Application of Raman spectroscopy and machine learning for Candida auris identification and characterization

ABSTRACT

Candida auris, an emerging fungal pathogen characterized by multidrug resistance and high-mortality nosocomial infections, poses a serious global health threat. However, the precise and rapid identification and characterization of C. auris remain a challenge. Here, we employed Raman spectroscopy combined with machine learning to identify C. auris isolates and its closely related species as well as to predict antifungal resistance and key virulence factors at the single-cell level. The average accuracy of identification among all Candida species was 93.33%, with an accuracy of 98% for the clinically simulated samples. The drug susceptibility of C. auris to fluconazole and amphotericin B was 99% and 94%, respectively. Furthermore, the phenotypic prediction of C. auris yielded an accuracy of 100% for aggregating cells and 97% for filamentous cells. This proof-of-concept methodology not only precisely identifies C. auris at the clade-specific level but also rapidly predicts the antifungal resistance and biological characteristics, promising a valuable medical diagnostic tool to combat this multidrug-resistant pathogen in the future.

IMPORTANCE

Currently, combating Candida auris infections and transmission is challenging due to the lack of efficient identification and characterization methods for this species. To address these challenges, our study presents a novel approach that utilizes Raman spectroscopy and artificial intelligence to achieve precise identification and characterization of C. auris at the singe-cell level. It can accurately identify a single cell from the four C. auris clades. Additionally, we developed machine learning models designed to detect antifungal resistance in C. auris cells and differentiate between two distinct phenotypes based on the single-cell Raman spectrum. We also constructed prediction models for detecting aggregating and filamentous cells in C. auris, both of which are closely linked to its virulence. These results underscore the merits of Raman spectroscopy in the identification and characterization of C. auris, promising improved outcomes in the battle against C. auris infections and transmission.

INTRODUCTION

The emergence of pathogenic fungi poses a serious threat to human health (1, 2). Candida auris, known as a “super fungus,” is an emerging multidrug-resistant pathogenic fungus that readily spreads among patients in healthcare facilities. Since first isolated in Japan in 2009, C. auris infections have been reported in at least 40 countries on 6 continents over the past decade (2, 3). Especially, C. auris has the potential to trigger large-scale outbreaks in the nosocomial setting (4, 5). C. auris can cause invasive infections in the bloodstream, central nervous system, and internal organs. The crude in-hospital mortality rate for C. auris is estimated to range from 30% to 60% (6, 7). The high mortality rate caused by C. auris is primarily attributed to its resistance to multiple antifungal agents. The majority of C. auris isolates exhibit resistance to fluconazole, and some display resistance to more than two antifungal agents (8). Another notable pathogenic trait of C. auris is its morphological diversity, including biofilm formation and yeast-filament switching, which augments the virulence of C. auris (9–11).

Currently, the origin and epidemiology of C. auris have not yet been elucidated. C. auris has been classified into five major discrete clades based on its genetic and genomic characteristics as well as the region of its initial isolation: the South Asia clade (I), the East Asia clade (II), the South Africa clade (III), and the South America clade (IV), and the Middle East (Iran) clade (V), and a possible sixth clade from Singapore and Bangladesh was reported (12). C. auris isolates exert clade-specific properties correlated to drug resistance and pathogenesis. Isolates belonging to clades I, III, and IV often cause invasive infections and nosocomial outbreaks, while clade II isolates are prone to colonize the ear (13). C. auris isolates exhibit fairly high rates of fluconazole resistance, while a majority of clade II isolates have susceptibility (14, 15). Currently, the diagnosis of C. auris in clinical laboratories primarily relies on conventional phenotypic methods and commercial identification systems, such as VITEK 2, BD Phoenix, and API20. Due to the close phylogenetic relationships, C. auris isolates are easily misidentified as Candida haemulonii, Candida lusitaniae, Candida guilliermondii, and Candida famata using the routine diagnostic methods (16, 17). Among these strains, misidentification occurs most frequently between C. auris and C. haemulonii (4). Molecular methods, such as PCR or real-time PCR, DNA sequencing, and matrix-assisted laser desorption ionization-time of flight mass spectrometry, provide more precise identification, but these methods usually necessitate time-consuming cell culturing and DNA and protein extractions and cannot directly identify C. auris at single-cell level from clinical specimens timely (18). Moreover, all the methods seem to have limitations in identifying C. auris isolates at clade level and can hardly assess drug resistance and virulence attributes of C. auris isolates promptly, potentially resulting in the improper use of antifungal agents or unnecessary interventions.

