Transcriptome and Proteome Exploration to Model Translation Efficiency and Protein Stability in Lactococcus lactis

This genome-scale study analysed the various parameters influencing protein levels in cells. To achieve this goal, the model bacterium Lactococcus lactis was grown at steady state in continuous cultures at different growth rates, and proteomic and transcriptomic data were thoroughly compared. Ratios of mRNA to protein were highly variable among proteins but also, for a given gene, between the different growth conditions. The modeling of cellular processes combined with a data fitting modeling approach allowed both translation efficiencies and degradation rates to be estimated for each protein in each growth condition. Estimated translational efficiencies and degradation rates strongly differed between proteins and were tested for their biological significance through statistical correlations with relevant parameters such as codon or amino acid bias. These efficiencies and degradation rates were not constant in all growth conditions and were inversely proportional to the growth rate, indicating a more efficient translation at low growth rate but an antagonistic higher rate of protein degradation. Estimated protein median half-lives ranged from 23 to 224 min, underlying the importance of protein degradation notably at low growth rates. The regulation of intracellular protein level was analysed through regulatory coefficient calculations, revealing a complex control depending on protein and growth conditions. The modeling approach enabled translational efficiencies and protein degradation rates to be estimated, two biological parameters extremely difficult to determine experimentally and generally lacking in bacteria. This method is generic and can now be extended to other environments and/or other micro-organisms.


Introduction
In the era of ''omics'', systems biology has emerged with the availability of genome-wide data from different levels, i.e. genome, transcriptome, proteome, metabolome [1,2]. This approach aims at integrating omics data, mainly through computational and mathematical models [3,4] so as to decipher biological systems as a whole [5]. The integration of transcriptomic and proteomic results is a huge challenge by itself. The literature usually exploits these two approaches as complementary tools and does not often provide a correct confrontation of the two datasets. Until now, only a few researchers, mainly interested in yeast physiology [6,7], have been working on this aspect and the results typically revealed modest correlations between those two datasets [8][9][10]. These weak correlations between transcript and protein levels can be the consequence of the involvement of post-transcriptional regulations [11], such as translation control and protein degradation as evidenced by Brockmann et al. [12]. Translation regulations are believed to be involved in protein level control but are generally studied at the level of controlling specific molecular mechanisms and not at the genome scale [13][14][15]. Although polysome profile analysis allows translation efficiencies to be experimentally determined for the various transcripts simultaneously, this technique has been only rarely used and almost exclusively for S. cerevisiae [16]. Protein stability can also influence intracellular protein level and the correlation between transcript and protein [10,17,18]. However protein stability is rarely studied at the genome scale and data are only available for S. cerevisiae [19,20]. Finally, the rate of protein disappearance due to protein dilution by cellular growth is also potentially involved in protein level modifications but this physical phenomenon is generally neglected. More generally, even if translation efficiency, protein degradation and dilution rate can all influence protein levels, these parameters are not usually studied simultaneously. The role of each parameter in a whole cellular adaptation process has not been elucidated and it is not clearly known today which parameter is preponderant and if the control is constant or not when environmental conditions are modified.
The aim of this study was to analyse the control of intracellular protein level taking into account all the parameters of this control, in a prokaryotic organism, the model of lactic acid bacteria, Lactococcus lactis. To achieve this purpose, transcriptomic and proteomic analyses were performed with cells from the same were found in higher quantity, while on the other hand, the cold shock associated protein CspE, ClpC and the adaptation related peroxidase Tpx, had decreased levels in response to growth rate increase. Besides those opposite punctual regulations, other proteins encoding important functions involved in stress protection such as ATPases or peptidases (excepting PepP), were present at constant levels, independently of the growth rate. This lack of general tendency observed here at proteomic level was also observed at transcriptomic level [21]. In contrast, a wide downregulation of genes involved in stress protection was observed in yeast when growth rate was increased [23,24]. Finally, one can notice that the two single phage-related proteins measured in those proteomics experiments showed significantly reduced levels at high growth rate. This last observation can be connected with the previously described massive down regulation of the expression of phage-related genes [21].
