Lactose and lactate concentration.

The lactose and lactate contents of cheeses were determined by high-performance liquid chromatography (HPLC), as previously described [20]. Briefly, a cheese suspension was prepared with finely ground cheese (7.5 g and 10 g for lactose and lactate, respectively) and 10 mL of distilled water. After incubation for 1 h at 50°C, it was homogenized for 2 min at 25000 rpm using a mechanical blender (Ultra-Turrax model T25, Ika Laortechnik).

For lactose measurement, 12.5 mL of Carrez I solution (150 g K₄(Fe(CN)₆), 3H₂O /L water), 12.5 mL of Carrez II solution (240 g Zn(CH₃COO)₂, 2 H₂O /L water) and 2.5 mL of 1 M NaOH were added to the suspension. After homogenization and incubation for 1 h at 25°C, the suspension was filtered through Whatman paper n°42 (GE Healthcare) and through a cellulose filter of 0.22 μm porosity (Minisart SP25, Sartorius). Finally, the filtrate was analyzed by HPLC.

For lactate measurement, 10 mL of 240 g/L trichloroacetic acid (TCA) and 10 mL of water were added to the suspension after cooling at 25°C. The mixture was incubated for 1h at 25°C and filtered through Whatman paper n°42 (GE Healthcare). Finally, the filtrate was analyzed by HPLC.

Free amino acids concentration.

A cheese suspension was prepared from 5 g of cheese and 45 mL of distilled water. It was homogenized for 2 min at 25000 rpm using a mechanical blender (Ultra-Turrax model T25, Ika Laortechnik) and incubated for 30 min at 50°C. After centrifugation for 30 min at 5000 x g and 4°C, the supernatant was filtered through Whatman paper n°42 (GE Healthcare). The free amino acids concentration was measured on the filtrate according to the ninhydrin method [21] and quantified using leucine as a standard.

Proteolysis index.

A cheese suspension was prepared from 10 g of cheese and 90 mL of distilled water. It was homogenized for 2 min at 25000 rpm using a mechanical blender (Ultra-Turrax model T25, Ika Laortechnik). This suspension was used for both total nitrogen and non-casein nitrogen content measure.

Total nitrogen content was measured directly on the cheese suspension using the Kjeldahl method according to standard NF EN ISO 8968–1.

The non-casein nitrogen content (soluble at pH 4.6) was measured according to the procedure described in standard NF ISO 27871. The cheese suspension was adjusted to pH 4.6 using 5N HCl and filtered through Whatman paper n°42 (GE Healthcare). The nitrogen content of this soluble fraction was measured using the Kjeldahl method according to standard NF EN ISO 8968–1.

The proteolysis index was calculated by dividing the soluble nitrogen content by the total nitrogen content and multiplying by 100, as previously described [22].

Lipolysis index.

The lipolysis index corresponds to the quantity of KOH (in mg) required to neutralize the free fatty acids contained in 1 g of cheese. It was determined by adapting the titrimetric measurement described by Mouillet et al. [23]. Briefly, 2.5 g of cheese was homogenized for 2 min at 25000 rpm using a mechanical blender (Ultra-Turrax model T25, Ika Laortechnik) in 22.5 mL of ethanol/diethyl ether (1:1 v/v) in order to extract fat content. The suspension was filtered through Watman paper No. 42 (GE Healthcare) and the acidity of the filtrate was measured by standard titration with 1M KOH in ethanol using phenolphthalein as a color indicator.

Extraction of DNA from cheese samples.

