Microbial Ecology Dynamics Reveal a Succession in the Core Microbiota Involved in the Ripening of Pasta Filata Caciocavallo Pugliese Cheese

Manufacture of cheese.

Caciocavallo Pugliese was manufactured with raw cows' milk (Friesian cow breed), using commercial freeze-dried (Sacco, Cadorago, Como, Italy) primary starter Streptococcus thermophilus (initial cell count of ca. 7.0 ± 0.2 log CFU ml⁻¹ of milk). The manufacture was carried out at the industrial plant Ignalat (vat of about 200 liters), located in Noci, Bari (Apulia region), Italy. Cheese making was carried out on 3 consecutive days (total of 3 batches), using cows' milk from 2 daily milkings. All the results were the averages for 3 batches, which were analyzed in triplicate (total of 9 samples analyzed). Raw cows' milk was heated at 37°C and inoculated with primary starter, and liquid calf rennet (10 ml 100 l⁻¹) was added. Coagulation took place within 30 min. The coagulum was first cut coarsely by hand, held under whey at 37°C for 2 h, and then reduced mechanically to particles of 1.5 to 2 cm. When the curd reached a pH of 5.25 (ca. 5 h, at room temperature), it was stretched in hot water (80°C). Cheeses were salted in brine (30% [wt/vol] NaCl) for 12 h. Ripening was at ca. 10°C and a relative humidity of 83% for 90 days. The weight of the cheese was approximately 1.5 kg. Raw cows' milk, curd immediately after coagulation, curd after 5 h of incubation (when the pH reached ca. 5.25), curd after stretching, and cheeses after 1 (C1), 3 (C3), 7 (C7), 15 (C15), 30 (C30), 45 (C45), 60 (C60), 75 (C75), and 90 (C90) days of ripening (after brine treatment) were collected from each batch. All samples were transported to the laboratory in thermal plastic bags under refrigerated conditions (ca. 4°C) and analyzed immediately (microbiological analysis and community-level catabolic profiles) or frozen (−80°C) (biochemical analysis and extraction of total bacterial genomic RNA).

Compositional, microbiological, and biochemical analyses.

Samples of milk, curd, and cheese were analyzed for protein (16), fat (17), moisture (oven drying at 102°C) (18), and salt (19). The pH was measured by a Foodtrode electrode (Hamilton, Bonaduz, Switzerland). The raw cows' milk used had the following composition: pH 6.7; fat, 3.8%; protein, 3.7%; lactose, 4.9%; and salt, 0.09%. No significant (P > 0.05) differences were found between the 3 batches.

Microbiological analyses were carried out as described previously (20). Twenty grams of sample was homogenized with 180 ml of sterile sodium citrate (2% [wt/vol]) solution. Presumptive mesophilic lactobacilli and lactococci were enumerated on MRS and M17 agar (Oxoid), respectively, under anaerobiosis at 30°C for 48 h. Presumptive thermophilic streptococci were enumerated on M17 agar (Oxoid, Basingstoke, Hampshire, United Kingdom) under anaerobiosis at 42°C for 48 h. Enterococci were counted on Slanetz and Bartley agar (Oxoid) at 37°C for 48 h. The number of yeasts was estimated at 30°C for 48 h on Sabouraud dextrose agar (SDA) (Oxoid) medium supplemented with chloramphenicol (0.1 g liter⁻¹). The number of molds was estimated on wort agar (Oxoid) at 25°C for 5 days. Total coliforms were counted using violet red bile lactose (Oxoid) at 37°C for 24 h. Except for enterococci, the media for plating bacteria were supplemented with cycloheximide at 0.1 g liter⁻¹.

The pH 4.6-insoluble and -soluble nitrogen fractions of the cheeses were analyzed by urea polyacrylamide gel electrophoresis (urea-PAGE) and reverse-phase high pressure liquid chromatography (RP-HPLC), as described by Andrews (21) and Gobbetti et al. (4), respectively. Total and individual free amino acids (FAA) from the water-soluble extracts were determined by a Biochrom 30 series amino acid analyzer (Biochrom Ltd., Cambridge Science Park, United Kingdom), as described by Di Cagno et al. (22).

Determinations of volatile components and volatile free fatty acids.

