Species identification and strain discrimination of fermentation yeasts Saccharomyces cerevisiae and Saccharomyces uvarum using Raman spectroscopy and convolutional neural networks

ABSTRACT

Reliable typing of yeast strains is of great importance to the alcoholic beverage industry to ensure a reliable fermentation process and high-quality products. Saccharomyces cerevisiae is the most used yeast species in wine, sake, and ale beer fermentation, whereas Saccharomyces uvarum is more commonly used for cider fermentation and, due to its cryotolerance, white wine production. We propose a promising method for species identification and strain discrimination of S. cerevisiae and S. uvarum using Raman spectroscopy in combination with convolutional neural networks (CNNs). Raman spectra collected from various S. cerevisiae and S. uvarum strains were accurately classified at the species level using random forest. Cultivation time and temperature did not significantly affect the spectral reproducibility and discrimination capability. An overall accuracy of 91.9% was achieved to discriminate 27 yeast isolates at the strain level using a CNN model. Raman-CNN further identified eight yeast isolates spiked in grape juice with an accuracy of 98.1%. Raman spectral signatures derived from diverse protein and lipid compositions may contribute to this discrimination. The proposed approach also precisely predicted the concentration of a specific yeast strain within a yeast mixture with an R² of 0.9913 and an average error of 4.09%. The entire analysis can be completed within 1 hour following cultivation and only requires simple sample preparation and low consumable cost. Taken together, Raman spectroscopy coupled with CNN is a robust, accurate, and reliable approach for typing of fermentation yeast strains.

IMPORTANCE

The use of S. cerevisiae and S. uvarum yeast starter cultures is a common practice in the alcoholic beverage fermentation industry. As yeast strains from different or the same species have variable fermentation properties, rapid and reliable typing of yeast strains plays an important role in the final quality of the product. In this study, Raman spectroscopy combined with CNN achieved accurate identification of S. cerevisiae and S. uvarum isolates at both the species and strain levels in a rapid, non-destructive, and easy-to-operate manner. This approach can be utilized to test the identity of commercialized dry yeast products and to monitor the diversity of yeast strains during fermentation. It provides great benefits as a high-throughput screening method for agri-food and the alcoholic beverage fermentation industry. This proposed method has the potential to be a powerful tool to discriminate S. cerevisiae and S. uvarum strains in taxonomic, ecological studies and fermentation applications.

INTRODUCTION

Yeast plays an important role in the fermentation of foods and alcoholic beverages, such as bakery products, beer, and wine. Yeast fermentation can also be used to produce various valuable secondary metabolites such as vitamins and antibiotics (1). Saccharomyces cerevisiae is the most prominent yeast starter in commercial wine fermentation as it exhibits high ethanol tolerance and excellent fermentative ability (2). S. uvarum is another natural and non-hybrid species that commonly involved in industrial fermentations (3). Both S. cerevisiae and S. uvarum are suitable for different fermentation processes because they demonstrate distinct fermentation profiles in grape must. For example, S. uvarum has a lower production of acetic acid and ethanol but more glycerol and succinic acid compared to that by S. cerevisiae (4–7). S. uvarum also generates volatile fermentative compounds such as phenylethanol and 2-phenylethyl acetate (7, 8). Due to its cryotolerance, S. uvarum is also useful for white wine, cider, and apple chica fermentations that are conducted at low temperatures (9–11). Furthermore, Saccharomyces strains within the same species present diverse oenological properties (7, 12). The ability to produce a quality beverage or food varies significantly from strain to strain for both S. cerevisiae and S. uvarum. Most modern winemakers select specific Saccharomyces strains for inoculation in grape must to produce an exclusive product with the desired characteristics. Thus, accurate identification of yeast strains is required to monitor the authenticity of final products in dry yeast production and to ensure a reliable and predictable fermentation process.

Various approaches have been attempted to discriminate yeast strains at interspecific and intraspecific levels. Traditional phenotypic methods classified yeast strains based upon the characterization of their morphological, physiological, and biochemical properties (13). These assays are time consuming and laborious, and the cultivation conditions may significantly affect the outcomes. Moreover, conventional methods are insufficient to differentiate closely related strains of the same species due to their physiological similarities (14). A variety of nucleic acid-based methods have been developed for typing Saccharomyces yeast strains, including restriction fragment length polymorphism (RFLP) analysis of chromosomal and mitochondrial DNA (15–17), ribosomal DNA sequencing (18, 19), randomly amplified polymorphic DNA (20), amplified fragment length polymorphism (21, 22), and microsatellite genotyping (23, 24). Although these methods are relatively accurate, most of them require sophisticated equipment, costly reagents, and highly trained personnel (25). Matrix-assisted laser desorption ionization-time of flight mass spectrometry (MALDI-TOF) is an alternative tool to differentiate yeast strains based on the analysis of proteomic profiling (26). However, the sample preparation is complicated and requires expensive chemicals. Therefore, a rapid, reliable, and easy-to-conduct approach to identify and discriminate yeast strains is of great demand.

Raman spectroscopy has emerged as a powerful fingerprinting method for rapid identification of microorganisms, including bacteria, molds, and yeast species (27). This technique records the energy changes of vibrational chemical bonds in molecules, thereby providing comprehensive information about the biochemical compositions (e.g., nucleic acids, proteins and lipids, etc.) of the whole cells (28). Raman spectroscopy can distinguish different cellular phenotypes based on subtle variations of their Raman spectral features in a nondestructive and label-free manner. Good reproducibility and high discriminatory capability of Raman spectroscopy have been validated in previous studies (28, 29). For example, Yan and coauthors achieved an accuracy of 88.4% for the discrimination of Listeria monocytogenes at the strain level using Raman spectroscopy coupled with chemometrics (30). A high accuracy of 95.64% ± 5.46% was obtained for the identification of 14 microbial species using single-cell Raman data (31). In addition, minimal sample preparation is needed, and a single Raman spectrum can be acquired within seconds, making it a suitable strategy for typing yeast strains in food industry.

