Abstract

   Double-Layer Steamed Milk Custard (DLSMC) is a famous traditional
   Chinese dessert. This study aimed to analyze the flavor and the changes
   in metabolites during different stages of DLSMC preparation, including
   raw buffalo milk, thermo-processing, first and second-layer milk skin
   formation. Electronic nose and electronic tongue were employed to
   preliminarily assess the overall flavor characteristics between these
   stages. The results indicated that DLSMC demonstrated increased
   sweetness, saltiness, and umami compared to raw buffalo milk, along
   with heightened levels of nitrogen oxides, methane, alcohols,
   aldehydes, ketones, and sulfur-containing compounds. Thus, the
   thermo-processing and second-layer milk skin formation were pinpointed
   as critical for flavor alterations. Additionally, headspace-gas
   chromatography-ion mobility spectrometry (HS-GC-IMS) detected a total
   of 46 volatile organic compounds (VOCs), with 8 compounds identified as
   key flavor components. Untargeted metabolomics revealed 98 differential
   metabolites, including amino acids, lipids, and nucleotides, that were
   significantly linked to changes in key flavor compounds. Indeed, the
   fluctuations in the levels of amino acids, lipids, and nucleotides play
   a crucial role in influencing flavor changes during the production of
   DLSMC made from buffalo milk.

Introduction

   In southern China, buffalo farming has a rich history and offers
   significant economic benefits through the production of milk, meat,
   leather, and draft power [[38]1]. As a significant component of global
   milk resources, buffalo milk has higher levels of fat than other milk
   [[39]2]. Previous research indicates that buffalo milk is a rich source
   of polyunsaturated fatty acids and amino acids, both are beneficial for
   human lipid metabolism [[40]3,[41]4]. Buffalo milk, with its higher fat
   and protein content, is more suitable for the production of high-fat
   dairy products compared to other types of milk [[42]5]. Buffalo milk
   mozzarella cheese was rich in various volatile flavor substances,
   including hexanal, heptanal, ethyl caproate, (E)-2-nonenal, furfural,
   and ethyl butyrate. Compared to mozzarella prepared from holstein milk,
   mozzarella prepared from buffalo milk exhibits a more complex aroma
   profile [[43]6]. Double-Layer Steamed Milk Custard (DLSMC) is renowned
   as a “Famous Chinese Snack” in Guangdong province, China. It represents
   a key part of Chinese traditional dairy products, reflecting the
   essence of classic dairy culture. The technique of making DLSMC was
   included in the seventh batch of provincial intangible cultural
   heritage representative items in Guangdong province in 2018. DLSMC,
   made from buffalo milk, eggs, and sugar, is characterized by the
   formation of two layers on the surface during production. Raw buffalo
   milk is boiled and cooled to form the first layer of milk skin, which
   is then pierced to pour out the remaining milk. After adding eggs and
   sugar, the mixture is stirred and steamed, resulting in a second layer
   of milk skin. The method primarily involves heating, cooling, and
   stirring. The changes caused by heat treatment—such as microbial
   killing, whey protein and casein cross-linking, fat oxidative
   degradation, Maillard reaction, and thermal denaturation—can alter the
   flavor components and metabolites in milk [[44]7–[45]10].

   A variety of analysis techniques have been widely used to detect
   volatile organic compounds (VOCs) responsible for the volatile flavor
   in milk and dairy products [[46]11–[47]14]. Electronic nose and
   electronic tongue are electronic instruments that mimic human sensory
   responses and provide objective analysis for flavor detection [[48]15].
   During the thermal treatment of milk, metabolites such as lipids and
   proteins can be easily oxidized to VOCs including aldehydes, alcohols
   and ketones [[49]16]. Compared to gas chromatography (GC) and gas
   chromatography-mass spectrometry (GC-MS), headspace-gas
   chromatography-ion mobility spectrometry (HS-GC-IMS) provides several
   advantages, including with no preprocessing, rapid response times, high
   sensitivity and convenient operation [[50]17]. Consequently, electronic
   nose, electronic tongue, and HS-GC-IMS were utilized to comprehensively
   analyze flavor formation in this study.

   In recent years, non-targeted metabolomics has become a powerful tool
   for exploring metabolic changes under various conditions, particularly
   in complex samples [[51]18,[52]19]. Non-targeted metabolomics detects
   differential metabolites under different heat treatment conditions and
   establishes predictive models to effectively distinguish between
   pasteurized milk and overheated milk [[53]20]. Additionally, it can be
   used to identify metabolites in buffalo milk across different varieties
   [[54]21]. Dairy samples contain numerous small molecule metabolites
   with a wide range of dynamic changes in concentration and polarity.
   Therefore, the high sensitivity and resolution of mass
   spectrometry-based metabolomics make it particularly suitable for
   analyzing the metabolites in buffalo milk and dairy products.
   UHPLC-Orbitrap Exploris 240 MS shows great potential in analyzing the
   composition of buffalo milk.

   This research was designed to investigate the variations in flavor and
   metabolites of DLSMC and analyzed correlation among them during the
   production process. In our study, the flavor and taste profiles at
   different production stages were analyzed using electronic nose and
   electronic tongue. The VOCs in the production process of DLSMC were
   determined by HS-GC-IMS. Non-targeted metabolomics technology using
   UHPLC-Orbitrap Exploris 240 MS was utilized to determine the
   alterations in metabolites during production. This study also explored
   the relationship between VOCs and metabolic compounds in DLSMC. This
   research proposes a novel perspective for investigating VOCs in DLSMC,
   providing a scientific basis for understanding the unique flavor and
   nutritional characteristics of DLSMC. Additionally, it establishes a
   theoretical framework for enhancing the processing methodology of
   DLSMC, as well as quality control throughout the processing.