Neutral volatile components (VOC) were determined by a purge-and-trap method (PT) coupled with gas chromatography-mass spectrometry (PT–GC-MS). Prior to PT, 3 g of cheese was mixed with 27 g of ultra-high-quality (UHQ) deionized water with an Ultra-Turrax 4 times for 40 s each time, separated by 10-s rests, in a glass flask with a narrow neck plunged into ice. Ten milliliters of this suspension was placed in a glass extractor connected to the PT apparatus (Tekmar 3000; Agilent Instruments, NY, USA). Extraction was performed with helium at a flow rate of 40 ml min⁻¹ on a Tenax trap at 37°C. The trap was desorbed at 225°C, and injection used a cryoconcentrator at −150°C. The chromatograph (model 6890; Agilent Instruments) was equipped with a DB5-like capillary column (RTX5 Restek; Agilent Instruments) with a 60-m length, 0.32-μm internal diameter, and 1-μm thickness. The helium flow rate was 2 ml min⁻¹; the oven temperature was 40°C during the first 6 min and then was increased at 3°C min⁻¹ to 230°C. The mass detector (MSD5973; Agilent Instruments) was used in electronic impact at 70 eV and in scan mode, at an atomic mass from 29 to 206. Quantification of compounds was expressed in arbitrary units of area.

For solid-phase microextraction (SPME) extraction of volatile free fatty acids (VFFA), each sample was analyzed twice with two different dilutions. A 400- or 1,000-μl portion of the suspension described above was poured into a 10-ml flask with 100 μl of 2 N sulfuric acid and 1,200 μl of UHQ water. The flask was sealed and placed in a bath at 60°C for 15 min. An SPME Carboxen-polydimethylsiloxane 75-μm fiber (Supelco, L'Isle d'Abeau, France) was introduced into the flask and held in the headspace for 30 min at 60°C. It was then removed and desorbed for 5 min in a splitless chromatograph injector at 240°C. The chromatograph (CE8160; Thermoquest, Les Ulis, France) was equipped with an FFAP column with a 30-m length, 0.53-μm diameter, and 1-μm thickness (Stabilwax DA; Restek, Lisses, France) and a flame ionization detector (FID). The helium flow rate was 6 ml min⁻¹; the oven temperature was 120°C for 1 min, was increased to 162°C at 1.8°C min⁻¹ and to 240°C at 10°C min⁻¹, and lastly was held at 240°C for 5 min. Concentrations of VFFAs were calculated from calibration curves established with external standards (Sigma, L'Isle d'Abeau, France).

Extraction of total bacterial genomic RNA.

Total RNA was extracted using the RiboPure bacterial kit (Ambion RNA; Life Technologies Co., Carlsbad, CA, USA), according to the manufacturer's instructions. The quality of the RNA was checked by agarose gel electrophoresis. The RNA concentration was measured in a NanoDrop ND-1000 spectrophotometer (NanoDrop Technologies, Rockland, DE). The purified RNA (100 ng; final volume, 20 μl) was incubated at 42°C for 2 min in 2 μl of genomic DNA (gDNA) wipeout buffer (7×; QuantiTect reverse transcription kit; Qiagen srl, Milan, Italy) and RNase-free water (final volume, 14 μl). The cDNA was obtained with the QuantiTect reverse transcription kit (Qiagen), according to the manufacturer's instructions. All reactions were set up in a Rotor Gene 6000 instrument (Corbett Life Science, New South Wales, Australia) equipped with a 36-well reaction rotor.

Pyrosequencing and data analyses.

Three cDNA samples, corresponding to the three batches, were pooled and used for 16S based bacterial diversity analysis. Bacterial diversity was assessed via pyrosequencing on a 454 FLX sequencer (454 Life Sciences, Branford, CT, USA) and was performed by Research and Testing Laboratories (Lubbock, TX), according to standard laboratory procedures. Primers targeting the V1-V3 region (Escherichia coli positions 27 to 519; forward, 28F [GAGTTTGATCNTGGCTCAG], and reverse, 519R [GTNTTACNGCGGCKGCTG]) of the 16S rRNA gene (23) were used. Pyrosequencing procedures were carried out based upon protocols from Research and Testing Laboratories.

Bioinformatics.