Total genomic DNA was obtained after casein solubilization and cell recovery from two grams of cheese by adapting the method described by Baruzzi et al. [24]. Cheese samples were mixed with 18 mL of sodium citrate solution (20 g/L trisodium citrate dihydrate) and the mixture was dispersed for 2 min at 24,000 rpm with a mechanical blender (Ultra-Turrax model T25, Ika Laortechnik). A second dispersion was performed after 10 min of incubation at room temperature. The mixture was then centrifuged at 6,400 x g for 10 min at 4°C, and the supernatant was removed. The casein pellet (containing the microbial cells) was resuspended in 5 mL of a Triton X-100 aqueous solution (2.5% v/v), vigorously shaken, heated in a water bath at 70°C for 10 min, centrifuged at 6,400 x g for 10 min at 4°C, and rinsed twice in physiological saline solution. The pellet was dissolved in a mixture of 270 μL of guanidium thiocyanate (4 M) in Tris-HCl (pH 7.5, 0.1 M) and 30 μL of sodium lauroyl sarcosinate (100 g/L), and transferred to a 2 mL tube containing 250 mg of 0.1 mm-diameter zirconium beads and 250 mg of 0.5 mm-diameter zirconium beads (Sigma, St-Quentin-Fallavier, France). Proteinase K treatment, bead-beating, phenol-chloroform extraction, RNase treatment and ethanol precipitation were then performed as previously described by Leclercq-Perlat et al. [25]. The DNA pellet was dissolved in 120 μL of Tris EDTA buffer (10 mM Tris-HCl, pH 8.0, 1 mM EDTA).

Extraction of RNA from cheese samples.

Total RNA was extracted from 500 mg cheese samples without the prior separation of microbial cells, as previously described [26]. Three separate extractions were performed for each cheese sample and pooled before library construction. The quality of total RNA was analyzed with a 2100 Bioanalyzer and RNA 6000 NANO chips (Agilent, Palo Alto, CA, USA).

DNA and cDNA library construction and SOLiD sequencing.

Library construction and SOLiD sequencing were performed at the INRA MetaQuant facility (Jouy-en-Josas, France). The DNA libraries were constructed from 3 μg of total DNA by using the SOLiD Fragment Library Construction Kit (Applied Biosystems, Bedford, MA, USA) and then barcoded with the SOLiD Fragment Library Barcoding Kit. DNA samples collected at day 1 did not allow to construct libraries that meet the quantity and quality criteria enabling to perform sequencing and, thus, were excluded from this study. The cDNA libraries were constructed from 200 to 500 ng total RNA using a SOLiD Whole Transcriptome Analysis Kit and were barcoded with the SOLiD Transcriptome Multiplexing Kit. The SOLiD ePCR kit and SOLiD Bead Enrichment Kit were used to process DNA and cDNA samples for sequencing, and the SOLiD 4 System was used for sequencing.

Reads processing and mapping.

Short sequence reads of 35 bp were mapped (a maximum of three mismatches were allowed) onto a reference composed of the nine sequenced strains of microorganisms (Table 1) using Bowtie software (version 0.12.7) [27].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Reference genomes used for mapping of the sequence reads.

https://doi.org/10.1371/journal.pone.0124360.t001

After the smart filtering of multiple reads, the numbers of reads mapped onto the reference were counted with METEOR software [28]. Only uniquely mapped reads were further analyzed. General library statistics are available in S1 Table.

DNA-Seq data analyses.

For each microorganism, sequencing coverage was estimated with the Lander/Waterman equation [29]: C = LN / G, where C is the coverage, L is the read length, N is the number of reads and G is the haploid genome length.

RNA-Seq data analyses.

Sequence reads mapping to the coding DNA sequence (CDS) features were retrieved from the raw dataset.

First, functional classification of the metatranscriptomic dataset was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG) annotations [30]. To do this, data were filtered to remove genes displaying an average of less than five reads per sample across the entire dataset (15 samples) and normalized according to the library size using custom scripts built under the statistical environment R (http://www.r-project.org/).

Second, differential expression analysis over time was done by comparing the number of mapped reads for each gene at two different time-points (n = 3 per ripening time). For this analysis, data (six samples for each comparison) were filtered: (i) genes with less than five mapped reads were eliminated; and (ii) genes whose replicates were heterogeneous were discarded (variation coefficient > minimum mean observed for the two groups of three replicates). Data normalization and determination of differentially expressed genes were then conducted using the Bioconductor DESeq2 package in the statistical environment R [31,32]. Raw p-values were adjusted for multiple testing using the Benjamini-Hochberg procedure [33], which assesses the False Discovery Rate. Gene transcripts with an adjusted p-value < 0.05 were considered to be differentially abundant between two ripening times.

Energy metabolism and iron homeostasis.