Materials and methods

Preparation method for Double-Layer Steamed Milk Custard

   Buffalo milk samples were collected from the buffalo breeding farm at
   the Buffalo Research Institute, Chinese Academy of Agricultural
   Sciences in Nanning, Guangxi, China. The method was derived from the
   traditional Double-Layer Steamed Milk Custard technique documented in
   Foshan Intangible Cultural Heritage Database
   ([55]https://foshanfy.mgdatatech.com/feiyi/project/show/id/38.html).
   Raw milk is boiled for 10 minutes and cooled at 25°C for 30 minutes,
   during which a first layer of milk skin forms. A small hole is then
   pierced in the skin, allowing the buffalo milk to be poured out. The
   first layer of milk skin sticks to the bottom of the bowl. After adding
   eggs and sugar in the standardized ratio of raw buffalo milk: egg:
   sugar = 10:1:1 (w/w/w), the mixture is steamed in a pot for 20 minutes,
   leading to the formation of a second milk skin. Samples were collected
   at four key stages of the production process: raw buffalo milk (group
   A), thermo-processing (100°C for 10 minutes, group B), first-layer of
   milk skin formation (25°C for 30 minutes, group C), and second-layer of
   milk skin formation (100°C 20 minutes, group D). Three samples were
   collected at each stage.

Detection of flavor by electronic nose

   The AIRSENSE PEN 3 portable electronic nose (Germany) contains 10
   sensors, including W1C, W5S, W3C, W6S, W5C, W1S, WIW, W2S, W2W, W3S.
   Ten sensors detect different types of VOCs, as shown in the [56]S1
   Table [[57]10]. Weigh 5 g ± 0.1 g of the sample and equilibrate it at
   40°C for 20 minutes before testing. The instrument parameters were
   configured as follows: flush time of 60 seconds, zero-point trim time
   of 5 seconds, measurement time of 180 seconds, and gas flow rate set to
   300 mL/min. Every sample underwent once tests. Three relatively stable
   data were selected for further analysis using the Winmuster software
   (version 1.6.2.22; Airsense Analytics GmbH).

Detection of taste by electronic tongue

   The electronic tongue detection was conducted using the Taste-Sensing
   System SA 402B (Intelligent Sensor Technology Co., Ltd., Atsugi,
   Japan). The electronic tongue is equipped with six sensors: CT0 for
   saltiness, CA0 for sourness, AAE for umami and aftertaste-U, AE1 for
   astringency and aftertaste-A, C00 for bitterness and aftertaste-B, and
   GL1 for sweetness. Following dilution with water at a 1:1 ratio and
   centrifugation at 4500 rpm for 20 minutes, both the upper fat layer and
   lower solids were removed. A 30 mL ± 0.5 mL sample was placed in a
   sample test cup. All samples were tested four times. The first test
   data was discarded, and the data from the last three tests were used
   for analysis.

Detection of volatile organic compounds by HS-GC-IMS

   VOCs were analyzed by HS-GC-IMS (FlavourSpec®-G.A.S. Dortmund Company,
   Germany) according to Zhao's study [[58]13]. A 2 g ± 0.05 g sample was
   placed in a 20 mL headspace vial, incubated at 80°C, and shaken at
   500 rpm for 20 minutes. A 500 μL headspace sample was injected at an
   injection needle temperature of 85°C. The GC was fitted with a
   chromatographic column MXT-WAX (polar, 30 m × 0.53 mm × 1.0 μm; RESTEK,
   USA) and operated at 60°C for separation. Nitrogen (purity ≥99.999%)
   was used as the carrier gas, with an initial flow rate of 2 mL/min. The
   GC was operated under the following conditions: the flow rate was
   maintained at 2 mL/min for 2 minutes, from 2 mL/min to 10 mL/min within
   8 min, from 10 mL/min to 100 mL/min within 10 min, from 100 mL/min to
   150 mL/min within 10 min. The drift gas flow rate was set at
   150 mL/min, while the column temperature was maintained at 60°C. A
   mixture of ketones, including 2-butanone, 2-pentanone, 2-hexanone,
   2-heptanone, 2-octanone, and 2-nonanone (C4-C9), was used to formulate
   the column-modified curve for calculating the retention index (RI). LAV
   software (v2.2.1) was utilized for analyzing the HS-GC-IMS data, while
   GC × IMS Library Search (v1.0.3) was employed for qualitative analysis
   within the GC-IMS library. VOCs were identified by comparing retention
   indices (RI) and drift times (DT) with reference data from the GC-IMS
   spectrum library.

Metabonomics analysis

   Weigh 200 ± 5 mg of the sample precisely, and add 800 µL of 80% aqueous
   methanol solution (methanol: water = 4:1 (v:v), HPLC grade) containing
   the internal standard solution L-2-chlorophenylalanine
   (0.02 mg/mL, ≥ 98%). The tissue sample is ground for 6 minutes at −10°C
   and 50 Hz using a cryogenic grinder, followed by ultrasonic extraction
   for 30 minutes at 5°C and 40 kHz. After that, the sample is placed at
   −20°C for 30 minutes and centrifuged at 13,000 g and 4°C for 15
   minutes. The supernatant is carefully collected for further analysis. A
   portion of the supernatant from each sample is combined to create a
   quality control (QC) sample. Analysis is performed on a UHPLC-Exploris
   240 system (Thermo Fisher Scientific Inc., Waltham, MA, USA).
   Separation of 3 µL samples is achieved using an ACQUITY UPLC HSS T3
   column (100 mm × 2.1 mm, 1.8 µm, Waters, Milford, USA), followed by
   mass spectrometric detection. The column is maintained at 40°C. Mass
   spectra are acquired in both positive and negative ion modes, covering
   a mass range of 70–1050 m/z. The sheath gas flow is set at 60 psi,
   auxiliary gas at 20 psi, and the auxiliary gas temperature at 350°C.
   Ion spray voltages are set to 3400 V for positive mode and −3000 V for
   negative mode, while the ion transfer tube is heated to 350°C. The
   primary and secondary mass spectrometric resolutions are set at 60,000
   and 15,000 respectively. Each sample were detected once in the data
   dependent acquisition (DDA) mode.