Transcriptomic and proteomic analyses were performed with cells collected simultaneously from the same fermentor; thus data can be strictly compared. Proteins and their corresponding transcript levels were compared individually. Transcriptomic data were already available [21] but were nevertheless re-processed so as to obtain concentrations rather than abundances (see Materials and methods). For proteomic data, concentration and abundance values are expected to be similar (see Materials and methods). For each growth rate, transcriptomic and proteomic mean values with their standard deviations are given in supplementary data (Table  S1). The mRNA/protein ratios were not constant for the different genes since, at a given growth rate, data were spanned among five orders of magnitude ( Figure 1C). These variations were linked both to protein and mRNA changes though protein concentrations were globally more spanned than mRNA concentrations (4 and 2 log of magnitude respectively; see figure 1A and 1B). mRNA/protein ratios were compared between two conditions using the lowest growth rate (0.09 h 21 ) as reference ( Figure 1D) and they globally increased with the growth rate. This tendency was confirmed when we analyzed similarly data corresponding to the maximum growth rate of 0.88 h 21 . These last data, also available in our group, were obtained in batch culture during the exponential growth phase, since this high growth rate could not be reached in continuous culture without any wash out of the cells from the chemostat [25].

Modeling of cellular process
What normally occurs in bacterial cells is the transcription of genes into mRNA, which are then translated into proteins that can be either diluted by growth or degraded ( Figure 2). Hence, protein concentration is determined not only by translation rates but also by dilution and degradation, therefore the following balance equation can be written: Rates of mRNA translation or protein degradation/dilution are assumed to be not constant and related to mRNA or protein concentration respectively. Biological rates were expressed as first order kinetics of their substrate concentration as previously postulated in L. lactis [26,27] but also in yeast strains [19,28]. Such modeling approach at the genomic scale is rare in the literature and dynamic experimental data allowing more elaborated kinetics to be hypothesized are not available. Making more complex those rate expressions would thus not make sense today. Dilution constant corresponds to the growth rate (m), degradation

Author Summary
This work is in the field of systems biology. Via an in-depth comparison of proteomic and transcriptomic data in various culture conditions, our objective was to better understand the regulation of protein levels. We have demonstrated that bacteria exert a tight control on intracellular protein levels, through a multi-level regulation involving translation but also dilution due to growth and protein degradation. We have estimated translational efficiencies and protein degradation rates by modeling. These two biological parameters are extremely difficult to measure experimentally and have not been previously determined in bacteria. We have found that they are growth rate dependent, indicating a fine control of translation and degradation processes. We have worked with the small genome bacterium Lactococcus lactis on a limited number of mRNA-protein couples but keeping in mind that this approach could be extended to other micro-organisms and biological phenomena. We have exhibited that mathematical modeling associated to experimental steady-states cultures is a powerful tool to understand microbial physiology.
constant (k0) is proportional to protein half-life (t 1/2 = ln(2)/k0) and k9 represents the translation efficiency. At steady state, the various concentrations are expected to remain constant, the time derivative of the protein concentration is equal to zero and the previous equation can be simplified and reorganized as follows: In the chemostat cultures at the various growth rates, cells are at steady state; similarly, during the exponential growth phase, cells are physiologically stable and are also considered to be at steadystate [29]. The previous observation, establishing a relationship between mRNA/protein ratios and the growth rate for these four steady states (see above), is in accordance with this last equation (2). 171 different mRNA and protein couples were available in each repetition of the various steady states (intersection of 308 couples in the 3 chemostat steady states and 191 in the batch). For only a few proteins were probes missing on the microarray; hence it was not possible to rebuild these couples. In order to estimate translation efficiency and protein degradation rate, the best mathematical solution to the equation (2) was sought, using numerical estimations performed on Matlab. The k9 and k0 values were postulated to be positive, in accordance with biological reality. Different solutions with k9 and/or k0 constant, directly or inversely proportional to m were investigated. Estimation of the best fitting solution was based on the least square criterion [30]. For the 171 couples, the mean sum of the squared residuals (difference between a ratio and its estimation) associated to every combination are given in Table 2. Considering the lower mean sum of the squared residuals, the best solution was obtained when both k9 and k0 were proportional to 1/m (k9 = a/m and k0 = b/m). Hence the equation (2) could be written as follows: The mRNA/protein ratios were thus linked to the growth rate (m) through a polynomial function of order two (m 2 ), which is consistent with the visual observation of the various curves (not shown). For each mRNA-protein couple the reliability of the two estimated constants a and b was evaluated by their associated R 2 . All regression coefficients are listed in Table 3. The mean linear coefficient (R 2 ) associated to this model was 0.8360.04. Finally, the consistency of our modeling approach was checked when removing the data of the batch exponential growth phase from the analysis. A high mean R 2 of 0.7760.12 was still obtained using chemostat data exclusively. On the contrary, data not at steady state coming from other growth phases in batch cultivation could not be included. Indeed when taking into account mRNA/protein values during growth deceleration or during stationary phase, R 2 was strongly affected and dropped to 0.2160.03 and 0.2460.04 respectively.
