chaperones, Lon, and I54N σ³². Test proteins were expressed using a low copy number pET29b vector. Further details can be found in the Supplemental Experimental Procedures.

Measurement of the Levels of Aggregated and Soluble Protein

Overexpression of desired PN components was induced by adding arabinose to E. coli harboring a plasmid containing the genes for the desired PN component(s) and a plasmid containing the gene for the desired test protein. After 1 h of overexpressing the PN component(s) at 30 °C, test protein overexpression was induced by adding IPTG. After 2 h at 30 °C, cells were lysed by sonication at 4°C. Portions of the lysates were centrifuged for 10 min at 13500×g at 4°C. Test protein in the supernatant and pellet were defined as the soluble and aggregated fractions, respectively. The total (i.e., not centrifuged), soluble, and aggregated fractions were analyzed by SDS-PAGE (12 or 15 % Biorad gel) and then stained with Coomassie blue (m-EcDHFR) or subjected to western blot analysis (m-MmCRABP1 and m-RA114) to quantify protein levels. Adapted-basal PN controls were included in all gels so that the total test protein expression levels under perturbed PN conditions and adapted-basal PN conditions could be directly compared. Further details can be found in the Supplemental Experimental Procedures.

Fits of Equations (1) and (2) and FoldEco to Test protein Overexpression Data

Equations (1) and (2) were fit to the data by linear regression. FoldEco was fit to the data using a least squares approach. Details can be found in the Supplemental Experimental Procedures.

HIGHLIGHTS.

• DnaK/DnaJ/GrpE, GroEL/GroES, and Lon dominate the E. coli proteostasis network.

• The effects of the proteostasis network components are largely additive.

• Modeling reveals relationships between chaperone efficacy and folding energetics.