The great majority of genes involved in energy production detected in our metatranscriptomic dataset were involved in the oxidative phosphorylation pathway (PATH:ko00190) (S2 Fig). As expected, the dynamics in the detection of transcripts involved in this pathway followed the changes in the community structure observed using microbial counts (Fig 1A). G. candidum genes encoding ATPases (EC:3.6.3.14 and 3.6.3.6), NADH dehydrogenase (EC:1.6.5.3) and cytochrome C oxidase (EC:1.9.3.1) accounted for the most frequently detected genes (S2 Table). Their maximum detection levels were observed between days 7 and 21. At day 1, it was possible to observe transcripts of genes encoding K. lactis ATPase (EC:3.6.3.6) and L. lactis ATPase (EC:3.6.3.14), NADH dehydrogenase (EC:1.6.99.3) and cytochrome D ubiquinol oxidase (EC:1.10.3.-), which reflected the early growth of these two microbial species. At the late stage of ripening (days 21 to 31), the detection of gene transcripts involved in the oxidative phosphorylation pathway from H. alvei, C. casei and, to a lesser extent, A. arilaitensis increased.

Metals ions, especially iron, are cofactors and/or components of various enzymatic systems involved in key microbial metabolisms (e.g., respiration). In cheese, iron often forms complexes with various curd components such as proteins (e.g., lactoferrin, ferritin and casein) and peptides, leading to low bioavailability [36]. Eukaryotic and prokaryotic microorganisms have different specific systems to transport metals. Numerous bacteria synthesize and excrete siderophores to trap iron [37]. However, some yeasts, including K. lactis, D. hansenii and G. candidum, express different siderophore transport systems to take advantage of siderophores present in their environment [38,39]. Another interesting feature is that fungal genes encoding iron high-affinity transport systems are generally induced under iron starvation conditions [38]. In our metatranscriptome, gene transcripts related to iron capture and/or transport were frequently detected (S2 Table), which may indicate a need for iron mobilization within the microbial community. For instance, feoB transcripts encoding a ferrous iron transport protein in L. lactis were mainly detected at day 1, transcripts of SIT1 encoding a ferrioxamine B transporter in G. candidum were essentially detected from day 7 to day 21, and transcripts of several iron-siderophore ABC transporters from C. casei (locus tag CCAS_01175, CCAS_03565, CCAS_00130, CCAS_05220, CCAS_05225, CCAS_12545, CCAS_12550, CCAS_12555, CCAS_10415) were mainly detected at day 31. Furthermore, sequence reads mapping to genes encoding high-affinity iron transport systems from both G. candidum (e.g., FTR1, FET3, FTH1), D. hansenii (e.g., FET3) and K. lactis (e.g., FTH1) were also detected in our conditions. Surprisingly, we did not observe bacterial transcripts for genes possibly involved in siderophore biosynthetic pathways, although A. arilaitensis, C. casei and B. aurantiacum are known to carry such pathways in their genome [26,40,41]. This could be due to an insufficient sequencing depth for cheese-surface bacterial species. A second hypothesis, which requires complementary analysis to be confirmed, could be that peptides released during the ripening process through casein hydrolysis would act as iron trappers, as suggested by [36], and be used as siderophore-like complexes by members of the microbial community.

Lactose consumption.

According to Fig 2, carbohydrate metabolism was mainly detected in the early stage of ripening. Lactose is the major carbohydrate compound present in cheese curd and is rapidly consumed during the process (Fig 3A). We identified numerous expressed genes encoding enzymes involved in lactose uptake and degradation (Fig 3B and S3A Fig). For example, L. lactis genes encoding components of the lactose-specific phosphoenolpyruvate-dependent transport system (PTS), including lacEF (EC:2.7.1.69) and ptsI (EC:2.7.3.9), were highly expressed at day 1. The PTS system allows the concomitant translocation and phosphorylation of carbohydrates in L. lactis. The detection of components from other carbohydrate-specific PTS systems such as celB (EC:2.7.1.69) and manXYZ (EC:2.7.1.69) might indicate a versatile use of these systems for lactose uptake in this bacterium, as previously suggested by Aleksandrzak-Piekerczyk [42]. Transcripts of the lacG gene encoding the 6P-β-Galactosidase (EC:3.2.1.85), as well as lacACD (EC:5.3.1.26; EC:2.7.1.144; EC:4.1.2.40) involved in Lactose-6-P degradation via the D-Tagatose-6-P pathway in L. lactis, were also detected, mainly during the early stages of ripening (days 1 to 7). The products from this pathway, i.e., β-D-Glucose and Glyceraldehyde-3-P, then undergo glycolysis and/or the pentose phosphate pathway. Transcripts involved in these pathways were also highly detected in L. lactis at the same time (S2B and S2C Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Lactose metabolism during surface-ripened cheese maturation.