Advanced multivariate analyses are required to decipher the minor variations among Raman spectral features from different microbiological samples. Principal component analysis (PCA) and hierarchical cluster analysis are the common unsupervised pattern recognition methods that can categorize the data set based on their spectral differences without prior knowledge about the samples (32). Supervised chemometrics have also been widely used for classification and discrimination in Raman spectroscopic analysis, such as partial least squares discriminant analysis (PLSDA), soft independent modeling of class analogy (SIMCA), random forest (RF), and partial least squares regression (PLSR). Machine learning methods are usually employed to perform complicated tasks for a large-scale spectroscopic data set. Convolutional neural network (CNN) is a state-of-the-art deep learning algorithm that has been extensively used in image reconstruction, video recognition, and natural language processing (33). The widespread success of CNN is attributed to its efficient computation, high accuracy, and remarkable tolerance to overfitting. CNN has recently been applied to one-dimensional Raman spectral data for mineral classification (34) and bacterial discrimination (31, 35, 36). It not only simplifies spectral analysis by combing automatic feature extraction with classification but also achieves superior performance compared to other conventional machine learning techniques, such as artificial neural networks and support vector machine (SVM) (37). Thus, it has the potential to be coupled with Raman spectroscopy to achieve strain-level identification of yeast.

Here, we reported a novel and reliable approach to discriminate yeast isolates of S. cerevisiae and S. uvarum at both the species and strain levels using Raman spectroscopy combined with chemometrics (i.e., PCA and RF) and machine learning (i.e., CNN). The robustness, quantitative capability, and practicality for real wine samples of the current approach were also evaluated. Raman spectroscopic-based CNN demonstrates great potential to be utilized as an alternative strain typing method for wine yeast.

RESULTS AND DISCUSSION

Raman spectra of different yeast strains

Establishment of a reference database is the first step for the identification of microorganisms using Raman spectroscopy. To build up a comprehensive database, a total of 289 Raman spectra were collected from 12 S. cerevisiae strains and 15 S. uvarum strains isolated from wine or wild sources and diverse geographical origins (Table 1). Figure 1 displays the mean Raman spectra of each strain and the major Raman peaks shown in the spectra. The mean spectra were calculated based upon all the Raman spectra of each strain after baseline correction, smoothing, and normalization. Raman spectral features at fingerprinting region (400–1,800 cm⁻¹) represent different vibrational modes of functional groups from biochemical compositions of the whole yeast cells, including polysaccharides, proteins, phospholipids, and nucleic acids (38). An overview of the observed Raman peaks and their tentative assignments is provided in Table 2.

TABLE 1.

Summary of Saccharomyces cerevisiae and Saccharomyces uvarum strains used in this study

Yeast strain	Source	Type
S. cerevisiae EC1118	Lallemand	Commercial wine yeast
S. cerevisiae M2	Lallemand	Commercial wine yeast
S. cerevisiae T73	Lallemand	Commercial wine yeast
S. cerevisiae VL3	Laffort	Commercial wine yeast
S. cerevisiae F15	Laffort	Commercial wine yeast
S. cerevisiae Pasteur Red	Red Star	Commercial wine yeast
S. cerevisiae SBV008	South Okanagan Pinot Noir spontaneous fermentation, Canada	Wine yeast
S. cerevisiae DV10	Lallemand	Commercial wine yeast
S. cerevisiae V1116	Lallemand	Commercial wine yeast
S. cerevisiae BGY	Lallemand	Commercial wine yeast
S. cerevisiae ICV-D254	Lallemand	Commercial wine yeast
S. cerevisiae S6U	Lallemand	Commercial wine yeast
S. uvarum 7D4	New Zealand winery isolate	Wine yeast
S. uvarum A4	New Zealand winery isolate	Wine yeast
S. uvarum P01A05	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum P07F02	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum P36C10	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum P45D09	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum CBS8711	Centraalbureau voor Schimmelcultures, France	Wine yeast
S. uvarum CBS8696	Centraalbureau voor Schimmelcultures, USA,	Wild yeast, Drosophila
S. uvarum CBS7001	Centraalbureau voor Schimmelcultures, Spain	Wild yeast, insect
S. uvarum A9	New Zealand winery isolate	Wine yeast
S. uvarum 7H2	New Zealand winery isolate	Wine yeast
S. uvarum 8B11	New Zealand winery isolate	Wine yeast
S. uvarum P01E01	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum P01H01	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast
S. uvarum P01G12	Central Okanagan Pinot Gris spontaneous fermentation, Canada	Wine yeast

Fig 1.

Mean Raman spectra of 12 Saccharomyces cerevisiae (a) and 15 Saccharomyces uvarum (b) strains. The mean spectrum was calculated based upon all the preprocessed Raman spectra of each strain. Major Raman peaks observed in both species were labeled in black, and peaks only shown in S. cerevisiae or S. uvarum were labeled in blue or red, respectively. Shaded regions indicate the visible spectral variations between different strains. The spectra are vertically shifted and displayed for clarity. S. cerevisiae (1–12): 1, BGY; 2, D254; 3, DV10; 4, EC1118; 5, F15; 6, M2; 7, Pasteur Red; 8, S6U; 9, SBV008; 10, T73; 11, V1116; 12, VL3. S. uvarum (13–27): 13, 7B11; 14, 7D4; 15, 7H2; 16, A4; 17, A9; 18, CBS7001; 19, CBS8696; 20, CBS8711; 21, P01A05; 22, P01E01; 23, P01G12; 24, P01H01; 25, P07F02; 26, P36C10; 27, P45D09.

TABLE 2.

