The strain was inoculated into Bennett's slant medium and cultured at a temperature of 30°C with a humidity of 50% for 5–7 days to harvest gray spores. A ring of spores was picked from Bennett's slope and inserted into 100 ml of M3G seed medium. The flasks were incubated at 30°C and 180 rpm on a shaker for 30 h. Subsequently, the seed culture was transferred to another 500-ml flask containing 100 ml of fermentation medium at an inoculation amount of 6.4% (v/v) and was cultured at 30°C and 180 rpm for 72 h.

Bennett's solid medium (pH 7.7) is comprised of 10 g/L glucose, 1 g/L beef extract, 2 g/L peptone, 1 g/L yeast extract, and 20 g/L agar. The M3G seed and fermentation medium contained 10 g/L ammonium sulfate, 1.36 g/L potassium dihydrogen phosphate, 0.8 g/L dipotassium phosphate, and 5 g/L yeast extract, and the initial pH was adjusted to 7.2 with ammonium hydroxide solution.

To evaluate the effects on the synthesis of ε-PL by high-yield strains, the fermentation parameters, such as pH, biomass, residual sugar (RG), and ε-PL yield, were measured every 24 h. The pH value was measured by using an FE20 pH meter (METTLER TOLEDO, Shanghai). The biomass concentrations were calculated from the dry cell weight. In brief, 8-ml culture aliquots were taken every 24 h and centrifuged at 6,000×g for 10 min, and the obtained residues were washed three times with sterile water and finally dried at 95°C to constant weight. The corresponding biomass was calculated from the dried samples. The supernatant was diluted 50 times with distilled water, and the residual glucose (i.e., extracellular) concentration in the fermentation broth was measured by an SBA-40E biosensor analyzer (Biology Institute of Shandong Academy of Sciences, Shandong, China). The ε-PL yield was determined by the spectrophotometer (Thermo Fisher, China). The fermentation broth cultured for a certain period was centrifuged at 8,000×g for 5 min, and the supernatant was appropriately diluted. Later, the diluted solution was mixed with 1 mM methyl orange solution at a ratio of 1:1, shaken at 30°C and 140 rpm for 30 min, and centrifuged at 4,000×g for 15 min. Of note, 1 ml of the supernatant was diluted to 50 ml with 0.1 M Na₂HPO₄-NaH₂PO₄ buffer (pH 6.6), and then the absorbance at 465 nm was measured (i.e., a pure buffer was used as a blank control).

Prior to the GC–MS analysis, the sample was derivatized by the method proposed by Bo et al. (2014). In brief, 50 μL of methoxy ammonium hydrochloride/pyridine solution (20 mg/ml) was added to the freeze-dried sample, vortexed to fully dissolve it, placed in a 40°C water bath, and shaken for 80 min. Then, 80 μL of N-methyl-N-(trimethylsilyl) trifluoroacetamide (MSTFA) was added, vortexed, and placed back in the water bath for another 80 min. Finally, after centrifuged at 10,000×g for 5 min, 100-μL supernatants of the derivatized sample were taken and placed at room temperature for 2 h.

The GC–MS system consisted of an Agilent 7890A GC system and an Agilent 5795C quadrupole mass selective detector (Agilent Technologies, Palo Alto, CA, USA). GC was performed on an HP-5 column (60 m × 0.250 mm × 0.25 μm, Agilent Technologies). A total of 1 μL of the derivatized sample was injected with a split ratio of 1:10. The temperatures of injector, GC interface, and ion source were set at 280, 250, and 280°C, respectively. The oven temperature was initially controlled at 70°C for 2 min, increased to 290°C with a gradient of 5°C/min, kept at 290°C for another 3 min, and finally dropped to 70°C. The carrier gas was helium at a constant flow of 1 ml/min. Mass spectra were recorded at two spectra per second with an m/z 50–800 scanning range.

Both identification and quantification of the mass spectral peaks were performed by MESCHEM software (Agilent Technologies). Peak deconvolution was previously performed by AMDIS-32. The NIST MS standard reference databases 2.0 were used for the identification of metabolites. The relative content of each metabolite was calculated by dividing its GC peak area by that of the internal standard (Liu et al., 2015). Also, the data were normalized as a percentage of that of the original strain for further analyses.