Sequence data were processed using Research and Testing Laboratory's in-house pipeline, described at http://www.researchandtesting.com/. Briefly, sequences were grouped using their barcodes, and any sequence that contained a low-quality barcode or that failed to be at least half the expected amplicon length (or 250 bp, whichever was shorter) was removed from the data pool. Sequences that passed the quality filter were denoised using an algorithm based on USEARCH (24) and then checked for chimeras using UCHIME (25). Finally, sequence data were separated into operational taxonomic units (OTUs) and annotated using USEARCH and a distributed. NET algorithm that utilizes BLASTN+. The output from BLASTN+ was then parsed, and the taxonomic information for each sequence was resolved using the following rules: sequences with more than 97% identity to well-characterized database sequences (<3% divergence) were resolved at the species level, those with 95% to 97% at the genus level, those with 90% to 95% at the family level, those with 85% to 90% at the order level, those with 80 to 85% at the class level, and those with 77% to 80% at the phylum level. Any match below this percent identity was discarded. In addition, the highest-scoring pair had to be at least 75% of the query sequence or it was discarded, regardless of identity.

Community level catabolic profiles (CLCPs).

Biolog Eco-Microplates (Biolog, Inc., Hayward, CA, USA) were used to determine bacterial CLCPs (26). Microplates contain 31 carbon sources grouped by chemical class (carbohydrates, carboxylic acids, polymers, amino acids and amines) and the control, without a carbon source, in triplicate. Ten grams of sample was homogenized with 90 ml of sterile sodium chloride (0.9% [wt/vol]) solution and centrifuged at 10,000 × g for 15 min at 4°C. The pellet was washed with sterile 50 mM Tris-HCl (pH 7.0) and again with sterile sodium chloride solution and then centrifuged again. The cell suspension was diluted (1:1,000) in sterile sodium chloride solution and dispensed (150 μl) into each of the 96 wells of the Biolog Eco-Microplates. Incubation was at 30°C in the dark, and color development was measured at 590 nm with a microplate reader (Biolog Microstation), every 24 h up to 120 h. Three indices were determined (27). Shannon's diversity (H′), indicating the substrate utilization pattern, was calculated as −∑pi ln(pi), where pi is the ratio of the activity of a particular substrate to the sums of activities of all substrate activity at 120 h. Substrate richness (S), measuring the number of different substrates used, was calculated as the number of wells with a corrected absorbance greater than 0.25. Substrate evenness (E) was defined as the equitability of activities across all utilized substrates: E = H′/log(S).

Statistical analyses.

Data were subjected to one-way analysis of variance (ANOVA), and pairwise comparison of treatment means was achieved by Tukey's procedure at a P value of <0.05, using the statistical software Statistica 7.0 for Windows. Data for VOC and VFFA analyses were subjected to permutation analysis using PermutMatrix (28).

Compositional and microbiological analyses.

Table S1 in the supplemental material shows the main chemical composition of Caciocavallo Pugliese during manufacture and ripening. As expected, the pH decreased from the curd immediately after coagulation (6.40 ± 0.04) to the curd after ca. 5 h of incubation (5.28 ± 0.02). An increase (5.62 ± 0.03) was found after curd stretching, followed by a decrease at the beginning of ripening (5.24 ± 0.01) until 30 days of ripening (5.09 ± 0.03). Then, the pH increased over time, reaching 5.36 ± 0.04 at 90 days and varying only slightly during ripening. The highest increase in the NaCl concentration was found at the beginning of ripening, even though it progressively rose over time. As expected, the cheese moisture decreased progressively during ripening, reaching a value of 38.1 ± 1.5% at 90 days. The concentration of fat and protein inversely followed the moisture trend. At 90 days, the cheese had 33.1% ± 1.5% fat and 23.8% ± 1.5% protein.

Presumptive mesophilic lactobacilli were already present in the raw cows' milk (4.2 ± 0.2 log CFU g⁻¹) (Table 1). Their numbers increased by ca. 1.4 log cycles after curd incubation and decreased after curd stretching (4.1 ± 0.2 log CFU g⁻¹). During ripening, presumptive mesophilic lactobacilli progressively increased, especially after 30 days (7.2 ± 0.1 log CFU g⁻¹), and attained the highest number during late aging (average value of ca. 8.5 log CFU g⁻¹). Compared to mesophilic lactobacilli, raw cows' milk contained a slight but significantly (P < 0.05) higher number of presumptive mesophilic lactococci, which decreased after curd stretching (4.1 ± 0.1 log CFU g⁻¹) and remained almost constant throughout ripening. As expected, a high number of presumptive thermophilic streptococci was found after coagulation and incubation of the curd (average value of ca. 8.7 log CFU g⁻¹). Although the number decreased by ca. 1 log cycle after stretching of the curd, it progressively increased up to 15 days (9.0 ± 0.3 log CFU g⁻¹). From 30 days onwards, the number decreased and attained the lowest value at the end of ripening (7.3 ± 0.1 log CFU g⁻¹). Enterococci were found in the raw cows' milk (2.6 ± 0.1 log CFU g⁻¹), and the number increased after 1 day of ripening (4.1 ± 0.2 log CFU g⁻¹) and remained almost constant for 15 days but decreased during late ripening (from 30 to 90 days). Yeasts and molds disappeared during ripening and manufacture, respectively. Total coliforms were detected in the raw cows' milk but progressively disappeared.