Tentative assignments of Raman peaks identified in the Raman spectra of S. cerevisiae and S. uvarum

Wave number (cm−1)	Tentative assignment	Compounds	Reference
429	Cholesterol and cholesterol ester	Lipid	(39)
535	S-S stretching of cysteine	Protein	(40)
620	C-C twisting of aromatic ring	Protein	(41)
640	C-S stretching and C-C twisting of tyrosine	Protein	(41)
720	Phospholipids	Nucleic acid	(42)
759	Tryptophan	Protein	(43)
776	Phosphatidylinositol	Lipid	(39)
827	Proline, hydroxyproline, tyrosine, PO2− stretching	Protein/nucleic acid	(44)
850	Single-bond stretching of valine and polysaccharides	Protein/carbohydrate	(45)
898	Monosaccharides (β-glucose), C-O-C skeletal mode of disaccharide (maltose) or adenine	Nucleic acid/carbohydrate	(40, 46)
904	C-C skeletal stretching	Protein	(47)
1002	Phenylalanine	Protein	(46)
1030	C-H in-plane bending mode of phenylalanine	Protein	(48)
1084	Phosphodiester groups	Nucleic acid	(49)
1264	Triglycerides	Lipids	(50)
1305–1344	Guanine, phospholipids, CH2 deformation, CH2 twisting, Amide III (α-helix)	Protein/lipid/nucleic acid	(51–54)
1448	CH2 deformation	Protein/lipid	(55)
1576	Ring stretching mode of adenine and guanine	Nucleic acid	(56)
1604	C = C in-plane bending mode of phenylalanine and tyrosine	Protein	(43, 46)
1654	Amide I (collagen)	Protein	(55)
1670	Amide I (anti-parallel β-sheet)	Protein	(40)
1747	C = O stretching	Lipid	(52)

Most of the Raman peaks are shown in both S. cerevisiae and S. uvarum and are labeled in black. The most prominent peak located at 1,448 cm⁻¹ is assigned to CH₂ deformation modes of proteins and lipids (55). The broad peak in the range of 1,305–1,344 cm⁻¹ arises from the overlapping signals from phospholipids and guanine in nucleic acids and proteins such as amide III (51–54). Raman signature for phenylalanine is the sharp peak at 1,002 cm⁻¹ (46). Raman signals at 1,030 cm⁻¹ and 1,604 cm⁻¹ are a result of the C-H bending mode and C = C bending mode of phenylalanine (43, 48). Additional protein peaks can be found at 535, 620, 640, 759, and 850 cm⁻¹ (38). The minor signal at 1,747 cm⁻¹ is associated with the stretching vibration of the C = O double bond of lipids such as polyhydroxybutyrate (52). Other peaks attributed to lipids are also observed, including the ones at 429 cm⁻¹ (cholesterol and cholesterol ester), 776 cm⁻¹ (phosphatidylinositol), and 1,264 cm⁻¹ (triglycerides) (39, 50). In addition, Raman peaks related to phospholipids in nucleic acids are located at 720, 1,084, and 1,576 cm⁻¹ (38). Not all Raman signals originate exclusively from one single compound but are rather due to several different types of biomolecules.

Upon further inspection, there were visible spectral variations between Raman spectra of S. uvarum and S. cerevisiae. Two peaks exclusively appeared in S. cerevisiae strains, including C-C skeletal stretching (904 cm⁻¹) and C = O stretching mode of amide I (1,654 cm⁻¹) (47, 55). In comparison, unique Raman peaks at 827 cm⁻¹ (amino acids and PO²⁻ stretch of nucleic acids), 898 cm⁻¹ (monosaccharides, disaccharides, or adenine), 1,576 cm⁻¹ (nucleic acids), and 1,669 cm⁻¹ (amide I in anti-parallel β-sheet) were solely observed in S. uvarum strains (40, 44, 56). The presence of unique Raman peaks provides evidence for the existence of species-specific compounds. S. cerevisiae and S. uvarum also varied in the relative intensities of several peaks. For instance, S. cerevisiae strains exhibit enhanced signal intensity at 429 and 1,604 cm⁻¹, while a higher intensity at 620, 640, and 1,002 cm⁻¹ are shown in S. uvarum strains. The relative peak intensities indicate the relative abundances of the corresponding chemical components in each species. Yeast strains within the same species have similar Raman spectral features due to sharing equivalent cellular components. Only subtle variations in the relative intensities of some regions can be visualized, such as the lipid peak at 1,264 cm⁻¹ and protein peak at ~1,604 cm⁻¹. It is challenging to differentiate Raman spectra from diverse strains by the naked eye and to predict the identity of unknown strains. Therefore, multivariate analyses are required to further decipher the imperceptible spectral variances and identify yeast strains at different taxonomic levels.

Identification of yeast strains at the species level using PCA and RF

We classified 12 S. cerevisiae and 15 S. uvarum isolates at the species level using Raman spectroscopy combined with PCA and RF. PCA is an unsupervised chemometric that can reduce the dimensionality of the spectra into a few uncorrelated variables named principal components (PCs) (57) while carrying most of the spectral variations. A PCA plot was firstly constructed for all yeast strains using the first three PCs to visualize their spectral differences and determine if they can be categorized (Fig. 2a). Two distinctive clusters were formed for the S. cerevisiae strains and S. uvarum strains, indicating that these two yeast species could be clearly separated based upon their Raman spectra. Supervised algorithm RF was further used to identify yeast strains at the species level. RF utilizes ensemble learning that combines multiple decision trees to make predictions or classifications (58). To avoid the potential overfitting of RF, the whole data set was randomly split, with 70% allocated to the training data set and 30% reserved for a hold-out test data set. A high accuracy of 98.9% was achieved to classify S. cerevisiae and S. uvarum on the test data set. Therefore, Raman spectroscopy combined with chemometrics achieved a reliable species-level identification for S. cerevisiae and S. uvarum yeast strains.

Fig 2.

Classification of S. cerevisiae and S. uvarum at the species level. (a) Principal component analysis was constructed using Raman spectra collected from 12 strains of S. cerevisiae and 15 strains of S. uvarum. The clear separation of two clusters indicates a reliable classification of two yeast species. (b) Comparison of the loading plot of the first PC and the average Raman spectrum of each species. The absolute value of coefficient explains the contributions of spectral features to the classification model. Spectral regions with high coefficient values are shaded in gray. An accuracy of 98.9% was achieved for the identification of S. cerevisiae and S. uvarum using RF.