(A) Lactose and lactate concentrations. (B) Expression dynamics of lactose degradation pathways in Lactococcus lactis and Kluyveromyces lactis. Read numbers were normalized (according to the library size) to 50,000 reads per sampling day. For each degradation step, a histogram represents cumulative read numbers when several genes were involved.

https://doi.org/10.1371/journal.pone.0124360.g003

K. lactis is also able to metabolize lactose, but this yeast uses a slightly different pathway [43]. Lactose is imported through a specific lactose permease (encoded by the gene LAC12) and is then metabolized by a β-Galactosidase (EC:3.2.1.23) encoded by LAC4 and enzymes involved in the Leloir pathway (EC:2.7.1.6; EC:2.7.7.12) and encoded by genes GAL1 and GAL7. The expression of all genes from K. lactis involved in this pathway was also observed in our dataset, their detection being the most important during the early stage of ripening (days 1 to 14) (Fig 3B).

G. candidum does not consume lactose [44], whereas D. hansenii can efficiently catabolize this disaccharide [22]. Although our analysis indicated the detection of transcripts corresponding to the KEGG galactose pathway (PATH:ko00052) in these species (S3A Fig), this mainly corresponded to the detection of genes encoding enzymes not necessarily specific to this pathway, such as phosphoglucomutase (EC:5.4.2.2), 6-phosphofructokinase (EC:2.7.1.11) and hexokinase (EC:2.7.1.1). Indeed, these enzymes might instead reflect the detection of glycolysis (PATH:ko00010, S3B Fig) and the pentose phosphate pathway (PATH:ko00030, S3C Fig). Consequently, these species are not likely to significantly contribute to the lactose degradation in our model cheese.

Together, these results highlighted that in our experimental conditions, one of the key functions sustaining the cheese-ripening process, namely lactose biodegradation, involved the active participation of both L. lactis and K. lactis and revealed a functional redundancy existing for this metabolism within the studied ecosystem, as suggested from previous work [45].

Lactate metabolism.

Lactate is of major importance in cheese making. It is produced from lactose present in milk by lactic acid bacteria (Fig 3A). Two lactate dehydrogenases transcripts of L. lactis, ldhA (EC:1.1.1.27) and dld (EC:1.1.2.4), were generally detected at day 1 (S2 Table), suggesting that L. lactis actively produces lactate at the early stage of ripening. Lactate exporter(s), responsible for lactate extrusion from the intracellular environment in L. lactis, have not yet been characterized. Carvalo et al. [46] proposed llmg_2513 CDS from L. lactis MG1363 as a good candidate to encode a lactate transporter based on its predicted protein sequence topology that is similar to known lactate transporters from other species (such as LldP from E. coli, YqkI from B. subtilis and JEN1 from S. cerevisiae), but failed to confirm this activity using the mutagenesis approach. LLACS3_00055 from L. lactis S3, used as starter culture in our experiment, the only CDS presenting a strong homology with llmg_2513 (89% sequence identity at the nucleic level), was not detected in our metatranscriptome and, thus, would probably not be responsible for lactate export. Several gene transcripts encoding transporters in L. lactis were although detected, mostly at day 1, including MFS and ABC family transporters that could be good candidates to facilitate the movement of small solutes such as lactate across cell membranes. However, none of them could be identified as a potential lactate transporter on the basis of sequence similarity with known sequences from other organisms.

Lactate degradation by the cheese microflora, principally yeasts, is a key driver for curd deacidification and occurred essentially during the two first weeks of ripening in our conditions (Fig 3A). Concomitantly, we detected reads for genes JEN1, encoding a lactate transporter, together with DLD2 (EC:1.1.2.4), encoding a lactate dehydrogenase, in both D. hansenii and G. candidum (S2 Table). It should also be mentioned that G. candidum DLD1 transcripts encoding a D-lactate dehydrogenase (EC:1.1.2.4) and CYB2 encoding a L-lactate dehydrogenase (EC:1.2.2.3) were also detected, generally between day 7 and day 21. In Saccharomyces cerevisiae [47,48] and D. hansenii—cultivated in a cheese-like medium [49]—the authors reported that DLD1 and CYB2 were induced by lactate, making them good candidates for lactate degradation.