Proteolysis.

The pH 4.6-insoluble and -soluble nitrogen fractions of the cheese were analyzed by urea-PAGE (Fig. 1A and B, respectively). α_s1-Casein (α_s1-CN) persisted to the end of ripening, and its main degradation probably began at 7 days. β-CN also persisted. The formation of protein bands with low electrophoretic mobility, which presumably corresponded to γ-CN, was evident. The urea-PAGE electrophoretogram of the pH 4.6-soluble fraction showed differences over time. Characteristic polypeptide bands appeared at 7 days of ripening and persisted during late ripening. Complementary information emerged with RP-HPLC analysis (Fig. 2). The number of peaks which were recognized and matched visually with the Unicorn program (Amersham Biosciences) varied during manufacture and ripening of Caciocavallo Pugliese. Seven (raw cows' milk) to 19 (30 days) peaks were detected, which decreased to 11 at 90 days. An increase in the area of hydrophobic and especially hydrophilic peptide peaks was found up to 30 days. These results agreed with the concentration of total free amino acids (FAA) (Table 2). As expected, raw cows' milk (105.4 ± 14 mg kg⁻¹) and curd after coagulation, incubation, and stretching had the lowest concentrations of FAA (140.4 ± 29 to 218.2 ± 18 mg kg⁻¹). Subsequently, the concentration of FAA markedly (P < 0.05) increased, showing the highest rate from 30 days (1,343 ± 42 mg kg⁻¹) to 75 days (2,335 ± 63 mg kg⁻¹). The FAA found at the highest concentrations (>100 mg kg⁻¹) were Ser, Cys, Val, Ile, Leu, Phe, His, Trp, Lys, Arg, and Pro.

Volatile components.

Volatile components (VOC) (95 in total) were identified by PT-GC/MS. VOC belonged to several chemical classes: aldehydes (17), alcohols (19), ketones (26), esters (14), sulfur compounds (7), and furans (8). The levels of 80 VOC significantly (P < 0.05) differentiated samples during manufacture and ripening (Fig. 3; also, see Table S2 in the supplemental material). Overall, aldehydes were the VOC found at the lowest levels. Acetaldehyde, octanal, nonanal, 2-methyl-propanal, 3- and 2-methyl-butanal, and benzaldehyde, which were found in the curd after coagulation and incubation, decreased after curd stretching, and their levels remained almost constant up to 15 days. Slight increases were found later. The levels of all the other aldehydes increased from 30 days onwards (45 days for heptanal and 60 for hexanal). The levels of 9 alcohols (methanol, 1-butanol, 1-pentanol, 1-hexanol, 1-heptanol, 2-propanol, 2-butanol, 2-pentanol, and 3-methyl-3-buten-1-ol) increased after 30 or 45 days of ripening and remained almost constant throughout ripening. 2-Methyl-1-propanol and 1-hexanol increased earlier. 1-Penten-3-ol was the only alcohol found at the highest level before ripening (curd after coagulation). Regarding the number of compounds identified, ketones (26 compounds identified) were the most abundant VOC. Most of them (e.g., 2-pentanone, 2-heptanone, 2-octanone, and 2-nonanone) were found at the highest levels from 30 to 90 days of ripening. The highest levels of 2-butanone and 2,3-octanedione and of 2,3-butanedione (diacetyl) and 2,3-pentanedione were found during manufacture and early ripening, respectively. Except for butyl butanoate and 2-methyl propyl-acetate, which were identified at the highest levels at the end of ripening, all the other esters increased from 15 (e.g., methyl hexanoate) and 30 (e.g., propyl hexanoate) days onwards. Almost all sulfur compounds increased after 7 days of ripening. Except for 2-ethyl- and 2,5-dimethyl-furans, which were identified at the highest levels before ripening, all the others furans increased after 15 (e.g., 2-methyl furan) and 30 (e.g., furfural) days onwards.