To explore the cellular compositions that may contribute to the identification, loading plot of the first PC (representing 59% of the total variations) was compared to the mean spectra of two yeast species (Fig. 2b). Raman spectral features with high loading values carry most of the spectral variations responsible for the discrimination. The Raman peak at 776 cm⁻¹ had a high positive PC1 coefficient value, and it was attributed to phosphatidylinositol, which is the parent lipid for a family of phospholipids in cell membranes that is important for cellular signaling and membrane trafficking (59). These data agreed with a previous study suggesting that S. cerevisiae had a lower content of phosphatidylinositol compared to S. uvarum (60). Raman signals from C-C stretching mode of lipids (~1130 cm⁻¹) (39) also had high loading values. Lipid compounds are important constituents of yeast cell membranes and are involved in a variety of cellular processes (61), particularly the maintenance of membrane fluidity at low temperature (62). S. uvarum has a higher concentration of unsaturated fatty acids compared to S. cerevisiae due to its cryotolerance, which is the major difference between these two species (63). Different yeast species contain diverse membrane lipid compositions with different chain lengths and saturation states, and characteristic Raman peaks in the lipid region enable sufficient discriminatory capacity. Examination of lipid metabolic profiles using Raman spectroscopy was previously used to differentiate three phylogenetically close Malassezia yeast species (64). Fatty acid methyl ester and phospholipid fatty acid analysis on cell wall phospholipids could be conducted to assess for a correlation with Raman spectral features and to further understand the role of lipids in yeast identification. High loading values were also observed in protein-associated peaks at 1,002 cm⁻¹ (phenylalanine), 1,225 cm⁻¹ (amide III), and 1,671 cm⁻¹ (amide I), suggesting that proteomic differences contributed significantly to differentiation of S. cerevisiae and S. uvarum (Fig. 2b). This was consistent with a previous study that unraveled massive variations in protein abundances and metabolic pathways between these two species (65). For example, S. uvarum had more abundant proteins involved in thiamine and sulfur amino acid biosynthesis, while S. cerevisiae strains exhibited higher levels of proteins involved in cell defense, virulence, and lipid metabolism. Another study also reported a different proteomic profiling of S. cerevisiae and S. uvarum strains in translation, glycolysis and amino acid metabolism due to their different degrees of cryotolerance (66).

Influence of cultivation conditions on Raman spectral reproducibility

Spectral reproducibility is a critical factor for the application of Raman spectroscopic-based species classification. Yeast strains may demonstrate different growth behaviors under diverse cultivation conditions, including culture media, incubation time, and temperatures. This may result in the variation of cellular compositions and further affect the reproducibility of Raman spectra. Therefore, we cultivated yeast strains with different time periods and at different temperatures to increase the variability of the data set and explore the robustness of Raman spectroscopic method. S. cerevisiae EC1118 and T73, S. uvarum A4, and CBS70001 were randomly selected as the representative strains of these two species. To test the influence of cultivation time, a total of 280 Raman spectra were collected from four strains incubated for 48 and 72 h at 25°C, respectively. In parallel, four strains were cultivated at 25°C and 30°C for 48 h, respectively, and 274 Raman spectra were collected to evaluate the influence of growth temperature. These spectra were firstly subjected to PCA (Fig. 3) to visualize their similarities and then classified using RF (Table 3) at different levels. As shown in Fig. 3a, all the Raman spectra from same species (S. cerevisiae or S. uvarum) with different growth times are grouped into one cluster and hardly distinguishable, indicating a satisfactory spectral reproducibility. Raman spectra from S. cerevisiae and S. uvarum strains cultivated at different temperatures were also segregated into two individual clusters, and the spectral reproducibility was confirmed by the overlap of Raman spectra from the same strain (Fig. 3b). For the identification at the species level using RF (Table 3), excellent accuracies of 99.3% and 99.6% were obtained for the strains cultivated with different time periods and at different temperatures, respectively. Identification accuracies were also both over 95% (i.e., 95.9% and 95.4%) for the strain-level discrimination, demonstrating a good classification of yeast strains without being affected by cultivation conditions. Taken together, there was no significant influence of cultivation time and temperature on Raman spectroscopic-based classification, demonstrating that Raman spectroscopy is a robust approach to identify yeast species.

Fig 3.

Influence of cultivation time (a) and temperature (b) on Raman spectral reproducibility. Two S. cerevisiae and two S. uvarum strains were randomly selected as the representative strains of each species. (a) Raman spectra were collected from representative strains cultivated for different time periods of 48 and 72 h at 25°C and displayed in a three-dimensional PCA plot. (b) PCA was performed with Raman spectra recorded from yeast strains cultivated at different temperature of 25°C and 30°C for 48 h. Individual clusters were formed for each species cultivated at various conditions, and no intraspecies separation was observed. The identification of these four yeast isolates at species and strain levels was further verified using RF (Table 3).

TABLE 3.

Identification accuracy of two S. cerevisiae and two S. uvarum strains cultivated at different conditions using random forest

Taxonomic level	Time (48 or 60 h)	Temperature (25°C or 30°C)
Species level	99.3%a	99.6%
Strain level	95.9%	95.4%

^a Overall identification accuracy is the percentage of correctly identified spectra during the validation.

Similar results were reported in our previous study that neither cultivation time nor incubation temperature influenced the discriminatory power of Raman spectroscopy for bacterial identification (35, 67). In another study, Van de Vossenberg and coauthors also demonstrated that water type and temperature did not significantly affect the differentiation between Legionella and other waterborne bacteria by using Raman spectroscopy (68). In comparison, surface-enhanced Raman scattering is another spectroscopic technique that can enhance Raman scattering signals by several orders of magnitude, but its application is hindered by relatively low spectral reproducibility, especially on biological samples (69). Although experimental studies are usually conducted under consistent and optimal cultivation conditions, no standardized process can be ensured in various real-world fermentation scenarios. We verified that Raman spectroscopy has the potential to tolerate cultivation variations and generate reliable results for the identification of yeast strains grown under diverse conditions.