Addressing these challenges, we proposed a swift and precise identification method for C. auris using Raman spectroscopy. Raman spectroscopy is a rapid, non-destructive, and label-free analytical technique (19, 20). It allows for the measurement of all molecular vibrations within a single microbial cell, providing valuable information on various cellular molecules, such as proteins, nucleic acids, lipids, and carbohydrates, which has been extensively applied in biological research (21–23). Nevertheless, the Ramanome (the full Raman spectrum of a single cell) consists of overlapping Raman peaks from all cellular molecules, thus retrieving biological information from the complex Ramanome is challenging. Here, the convolutional neural network (CNN) algorithm which exhibited superior performance with no need for special pre-processing steps compared to other machine learning algorithms, such as random forest, support vector machine, and K-nearest neighbor, was employed to construct artificial intelligence models (24). Additionally, the CNN model can automatically extract common features from all spectra under each label for learning and exhibits heightened sensitivity in identifying mixture components (25). Machine learning algorithms are capable of streamlining the classification, clustering, and prediction tasks for high-dimensional data sets, making them highly advantageous for interpreting the biological significance of Raman spectral data (26).

In this study, we devised a novel proof-of-concept methodology that combines Raman spectroscopy with CNN, a machine learning technique, to precisely identify C. auris and its closely related species at the single-cell level, providing a rapid and accurate diagnostic strategy for C. auris infection. This approach is capable of not only precisely identifying C. auris but also differentiating among its various phylogenetic clades, achieving an average accuracy of 93.33%. Moreover, we successfully identified mixed C. albicans and C. auris in urine with an accuracy of 98%. Additionally, multiple CNN models were constructed to accurately predict three critical biological characteristics associated with C. auris infection: antifungal resistance, aggregation, and filamentous morphology.

RESULTS

Using Raman spectroscopy to identify C. auris isolates at single-cell level

The swift and widespread prevalence of C. auris infection can be attributed, in part, to the lack of a fast, accurate, and reliable diagnostic tool. To address this challenge, we employed a deep learning algorithm, convolutional neural networks, to construct an identification model aimed at identifying C. auris isolates based on the single-cell Raman spectrum. A total of 3,652 single-cell Raman spectra were acquired from 42 strains of 6 closely related Candida species. The average Raman spectra of 42 Candida strains are shown in Fig. 1A. We categorized the data set into nine distinct labels based on species or clades of strains, including C. albicans, C. lusitaniae, C. haemulonii, C. duobushaemulonii, C. pseudohaemulonii, and the four major clades of C. auris (I, II, III, IV) to facilitate model training (Fig. 1B). To address potential concerns about the impact of culture conditions on fungal identification accuracy with this method, we tested five different growth conditions varying in medium, temperature, and incubation time to evaluate the influence of growth conditions on Raman spectra. As shown in Fig. S1, the variance in culture conditions had a minimal impact on the spectral data, suggesting a degree of intrinsic stability in the Raman spectra of fungi.

Fig 1.

The average Raman spectra of C. albicans, C. lusitaniae C. haemulonii, C. duobushaemulonii, C. pseudohaemulonii, and four C. auris clades. (A) The average Raman spectra of 42 Candida strains. (B) The average Raman spectra of the four clades of C. auris, C. albicans, C. lusitaniae, C. haemulonii, C. duobushaemulonii, and C. pseudohaemulonii. The solid line represents the average Raman spectrum, and the shaded area indicates the standard deviation.