The model (1) states that translation rate is proportional to the concentration of mRNA species which assumes that translation is mRNA-limited. An alternative hypothesis, would be the saturation of the ribosome with mRNA, as previously postulated in E coli  and k0 [protein] where k9 is the translation efficiency, m the growth rate and k0 the degradation rate constant. doi:10.1371/journal.pcbi.1000606.g002 [31], which implicitly suggests competition of any specific mRNA with all others to be the determining factor in synthesis of the corresponding protein. We have thus tested the model (2) with mRNA abundances rather than concentration values. The modeling approach was robust since similar k9 and k0 dependencies to the growth rate (k9 = a/m and k0 = b/m) were obtained ( Table 2; data in italic). However the expression of individual protein/mRNA as a function of m 2 (3) had generally lower R 2 (only 48 couples with R 2 $0.90 compared to 130 with concentrations), indicating the modeling approach with abundances was less satisfactory than with concentrations.

Model predictions and biological relevance
In order to carry on our modeling approach, the data corresponding to mRNA concentrations were filtered and only couples with R 2 $0.90 (130) were retained for further analyses (Table 3). It could be noticed that among the 41 eliminated couples, 15 displayed non monotonous evolutions of their mRNA/ protein ratios against m 2 and 20 had very low mRNA/protein ratios, which were thus more sensitive to errors.
The k9 and k0 coefficients were numerically calculated for each protein and in each growth condition from the values of a and b (Table 2) by the relation k9 = a/m and k0 = b/m. Value distributions ( Figure 3) demonstrate the wide variability of k9 and k0 among proteins but also between growth conditions. The k0 decreases when growth rate increases which is consistent with the general idea that protein degradation is high in stationary phases [32]. Protein half-lives were calculated and median values of t 1/2 were respectively 23, 61, 119 and 224 min for 0.09, 0.24, 0.47 and 0.88 h 21 . These values are in the same order of magnitude as those obtained recently for S. cerevisiae (mean 43 min at a growth rate of 0.1 h 21 [19]). Like k0, the translation efficiency k9 is also expressed as 1/m function, indicating that, when the degradation process increases at low growth rate, the translation efficiency is also increasing in order to attenuate this negative biological effect.
Due to the restricted size of the dataset but also to the nonuniform distribution of detected proteins in the various functional categories, it was not possible to use statistical tests to rigorously determine functional enrichments in extreme values of k9 or k0. However, among the 15 genes that are translated the most efficiently (highest a values in Table 3), one can notice the overrepresentation of genes involved in major cellular processes: the Tig chaperone [33] and proteins involved in replication (HslA, which can unwind DNA and plays a role in its supercoiling, [34]), and translation (ribosomal proteins: RplA, RplF, RplK, RplN, RpsT). Carbon metabolism is also represented by 7 proteins (GapB, EnoA, FbaA, Pmg, Pyk, TpiA, Ldh), all belonging to glycolysis which is the major metabolic pathway for energy production in L. lactis. The extremely stable proteins correspond to null values of b, and consequently k0, were represented by a group of 26 proteins. Remarkably, half of them were related to stress responses: ClpE protease, PepC and PepP peptidases, three reductases that are usually linked to oxidative stress (AhpC, TrxB1, YpjH), but also MurF, involved in parietal structure, YtgH, which is homolog to Staphylococcus aureus alkaline stress protein [35], YtaA and YahB two hypothetical protein sharing homologies with E. coli universal stress protein Usp [36], YuhE, whose E. coli homologue is involved in copper resistance [37], and two cysteine desulfurases (YeiG and YseF) whose corresponding genes in E. coli are involved in oxygen and copper stress responses [38]. Moreover, those extremely stable proteins are rather in the last third for translation efficiency. Thus L. lactis may limit degradation of stress-related proteins so as to maintain a minimal pool ready to use in case of emergency, which is biologically relevant.