Data analysis

   The raw data was processed using the Progenesis QI metabolomics
   software (Waters Corporation, Milford, USA), which performed baseline
   filtering, peak identification, integration, retention time correction,
   and peak alignment. The MS and MS/MS mass spectrometry information were
   then compared with public databases such as Human Metabolome Database
   (HMDB) ([59]http://www.hmdb.ca/), Metlin
   ([60]https://metlin.scripps.edu/), as well as a self-built database.
   Following the database search, the resulting matrix data was uploaded
   to a cloud platform ([61]https://cloud.majorbio.com) for further
   analysis. During data preprocessing, the 80% rule was applied to
   exclude missing values. To mitigate errors from sample preparation and
   instrument variation, the response intensity of key spectral peaks was
   normalized using the sum normalization method. Variables with a
   relative standard deviation (RSD) greater than 30% in QC samples were
   removed, and log10 transformation was performed on the data.

   Multivariate statistical methods, including unsupervised principal
   component analysis (PCA) and orthogonal partial least
   squares-discriminant analysis (OPLS-DA), were conducted using the
   SIMCA-P 14.1. Differential metabolites were selected based on the
   variable importance in projection (VIP) value from the OPLS-DA analysis
   and p-values obtained from the Student's t-test. Metabolites with
   VIP > 1 and p < 0.05 were classified as differential. Metabolic
   pathways for the differential metabolites were identified through
   annotation using the Kyoto Encyclopedia of Genes and Genomes (KEGG)
   database ([62]https://www.kegg.jp/kegg/pathway.html).

Results and discussion

Electronic nose analysis

   The response intensity of the electronic nose sensor reveals the flavor
   profile characteristics of various samples, enabling effective
   differentiation among them [[63]22]. The response values are shown in a
   radar map ([64]Fig 1a). Notable differences were observed among the
   samples, indicating that various steps influenced the flavor formation
   of DLSMC. The response values of group A differ from those of groups B,
   C, and D. The response intensity of the sensors in group B changed to
   varying degrees, with W1W and W2W increasing, indicating that the
   content of sulfur compounds has increased. Sulfur compounds contribute
   to the cooked flavor, primarily due to the thermal denaturation of
   β-lactoglobulin and milk fat globule membrane proteins, which generates
   thiol groups (-SH). Additionally, thermal degradation of thiamine
   compounds may occur when exposed to dicarbonyl compounds, including
   diacetyl and 2,3-pentanedione. The response intensity of W5S, W1S, W1W,
   W2S, and W2W exhibited an increase in group D, signifying that during
   the second layer of milk skin formation, VOCs including nitrogen
   oxides, inorganic sulfides, alcohols, aldehydes, and organic sulfur
   compounds were generated and amalgamated to create the characteristic
   flavor of DLSMC. This may be associated with changes in metabolites or
   the release of VOCs due to heat treatment. PCA can extract information
   from sensor response values, utilizing new variables to represent data
   features [[65]23]. In [66]Fig 1b, PC1 and PC2 explained 77.2% and 10.2%
   of the total variance, respectively, capturing the majority of the
   data's variability. The spatial distribution between groups A, B, and C
   was similar. Among them, some response values of samples in groups B
   and C overlapped, while there was no overlap in response values between
   the first three groups and group D. Group D gathers only along the
   negative axis of the PC1 and is associated with the sensors W5S, W1S,
   W1W, W2S, W2W, and W3S ([67]Fig 1c). There was some overlap among
   samples A, B, and C, indicating that there was some correlation between
   these samples and the sensors W1C, W3C, W6S, and W5C. The analysis
   revealed that during the preparation of DLSMC, thermo-processing and
   the second-layer milk skin formation were the critical stages for
   flavor compound changes, while the flavor changes during the formation
   of the first milk skin were relatively minor.

Fig 1. The response results of the electronic nose and electronic tongue.

   [68]Fig 1
   [69]Open in a new tab

   Radar chart of electronic nose (a) and electronic tongue (d) responses.
   Principal component analysis (PCA) score plot for electronic nose (b)
   and electronic tongue (e). PCA loading plot for electronic nose (c) and
   electronic tongue (f) data.

Electronic tongue analysis

   Electronic tongue was utilized to evaluate the taste characteristics of
   the samples during the production process. The radar map is shown in
   [70]Fig 1d. All taste indicators except sourness were valid, as the
   sourness of different groups was below the tasteless (−13) and was thus
   deemed to have no sourness [[71]24]. The primary taste components of
   the samples included sweetness, bitterness, astringency, umami,
   richness, and saltiness, with sweetness being the most dominant across
   all samples. Group D exhibited comparatively elevated levels of
   saltiness, umami, and sweetness, while demonstrating a reduced level of
   bitterness compared to groups A, B, and C. As a result, the DLSMC
   formed after processing is saltier, fresher, sweeter, and less bitter.

   Excluding sourness values, the value of other sensors was applied for
   PCA, and the score plot and loading plot were shown in [72]Figs 1E and
   [73]1F. The first two principal components, PC1 and PC2, contribute
   58.7% and 29.0% of the total variation. Group D, clustered in the
   second quadrant, is associated with sweet and savory flavors. The
   distances between the three groups A, B, and C were very close,
   indicating that the taste profiles of these three groups are relatively
   similar. Group A, representing raw buffalo milk, was located in the
   third quadrant and was associated with umami. These results indicated
   that the production process primarily influences bitterness, umami,
   richness, saltiness, and sweetness. These changes primarily occurred
   during the formation of the second milk skin.