A 10-fold cross-validation was utilized to select the optimal CNN model with the highest identification accuracy, and its specificity and sensitivity were further verified by the receiver operating characteristic (ROC) curve analysis with a value of 0.99 based on the Raman spectra (Fig. S2). An independent testing data set, including 900 Raman spectra from 9 distinct clinical isolates not part of the training data set, was used for the model validation. As shown in Fig. 2A, an average accuracy of 93.33% (840/900) was achieved to identify the four C. auris clades, C. albicans, C. lusitaniae, C. haemulonii, C. duobushaemulonii, and C. pseudohaemulonii by the CNN model. Notably, the rate of accurate detection reached 100% (100/100) for C. albicans, 99% (99/100) for C. auris clade II, 96% (96/100) for C. lusitaniae and C. auris clade IV, 95% (95/100) for C. haemulonii, and 90% (90/100) for C. duobushaemulonii and C. auris clade I. For C. pseudohaemulonii and C. auris clade III, the classification accuracy was slightly lower at 88% (88/100) and 86% (86/100), respectively. Despite these variations, the accuracy of each category exceeded 85%, suggesting that the combination of Raman spectroscopy with the CNN is a reliable method for exactly identifying C. auris isolate at both the species level and clade levels.

Fig 2.

Classification results of the CNN model. (A) Identification accuracy of the CNN model for C. auris clades and closely related species. The confusion matrix shows the percentage of accurate prediction for each category. An average identification accuracy of 93.33% was achieved. (B) Identification accuracy of the CNN model in the simulated clinical sample. Total 100 single cells from the mixture of C. auris and C. albicans in artificial urine were identified by both the ITS sequencing and Raman spectroscopy. The white boxes indicated discrepancies between sequencing results and Raman results.

In a clinical setting, Candida infections commonly exist in a mixture of multiple pathogens, predominantly with C. albicans (27). To further verify the practical applicability of our method, we endeavored to precisely identify individual pathogens within a mixture of C. auris and C. albicans in artificial urine. Comparing the sequencing identification results with the Raman identification results, we found that only 2 out of 100 single cells were misidentified, achieving an accuracy of 98% (98/100) (Fig. 2B).

Furthermore, we employed the occlusion-based Raman spectra feature extraction (ORSFE) method to capture the specific Raman peaks of each Candida species and C. auris clades. The Raman peaks with the highest contributing values for each category are indicated in Fig. S3, and the corresponding chemical assignments are detailed in Table S1. Notably, the feature peak at 1,450 cm⁻¹ (collagen) for C. albicans, the peak at 1,339 cm⁻¹ (collagen and nucleic acid) for C. lusitaniae, the peak at 1,659 cm⁻¹ (collagen-like protein) for C. haemulonii, the peak at 1,004 cm⁻¹ (collagen) for C. duobushaemulonii, the peak at 716 cm⁻¹ (lipid) for C. pseudohaemulonii, the peak at 780 cm⁻¹ (Uracil) for C. auris clade I, the peak at 1,453 cm⁻¹ (protein) for C. auris clade III, and the peak at 1,005 cm⁻¹ (protein) for C. auris clade IV were identified. C. auris clade II contains three important feature peaks at 1,066 cm⁻¹ (collagen), 1,083 cm⁻¹ (lipid), and 1,126 cm⁻¹ (lipid and protein). These unique molecular signatures imply their potential involvement in distinct biological processes, contributing to species differentiation.

Predicting antifungal sensitivity of C. auris at single-cell level

Timely detection and accurate estimation of drug-resistant isolates are pivotal for the effective treatment and prevention of C. auris infections. However, few methods have successfully integrated rapid diagnosis with an assessment of drug resistance properties. To bridge this gap, four clinical isolates from clade I and another four clinical isolates from clade III were used to construct the CNN models of fluconazole and amphotericin B, respectively (Table S2). As shown in Fig. 3A, the predicted accuracy for fluconazole resistance reached 99% (198/200), with only 2 out of 200 single cells of C. auris deviating from the broth dilution outcome. Similarly, the predicted accuracy of amphotericin B resistance reached 94% (188/200). To demonstrate that the identification results are based on genuine Raman spectral differences in drug resistance phenotypes rather than differences between species, we used three fluconazole-resistant strains and three susceptible strains to validate this model (Table S2). Notably, all these strains were derived from the same strain, BJCA001, through an experimental evolution strategy (28), The classification accuracy of strains with different resistance phenotypes of the same strain was 100% (200/200) (Fig. 3B). These results underscore the high feasibility of this approach for detecting drug-resistant isolates of C. auris.

Fig 3.