Strain-level discrimination of yeast isolates using CNN

Strain-level identification of yeast isolates is crucial for dry yeast production and selection of appropriate starter for industrial wine fermentation. Raman spectral features of yeast strains within the same species exhibited high similarity in terms of signal locations and relative intensities as seen in Fig. 1. Raman spectra from 27 different strains were not clearly separated from each other in PCA (Fig. S1). Therefore, a more rigorous classification algorithm is necessary to interpret the subtle intraspecies spectral variations. We utilized a state-of-art deep learning technique named CNN to further discriminate yeast isolates at the strain level, and the simplified architecture is shown in Fig. 4. This CNN is composed of four convolutional layers for feature extraction and two fully connected layers to perform classification (details in Materials and Methods). Unlike previous studies for image analysis (70, 71), pooling layers were not included in this study to preserve the exact locations of Raman features and avoid data loss. For each input spectrum, a probability distribution across 27 yeast strains was generated in the output layer, and the one with maximum value was considered as the predicted class. The stability of CNN was evaluated using a 10-fold cross-validation.

Fig 4.

Schematic architecture of the proposed CNN for the discrimination of yeast strains. The CNN contains an initial input layer, four convolutional layers to extract spectral features, two fully connected layers for classification, followed by an output layer. Each convolutional layer consists of different number of filters with a kernel size of 2 × 1, and ReLU is used as the activation function. Two fully connected layers include 256 and 50 neurons, respectively. For each input spectrum, the probability distribution of 27 strains was generated in the output layer (Softmax activation function), and the one with the highest value was considered as the predicted strain. Model parameters were determined via grid search. Conv, convolutional layer; FC, fully connected layer; ReLU, rectified linear unit.

The classification results for all the yeast strains are summarized in the confusion matrix in Fig. 5. The values on the diagonal indicate the correct recognition rate for each strain, while the incorrect predictions are exhibited in the off-diagonal area. An overall accuracy of 91.9% was achieved on the 27-class task, i.e., 91.9% of the total spectra were assigned to the correct yeast strain during cross-validation. More details about spectral number and classification specificity for each strain are listed in Table S1. The identification accuracies for individual yeast strains ranged from 80.7% for S. uvarum 7H2 to 99.1% for S. cerevisiae V116. Three S. uvarum strains and six S. cerevisiae strains obtained an identification accuracy >95%, indicating a good separation of these strains from others. Eleven strains demonstrated sufficient accuracy ranging from 90%–95%. Only seven strains showed a relatively lower identification accuracy compared to other strains, where less than 90% of the spectra were correctly classified. The lowest accuracy of 80.7% was obtained for the classification of S. uvarum 7H2, and most of the incorrectly classified spectra were assigned to S. uvarum 8B11, P01E01, and P01G01. The majority of misidentification occurred among the strains belonging to the same species aside from the interspecific misidentification of S. uvarum 7D4 (5.5%), S. cerevisiae M2 (2%), Pasteur Red (1.3%), and SBV008 (1%). This was in accordance with the taxonomic relation that interspecific compositional variations are generally more apparent than that of intraspecies. The average identification accuracy at the strain level was higher for S. cerevisiae (93.8%) than S. uvarum (90.3%), indicating a more evident homogeneity among the selected strains of S. uvarum (3, 72). To further verify the prediction ability of the proposed CNN, a hold-out test set (i.e., not included in the training process) of eight yeast strains (30 spectra per strain) were constructed and subjected to the pretrained CNN. An overall accuracy of 89.2% was achieved with the accuracy ranging from 76.7% to 96.7% for each strain (Table S2). The results indicated a reliable prediction ability of the CNN for the hold-out test set (i.e., not included in the training process) and a low possibility of overfitting. Based on the results of the cross-validation and the independent test, Raman spectroscopy combined with CNN represents an reliable approach for the identification of yeast isolates to the strain level.

Fig 5.

Classification results of 27 yeast isolates at the strain level using the proposed CNN. Values exhibited on the diagonal represent the identification accuracy (i.e., the percentage of correctly identified spectra during cross-validation) for each strain. Misidentification rates are displayed on the off-diagonal region. The average strain-level accuracy is 91.9%. Values of 0 are not shown in the confusion matrix.

Employing CNN to interpret Raman spectroscopic data improves the classification sensitivity. CNN is one of the most popular deep learning algorithms that have been extensively applied to image classification and voice recognition (73). Only a few studies have been conducted in adapting CNN to one-dimensional Raman spectral data for microbial identification. For example, Ho and others classified 30 common bacterial and yeast pathogens from different genera or species with an overall accuracy of 82% based upon Raman spectra and CNN (36). In another study, the recognition rate for 14 strains containing bacteria and fungi via CNN was 95.64% (31). The average accuracy of 91.9% achieved in the current study further validated the power of Raman spectroscopic-CNN to discriminate yeast at the strain level. A variety of chemometrics have been used for the classification of Raman spectral data, such as SIMCA and PLSDA. However, these methods usually require an explicit model to capture the features, and valuable spectral information might be lost during the reduction of dimension or selection of appropriate variables (74). CNN combines feature extraction and classification in one architecture that automatically learns the underlying relationship from training data sets with little human supervision and achieves a high accuracy to predict unknown samples. Previous studies reported that CNN outperformed PLSR and SVM for Raman spectroscopic-based species identification (35, 36). Furthermore, the present CNN takes into consideration the whole spectral fingerprint, and no prior knowledge of yeast strain is mandatory. It can be generalized to identify other microorganisms with minor modifications.

The widespread diversity among both S. cerevisiae and S. uvarum strains has been reported in terms of physiological and morphological properties, such as growth temperature, response to drugs, and fermentative performance (3, 75, 76). A genome-wide polymorphism survey of 63 S. cerevisiae strains indicated significant genomic variations of strains collected from different ecological niches and locations (75). The genetic diversity of S. uvarum has also been investigated at the strain level. Almeida and coauthors studied a set of S. uvarum strains from five continents using whole-genome sequencing and revealed a global genetic divergency among some strains of this species (3). McCarthy and others also reported a highly diverse S. uvarum population isolated from a Canadian winery (10). Different cellular components could arise based on genetic differentiation and result in the characteristic Raman spectral profiles. Six S. cerevisiae wine yeasts were identified using Raman spectroscopy with an average accuracy of 91.5% (77). Based on the visualization of mean spectra of each strain, variations in spectral regions at 1,200–1,300 cm⁻¹ (proteins and lipids) and 1,600–1,800 cm⁻¹ (lipids) may contribute to the strain-level differentiation (Fig. 1). As aforementioned, lipids are vital metabolic energy source and building blocks of yeast membranes that are responsible for growing at different temperatures. Yeast strains demonstrate complex lipid metabolism, and fatty acid profiles were used to distinguish S. cerevisiae and S. uvarum strains (78). Proteins are also important molecular components that contribute to the yeast phenotypic complexity as they play a central role in signaling, transportation and metabolism (79). Both S. cerevisiae and S. uvarum isolates were differentiated at the strain-level based upon their proteomic variations (65). Therefore, Raman spectroscopic-based metabolomic profiling can assist in understanding the molecular basis for various physiological traits of yeast strains.