Protein and amino acid degradation.

Proteins and, more precisely, caseins, are major carbon and energy sources for microbial species living on cheese, together with lactose, lactate and lipids [6]. Proteolysis, which refers to the cleavage of caseins into small peptides and, ultimately, free amino acids by microbial proteases and peptidases, took place progressively and steadily in cheese, as attested to by the evolution of the proteolysis index and free amino acid concentrations (Fig 4A). Because a specific category for proteolysis does not exist in the KEGG database, we manually built two enzyme classes (Proteases and Peptidases) based on gene product annotations present in the reference genomes, and analyzed their expression profile during cheese ripening (Fig 4B and 4C). Normalized expression data for all these genes along with their annotation is given in S3 Table. These results suggested that G. candidum was the major contributor to proteolysis in this simplified surface-ripened cheese and supported other studies that indicated that proteolysis mainly occurs during the first three weeks of soft cheese ripening [17,50].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Protein degradation during surface-ripened cheese maturation.

(A) Proteolysis and free amino acid concentration. Expression data observed for genes encoding proteases (B) and peptidases (C). Read numbers were normalized (according to the library size) to 50,000 reads per sampling day. SE: Staphylococcus equorum. BA: Brevibacterium aurantiacum. AA: Arthrobacter arilaitensis. HA: Hafnia alvei. CC: Corynebacterium casei. LL: Lactococcus lactis. KL: Kluyveromyces lactis. DH: Debaryomyces hansenii. GC: Geotrichum candidum.

https://doi.org/10.1371/journal.pone.0124360.g004

In the genome of G. candidum, only one predicted CDS encoding a putative protease contains a signal sequence indicating a possible extracellular localization according to SignalP [51] and the PSORT server [52]. However, we didn’t detected any read mapping to this CDS in our metatranscriptome. Similar genome analysis revealed that there is no extracellular protease encoding genes predicted in the D. hansenii and K. lactis available genomes, but contrary to Y. lipolytica, another yeast commonly found at the surface of cheese. Furthermore, extracellular protease activity assay performed using the strains used for our experimental cheese production supported those predictions (S4 Fig). Thus, we hypothesized that there were three possible non-exclusive ways for G. candidum to utilize caseins. First, it may directly uptake casein-derived peptides present in the extracellular environment, freed from the hydrolytic activities of both rennet used for milk coagulation [53] and L. lactis used as starter culture [54]. In our metatranscriptomic dataset, this was supported by the high number of reads observed for OPT2 and GAP1 genes of G. candidum (S2 Table), mainly between days 7 and 21, encoding an oligopeptide transporter and a general amino acid permease, respectively, which are known to contribute to the amino acid and protein uptake in yeasts [55,56]. Furthermore, several transcripts reflecting the proteolytic activity of L. lactis were detected earlier in the ripening kinetics, between days 1 to 7, namely pepA, pepDB, pepC, pepN, pepM, pepT, pepV and pepX encoding various peptidases (EC:3.4.11.7; EC:3.4.11.23; EC:3.4.22.40; EC:3.4.13.-; EC:3.4.11.2; EC:3.4.11.18; EC:3.4.11.4; EC:3.4.14.11) involved in the intracellular cleavage of small peptides into free amino acids [54] as well as genes optA, optC and optF encoding an oligopeptide ABC transporter system that has been shown to be responsible for casein-derived peptide transport for amino acid supply in L. lactis [57]. Second, G. candidum may internalize caseins to the vacuole by endocytosis and degrade them using vacuolar proteases and peptidases. In agreement with this hypothesis, expression data revealed the important detection as of day 7, of several genes involved in endocytosis in G. candidum, among which the most frequently detected ones were EDE1 encoding an endocytic protein [58], SAC6 encoding fimbrin [59], ALY2 encoding an alpha arrestin [60], and PIL1 and LSP1 encoding primary components of eisosomes [61]. Among the most frequently transcribed genes encoding proteases and peptidases, CPS1 (EC:3.4.17.4), PRC1 (EC:3.4.16.5) and PEP4 (EC:3.4.23.25) encode enzymes known to be active in the vacuole [62]. The third possible way that casein can be used by G. candidum is the liberation of cellular proteases and peptidases in the extracellular environment during cell lysis, e.g., the highly expressed metalloendopeptidase encoded by the PRD1 gene. Although no data in our metatranscriptomic dataset enable us to support this hypothesis, this mechanism has already been suggested for a cheese isolate of D. hansenii [63].