Biological determinants of translation efficiency and protein stability were investigated through correlation studies. Correlations providing a Spearman coefficient (R Spearman ) with associated pvalues lower than 0.05 were considered as significant. The codon adaptation index (CAI) positively correlates with k9 (R Spearman = 0.57). Since CAI directly reflects translation efficiency during the elongation step [39], this result validates our translation efficiency estimations. Translation efficiency is also tightly related to the amino acid composition of proteins. A negative correlation of k9 was obtained with tyrosine, cysteine, histidine, aspartic acid and isoleucine frequencies while lysine and alanine richness had a positive influence ( Table 4). The amino acids the most used have a positive influence on k9 whereas those with a negative effect are the less frequent ones ( Table 4). The single exception is for isoleucine, but since it is the limiting nutrient it is not surprising to find it negatively correlated with translation efficiency, despite its high frequency in L. lactis proteins. This amino acid bias, together with the codon bias (CAI), shows that translation efficiency is strongly dependent of the gene sequence. This optimized functioning state is probably the result of a long evolutionary process. Finally it was found that translation efficiency is affected by protein length: the longer the protein, the more k9 decreases (R Pearson of 20.18). This negative correlation with length has already been reported for yeast [19] and can possibly be explained by a decrease of the ribosome density on long mRNA as previously shown for S. cerevisiae [40]. The only apparent correlation emerging for protein degradation constants k0 is a negative influence of cysteine richness (Table 4).

Controlling mechanisms
Degradation and dilution by growth are both involved in protein disappearance and are competitive reactions. The Table 2. Mean sum of squared residuals associated to different solutions to solve equation (2).   degradation and dilution constants, k0 and m, can be directly compared. The k0 is higher than m at low growth rate but becomes lower after a critical value of 0.39 h 21 (Figure 4). The role of the degradation may thus be major at low growth rate while dilution may become the main phenomenon at fast growth.   More generally, variations in protein concentration between two conditions can be related to changes in the three rates: protein synthesis, degradation and dilution. In order to better understand this regulatory node and identify which are the major controls, the quantitative involvement of the different actors was analyzed. Regulation coefficients corresponding to the protein level control were estimated with a method based on the one developed on S. cerevisiae [41,42]. Derivation of equation (2) leads to the following relationship: . The values of r t were used to elucidate the nature of the control and are given in supplementary data (Table   S2). If r t #0, protein disappearance is the major controlling mechanism; if r t $ ? 1, it is translation; and if 0 ? , r t ? , ? 1, the control of protein is shared. The nature of the control for a given protein and its strength differed in the various the growth rate intervals. However a constant control by disappearance was observed for 6 proteins distributed all over the metabolism (Als, GreA, LplL, PyrC RplQ, ThrC, see Table 3 for associated functions). Inversely the unknown protein YpdC was the single one constantly controlled by translation process. Independently of the growth rate, protein levels are mostly controlled by disappearance    (Table 5). Translation control strongly decreases, and protein level control becomes less specific and more and more shared with increasing growth rates. Similar conclusions were valid when r d was used for the control analysis instead of r t (data not shown).

Discussion
The comparison of mRNA and protein ratios revealed a strong heterogeneity among genes but also for a given gene, at different growth conditions. Variability among genes has recently been reported for the model yeast S. cerevisiae but these ratios remained constant between the two studied conditions, i.e. a rich and a poor media [43]. Though lacking in this publication, the maximum growth rates of S. cerevisiae were estimated to be 0.46 and 0.35 h 21 respectively in a rich and a poor media (Parrou J.L., personal communication). Thus it is postulated that the growth rate difference between these two conditions was too small to induce changes in mRNA/protein ratios.