The mathematical model and the reliability verification of the model are important guarantees for the accuracy of later data analysis. Also, cross-validation is a practical and accurate method to test whether the model is effective (Han and Yuan, 2009). Therefore, the principal component analysis (PCA) and the partial least squares analysis (PLSA) by SIMCA-P 11.5 were used. The spectra were reconstructed by peak area integration (Lv et al., 2016), and after being centralized and orthogonal signal correction (OSC), the PCA was performed. The raw data were converted from multiple metrics into comprehensive metrics, and the clustering effect was previewed by a score plot. Besides, the loading plot was used to find biomarkers. Then, the PLSA was used for the verification analysis. Also, the sample grouping, biomarker, and contribution rate of the metabolites were verified by the score plot, the loading plot, and the variable importance plot (VIP), respectively. The data were intercepted for the hierarchical cluster analysis (HCA) for metabolites with a threshold greater than 0.5. Meanwhile, the data were statistically analyzed by SPPSS (SPSS Inc., Chicago, USA) software version 20.0.

Aspartokinase (ASK) activity was detected by measuring the amount of aspartyl-β-hydroxamate formed as described by Shiio and Miyajima (Shiio and Miyajima, 1969). The assay mixture contained in a 0.5-ml volume is as follows: 100 mM Tris–H₂SO₄ (pH 7.0), 600 mM (NH₄)₂SO₄, 600 mM hydroxylamine–KOH (pH 7.0), 10 mM ATP, 10 mM MgSO₄, 10 mM aspartic acid–KOH (pH 7.0), and 40 μg/ml ASK. The reaction mixtures were incubated for 10 min at 30°C, and 0.75 ml of ferric chloride solution (i.e., 10% FeCl₃·6H₂O and 3.4% trichloroacetic acid in 0.7 mol/L HCl) was added. After centrifugation, the A₅₄₀ was measured in the supernatant. The background activity was measured in the absence of aspartic acid. Kinetic assays were performed under conditions identical to those described earlier, except that the enzyme concentration (i.e., 20 μg/ml) was reduced to enable the measurement of steady-state kinetic parameters. All assays were carried out under linear conditions. The amount of aspartyl-β-hydroxamate was calculated from the molar extinction coefficient of 600. One unit of enzyme activity is defined as the amount of enzyme catalyzing the formation of 1 μmol aspartyl β-hydroxamate per minute at 30°C. The protein concentrations were determined using a Bio-Rad Protein Assay Kit (Bio-Rad, Hercules, CA, USA). Bovine serum albumin was used as the standard protein.

The activity of HSD in both strains was measured in the reverse reaction by determining the initial rate of increase of A₃₄₀ at room temperature (Hamano et al., 2007). The standard assay mixture contained 0.2 ml of 500 mM Tris–HCl buffer containing 5 mM ethylenediaminetetraacetic acid (EDTA), pH 8.4; 0.1 ml of 100 mM l-homoserine, pH 7.0; 0.1 ml of 4 mM triphosphopyridine nucleotide disodium salt (NADP-Na₂) solution, 0.5 ml of distilled water, and 0.1 ml of cell-free extract or purified enzyme. For crude extracts, the absorbance at 340 nm was measured against a blank control containing all the components except the substrate. The specific activity of HSD was expressed as micromoles of nicotinamide adenine dinucleotide phosphate (NADPH) formed per minute per milligram of protein.

Polylysine synthetase (PLS) activity follows the method proposed by Kawai et al. (2003). The activity of enzyme preparation to synthesize ε-PL from l-lysine is defined as ε-PL-synthesizing activity. The reaction mixture for the assay of ε-PL-synthesizing activity contained 74 kBq/ml l-lysine, 5 mM ATP, 5 mM MgCl₂, 100 mM Tris–HCl buffer (pH 8.0), 5 mM dithiothreitol (DTT), and 3 μL of enzyme preparation (i.e., 3–15 μg of protein) in a final volume of 15 μL. The reaction was carried out at 30°C for 30 min. Of note, 10 μL of the reaction mixture was taken and subjected to the 0.25% sodium tungstate dihydrate (W) and 5% trichloroacetic acid (TCA) treatment, followed by the radioactivity assay of ε-PL in a similar way as described earlier.