Variations in volatile flavor compounds during process of Double-Layer
Steamed Milk Custard

Identification of volatile organic compounds via HS-GC-IMS analysis.

   HS-GC-IMS can obtain comprehensive ion migration spectra of samples. 64
   signal peaks were detected during the process ([74]Table 1), of which
   46 were identified as known signals and 18 as unknown signals. Known
   compounds include aldehydes, alcohols, ketones, esters, and ethers. To
   understand VOC changes between samples at different stages, a 3D
   topographic plot was displayed ([75]Fig 2a). Significant differences in
   peak positions and intensities were observed across processing stages,
   indicating dynamic changes in volatile profiles during preparation. The
   two-dimensional difference analysis ([76]Fig 2b) revealed that the
   measurement run time during DLSMC production primarily occurred between
   200–1,200 s. Every spot flanking the RIP peak corresponds to a VOC,
   with the color reflecting its concentration [[77]25]. The signal peaks
   in group A were distributed between 200–800 s. Compared to group A,
   group B exhibited peak concentration within 200–450 s, while group C
   showed similar volatile signatures to group B. Group D showed the most
   significant changes across all stages, exhibiting more spots and darker
   red. This result indicates that the most significant divergence from
   the raw buffalo milk occurred during the second layer of milk skin
   formation, with a significant increase in various volatile compounds.

Table 1. The signal peak of VOCs detected by HS-GC-IMS.

   No Compound CAS number Formula Molecular weight Retention Index
   Retention time [sec] Drift time [a.u.]
   1 1 / / / 748.40 212.52 1.23
   2 Dimethyl sulfide 75-18-3 C[2]H[6]S 62.1 773.70 224.85 0.96
   3 2-Methylpropanal 78-84-2 C[4]H[8]O 72.1 798.40 237.57 1.09
   4 2 / / / 815.60 246.86 1.18
   5 Propanal 123-38-6 C[3]H[6]O 58.1 822.20 250.51 1.14
   6 Ethyl formate 109-94-4 C[3]H[6]O[2] 74.1 822.90 250.90 1.08
   7 3 / / / 827.10 253.26 1.20
   8 Acetone 67-64-1 C[3]H[6]O 58.1 854.50 269.19 1.12
   9 4 / / / 881.50 285.91 1.20
   10 2-Methyl-2-propanol 75-65-0 C[4]H[10]O 74.1 883.00 286.84 1.14
   11 5 / / / 884.70 287.91 1.27
   12 6 / / / 887.30 289.59 1.30
   13 2-Propanol 67-63-0 C[3]H[8]O 60.1 889.00 291.42 1.22
   14 Ethyl acetate (M) 141-78-6 C[4]H[8]O[2] 88.1 891.10 292.07 1.10
   15 Ethyl acetate (D) 141-78-6 C[4]H[8]O[2] 88.1 891.70 292.47 1.34
   16 2-Butanone (D) 78-93-3 C[4]H[8]O 72.1 908.90 303.91 1.25
   17 2-Butanone (M) 78-93-3 C[4]H[8]O 72.1 909.20 304.06 1.06
   18 7 / / / 918.60 310.49 1.33
   19 8 / / / 918.70 310.59 1.09
   20 3-Methylbutanal 590-86-3 C[5]H[10]O 86.1 920.90 312.14 1.40
   21 9 / / / 922.30 313.11 1.22
   22 10 / / / 922.90 313.53 1.29
   23 Diacetyl 431-03-8 C[4]H[6]O[2] 86.1 927.10 316.43 1.18
   24 Ethyl propionate 105-37-3 C[5]H[10]O[2] 102.1 936.40 322.11 1.14
   25 11 / / / 958.70 339.53 1.04
   26 Propyl acetate (D) 109-60-4 C[5]H[10]O[2] 102.1 977.40 354.00 1.48
   27 Propyl acetate (M) 109-60-4 C[5]H[10]O[2] 102.1 977.40 354.00 1.16
   28 4-Methyl-2-pentanone 108-10-1 C[6]H[12]O 100.2 984.70 356.68 1.18
   29 2-Pentanone (D) 107-87-9 C[5]H[10]O 86.1 987.20 361.80 1.37
   30 2-Pentanone (M) 107-87-9 C[5]H[10]O 86.1 989.30 363.47 1.12
   31 Ethyl butyrate 105-54-4 C[6]H[12]O[2] 116.2 993.60 363.41 1.20
   32 12 / / / 1002.40 375.27 1.09
   33 3-Methyl-2-pentanone 565-61-7 C[6]H[12]O 100.2 1014.20 384.07 1.18
   34 13 / / / 1014.70 387.56 1.04
   35 1-Penten-3-one 1629-58-9 C[5]H[8]O 84.1 1017.30 387.42 1.10
   36 Methyl 2-methylbutyrate 868-57-5 C[6]H[12]O[2] 116.2 1023.40 393.94
   1.19
   37 14 / / / 1023.80 397.00 1.39
   38 Ethyl 2-methylbutyrate 7452-79-1 C[7]H[14]O[2] 130.2 1024.60 395.30
   1.23
   39 2-Butanol (D) 78-92-2 C[4]H[10]O 74.1 1026.90 397.81 1.32
   40 2-Butanol (M) 78-92-2 C[4]H[10]O 74.1 1027.50 398.42 1.15
   41 15 / / / 1032.30 406.03 1.04
   42 16 / / / 1041.00 415.38 1.11
   43 Ethyl 3-methylbutanoate 108-64-5 C[7]H[14]O[2] 130.2 1041.00 413.50
   1.26
   44 2-Methyl-1-propanol 78-83-1 Formula 74.1 1099.70 486.81 1.17
   45 Diallyl sulfide 592-88-1 / 114.2 1136.40 546.49 1.13
   46 17 / C[2]H[6]S / 1136.60 546.32 1.07
   47 Ethyl crotonate 623-70-1 C[4]H[8]O 114.1 1151.90 573.77 1.18
   48 1-Penten-3-ol 616-25-1 / 86.1 1168.60 604.72 0.94
   49 18 / C[3]H[6]O / 1188.30 644.39 1.66
   50 2-Heptanone (D) 110-43-0 C[3]H[6]O[2] 114.2 1188.90 644.79 1.63
   51 2-Heptanone (M) 110-43-0 / 114.2 1189.10 645.19 1.26
   52 Heptanal 111-71-7 C[3]H[6]O 114.2 1190.30 647.67 1.34
   53 4-Methyl-2-pentanol 108-11-2 / 102.2 1211.10 675.85 1.29
   54 3-Methyl-1-butanol (M) 123-51-3 C[4]H[10]O 88.1 1213.90 679.50 1.25
   55 3-Methyl-1-butanol (D) 123-51-3 / 88.1 1214.90 680.89 1.50
   56 (E)-2-Hexenal 6728-26-3 / 98.1 1257.00 739.19 1.18
   57 1-Pentanol 71-41-0 C[3]H[8]O 88.1 1259.10 742.15 1.26
   58 3-Heptanol 589-82-2 C[4]H[8]O[2] 116.2 1291.50 790.62 1.33
   59 2-Methylpyrazine (D) 109-08-0 C[4]H[8]O[2] 94.1 1292.20 791.79 1.41
   60 2-Methylpyrazine (M) 109-08-0 C[4]H[8]O 94.1 1292.20 791.79 1.07
   61 6-Methyl-5-hepten-2-one 110-93-0 C[4]H[8]O 126.2 1293.80 794.28 1.17
   62 3-Octanol 589-98-0 / 130.2 1390.90 967.09 1.40
   63 Hexyl butyrate 2639-63-6 / 172.3 1395.40 976.01 1.49
   64 Acetic acid 64-19-7 C[5]H[10]O 60.1 1494.20 1192.46 1.06
   [78]Open in a new tab