Prediction of antifungal sensitivity in C. auris. (A) Accuracy of the fluconazole and amphotericin B sensitivity prediction model. FLC-R indicates resistance to fluconazole; FLC-S indicates susceptibility to fluconazole; AMB-R indicates resistance to amphotericin B; AMB-S indicates susceptibility to amphotericin B. The predicted label shows results from the model prediction, while the true label shows results from broth dilution method. (B) Accuracy of the fluconazole sensitivity prediction model of different phenotypes of BJCA001. BJCA001-R indicates resistance to fluconazole; BJCA001-S indicates susceptibility to fluconazole. (C) Raman intensity of drug-resistant and susceptible strains at different peaks. The Raman intensity of fluconazole resistant and fluconazole susceptible at peaks 937 and 1,005 cm⁻¹. The Raman intensity of amphotericin B resistant and amphotericin B susceptible at peaks 1,005 cm⁻¹. The red dot represents the average Raman intensity, while the embedded box plot indicates the median and the first and third quartiles of the spectrum. The whiskers denote the minimum and maximum spectral values within 1.5 times the interquartile range from the first and third quartiles. Statistical significance in spectral differences between C. auris with different resistance levels was determined using a two-tailed t-test. **P ≤ 0.01. ****P ≤ 0.0001.

Next, we sought to pinpoint the representative Raman peaks crucial for the accurate antifungal resistance prediction using the ORSFE method. For predicting fluconazole susceptibility, the pivotal peaks were located at 718, 781, 860, 937, and 1,005 cm⁻¹. Peaks at 856, 935, 1,005, and 1,452 cm⁻¹ played a significant role in predicting amphotericin B sensitivity (Fig. S4 and S5). As illustrated in Fig. 3C, the peaks at 937 and 1,005 cm⁻¹ associated with protein exhibited significantly higher intensities in fluconazole-susceptible strains, and the intensity of the peak at 1,005 cm⁻¹ (protein) was significantly increased in amphotericin B-susceptible strains (29, 30).

Predicting the aggregative phenotype of C. auris at single-cell level

A characteristic phenotypic feature of C. auris is its multicellular aggregating form, which is believed to be related to defects in cell division and significantly influences its biofilm formation, drug resistance, virulence, and ability to evade the immune system (10, 31, 32). A total of 447 single-cell Raman spectra from 2 aggregative-form strains (SJ02, A103) and 2 non-aggregative-form strains (SJ01, XM03-1) were acquired to build a training data set. As shown in Fig. 4A, despite similar patterns in the average Raman spectra between aggregative and non-aggregative cells, changes in Raman intensities at the same Raman shift exhibit obvious differences. The t-distributed stochastic neighbor embedding (t-SNE) clustering analysis indicated a clear separation between the aggregative-form and non-aggregative-form strains into two distinct clusters (Fig. 4B). This separation was further confirmed by a principal component analysis (PCA), where the first two PCs accounted for 91.23% of the original variance, and the two phenotypic cell types were completely separated (Fig. S6A). The loading of PC1 revealed 13 spectral peaks with significant differences between aggregative cells and non-aggregative cells (Fig. S6B), with the corresponding assignment of feature Raman peaks depicted in Table 1. Notably, the intensity of the 722 cm⁻¹ peak (DNA) exhibited the most substantial difference between the two groups, which was higher in the spectrum of non-aggregative cells compared to aggregative cells (33). These spectral variances suggested that different compositions and content of biomolecules, such as nucleic acid, proteins, and lipids, might play a critical role in contributing to the diverse phenotypic characteristics.

Fig 4.

The aggregative-form characteristics of C. auris single-cell Raman spectra. (A) The average Raman spectra of aggregative and non-aggregative cells of C. auris. The heatmap illustrates the spectral differences between the two groups. The colors from light blue to orange, corresponding the values of 0 to 0.4, indicate higher Raman peak intensity in non-aggregative cells of C. auris, while the colors from light blue to dark blue, corresponding the values of 0 to −0.1, indicate higher intensity in aggregative cells of C. auris. (B) t-SNE analysis between aggregative and non-aggregative strains in C. auris. The aggregative strains (SJ02, A103) and non-aggregative strains (SJ01, XM03-1) were sorted into two separate cluster. (C) The accuracy of aggregative-form prediction model. The predicted label represents the predicted results from the model, while the true label represents the actual outcome from the observation of cell culture. The model achieves an average predictive accuracy of 100%. (D) Cellular images of aggregative and non-aggregative cells of C. auris. The scale bar is 10 μm.