The activities of glucose-6-phosphate dehydrogenase (G6PDH), phosphoenolpyruvate carboxylase (PEPC), succinate dehydrogenase (SDH), pyruvate kinase (PK), and hexokinase (HK) were determined using the kit from Suzhou Comin Biotechnology Co., Ltd., China.

The difference between the original strain TUST and the high-yield strain 6#-7 in the fermentation performance of ε-PL production was compared by shake flask fermentation, which provided a foundation of the data for the exploration of ε-PL biosynthesis mechanisms.

The fermentation characteristics of the wild-type strain TUST and the mutant strain 6#-7 are shown in Figure 1. With the increase in fermentation time, the pH value of the fermentation broth showed a downward trend due to the production of primary metabolic organic acid. During 24–48 h, the pH value decreased obviously in the early stages of the fermentation due to the accumulation of various acids (Table 1; Supplementary Table 1), while the downward trend became slower during 48–72 h. In addition, the pH value of the fermentation broth of strain 6#-7 was slightly larger than that of the starting strain TUST during the whole fermentation process. As for RG, it also showed a downward trend indicating that both strains were continuously consuming the carbon source. Also, there was no significant difference in the sugar consumption rate between the two strains. In addition, the biomass of strain TUST reached a maximum of 7.72 g/L at 72 h and then decreased. However, the biomass of strain 6#-7 kept increasing and reached 8.48 g/L at 96 h. During 24–48 h, there was less difference in the amount of ε-PL between the two strains, while in the middle and late stages of fermentation 6#-7 produced more ε-PL than that of TUST. For both strains, the amount of ε-PL reached a maximum at 72 h, and the yield of strain 6#-7 reached 0.92 g/L, which was much larger than the original strain.

A total of 101 metabolites including 21 amino acids, 18 sugars, 6 alcohols, and their derivatives were detected and quantified by GC–MS. The data discussed in this section are shown in Table 1, and the relative contents of other intercellular metabolites are shown in Supplementary Table 1.

As shown in Table 1, after 48-h fermentation, the contents of amino acids, especially for the ε-PL precursors such as asparagine (Asn), aspartic acid (Asp), and lysine (Lys), in the high-yield strain 6#-7 were larger than those in the original strain TUST, indicating that the high-yield strain can produce more products in key steps of the ε-PL synthesis pathway. Among the 18 sugars, the glucose content in strain 6#-7 was significantly lower than that in strain TUST, indicating that 6#-7 has a stronger glucose-consumption ability, which is conducive to the production of ε-PL yield (Pan et al., 2019). Also, maltose content in the branched-chain ε-PL metabolic pathway was lower in 6#-7, which benefits the accumulation of ε-PL. On the contrary, increased trehalose during the initial growth stage of 6#-7 would enhance its resistance to ε-PL (Sanchez et al., 2010). The contents of 28 organic acids, hexadecenoic acid, octadecanoic acid, and other fatty acids in strain 6#-7 were significantly lower than those in strain TUST, indicating that the fatty acid branch metabolism pathway in 6#-7 is weakened and thus the consumption of acetyl-CoA (Accoa) is reduced. Also, all of these benefit the accumulation of ε-PL in 6#-7. The contents of other compounds, such as norvaline, sedoheptulose, dodecanol, and acetamide, did not show any obvious difference between the two strains.

Unsupervised (i.e., PCA) and supervised (i.e., PLSA) approaches were used to classify the data set of different strains. The values of R² and Q² represent the accuracy of fit and the predictability of the model, respectively. It is considered that if R² and Q² are greater than 0.9, the model is valid, and generally, it is reasonable if the values are greater than 0.5. The main statistics are shown in Supplementary Table 2. It shows that the values of R² and Q² are close to 1.0 except for the values of Q² of PCA, indicating that the model is credible.