Quantification of a targeted yeast strain from a yeast mixture

Multiple yeast strains might co-exist in the environment and fermentation process, especially for spontaneous fermentations (80), making the determination of a specific strain from the yeast flora challenging. Here, we evaluated the sensitivity and selectivity of Raman spectroscopy to identify a targeted strain from yeast mixture. S. cerevisiae EC1118 was selected as the target as it is one of the most popular commercial strains used in the wine industry. S. cerevisiae EC1118 was mixed into a yeast cocktail containing three yeast strains with different concentrations ranging from 10⁵ to 10⁸ CFU/mL (details in Materials and Methods). A CNN including two convolutional layers, one pooling layer, and six fully connected layers was established to predict the ratio of EC1118 within the yeast mixture. CNN was evaluated using 10-fold cross-validation, and the red circles in Fig. 6 represent the predicted concentration generated by CNN during validation. The average error of the predicted concentrations compared to the actual values was 4.09%, suggesting a precise quantitative capability of CNN. Other evaluation parameters were also calculated and listed in Fig. 6, including coefficient of determination (R²), root mean square error of prediction (RMSE), and residual prediction deviation (RPD). Since R² >95%, RPD >3, and RMSE <1 are generally regarded as the criteria for a good regression model (67, 81), our Raman spectroscopic-CNN achieved excellent performance and indicated a reliable correlation between the concentration of the targeted strain and its corresponding Raman spectral features. It has the potential to become an effective tool to track the fate of a specific yeast strain from complex microbial mixtures.

Fig 6.

Prediction for the concentration of a target yeast strain within a yeast mixture using CNN. S. cerevisiae EC1118 was suspended in a yeast cocktail of S. uvarum 7D4, P45D09, and A9 (10⁷ CFU/mL in total) at different concentrations of 10⁵, 5 × 10⁵, 10⁶, 5 × 10⁶, 10⁷, 2 × 10⁷, and 10⁸ CFU/mL, respectively. The performance of CNN was evaluated using 10-fold cross-validation, and the red circle represents the predicted concentration generated by CNN.

Identification of yeast strains from grape juice

It is of practical interest to detect a yeast strain from real wine fermentation samples with complex nutritional components and microbial populations. To further evaluate the practicability of the current method, identifying yeast in grape juice was conducted. Pinot Gris grape juice was obtained from a local winery. Eight yeast strains (four S. cerevisiae and four S. uvarum) were randomly selected and inoculated into the grape juice. Raman spectra were collected following the similar procedures to those used for identifying yeasts cultivated from pure medium. CNN was applied to classify these eight yeast isolates at the strain level, and the results are summarized in Fig. 7. During the cross-validation, the overall strain-level identification accuracy was 98.1% with the range from 94.1% to 100% for individual strains. All the incorrect identifications were among the strains belonging to the same species and no interspecific misidentification occurred. Thus, our Raman-CNN was able to discriminate different yeasts at the strain level from grape juice with a high accuracy. The accuracy was higher compared to that of identifying yeasts from pure culture (i.e., 91.9%), possibly due to a higher number of strains that were included in the pure culture data set (i.e., 27 strains). More yeast strains should be included to construct a more comprehensive database in the future. Raman spectroscopy was validated to be effective for microbial identification from complicated sample matrices, including milk (82), apple juice (83), and urine samples (84). Therefore, it has the potential to monitor the fermentation kinetics of Saccharomyces inoculum during wine fermentation.

Fig 7.

Classification results of eight yeast isolates spiked in grape juice at the strain level using CNN. Values exhibited on the diagonal represent the identification accuracy (i.e., the percentage of correctly identified spectra during cross-validation) for each strain. Misidentification rates are displayed on the off-diagonal region. The average strain-level accuracy was 98.1%. Values of 0 are not shown in the confusion matrix.

Advantages of Raman spectroscopy combined with CNN

We compared the performance of the current Raman-CNN approach to other identification methods used for yeast strains (Table 4). Traditional microscopic examination and biochemical assays are usually of limited value to discriminate different strains within the same species due to their high phenotypic similarities (14, 85). Our Raman spectroscopic-CNN successfully discriminated 27 yeast strains with a high accuracy of 91.9%, indicating an excellent identification resolution. DNA-based techniques (i.e., PCR and sequencing) employed to discriminate yeast strains can obtain accurate results but require substantially longer time, complicated sample preparation, and expensive reagents. For example, RFLP takes approximately 30 h (86), and sequencing-based methods take 3–4 days (25). In comparison, Raman spectra can be collected within seconds, and the entire identification process described in this study, including simple sample preparation, Raman data collection, and spectral analysis, could be completed in ~50 min after harvesting yeast colonies. With the development of Raman optical tweezers, future work should aim to analyze a single yeast cell directly from wine samples without cultivation to further improve sensitivity and reduce the turnaround time. In contrast to DNA-based molecular assays, Raman spectroscopic analysis is also cost-effective because no chemical reagents and only minimal consumables (e.g., pipette tips and centrifuge tubes) are required. The present methodology has the potential to be a real-time screening tool for S. cerevisiae and S. uvarum strains in food industry as it is rapid and non-destructive and permits in situ analysis by using a portable device.

TABLE 4.