TABLE 1.

Distribution of differential Raman peak positions associated with aggregative form and non-aggregative form of C. auris

Wavenumber (cm−1)	Assignment/wavenumber (cm−1)	Reference
722	DNA	(33)
761	Ring breathing tryptophan (protein) (760)	(30)
882	Valine (protein)	(34)
960	Symmetric P-O stretching mode (phosphate)	(35)
1,007	Phenylalanine (protein) (1,008)	(36)
1,111	Saccharide (1,112)	(37)
1,132	Palmitic acid and fatty acid (lipid) (1,131)	(37)
1,158	C-C/C-N stretching (protein)	(30)
1,282	Amide III (protein)	(38)
1,347	Ring plus C-H	(39)
1,439	CH3, CH2 deformation (protein)	(40)
1,469	Adenine and guanine	(41)
1,609	NH2 (Cytosine)	(42)

Similarly, we developed a CNN prediction model for the aggregative phenotype. To assess the model’s accuracy, cells recovered from an independent testing data set consisted of a mixture of the four isolates (SJ01, SJ02, A103, and XM03-1). The 200 single cells were isolated from the cell mixture to acquire Raman spectra and subsequent cell culture and cellular phenotype observation. All cellular phenotypes were consistent with the predictions made by the CNN model, and the model’s accuracy reached 100% (200/200) (Fig. 4C and D), emphasizing the robustness of our CNN model in precisely predicting the aggregation phenotype of C. auris at the single-cell level.

Utilizing the single-cell Ramanome to screen the filamentation-competent cells of C. auris

Filamentation is a feature of C. auris isolates, which is associated with both virulence, immune response, and antifungal resistance (9, 43, 44). In C. auris, some typical yeast cells can acquire the ability to form filaments, which are termed “filamentation-competent yeast (FC yeast).” These FC yeast cells can develop robust filamentous cells for many generations under environmental stimuli, such as low temperatures (43). However, as the temperature rises, they revert to a yeast phenotype and become indistinguishable from typical yeast cells. The yeast-filamentation switch in C. auris occurs at low frequency, making the filamentation-competent cells extremely challenging to screen and identify in clinical and routine laboratories (43). Given this challenge, we sought to discriminate between the two distinct phenotypes of C. auris using single-cell Raman analysis. A total of 416 single-cell Raman spectra were collected from typical yeast and FC yeast cells of three isolates (NSICU4, RICU4, RICU13) derived from C. auris clade III, which differ solely in their filamentation capabilities. As illustrated in Fig. 5A, the average spectra of FC yeast cells and typical yeast cells exhibited discernible differences. We then developed a CNN model and used 200 independent testing Raman spectra to evaluate the sensitivity and specificity of the morphological identification (Fig. 5B). The morphological identities of single cells subjected to CNN model prediction were verified by incubation at 25°C for 3 days (Fig. 5C). For typical yeast cells, the prediction accuracy reached 100% (100/100). For FC yeast cells, approximately 6% (6/100) of cells were mistakenly predicted as typical yeast cells. Therefore, the average prediction accuracy of the CNN model was 97% (194/200), demonstrating its effective capacity for detecting the FC yeast cells and predicting the potential phenotypic switching of C. auris (Fig. 5B and C).

Fig 5.

The FC yeast and typical yeast features of C. auris single-cell Raman spectra. (A) Raman spectra of FC yeast and typical yeast forms of C. auris. The solid line represents the average spectrum, and the shaded area represents the standard deviation. (B) The accuracy of FC yeast and typical yeast forms prediction model. The horizontal axis represents the predicted labels from the model validation, while the vertical axis represents the actual phenotypes validated by cell culture under 25°C. (C) Representative cellular images of cells recovered from single cells after acquiring Raman spectrum. Once measuring the Raman spectrum, the single cell was isolated by optical tweezers and incubated at 25°C for 3 days for phenotypic verification. The scale bar is 10 μm. (D) The loading of PC1 and PC3 by PCA analysis. (E) Box plots showing Raman intensity of peaks at 860, 1,002, and 1,667 cm⁻¹ for FC yeast and typical yeast forms. The middle line represents the median, the box represents the first and third quartiles, and the vertical line represents the minimum and maximum within 1.5 times the interquartile range from the first and third quartiles. Each point represents the Raman intensity of an individual cell of C. auris. Statistical significance in the differences between FC yeast and typical yeast strains was determined using a two-tailed t-test. ****P ≤ 0.0001.