The scores and loadings represent the discrete (Jackson, 1980) and central tendency of the data (Kettaneh et al., 2005). Specifically, discrete tendency reflects the differences among the variables, while central tendency reflects the similarity among the variables (Wold et al., 1987).

Each experiment was repeated five times to ensure the reproducibility of the results, and the data were initially classified by the PCA. As shown in Figure 2A, the data of the same group were centralized, indicating that the model has good testing repeatability. The distribution of each group, except for TUST at 12 and 24 h, was discrete from other groups, meaning there were significant differences in the metabolisms between TUST and 6#-7 during fermentation periods. The data of TUST at 12 h and 24 h were clustered, indicating that the metabolism was stable for TUST at the initial stage. The loading plot was used to find biomarkers. The farther a metabolite is away from the origin of coordinates, the greater contribution it makes (Teague et al., 2007). As shown in Figure 2B, the metabolites with a high contribution rate during the whole fermentation process were d-(+)-trehalose, d-glucose (d-Glc), d-galactose, butanoic acid, l-alanine (l-Ala), oxalic acid, glutamic acid (Glu), glycerol, and l-valine.

To further confirm the results of PCA, the PLSA was also used for the statistical analysis (Wang, 1999). As shown in Figure 2C, the results were consistent with those of the PCA, indicating that the metabolisms of TUST and 6#-7 were significantly different during fermentation periods. As shown in Figure 2D, there were other major metabolites, such as palmitic acid and stearic acid, besides those found in the PCA.

Under the PLSA model, a multidimensional analysis of differential metabolite screening can be performed by the VIP analysis. The VIP coefficients reflect the contribution of each metabolite to the PLSA models; a higher VIP value indicates that the metabolite has a larger contribution (Lv et al., 2016). In this study, a VIP value of 0.5 was used as the cutoff value for statistical significance, and thus more potential biomarkers can be screened.

As shown in Supplementary Figure 2, biomarkers extracted from the loading plot can be sorted according to the contribution rate as follows: d-Glc > d-galactose > d-(+)-trehalose > stearic acid > palmitic acid > glycerol > l-Ala and oxalic acid. Among them, d-Glc, d-galactose, and d-(+)-trehalose are the main intermediates related to sugar metabolism. Stearic acid and palmitic acid are the important products of the fatty acid metabolism pathway. l-Ala and oxalic acid are closely related to cell center carbon metabolism.

Hierarchical cluster analysis (HCA) was performed to further validate the prediction accuracy of the PCA model. Heatmap generated from HCA based on the first principal component is shown in Supplementary Figure 3. The samples have obvious distinctions and can be sorted according to the strains, indicating that the HCA results are consistent with those of PCA.

The normalized relative contents of differential intracellular metabolites between TUST and 6#-7 are shown in Figure 3. It shows that strain 6#-7 has a strong ability to absorb glucose. Also, the weakening of the fatty acid metabolism pathway liberates more Accoa to the tricarboxylic acid cycle (TCA), which can promote the TCA pathway and increase the accumulation of Oaa. Then, the pathway from Oaa to l-Lys is enhanced. This provides sufficient energy for the synthesis of Oaa and precursor l-Lys, which ultimately leads to higher ε-PL productivity (Zhang et al., 2018).

The enhancement of the diaminopimelate pathway increases the content of threonine (Thr), homoserine (Hse), and isoleucine (Ile), and that promotes the accumulation of Lys, which is the precursor of ε-PL (Xiang et al., 2020). In addition, it is presumed that there may be a pathway that can compensate for the loss of intermediate metabolites in the pentose phosphate pathway (PPP) and allow the PPP to be restored, which in turn provides a large amount of NADPH for ε-PL synthesis. Moreover, the increase of α-ketoglutarate (Akg) and succinic acid (Succ) promotes the production of proline (Pro) and Glu, which enhances the ability of the strain itself to resist ε-PL (Brandriss and Falvey, 1992; Guan et al., 2013). Thus, the ability of the mutagenized strain to synthesize ε-PL is enhanced.