Comparisons of the proposed Raman-CNN with other methods used for yeast identification

Criteria	Raman-CNN	PCR-based methods	DNA sequencing	Biochemical tests	Microscopic examination	MALDI-TOF mass spectrometry
Analysis time	Seconds to minutes	Hours to a day	Several days	Several days	Minutes to hours	Minutes
Sample preparation	Minimal	DNA extraction and purification	DNA extraction and purification	Culturing and sample preparation	Cell staining	Minimal
Non-destructive	Yes	No	No	No	Yes	No
Sensitivity	High	High	High	Moderate to low	Moderate	High
Identification resolution	Strain level	Species to strain level	Strain level	Species to strain level	Species level	Species to strain level
Sample volume	Minimal	Small amount of DNA	Small amount of DNA	Large	Minimal	Small amount
Real-time monitoring	Possible	No	No	No	No	No
Cost	No reagent and minimal consumables, primarily driven by the instrument	Moderate, including the equipment, reagents, and consumables	High, driven by the equipment, reagents, and consumables	Moderate, including culture media, reagents, and consumables	Low, emphasizing on microscope and staining reagents	Minimal consumables, primarily driven by the instrument and reagents
Portability	Possible	Possible	Possible	Possible	Possible	No
References	(13, 14, 16, 18, 20, 85, 87–90)

Conclusion

We developed a rapid and reliable approach of using Raman spectroscopy coupled with machine learning to achieve accurate discrimination of S. cerevisiae and S. uvarum isolates at both the species and strain levels. Strains from S. cerevisiae and S. uvarum displayed slightly different Raman spectral profiles that reveal the diversity of cellular biochemical signatures. Identification of these two species were achieved using RF and mainly attributed to the variations in their protein and lipid compositions. The spectral reproducibility and model robustness were validated by the distinct classification of yeast strains cultivated under various conditions. By utilizing a state-of-art CNN model, 27 yeast isolates were discriminated at the strain level with a high accuracy of 91.9%. Raman-based CNN also achieved an accuracy of 98.1% for the identification of eight yeast strains spiked in grape juice, indicating the potential to be applied for wine fermentation. This approach also demonstrated strong capability (i.e., R² = 0.9913) to predict the concentration of a specific yeast strain within a yeast mixture based upon the corresponding Raman spectral features. The entire process could be completed within an hour after harvesting yeast cells with no requirement of expensive reagent or highly trained personnel. With the miniaturization of Raman spectrometer and fast cloud data service (e.g., Internet of Things), this approach has the potential to allow real-time monitoring of yeast flora in industrial fermentation process and thus assure more desirable wine products. The knowledge from this study provides new insights into strain typing of biotechnologically important microbes.

MATERIALS AND METHODS

Yeast strains and sample preparation

A total of 12 strains of S. cerevisiae and 15 strains of S. uvarum (Table 1) were included in this study. Commercial wine starter yeast strains were purchased from different suppliers, and noncommercial yeast isolates were obtained from culture collections or wineries from Canada (91), USA, France, Spain, and New Zealand (72). The detailed information is summarized and shown in Table 1. Yeast strains were identified and characterized using the conventional PCR and sequencing as previously described (72, 91). All the strains were stored in 15% (vol/vol) glycerol stocks at –80°C for long-term storage. Yeast strains were aerobically cultured on Synthetic Complete agar at 25°C for 48 h. Half-loop (10-µL inoculation loop, VWR) of cell mass harvested from agar plates was suspended in 1 mL of sterile double-deionized water (ddH₂O), followed by centrifugation at 5,000 × g for 2 min. The supernatant was removed, and the cell pellet was washed twice using ddH₂O and finally resuspended in 100 µL of ddH₂O. Two microliters of each yeast suspension were transferred onto a gold-coated microscope slide (Sigma-Aldrich, St. Louis, MO, USA) and dried at 22°C for 20 min before spectral collection.

Raman spectral collection

Raman spectra were recorded using a Renishaw Raman system (Renishaw, Gloucestershire, UK) equipped with a Leica microscope and 785-nm near-infrared diode laser. The gold-coated microscope slide was mounted onto an x/y/z motorized microscope stage. A laser power of ~25 mW was focused on the sample through a 50× Nikon objective (NA = 0.75). The scattered Raman signals were delivered to a spectrometer equipped with a diffraction grating of 1,200 lines/mm and detected using a charge-coupled device (578 × 385 pixels) cooled at −60°C. Raman spectra were collected over a wavenumber range of 400–1,800 cm⁻¹ with a resolution of 1 cm⁻¹. The integration time for each spectrum was 10 s. The operation of Raman spectroscopic system was controlled using WiRE 3.4 software (Renishaw, Gloucestershire, UK). Three biological replicates of each yeast strain were analyzed.

Effect of cultivation conditions on spectral reproducibility

To evaluate the influence of cultivation conditions on spectral reproducibility, we investigated the discrimination capability of Raman spectroscopy using yeast strains cultivated at different temperatures and time periods. Four yeast strains, S. cerevisiae EC1118, S. cerevisiae T73, S. uvarum CBS70001, and S. uvarum A4 were randomly selected as the representative strains for two species. Each strain was cultivated at 25°C for 48 h and 72 h, respectively, and then assessed by Raman spectroscopy. In parallel, Raman spectra were also collected from these strains grown at two different temperatures (i.e., 25°C and 30°C) for 48 h to investigate the effect of cultivation temperature on spectral reproducibility.

Identification of yeast strains from grape juice

We applied Raman spectroscopy to identify yeast strains in grape juice so as to investigate its feasibility to be used in wine fermentation. Pinot Gris grape juice was obtained from Okanagan Crush Pad winery and kept frozen at −20°C and thawed at 4°C before use. The properties of grape juice are available in Table S3. The Pinot Gris grape juice was inoculated with eight yeast strains, namely, S. cerevisiae EC1118, F15, M2, and DV10 and S. uvarum 7D4, A4, CBS8711, and P36C10, which were randomly selected as the representative strains. For each strain, colonies were collected from agar plates and inoculated into grape juice with a final concentration of ~10⁷ CFU/mL. Grape juice was then incubated at 25°C with shaking at 180 rpm for 2 h. To collect the yeast cells, 1 mL of grape juice was centrifuged at 5,000 × g for 5 min, followed by removing the supernatant and washing with ddH₂O three times. Raman spectral collection was then conducted as aforementioned.