The levels of all volatile free fatty acids (VFFA) increased throughout ripening (Fig. 3; also, see Table S2 in the supplemental material). In particular, acetic (8.1 ± 0.4 to 26.7 ± 1.1 ppm) and hexanoic (8.9 ± 0.2 to 23.4 ± 0.7 ppm) acids were found at the highest concentrations, followed by butyric acid (3.7 ± 0.1 to 16.9 ± 0.5 ppm). The permutation analysis based on VOC and VFFA (Fig. 3) distributed the samples into two major clusters (A and B). Cluster A included curds after the various manufacturing steps and cheeses from 1 to 15 days of ripening, which contained the lowest levels of most of the VOC and VFFA. Cluster B contained cheeses from 30 days of ripening onwards. Within cluster B, subclusters B1 and B2 further differentiated the cheeses based on the time of ripening (from 30 to 75 and at 90 days, respectively).

Structure and changes of the microbiota.

Pyrosequencing of 16S rRNA was used to characterize metabolically active bacteria. No significant (P > 0.05) differences were found between the three batches analyzed. After pyrosequencing, a total of 197,385 raw sequence reads of 16S rRNA gene amplicons were obtained (average length, 515 bp). The calculated alpha diversity parameters are reported in Table S3 in the supplemental material. The highest values of Chao1 richness and the Shannon diversity index were found in raw cows' milk (50.2 and 2.15, respectively). Further, the indices fluctuated according to technology steps. After these variations, Chao1 richness and the Shannon diversity index slightly increased from day 7 onwards. Good's estimated sample coverage (ESC) was above 99% for all the samples, indicating a satisfactory description of the microbial diversity. The community structure was also analyzed using three phylogeny-based beta-diversity measures (see Fig. S1 in the supplemental material). The two principal coordinate analyses (PCoA) clearly differentiated raw cows' milk, which was characterized by higher variability and was spread in the right part of the plot. Curds at various technological steps (coagulation, incubation, and stretching) and cheeses at early ripening (1 to 15 days) tended to form a separate subgroup. As the time of ripening proceeded (30 days onwards), cheeses grouped together.

The bacterial sequences from RNA which were assigned to bacterial phyla and their relative abundance varied during cheese manufacture and ripening (data not shown). RNA from raw cows' milk included mainly Firmicutes (53%) and Proteobacteria (39%), followed by Bacteroidetes (7.8%) and, at low frequency, Actinobacteria and Fusobacteria (0.06% and 0.03%, respectively). From curd after coagulation onwards, the phylum Firmicutes dominated (ca. 99 to 100%). The phylum Proteobacteria appeared at a very low frequency only after 30 and 75 days of ripening (0.04 to 0.03%, respectively).

The distribution of the OTUs that were classified at genus level is reported in Fig. 4. Although 35 OTUs were identified in raw cows' milk (Fig. 4A and B), only 8 had a relative abundance higher than 0.5% in at least one sample (Fig. 4B). Lactococcus (31.9%), Acinetobacter (30.5%), Streptococcus (16.2%), Chryseobacterium (7.5%), Lactobacillus (4.6%) Yersinia (4.2%), Pseudomonas (2.9%), and Hafnia (0.55%) were the main genera found in raw cows' milk. Subdominant genera belonging to the phyla Proteobacteria (Sphingomonas, Methylophilus, Citrobacter, Enterobacter, Escherichia, Klebsiella, Serratia, Rahnella, Moraxella, Paracoccus, and Stenotrophomonas) (0.40 to 0.029%), Firmicutes (Staphylococcus, Clostridium, Carnobacterium, and Leuconostoc) (0.17 to 0.029%), Bacteroidetes (Epilithonimonas, Arcicella, Flavobacterium, and Sphingobacterium) (0.086 to 0.057%), Fusobacteria (Fusobacterium) (0.029%), and Actinobacteria (Mycobacterium) (0.057%) were also found at very low frequency (Fig. 4A). As expected, the bacterial profile of the curd after the inoculum of the primary starter (St. thermophilus) and coagulation markedly changed and became dominated by Streptococcus (100%) (Fig. 4B). At this manufacturing step, all the other genera were not detectable. Only in the curd after incubation was Streptococcus accompanied by Lactococcus (0.22%) and Lactobacillus at very low frequency (0.04%). From the curd after stretching to 15 days of ripening, Streptococcus still dominated (ca. 99%) and Lactobacillus was present in low numbers. In agreement with the numbers of presumptive thermophilic streptococci and mesophilic lactobacilli (Table 1), Streptococcus and Lactobacillus, respectively, decreased (80.7%) and increased (19.2%) at 30 days of ripening. This trend maintained throughout ripening, and at 90 days the bacterial profile of the Caciocavallo Pugliese was characterized by Streptococcus (79.2%), followed by Lactobacillus (20.1%). Other genera, occasionally and in low abundance, were also present. In particular, Paracraurococcus, Pseudoalteromonas, and Rhodococcus (after 30 days) and Azospirillum and Gelria (after 75 days) were found.