   Note: The signal peaks of VOCs after headspace extraction were detected
   successively due to their different mobility, and some single compounds
   may produce multiple signals of dimers depending on their
   concentration. (M) monomer, (D) dimer.

Fig 2. VOCs identified by HS-GC-IMS.

   [79]Fig 2
   [80]Open in a new tab

   (a) 3D topographic plot; The horizontal coordinate of the signal peak
   indicates the drift time of VOC, the vertical coordinate indicates
   measurement run time, and the volume of the signal peak indicates the
   concentration of VOC. (b) 2D topographic plots; The color spots on
   either side of the reactive ion peak (RIP) represent the response
   signal peaks of the substance, and their shades indicate the
   concentration of that substance. Using group A as a control, blue spots
   signify that the flavor compounds are below the values of group A,
   while red spots indicate higher values than group A. (c) Gallery plot.
   “M” represents monomer, and “D” stands for dimer. (d) PLS-DA score
   plots and (e) VIP values of samples during different processing.

Volatile organic compounds in different processing stages of Double-Layer
Steamed Milk Custard.

   All peaks were selected to compare VOCs across different samples. The
   Gallery Plot ([81]Fig 2c) displays comprehensive VOC information and
   illustrates the dynamic changes in VOCs. Esters and alcohols were the
   most abundant VOCs detected during the production of DLSMC, followed by
   ketones and aldehydes. 25 signal peaks were detected across all 4
   groups, with 18 VOCs identified, including 6 ketones, 6 esters, 3
   alcohols, 1 ether, 1 acid, and 1 aldehyde, corresponding to region (Ⅰ)
   of the plot. Ketones are typical VOCs in various types of milk, with
   acetone being an important carbonyl compound that imparts a slightly
   sweet fruity flavor. 2-Pentanone and 2-heptanone are characterized by
   sweet and creamy flavors, respectively. Ketones primarily derive from
   the β-oxidation series reactions of saturated fatty acids in milk fat
   [[82]26]. Esters are common VOCs in milk and dairy products,
   potentially arising from the esterification reaction of fatty acids and
   free alcohols released by hydrolysis in the lactation system under the
   action of endogenous esterase. Aldehydes mainly originate from lipid
   oxidation in buffalo milk [[83]27].

   Methyl 2-methylbutyrate, 4-methyl-2-pentanone, ethyl butyrate,
   2-methylpropanol, 1-pentanol, 6-methyl-5-hepten-2-one, 2-methylpyrazine
   (M), 3-methyl-1-butanol (M), 4-methyl-2-pentanol, 3-methyl-1-butanol
   (D), 3-methylbutanal, and 3-heptanol are present in raw buffalo milk,
   with their contents decreasing after thermo-processing, marked as
   region (Ⅱ). 3-Heptanol and 4-methyl-2-pentanol are only found in raw
   buffalo milk, indicating that these substances may be the
   characteristic VOCs of raw buffalo milk. Region (Ⅲ) represents flavor
   compounds produced or significantly increased during thermo-processing.
   A minimal quantity of diacetyl was detected in raw buffalo milk (group
   A), whereas an increase in diacetyl content was observed in group B.
   Previous research has indicated that diacetyl exists in raw buffalo
   milk at 65.2 mg/kg [[84]28]. And diacetyl is a compound characterized
   by its butter-like flavor, which has been demonstrated to form in milk
   as a result of heat treatment [[85]29]. Diacetyl, acetone, and
   2-butanone may originate from lipid oxidation or the degradation of
   reducing sugars into carbonyl compounds [[86]30].