Quantification of a specific strain from a yeast mixture

As various yeast strains may co-exist in natural conditions, we investigated the ability of Raman spectroscopy to selectively identify a particular yeast strain within a mixture. Commercial strain S. cerevisiae EC1118 was selected as the targeted strain because it is commonly used in wine fermentation. Cell suspensions of three randomly selected S. uvarum strains 7D4, P45D09, and A9 were mixed in equal biomass to form a yeast cocktail with a concentration of 10⁷ CFU/mL. EC1118 was added to this mixture at different concentrations of 10⁵, 5 × 10⁵, 10⁶, 5 × 10⁶, 10⁷, 2 × 10⁷, and 10⁸ CFU/mL to generate a series of new interspecific mixtures. The absolute cell concentration of each yeast strain was determined using a conventional plating assay on synthetic complete agar plates. Raman spectra were recorded from these mixtures and analyzed to predict the concentration of EC1118 in the mixture based upon the corresponding spectral profiles.

Spectral reprocessing, chemometrics, and machine learning

As the efficiency of Raman scattering is inherently low (~10⁻⁸ probability), subtle spectral variations are easily masked by multiple background noises, such as cosmic noise, intrinsic fluorescence of biological samples, and noises from instrument and environments (92). Therefore, raw Raman spectra were preprocessed before further analysis. First, baseline correction was conducted using Vancouver Raman Algorithm software based upon a five-order polynomial fitting (93). Raman spectra were subsequently smoothed by OMNIC software (Thermo Fisher Scientific, Waltham, MA, USA) using a five-point Savitzky-Golay filter. Finally, spectral normalization was performed based upon the maximum peak at 1,448 cm⁻¹ to minimize spectral variation caused by the variations of cell density. Normalization and the following analyses were conducted by Matlab 2018b (Mathworks, Inc., Natick, MA, USA).

PCA and RF were employed to identify S. cerevisiae and S. uvarum to the species level. Raman spectra collected from all the yeast strains were subjected to PCA, and a score plot was generated using the first three PCs to visualize the separation of S. cerevisiae and S. uvarum strains. We also constructed a loading plot of the first PC, where the absolute coefficient value represents the weight of each wavenumber in PC1 and further indicates the contribution of the corresponding biochemical compositions to species identification. RF was further applied to classify S. cerevisiae and S. uvarum species. The number of trees was 50. To assess the performance of RF, the entire data set was randomly divided into 70% as the training data set and 30% as a hold-out test data set to mitigate overfitting. Identification accuracy is the percentage of correctly assigned Raman spectra out of the total tested spectra.

The influence of cultivation conditions on spectral reproducibility and discriminatory capability was investigated using PCA and RF. We created a three-dimensional PCA plot using Raman spectra obtained from four representative yeast strains incubated for different time periods. Similarly, another PCA plot was established to illustrate the relationships of yeast strains cultivated at different temperatures. We further conducted RF to identify these four yeast strains at both the species and strain levels based on the Raman spectra collected with different time periods or at different temperatures. The identification accuracies were calculated as aforementioned. Spectral reproducibility can be verified if the spectra collected from the same strain under different conditions are overlapped in the PCA plot. Discriminatory ability was evaluated based upon the classification accuracy of the yeast isolates using RF.

CNN was applied to further discriminate yeast isolates to the strain level. CNN is a state-of-art machine learning technique that has recently been adapted to the classification of spectral data (73). The proposed CNN in the current study consists of an initial input layer, four convolutional layers for feature extraction, two fully connected layers for classification, and an output layer (Fig. 4). Four convolutional layers include 64, 128, 32, and 16 convolutional filters with kernel size of 2 × 1 and a rectified linear unit activation function, respectively. The extracted features are flattened and transferred to two fully connected layers containing 256 and 50 neurons, respectively. Softmax is used as activation function in the output layer. The model was trained for 100 epochs with an initial learning rate of 0.001. One-dimensional Raman spectra from yeast strains were used as the input of CNN. For each spectrum, the probabilities across 27 stains were calculated, and the one with the maximum value was considered as the predicted strain. The hyperparameters of CNN were tuned using grid search algorithm, and the details are available in Table S4. We assessed the performance of CNN using a 10-fold cross-validation. Briefly, 90% of the data were randomly selected to train the model, and the remaining 10% was employed as the test data set in each validation. The validation process was repeated for 30 times to evaluate the stability of the model, and the overall accuracy was defined as the average accuracy of each strain over all the validations. A genuine hold-out test data set was included to further validate the high performance of the constructed CNN. Specifically, eight representative yeast strains (i.e., S. cerevisiae EC1118, F15, M2, and DV10 and S. uvarum 7D4, A4, CBS8711, and P36C10) were randomly selected out of the 27 yeast strains, and an independent batch of 30 Raman spectra were collected from each strain with the same procedures as aforementioned. The spectra were subjected to the cross-validated CNN, and the average identification accuracy was also calculated.

To predict the concentration of a specific yeast strain within a yeast mixture, another CNN was established and validated using 10-fold cross-validation. The structure of the CNN is shown in Fig. S2. It contains an input layer, two convolutional layers, one max-pooling layer, six fully connected layers with different numbers (i.e., 2,048, 1,024, 256, 128, 32, and 1) of neurons, and an output layer. The numbers of the filters in two convolution layers are 32 and 64, respectively. The kernel size of both convolution and pooling layers is set to 5 × 1. Raman spectra collected from ratio-defined yeast mixture were served as input, and the relative concentrations of the targeted strain within the mixture were predicted in the output layer. We determined the average error of the predicted ratio compared to the actual values, R², RMSE, and RPD to assess the reliability of CNN.

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/aem.01673-23.

Supplemental material. aem.01673-23-s0001.docx.

Fig. S1 and S2 and Tables S1 to S4.

ASM does not own the copyrights to Supplemental Material that may be linked to, or accessed through, an article. The authors have granted ASM a non-exclusive, world-wide license to publish the Supplemental Material files. Please contact the corresponding author directly for reuse.