The combination of two modeling approaches, one based on biological knowledge and the other on experimental data fitting, has enabled translation efficiency and protein degradation rate to be determined for each protein, phenomena which have been shown to be protein specific and growth rate dependant. The positive correlation of translation efficiency with codon bias in L. lactis is consistent with the results obtained for the yeast, though translation efficiencies have been calculated differently [12]. The presence of genes related to major cellular processes essential for growth were marked among the best translated. This finding corroborates what was found in archaebacteria for ribosomal proteins [44]. In L. lactis, the growth-rate dependant variations in translation efficiency are probably not related to changes in the amount of intracellular ribosomes if the constant ratio between mRNA and ribosomal RNA (see Material and methods) is taken into account. However one has to bear in mind that rRNA does not necessarily means assembled and/or active ribosomes. It is known for example that E. coli ribosome activity can be modulated by the inter-conversion between a functional 70S and a dimerized 100S inactive form [45]. To resolve this question, it will be necessary to investigate genome-wide ribosomal activity via polysome distribution which would provide key information to decipher the regulatory processes controlling translation. Polysome profile technology is already available for yeast but may be difficult to adapt to bacteria due to the co-localisation of transcription and translation in the cytoplasm.
Protein half-lives for the whole genome have never been determined nor estimated in any bacteria and data are only available in the literature for S. cerevisiae. However studies disagreed in terms of average half-life values: 31 h for Pratt et al. against 43 min for Belle et al. [19,20]. Those differences could be explained by methodological reasons since one study used pulse chase experiments [20] whereas the other one consisted in a direct measurement of each epitopetagged proteins [19]. In our study, for L. lactis, protein median half-lives ranged from 23 to 224 min. These low values are in good agreement with most recent values obtained for S. cerevisiae [19] and indicate that protein degradation is considerably more rapid than was once believed. Degradation rates in L. lactis were negatively correlated to cysteine content in proteins. In yeast, stable proteins were previously found to have a higher valine density whereas unstable ones are enriched in serine [19]. It is difficult to strictly compare those results since amino acid bias may be species specific and reflect the particularities of proteases involved in protein degradation. The negative correlation with cysteine could nevertheless be related to the potential formation of disulfide bridges known for stabilizing proteins [46]. The current work also revealed the presence of stress related proteins among the most stable. This last observation differs from results obtained in yeast indicating that ribosomal proteins and enzymes from amino-acids metabolism have the higher half-life [19]. This high stability of stress protein together with the lack of global transcriptional stress response observed in L. lactis when the growth rate is changed clearly underlines differences of stress adaptation mechanisms between the two micro-organisms.
Protein degradation exerts a major role in the cellular adaptation process since protein half-live data depend on the growth rate (1/m function). Moreover, the degradation rate is even higher than dilution rate at low growth rate (Figure 4). Considering that protein degradation is an ATP consuming process [47], high protein degradation at slow growth rate may contribute to the increase of maintenance energy that is generally observed in such conditions [48]. Like protein degradation, translation efficiency is also increased at slower growth rates. Effects of translation efficiency and protein degradation are thus antagonist and this mode of regulation is probably dedicated to attenuate biological changes. Inversely, proteins with the lowest degradation rates also corresponded to low translation efficiencies. The analysis of the regulation involved in the control of protein concentrations demonstrated that it is not constant in the different ranges of growth rate. At low growth rates, disappearance seems to be the main controlling mechanism, which could be attributed to high degradation rate. At high growth rate, the control becomes more complicated with some proteins regulated at the level of synthesis, disappearance or both (shared control). This increased complexity is consistent with cells approaching their maximum growth performance.
With this modeling approach, we have estimated translational efficiencies and protein degradation rates. These two biological parameters are extremely difficult to measure experimentally and have even never been previously determined in bacteria. The method was based on an in depth comparison of proteome and transcriptome data and was developed with the small genome bacterium L. lactis on a limited number of mRNA -protein couples (171). It will be possible in the future to broaden these couples since other proteomic methodologies, such as the APEX technology [8], allow more proteins to be detected. The approach remains generic and can be applied to all microorganisms. Modelling equations were solved because steady-states cultures were used: chemostat fermentation technology enabling steady states to be studied has thus proved to be a powerful tool to understand microbial physiology. We have demonstrated that bacteria exert a sharp control on intracellular protein levels, through a multi-level regulation involving three growth rate dependant actors: translation, dilution and degradation. Here, the growth rate was changed via chemostat cultures, but such growth rate modifications are also encountered in nature when cells have to face new environments. In this case, the adaptation process involves growth rate adaptation as well as other specific metabolic adaptations. It remains to be determined how the protein control is exerted in such natural environment.