The next step following initial proteolysis is amino acid biodegradation, which is generally linked to the cheese matrix alkalinization and volatile compound production known to play an important role in aroma perception [5,6,64–66]. The most dominant amino acids composing caseins are glutamate, proline, leucine, lysine, aspartate, valine, serine, tyrosine and isoleucine. As shown in Fig 5, among the complete list of amino acid metabolic pathways, those responsible for the metabolism of these dominant amino acids are also the most detected in our metatranscriptome. G. candidum accounted for the majority of the expression data observed regarding most amino acid metabolism. For example, genes involved in glutamate catabolism were highly detected in this organism. This included genes encoding the NAD-dependent glutamate dehydrogenase (GDH2, EC:1.4.1.2) and the NADP-dependent glutamate dehydrogenase (GDH3, EC:1.4.1.4) responsible for the deamination of this amino acid to generate 2-oxoglutarate, which is then supplied to the TCA cycle, as well as genes encoding the glutamate decarboxylase (GAD1, EC:4.1.1.15), degrading glutamate into 4-aminobutyrate (GABA), which is then converted into succinate by the enzymes 4-aminobutyrate aminotransferase (EC:2.6.1.19) and NAD(P)-dependent succinate semialdehyde dehydrogenase (EC:1.2.1.79), encoded by genes UGA1 and UGA2, respectively, which also feed the TCA cycle. GDH2 induction has already been observed in G. candidum by RT-qPCR in Reblochon-type cheese (a French surface-ripened cheese) at the end of ripening, and this gene was proposed as a biomarker for amino acid catabolism [67]. However, G. candidum was not the only microorganism responsible for amino acid catabolism in our experimental surface-ripened cheeses. Transcripts involved in glycine, serine and threonine metabolism (PATH:ko00260) as well as the valine, leucine and isoleuline degradation pathway (PATH:ko00280) were also highly detected in L. lactis at day 1.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Gene expression related to amino acid metabolism.

For each pathway, the heatmap represents the expression dynamics over time (cumulative number of normalized reads per pathway) using a gray scale bar from 0 read in white to 500 reads in black. For seven pathways, histogram charts detail this dynamic per microbial species. CC: Corynebacterium casei, HA: Hafnia alvei, AA: Arthrobacter arilaitensis, BA: Brevibacterium aurantiacum, SE: Staphylococcus equorum, LL: Lactococcus lactis, KL: Kluyveromyces lactis, DH: Debaryomyces hansenii, GC: Geotrichum candidum.

https://doi.org/10.1371/journal.pone.0124360.g005

Lipid metabolism.

Lipolysis refers to the hydrolysis of triglycerides and results in the liberation of free fatty acids (FFAs) that are important precursors of catabolic reactions that produce volatile compounds that contribute to cheese quality and flavor [4,66,68]. In our metatranscriptomic data, we observed that lipid metabolism pathway expression approximately followed the same dynamics as amino acid metabolism, i.e., a strong increase from day 1 to day 7 followed by a global stability or a slow decrease during the next three weeks (Fig 2). This is in agreement with the dynamics observed by Lessard et al. [17] on a Camembert-type cheese. Furthermore, in our study, the lipolysis index steadily increased along the ripening process, which indicates a regular lipolysis of the food matrix (S5 Fig). G. candidum once again accounted for the most highly expressed genes detected for this metabolism and is likely to be the major contributor to cheese lipolysis in our conditions. However, we also detected two genes encoding putative esterases from L. lactis (locus tag: LLACS3_11125 and LLACS3_11440), mostly at day 1, which may contribute to the initial lipolysis. Indeed, it has been demonstrated that esterases produced by lactic acid bacteria could degrade milk fat into FFAs in hard cheese [69]. Subsequently, few lipase (EC:3.1.1.3) or esterase (EC:3.1.1.13) encoding genes from G. candidum (i.e., ATG15, TGL1, ROG1) were detected, and their expression level increased along with the ripening time. We also detected several transcripts involved in free fatty acid import including FAA1, FAA2 and FAA4 encoding long-chain fatty acyl-CoA synthetases (EC:6.2.1.3) and PXA1 (synonym: PAT2) and PXA2 (synonym PAT1) encoding peroxisomal fatty acyl-CoA ABC transporters. Furthermore, genes involved in the peroxisomal version of the β-oxidation pathway, namely POX1 (EC:1.3.3.6), CTA1 (EC:1.11.1.6), FOX2 (EC:4.2.1.119), ECI1 (EC:5.3.3.8), SPS19 (EC:1.3.1.34) and POT1 (EC:2.3.1.16), were highly detected from day 7 to day 31.