   In DLSMC, 60 peaks were detected. After the second-layer milk skin
   formation, the concentrations of propanal, ethyl acetate (D),
   1-penten-3-ol, 3-octanol, 2-methyl-2-propanol, heptanal, 2-heptanone
   (D), diallyl sulfide, 2-methyl-1-propanol, 2-methylpyrazine (D), and
   hexyl butyrate increased, corresponding to region (Ⅳ). Ketones are
   mainly produced by fat degradation during thermal processing and
   decarboxylation reactions after the oxidation of free fatty acids to
   β-ketoacids. These compounds are important flavor contributors in dairy
   products [[87]13,[88]26]. 2-Heptanone, produced from β-keto acids, is
   the most common compound in the heat treatment of milk. Compared to raw
   milk, ultrahigh-temperature sterilization also significantly increased
   the levels of methyl ketones, including 2-heptanone, leading to the
   formation of oxidized flavors in the milk [[89]31]. Aldehydes, which
   may result from milk fat oxidation, contribute to milk flavor due to
   their strong volatility, high concentration, and low threshold.
   Heptanal is primarily formed by the oxidation of linoleic and
   arachidonic acids, while octanal and nonanal are primarily formed by
   the oxidation of oleic acid. Most of these compounds may result from
   the Maillard reaction or the thermal degradation of fat during making.
   Leucine and isoleucine produce 3-methylbutyral and 2-methylbutyral,
   respectively, through Strecker degradation, providing a fatty aroma and
   nutty taste [[90]32]. Ester compounds, especially fatty acid ethyl
   esters, exhibit fruity and sweet flavors that can reduce the bitterness
   brought by fatty acids and amines [[91]33]. As shown in the electronic
   tongue results, the bitterness of group D was the lowest, possibly
   resulting from a rise in the variety and concentration of ester
   compounds. This increase reduced the bitterness, enriched the flavor
   profile of DLSMC, and optimized its sensory quality.

The key flavor compounds during different processing.

   The PLS-DA model was established ([92]Fig 2d), and the combined
   variance explained by the top three PCs was 96.7%, providing a strong
   basis for the model's discriminative ability. The 3D score map
   illustrated distinctions among the four groups, indicating differences
   in the types and contents of aroma compounds across them. The VIP value
   can further evaluate and classify the parameters, distinguishing
   different groups. In total, 8 compounds with VIP > 1 were identified as
   key VOCs for distinguishing aroma compounds among the groups ([93]Fig
   2e). VOCs such as 2-butanone, 2-heptanone, 3-methyl-1-butanol, acetone,
   2-methylpyrazine, 2-butanol, 2-pentanone, and 3-heptanol significantly
   impacted the flavor characteristics of DLSMC. Ketones, which account
   for 62.50% of all key flavor compounds, are considered the primary
   contributors to the creamy and sweet taste in dairy products [[94]13].
   2-Butanone has been identified as the compound responsible for the egg
   flavor. This oxygenated compound is primarily formed as a result of
   lipid oxidation [[95]34]. After thermal treatment, the content of
   2-butanone in the sample changed the most, making it a key compound for
   the study of DLSMC made from buffalo milk. Ketones are mainly derived
   from the β-oxidation reactions of saturated fatty acids in milk fat.
   This suggests that the change in flavor substances in DLSMC may be
   associated with fatty acid oxidation.

Changes in metabolites during the processing of Double-Layer Steamed Milk
Custard

Analysis of metabolite differences and pathways during processing of
Double-Layer Steamed Milk Custard.

   Raw data undergoes pre-processing steps such as identification,
   filtering, complementing, normalization, and logarithmic
   transformation, which can eliminate or reduce errors caused by
   experiments and analysis. The VIP value can be used to identify
   metabolites that significantly contribute to differences between
   groups. According to OPLS-DA, criteria of VIP > 1 and p < 0.05 were
   used to search for metabolites that were significantly differentially
   expressed between groups A and D. Overall, 913 distinct metabolites
   were identified, including amino acids, nucleotides, lipids, and other
   metabolites. For additional details, please consult the supplementary
   material ([96]S3 Fig). We investigated buffalo milk metabolic pathways
   during processing, potentially involving amino acid metabolism, lipid
   metabolism, and glycolysis/gluconeogenesis.

   For a deeper analysis of the potential metabolic pathways involved in
   the flavor formation of DLSMC production, KEGG database was used to
   analyze the functions of pathways related to differential metabolites.
   As shown in the [97]S2 Table, a total of 913 different metabolites were
   enriched in 60 metabolic pathways. There were 19 significantly enriched
   metabolic pathways with an impact value greater than 0.1 ([98]Fig 3).
   The results showed that the effects of nucleotide metabolism were the
   most significant. Additionally, metabolic pathways such as
   phenylalanine, tyrosine and tryptophan biosynthesis,
   glycerophospholipid metabolism, purine metabolism, tryptophan
   metabolism, pyrimidine metabolism, glycine, serine and threonine
   metabolism were also significantly enriched. These pathways involve
   nucleotides, amino acids, and lipids, suggesting that these compounds
   may primarily contribute to the significant differences in VOCs from
   buffalo milk to DLSMC.

Fig 3. Pathway enrichment analysis of differential metabolites of
Double-Layer Steamed Milk Custard during different processing.

   [99]Fig 3
   [100]Open in a new tab

   Colors (varying from blue to orange) denote increased significance
   levels, and the horizontal coordinate indicates the impact value of the
   pathway.

Analysis of metabolite differences during the processing of Double-Layer
Steamed Milk Custard.

   To better visualize sample relationships and variations in metabolite
   expression, the relative values of metabolites across different groups
   were used as metabolic levels, and 98 metabolites enriched in pathways
   were analyzed using hierarchical clustering. Data differences are
   represented using color gradients ([101]Fig 4.).

Fig 4. Changes in metabolites during the processing of Double-Layer Steamed
Milk Custard.