Growth conditions
Lactococcus lactis ssp. lactis IL1403, whose genome has been entirely sequenced [22], was grown as previously described [21]. Briefly, three different growth rates have been studied, namely 0.09, 0.24 and 0.47 h 21 during anaerobic chemostat cultures (under nitrogen atmosphere and regulated pH) on a chemically defined medium limited by isoleucine concentration. For each steady-state, samples have been harvested in at least quadruplicate with a minimum delay of five doubling time between each sampling.

Transcriptomic data reprocessing
Transcriptomic data (geo platforms GSE10256 [21] for chemostat culture and GSE12962 for batch exponential phase) were already available. Briefly, these transcriptomic analyses had been obtained with a constant amount (10 mg) of total RNA (mRNA, ribosomal RNA and transfer RNA) labeled by retrotranscription ( 33 P) and hybridized on nylon membrane as previously described [49]. Three independent biological repetitions were used. These transcriptomic data had been normalized by all spots' mean intensity and thus corresponded to mRNA abundances. They were reprocessed here in order to calculate mRNA concentrations with the method previously described [49]. Raw data were first standardized by the all spots' mean intensities of the reference membrane (and not with its proper membrane) in order to eliminate the bias of the radioactivity level between the various repetitions and then corrected by total RNA concentration in order to take into account changes in intracellular RNA yield in the cells. Consistent with previous results [50], this yield increased significantly with the growth rate in L. lactis (3.5860.39, 4.9260.52, 7.3460.28 and 11.0660.23 g for 100 g cell dry weight at m = 0.09, 0.24, 0.47 and 0.88 h 21 respectively). Since the amount of RNA to perform transcriptomic analysis is maintained constant in order to avoid retro-transcription labelling bias, these RNA yield changes are completely hidden by the technology.
The total raw intensity of the membrane without any normalisation represents the amount of mRNA in the RNA sample used for transcriptomic analysis (10 mg). This total intensity was constant at each growth rate and lower than the saturation threshold (mean value of 16606584, 14626383, 13896366, 14746367, 14966425 at m = 0.09, 0.24, 0.47 and 0.88 h 21 respectively). Thus, it can be deduced that the ratio mRNA/total RNA was constant and assuming that ribosomal RNA is the major component of total RNA we can postulate that the fraction mRNA/ribosomal RNA is independent of the growth rate.

Proteomic analyses
For each condition, three repetitions were performed with independent cultures, extractions and electrophoresis. Bacteria were harvested from the cultures and cell pellets were washed twice with ice-cold 200 mM Na-phosphate, pH 6.4 and resuspended in 4 ml of 20 mM Na-phosphate buffer, pH 6.4, 1 mM EDTA, 10 mM tributylphosphine, a cocktail of protease inhibitors (P8465; Sigma Aldrich, St Louis, MO) 20-fold diluted and catalase 40 U/ml (C3155; Sigma Aldrich, St Louis, MO) to limit isoform formation. The cell suspension (approximately 35 units of optical density at 600 nm [OD600]/mL,) was transferred to the pre-cooled chamber of a BASIC Z cell disrupter (Celld, Warwickshire, United Kingdom) and was subjected to a pressure of 2,500 bars. The suspension was centrifuged at 5,0006g for 20 min at 4uC to remove unbroken cells and large cellular debris. The supernatant was collected and centrifuged at 220,0006g for 30 min at 4uC. The total protein concentration in the resulting supernatant (cytosolic fraction) was determined with the Coomassie protein assay reagent (Pierce, Rockford, IL) using bovine serum albumin as standard and was included between 1 and 2 mg/mL. The cytosolic fraction was aliquoted and stored frozen at 220uC.