In general, based on the metabolomic analysis, it was found that the intracellular metabolism of the strain significantly changed after mutagenesis, which is consistent with the results of PCA and PLSA. Moreover, combined with Figure 1, it is indicated that the difference in metabolism between the two strains occurred earlier. Also, except for the normalized contents of Thr, Ile, and Lys that showed the biggest difference between TUST and 6#-7 at 48 h, the contents of all the other 9 metabolites showed the biggest difference at 12, 24, or 36 h. This is because of Thr, Ile, and Lys which are very close to ε-PL in the metabolic network, and they showed the biggest difference when the strain began to produce ε-PL in large quantities (Figures 1, 4). On the contrary, the other metabolites are far away from or not that close to ε-PL in the metabolic network, and they showed the biggest difference mainly during the growth stage of the strain.

Based on the findings of main metabolites by the abovementioned loading plot analysis and the analysis of related metabolic intermediates in the metabolite network, a possible metabolic regulation mechanism of the high-yield mutagenized strain was revealed. First, the transport and absorption capacity for Glc of strain 6#-7 is improved. Second, the TCA pathway is promoted and that can increase the accumulation of the precursor Oaa. Third, the diaminopimelate pathway is enhanced to produce more Lys, which is an essential precursor of ε-PL synthesis. Finally, the increase of trehalose, Glu, and Pro makes the strain 6#-7 more resistant to ε-PL.

The enzyme plays an important role in the growth and metabolism of microorganisms (Teague et al., 2007); thus, the activities of enzyme in metabolic pathways associated with ε-PL synthesis were measured to gain further insight into the biosynthesis mechanism of ε-PL in the high-yield strain 6#-7 at the enzymatic level.

The activities and normalized activities of the key enzymes are shown in Supplementary Figure 4 and Figure 4, respectively. During the metabolic synthesis in microorganisms, the PPP mainly provides NADPH (Kind et al., 2013). Increased G6PDH activity indicates the enhancement of the PPP that can provide a large amount of NADPH for the synthesis of ε-PL (Hamano, 2010). The activity of G6PDH increased in strain 6#-7 benefits the production of ε-PL. The activity of PK in the high-yield strain 6#-7 is strengthened, indicating that the glycolysis pathway (EMP) pathway in strain 6#-7 is still unblocked which can provide sufficient pyruvate (Pyr) (Calmettes et al., 2015; Alves-Filho and PåLsson-Mcdermott, 2016).

The increase in the activity of PEPC in strain 6#-7 benefits the accumulation of Oaa (Atsushi and Shiio, 2016). Besides, the increasing activity of PEPC can also provide more NADPH for ε-PL synthesis. Also, the enhancement of ASK activity can provide more β-aspartic acid (β-Asp) in strain 6#-7 for the synthesis of Lys (Wittmann and Becker, 2007), which is the precursor of ε-PL. Meanwhile, the decrease in the activity of HSD leads to less Hse to the branched-chain pathways in strain 6#-7 (Hamano, 2010). All of these promote the production of ε-PL in the high-yield strain 6#-7. In addition, the activities of HK, SDH, and PLS change a few in strain 6#-7 as compared with the original strain.

Based on the results of metabolomic and enzymatic analyses, a possible metabolic regulation mechanism of the high-yield mutagenized strain was revealed (Figure 5). First, the transport and absorption capacity for Glc of strain 6#-7 is improved. Second, the activity of enzymes benefits ε-PL synthesis, such as G6PDH, PK, PEPC, and ASK, is strengthened. On the contrary, the activity of HSD in the branched-chain pathways is decreased. Third, the increase of trehalose, Glu, and Pro makes the strain 6#-7 more resistant to ε-PL. Thus, the ability of the mutagenized strain to synthesize ε-PL is enhanced, and strain 6#-7 can produce more ε-PL compared with the original strain. To analyze a mutant strain, genomic sequencing is also a useful method (Wang et al., 2015; Hwang et al., 2019). Thus, genomic sequencing was conducted in another study by our laboratory that further improve the production of ε-PL using genetic engineering (i.e., data were not shown in this study). Also, the results obtained by the genetic analysis are highly consistent with those obtained by metabolomic and enzymatic analyses in this study.