The abundance of OTUs from RNA samples (Firmicutes only) is reported in Table 3, with taxonomic details up to species level when such assignment was possible. Lactococcus lactis (25.6%) and St. thermophilus (11.5%) were the most abundant species in raw cows' milk, followed by Lactococcus sp. (5.62%), Streptococcus sp. (3.2%), Streptococcus parauberis (1.34%), Lactobacillus kefiranofaciens (1.28%), and Lactobacillus plantarum (1.2%). Lactobacillus paracasei, L. casei, and Lactobacillus sp. were also found at very low frequency. After the addition of the primary starter, St. thermophilus (82.6%) and Streptococcus sp. (16.8%) dominated the curd. They persisted at the same ratio up to 15 days of ripening. After 30 days, the microbiota changed. The relative abundance of St. thermophilus decreased (ca. 65%) and remained almost constant. At the same time, Streptococcus sp. also decreased slightly, whereas L. paracasei became the dominant species of mesophilic lactobacilli, which persisted up to 90 days of ripening (11%). Although less abundant (ca. 3%), the L. casei group and Lactobacillus sp. accompanied L. paracasei and stably persisted.

Changes in the community level catabolic profiles.

The relative utilization of carbon sources (amino acids, carbohydrates, carboxylic acids, amines, and polymers) by the bacterial community varied during manufacture and ripening of Caciocavallo Pugliese (Fig. 5). Carbohydrates and amino acids followed by carboxylic acids were the chemical classes mainly used over time. After curd incubation, the most intense degradation of all the chemical classes occurred. The ability to use the carbon sources progressively decreased after curd stretching until 7 days of ripening. After 15 and especially 30 days, the catabolic activity increased and almost stabilized over time. Catabolic profiles were also determined using the indices H′, S, and E (see Table S4 in the supplemental material). Raw cows' milk showed the lowest values of substrate utilization (H′) (2.75 ± 0.2) and substrate richness (S) (13.33 ± 1.6). These indices significantly (P < 0.05) increased during cheese manufacture and ripening, and the highest values were found after curd incubation (3.08 ± 0.1 and 16.87 ± 1.9, respectively). The index E, a measure of the statistical significance (equitability) of the values of H′ and S, confirmed the significant (P < 0.05) differences over time.

Correlations among microbiota, proteolysis, catabolic profiles, and volatile components.

All the OTUs were used to find correlations. Only positive correlations (P < 0.05; r > 0.6) were further detailed. The concentration of total FAA and the abundance of Lactobacillus sp. and L. paracasei were correlated (P < 0.05). Also, the concentration of Asp, Glu, Leu, Phe, and Val was mainly correlated with L. paracasei (r > 0.80). The level of Lactobacillus sp. and L. paracasei was also correlated with the area of hydrophilic peptide peaks, whereas the abundance of St. thermophilus was positively correlated with the area of hydrophobic peptide peaks. St. thermophilus was also positively correlated with the substrate richness (S index) and with the utilization of amines and carbohydrates (P < 0.05). Regarding VOC, the abundance of L. paracasei was positively correlated with several aldehydes (e.g., propanal, butanal, nonanal, 2,4-hexadienal, 3-methyl-butanal, and benzaldehyde), alcohols (2-propanol, 2-butanol, 2-pentanol, 2-heptanol, and 3-methyl-3-buten-ol), ketones (2-hexanedione, 2-methyl 3-pentanone, cyclopentanone, 1-phenyl-ethanone, and 4-heptanone), esters (isopropyl acetate and 3-methyl-butyl acetate), sulfur compounds (carbon disulfide and thiophene), and all furans. St. thermophilus was correlated only with acetaldehyde, 2,3-butanedione (diacetyl), and 2,3-pentanedione. Permutation analysis based on the above parameters clustered cheeses throughout ripening (Fig. 6).