2-Dimensional electrophoresis. A volume of cytosolic fraction corresponding to 350 mg or 500 mg (for basic gels) of proteins was incubated with nuclease (benzonase, Novagen 70664-3; 25 U for 100 mL of cytosolic fraction) for 30 min at 37uC and then chilled on ice and precipitated with 75% (vol/vol) methanol. The protein pellet was resuspended in 500 mL (for pH 4.5-5.5 and 5-6 gels) or 100 mL (for pH 6-11 gels) of isoelectric focusing (IEF) buffer 1, consisting of 7 M urea, 2 M thiourea, 4% CHAPS{}, 100 mM dithiothreitol or 4 mM tributylphosphine and DeStreak (1.2% v/v, Amersham Biosciences, GE Healthcare) (for basic gels), and 0.5% pH 4.5 to 5.5 or 5 to 6 or 6 to 11 immobilized pH gradient (IPG) buffer (Amersham Biosciences, GE Healthcare). The sample was loaded on 24 cm pH 4.5 to 5.5 or 5 to 6 IPG strip (Amersham Biosciences, GE Healthcare) which was previously rehydrated at 50 V for 11 h. IEF was carried out for 65,000 V.h at a maximum of 8,000 V, using the Protean II IEF cell (Bio-Rad, Hercules, CA). Analysis of basic proteins was performed with 18 cm pH 6-11 IPG strip. After passive rehydration of the strip in buffer 1, the protein sample was loaded on sample cups and IEF was carried out for 20,000 V.h at a maximum of 3,500 V using the IPGphor device (Amersham Biosciences, GE Healthcare). Before the second dimension, IPG strips were incubated for 15 min with shaking in 150 mM Tris-HCl pH 8.8, 0.1% w/v SDS. The IPG strip was then positioned on sodium dodecyl sulfate-polyacrylamide gels, using 1% low-melting-point agarose in 150 mM Tris-HCl, pH 8.8. Second-dimension electrophoresis was performed on 12% polyacrylamide gels (24 by 20 by 0.1 cm) in 25 mM Tris, 192 mM glycine, 0.1% sodium dodecyl sulfate, pH 8.3, using the Ettan-Dalt II apparatus. Electrophoresis was run at 1 W/gel for 16 h at 15uC. The gels were stained with BioSafe colloidal Coomassie blue (Bio-Rad) for 1 h and destained with three successive washes in deionized water.
Images files were recorded at 65536 gray levels (16 BitsPer-Pixel). Image manipulation and analysis were performed with Samespot V2 software (Nonlinear Dynamics). Protein abundances were given using arbitrary units which correspond to spot volumes and which were calculated as follows: spot area x spot pixel intensity -background intensity.
Protein identification. Protein identification was carried out at the PAPPSO platform (INRA, Jouy-en-Josas) using MALDI-TOF mass spectrometry (MS). Protein spots were excised from Coomassie blue-stained gels and in-gel digested with trypsin. Gel pieces were placed in Eppendorf tubes and washed with 30 mL 25 mM ammonium carbonate, 50% acetonitrile. The supernatants were discarded and gel pieces were dried at 37uC for 15 min. The gels were rehydrated with 20 mL 50 mM ammonium carbonate containing 100 ng of porcine trypsin (Promega, Madison, WI, USA). The solutions were incubated overnight at 37uC. The supernatants containing peptides were directly analyzed by MALDI-TOF Mass spectrometry on a Voyager DE STR Instrument (Applied Biosystems, Framingham, CA, USA). The a-cyano-4-hydroxycinnamic acid matrix was prepared at 4 mg/mL in 0.1% TFA, 50% acetonitrile. An equal volume (1 mL) of matrix and sample were spotted onto the MALDI-TOF target plate. Spectra were acquired in the reflector mode with the following parameters: 2,000 laser intensity, 20 kV accelerating voltage, 62% grid voltage, 120 ns delay. The mass gates used were 840-3500 Da. Internal calibration was performed by using the trypsin peptides at 842.5 and 2,211.1 Da. Database searches were conducted with the MS-Fit software (http:// prospector.ucsf.edu) either on an L. lactis-specific protein database.