   [102]Fig 4
   [103]Open in a new tab

   Milk is a complex food matrix that is prone to lipolysis and oxidation
   under heat treatment [[104]35]. Linoleic acid derived 9,10-DiHOME,
   13-HODE, 9(S)-HpODE, and α-linolenic acid derived 9(S)-HOTrE initially
   decreased during the production process before gradually increasing and
   stabilizing, while linoleic acid derived 9,10,13-TriHOME and
   9,12,13-TriHOME were highly expressed in groups C and D. HpODE is a
   chemically unstable primary oxidation product of linoleic acid, while
   stable secondary oxidation products, such as HODE and TriHOME can be
   generated through non-enzymatic reduction [[105]36]. Dietary HODEs can
   be directly absorbed, leading to increased levels of HODEs in the body,
   but the extent to which exogenous HODEs contribute to the physiological
   effects of HODEs is unclear [[106]37]. Stirring and heat treatment may
   cause partial disruption of the milk fat globule membrane (MFGM),
   releasing components such as triglycerides, phospholipids, and proteins
   [[107]38]. The protein and phospholipid components released after MFGM
   rupture are more easily digested and absorbed by the human body
   [[108]39].

   Phosphatidylserine (PS), phosphatidylethanolamine (PE),
   phosphatidylglycerol (PG), and phosphatidic acid (PA) were observed to
   gradually decrease in [109]Fig 4. With the heat treatment, the
   oxidization of milk lipids was aggravated and more oxidized fatty acids
   were released [[110]40]. DLSMC contains a rich variety and high levels
   of lysophosphatidylcholine (LPC) compounds. Some may have been formed
   by oxidation during heat treatment, while others may have resulted from
   the addition of eggs. Eggs also contain lysophospholipids, namely
   lysophosphatidylcholine (LPC) and lysophosphatidylethanolamine (LPE)
   [[111]34]. Docosahexaenoic acid (DHA) is highly concentrated in the
   brain and its deficiency is associated with several neurological
   disorders including Alzheimer's disease. The enrichment of
   docosahexaenoic acid (DHA) in the brain is highly dependent on the
   molecular carrier of dietary DHA, with studies showing that LPC is more
   effective than phosphatidylcholine or triacylglycerol [[112]41].

   Free amino acids can be absorbed and utilized by the human body, and
   their content and composition can partly reflect the nutritional value
   of food. According to the results, most free amino acids were present
   at higher levels in DLSMC. Aspartic acid, an MSG-like component,
   imparted an umami flavor to DLSMC, while serine and alanine contributed
   a pleasant sweet taste [[113]42]. This may be one of the reasons why
   the DLSMC is sweeter. Aromatic amino acids such as tryptophan,
   phenylalanine, and tyrosine. Shikimic acid, an important intermediate
   in the biosynthesis of aromatic amino acids, gradually increases in
   content during the production process. Tryptophan is an essential amino
   acid that must be ingested through food and is a precursor to the
   synthesis of various biologically important compounds. In the body,
   tryptophan is converted to 5-hydroxytryptophan (5-HTP) and serotonin
   (5-HT), playing a crucial role in regulating adaptive responses to
   environmental changes, including sleep, cognition, and feeding behavior
   [[114]43]. Cysteine in whey protein is a precursor to glutathione, an
   important antioxidant in the body, and taurine, which can clear free
   radicals in the human body [[115]44]. Pyroglutamate, a metabolite of
   the glutathione cycle, is highly expressed in DLSMC. It can participate
   in energy metabolism, provide energy, and has antioxidant,
   anti-inflammatory, and other immune functions.

   In addition to the aforementioned metabolites, we also identified
   various nucleotides involved in nucleotide metabolism among the
   differential metabolites. Exogenous nucleotides (NTs) were essential
   nutrients under specific physiological conditions, including immune
   responses, hunger, rapid growth, and aging. Exogenous NTs can be
   absorbed by human tissues, thereby reducing the body's need for de novo
   synthesis or salvage synthesis [[116]45]. 5'-Cytidine monophosphate
   (5’-CMP) is a flavor nucleotide that, as a single product, does not
   have umami. However, when mixed with sodium glutamate, its umami is
   much greater than the sum of the umami of the individual products
   [[117]46]. Although the content of 5’-CMP gradually decreased and the
   umami taste slightly diminished during the production process, the
   umami taste of DLSMC was enhanced by the addition of free amino acids.
   The combined action of various umami substances made DLSMC rich in
   flavor. Guanosine and inosine are two nucleoside compounds that
   gradually increase in expression during processing. Guanosine is formed
   by the combination of guanine and ribose and can be phosphorylated to
   become guanosine monophosphate (GMP), cyclic guanosine monophosphate
   (cGMP), guanosine diphosphate (GDP), or guanosine triphosphate (GTP).
   These compounds have various physiological functions in the body,
   including participating in RNA synthesis and metabolism, providing
   energy, and affecting the immune system. Inosine, formed by combining
   hypoxanthine with ribose, also has various physiological functions in
   the human body, including participating in RNA synthesis and
   metabolism, providing energy, and affecting the immune system.

   The addition of sucrose significantly enhanced the sweetness of DLSMC.
   Glucose, trehalose, and lactulose also contributed to its sweetness.
   UDP-glucose exhibited a gradual decrease while trehalose demonstrated a
   gradual increase. This is because UDP-glucose may generate alginose
   through Starch and sucrose metabolism under conditions such as
   temperature. Furthermore, lactulose exhibited high expression after
   heat treatment. Lactulose is an isomer of lactose, and its
   concentration can be used as an indicator of heat damage to describe
   milk heat treatment. Lactulose is not metabolized by oral bacteria and
   is therefore not cariogenic. Additionally, it can regulate the
   intestinal microbial community and prevent the growth of harmful
   spoilage bacteria [[118]47]. 3-phosphoglyceric acid is a link in the
   glycolysis process, containing a high-energy phosphate bond, and its
   expression gradually decreases during processing. Moreover,
   3-phosphoglyceric acid is an intermediate in serine synthesis. Buffalo
   milk, rich in protein and sugar, undergoes the Maillard reaction during
   heat treatment, which leads to the production of various VOCs.