Brief description of differentially expressed genes between D1 and D14.

On the basis of the comparative analysis of D1 vs. D14, two genes related to glycolysis, encoding glyceraldehyde-3-phosphate dehydrogenase (EC:1.2.1.12) and enolase (EC:4.2.1.11), could be proposed as common biomarkers of microbial species activity at the early stage of the ripening process. Interestingly, they were more abundant at D1 in K. lactis and L. lactis, in agreement with an intense development and metabolic activity for these facultative anaerobic microorganisms during the first days of ripening. As mentioned above, both species are involved in lactose degradation within the cheese matrix. This was exactly the opposite for the cheese-surface aerobic yeasts G. candidum and D. hansenii for which both genes were more expressed at D14. Mitochondrial genes of yeasts related to the respiration chain, e.g., COB and COX, followed the same variation pattern.

Regarding G. candidum, genes involved in protein degradation (e.g., PRD1, PEP4) and in glutamate—the most abundant amino acid—degradation (e.g., GDH2) were more frequently detected at D14 when compared to D1 and could thus be proposed as biomarkers for proteolytic activity and amino acid catabolism. GUT1 encoding glycerol kinase (EC:2.7.1.30) involved in lipid metabolism followed the same trend and could be used as a biomarker for lipolysis.

A striking feature of L. lactis is the occurrence of several stress-related genes (e.g., sodA, dnaK, clpB) with high expression levels at D1 (S4 Table). This might be related to several possible stresses such as the shift from anaerobic (lactic acid production during the milk acidification and coagulation phase) to aerobic conditions during the ripening process, the osmotic pressure induced by salting or variations in the pH. Indeed, transcriptomic analysis of L. lactis has already revealed that sodA gene overexpression is associated with oxidative stress response during milk fermentation [74], and that genes encoding chaperones induction reflects the growth arrest of L. lactis in cheese observed almost 24h post-inoculation [75]. In our study, model cheeses were analyzed as a whole, thus including both surface and core parts. However, local conditions might have a great influence on physiochemical stresses encountered by L. lactis and thus should be taken into account for data interpretation.

Brief description of differentially expressed genes between D14 and D31.

The comparative analysis of D14 vs. D31 revealed a shift of the metabolism of G. candidum towards peptide (OPT2, PRD1, LAP3, CPS1) and amino acid (GAP1, AAT2, CAR2) catabolism/transport compared to D1/D14. Another interesting feature is the detection of genes related to iron capture and/or transport, higher in D14 compared to D31 for G. candidum (SIT1, FET3, FTR1), but lower in C. casei (iron-siderophore ABC transporter). This indicated that iron capture is crucial for microbial species living on the cheese surface, as demonstrated by Monnet et al. [26]. In C. casei, we observed the highest abundance at D31 of CCAS_05260 encoding a 2-methylcitrate dehydratase (EC:4.2.1.79), an important enzyme involved in the methylcitrate pathway. In H. alvei, gene transcripts involved in peptide/amino acid transport (meoA, tdcC) and degradation (speA, pepP, hlfB) were also more abundant at D31, revealing the catabolic activity of this species at the end of ripening. Finally, two transcripts involved in sulfur metabolism, encoding a putative sulfate permease (GECA07s04267g) and a methionine synthase (MET6), were more highly detected at D31 in G. candidum. They could be involved in sulfur metabolism through sulfur recycling from various sulfur compounds, including volatile ones, which are released in great quantity at the surface of ripened cheeses [7].