The few spots which could not be identified by MALDI-TOF were analysed by LC-MS/MS using an Ultimate 3000 LC system (Dionex, Voisins le Bretonneux, France) connected to a linear ion trap mass spectrometer (LTQ, Thermo Fisher, USA) by a nanoelectrospray interface to realize the separation, ionisation and fragmentation of peptides, respectively. The supernatant of trypsin hydrolysis was transferred to a new tube and the gel pieces were extracted with a) 25 mL of buffer B (50 mM ammonium carbonate) and b) two times buffer C (Formic acid 0.1% acetonitrile 50%). For each extraction, the gel pieces were incubated for 15 min at room temperature while shaking. The supernatants of each extraction were pooled with the original trypsin digest supernatants and dried for 2 h in a Speed-Vacuum concentrator. The peptides were then re-suspended in 25 mL of precolumn loading buffer (0.08% TFA and 2% ACN in water). LC-MS/MS analysis was performed on an Ultimate 3000 LC system (Dionex, Voisins le Bretonneux, France) connected to linear ion trap mass spectrometer (LTQ, Thermo Fisher, USA) by nanoelectrospray interface for separation, ionisation and fragmentation of all peptides. Four mL of tryptic peptide mixtures were loaded at flow rate 20 mL/min onto precolumn Pepmap C18 (0.365 mm, 100 Å , 3 mm; Dionex). After 4 min, the precolumn was connected with the separating nanocolumn Pepmap C18 (0.0756150 mm, 100 Å , 3 mm, Dionex) and the gradient was started at 300 nL/min. All peptides were separated on the nanocolumn using a linear gradient from 2 to 36% of buffer B, over 18 min. Eluting buffer A: 0.1% Formic acid, 2% acetonitrile and eluting buffer B: 0.1% Formic acid, 80% acetonitrile. Including the regeneration step, each run was 50 min in length. Ionization was performed on liquid junction with a spray voltage of 1.3 KV applied to non-coated capillary probe (PicoTip EMITER 10 mm ID; New Objective, USA). Peptides ions were analysed by the Nth-dependent method as follows: (i) full MS scan (m/z 300-2000), (ii) ZoomScan (scan of the 3 major ions), (iii) MS/MS on these 3 ions with classical peptides fragmentation parameter: Qz = 0.25, activation time = 30 ms, collision energy = 40%. Proteins identifications were performed with Bioworks 3.3 software. The raw data were converted and filtered in peak lists with default data generation parameters for LTQ mass spectrometer. All peak lists of precursor and fragment ions were matched automatically against a Lactococcus lactis IL 1403 protein database. The Bioworks search parameter included: trypsin specificity with one missed cleavage, variable oxydation of methionine and the mass tolerance was fixed to 1.4 Da for precursor ion and 0.5 Da for fragment ions. The search results were filtered using Bioworks 3.3. A multiple threshold filter applied at the peptide level consisted of the following criteria: Xcorr magnitude up to 1.7, 2.5 and 3.0 for respectively mono-, di-and tri-charged peptides; peptide probabilities lower than 0.01; DCn greater than 0.1 and only the first match result for each identified peptide.
Statistical treatment. Raw spot volumes were normalized by the mean intensity of the corresponding gel. A total of 542 spots corresponding to 352 different proteins were detected. Some of the spots corresponded to proteins mixture and were not considered. The intensities of spots corresponding to protein isoforms in a same gel were summed so as to represent the level of a single protein independently of post-transcriptional modifications. 15 proteins identified both on 4.5-5.5 and 5-6 pH ranges displayed very different amounts. We considered that the best protein level estimation was given by the highest signal.
Since total protein concentrations remain stable whatever the growth rate (4266 g protein per 100 g cell dry weight), the abundance data are considered to be equivalent to concentrations. Ratios were calculated using the slowest growth phase as a reference. The statistical significance of ratios were evaluated using Student test and False Discovery Rates (FDR, calculated according to Benjamini-Hochberg method [51]) calculated with R free statistical software. Proteins with ratio associated to a False Discovery Rate (FDR) lower than 20% were considered as differentially regulated (see Table 1).

Mathematical treatments
R 2 calculations and equations resolution were perform with MATLAB software.

Correlation calculations
Correlations were estimated using R free statistical software to calculate Spearman rank correlation coefficient and the associated p-value.