Relevance analysis of key volatile organic compounds and differential
metabolites

   The formation of VOCs during processing is a complicated process, and
   differential metabolites may be potential factors affecting the flavor
   of DLSMC. The correlation between 8 key VOCs with VIP > 1 and 98
   different metabolites is shown in [119]Fig 5. The results showed that
   LPC has strong positive correlation with 2-pentanone, 2-heptanone and
   acetone, while PG, PS, PA showed negatively correlation with
   2-pentanone, 2-heptanone and acetone. The oxidation product of linoleic
   acid and α-linolenic acid were negative correlation with 2-butanone and
   2-butanol and have positive correlation with 3-methyl-1-butanol,
   3-heptanol and 2-methylpyrazine. Many aromas typically found in milk
   are generated from short- and medium-chain fatty acids present in milk
   fat. Unsaturated fatty acids are transformed into aldehydes, acids, and
   alcohols, whereas the free fatty acids are transformed into esters
   [[120]48]. Amino acids also had a significant influence on the flavors.
   The most amino acids were positively correlated with 2-pentanone,
   2-heptanone, acetone. This may result from the production of VOCs, such
   as ketones, in a complex series of non-enzymatic browning reactions
   between reducing sugars and free amino acids during the Maillard
   reaction [[121]49]. Glycogen was also an important volatile flavor
   precursor substance. Sucrose and glucose were negatively correlated
   with 2-methylpyrazine. Pyrazines, important products of the Maillard
   reaction, impart a sweet and egg-like aroma to DLSMC [[122]34].

Fig 5. Relationship between key volatile organic compounds and differential
metabolites.

   [123]Fig 5
   [124]Open in a new tab

   In summary, the results indicate that amino acids, lipids, and
   glycogen, as precursors, significantly influence the VOCs of DLSMC.
   Throughout the DLSMC production process, not all metabolites were
   associated with VOCs, and some showed limited or no involvement in the
   formation of key volatile compounds. Generally, the reason for the
   flavor formation during production of buffalo milk Double-Layer Steamed
   Milk Custard was the combination of physicochemical actions such as
   oxidative decomposition of lipids and proteins, and Maillard reaction
   under high temperature. The unique flavor of DLSMC results from complex
   interactions among a wide variety of compounds originating from
   numerous chemical and biochemical reactions involving proteins, fats,
   and carbohydrates.

Conclusion

   This study comprehensively investigated the flavor formation and
   metabolic changes during the production of DLSMC. The results revealed
   that thermo-processing and the formation of the second milk skin layer
   are critical stages for flavor formation. Sulfur compounds, ketones,
   esters, and aldehydes were the main VOCs during the processing of
   DLSMC. HS-GC-IMS identified 46 VOCs, and selected 8 key VOCs (VIP > 1):
   2-butanone, 2-heptanone, 3-methyl-1-butanol, acetone, 2-methylpyrazine,
   2-butanol, 2-pentanone, and 3-heptanol. Metabolomics analysis revealed
   differentially expressed metabolites, which were enriched in pathways
   related to nucleotide metabolism, amino acid metabolism, and lipid
   metabolism, significantly contributing to changes in key compounds. The
   Maillard reaction, lipid oxidation, and protein degradation under heat
   treatment were identified as the primary drivers of flavor changes.
   These findings provide insights into the flavor formation mechanisms
   and highlight the importance of specific processing stages in
   optimizing the sensory quality of DLSMC. However, the characteristic
   flavor compounds in DLSMC and the specific mechanisms underlying their
   changes require further study.

Supporting information

   S1 Fig. HS-GC-IMS 2D topographic plots of the ketone mixture (C4-C9).

   (TIF)
   [125]pone.0331277.s001.tif^ (363.8KB, tif)
   S2 Fig. Calibration curve of the ketone mixture (C4-C9).

   (TIF)
   [126]pone.0331277.s002.tif^ (40.6KB, tif)
   S3 Fig. Differential metabolites between groups A and D.

   (TIF)
   [127]pone.0331277.s003.tif^ (55.3KB, tif)
   S1 Table. Sensor properties of electronic nose.

   (DOCX)
   [128]pone.0331277.s004.docx^ (13.4KB, docx)
   S2 Table. KEGG topology analysis.

   (DOCX)
   [129]pone.0331277.s005.docx^ (22.6KB, docx)
   S1 Dataset. Electronic nose, electronic tongue, and HS-GC-IMS.

   (XLSX)
   [130]pone.0331277.s006.xlsx^ (24.4KB, xlsx)
   S2 Dataset. Non-targeted metabolomics data.

   (DOCX)
   [131]pone.0331277.s007.docx^ (12.9KB, docx)

Abbreviations

   DLSMC
          Double-layer steamed milk custard

   VOCs
          Volatile organic compounds

   HS-GC-IMS
          Headspace-gas chromatography-ion mobility spectrometry

   PCA
          Principal component analysis

   OPLS-DA
          Partial least squares-discriminant analysis

   VIP
          Variable importance in projection

Data Availability

   Metabolomic raw data were uploaded ( ID: MTBLS12717) (MetaboLights:
   [132]https://www.ebi.ac.uk/metabolights/). Please refer to
   [133]https://www.ebi.ac.uk/metabolights/MTBLS12717.

Funding Statement

   This research was financially supported by Ji Hua Laboratory
   (X220991UZ230) and National Natural Science Foundation of China (No.
   32472869). The funders had no role in study design, data collection and
   analysis, decision to publish, or preparation of the manuscript.
   Professor Qingyou Liu is supported by the Ji Hua Laboratory
   (X220991UZ230) and the National Natural Science Foundation of China
   (No. 32472869).

References