To further investigate the chemical disparities between the two phenotypes, we applied the PCA method to the Raman spectra of the two morphologies to elucidate the differential peaks of the first and third principal components, which exhibited the highest variability (Fig. 5D; Fig. S7). The corresponding chemical assignments are detailed in Table S3. Specifically, the peak at 860 cm⁻¹ (lipids) in the FC yeast cells was notably higher compared to that in the typical yeast cells. Additionally, there was a substantial discrepancy between the two peaks at 1,002 and 1,667 cm⁻¹, which signifies protein (Fig. 5E). These findings underscore evident disparities in the chemical composition between the typical yeast cells and FC yeast cells, providing clues for understanding the molecular mechanisms underlying the filamentous and yeast phenotypes in relation to the pathogenesis and epidemiology.

DISCUSSION

In clinical practice, the rapid and accurate identification of pathogenic microorganisms, alongside the precise characterization of their attributes, such as drug resistance, adhesion, and virulence, is pivotal for the effective diagnosis and treatment of infectious diseases. In this study, we employed Raman spectroscopy in combination with machine learning to identify C. auris isolates and predict their biological characteristics at single-cell resolution. This innovative approach not only facilitates the precise diagnosis of C. auris infections, addressing the challenge of misidentification in clinical settings, but also promptly predicts the antifungal resistance and distinctive morphologies of C. auris. Importantly, this approach does not require complex sample preparation, time-consuming culturing or labeling, providing a potential advantage in the rapid, accurate detection and characterization of C. auris isolates in clinical settings, such as blood, urine, or skin swabs. Furthermore, considering the non-destructive, versatile, and adaptable features of Raman spectroscopy, this approach could be seamlessly integrated with other techniques, such as cell culture, genome sequencing, transcriptome, proteome, and metabolome analyses, to enhance overall performance and functionality.

Recently, there has been a surge in research focusing on the multidrug resistance and virulence factors of C. auris due to its association with high mortality rates and the prevalence of outbreaks in healthcare settings. In our study, we developed machine learning models tailored to detect antifungal resistance and differentiate two distinctive phenotypes of C. auris based on the single-cell Raman spectrum. Notably, our models accurately predicted drug resistance to two common antifungals, fluconazole and amphotericin B, with accuracies of 99% and 94%, respectively. This advancement underscores the significant potential of Raman spectroscopy in characterizing drug resistance directly from the single cells in clinical specimens, which could greatly enhance drug resistance analysis of pathogenic microorganisms. Moreover, we constructed prediction models for detecting aggregating and FC yeast cells within C. auris populations. Interestingly, the aggregating-form prediction model achieved an impressive accuracy of 100%. The identification accuracy of FC yeast cells was 94%, associated with the fact that a small fraction of FC yeast cells of C. auris might revert to typical yeast cells for some unknown reasons (43). These results further highlight the merit and potential of Raman spectroscopy in the prevention and treatment for C. auris infection.

Understanding and unraveling the intricate molecular mechanisms of drug resistance and pathogenesis in C. auris infection are of paramount importance. In the current study, we employed the ORSFE and PCA to infer the molecular differences from the Raman spectral signatures. In the spectral fingerprints related to amphotericin B sensitivity, we observed significantly higher protein-related Raman intensity in drug-susceptible strains compared to drug-resistant strains, which might be linked to mutations in the ERG6 gene (2). For fluconazole, drug-resistant strains exhibited lower protein-related peak intensity, indicating that reduced metabolic changes could play a major role in the response to antibiotics in resistant strains of C. auris. The distinctive cell aggregation observed in strains SJ02 and A103 was related to the amplification of ALS4 gene (10). This amplification might significantly increase the Raman peak intensity of proteins (882 cm⁻¹) (34). Furthermore, the heightened presence of phosphates (960 cm⁻¹), saccharide (1,111 cm⁻¹), lipids (1,132 cm⁻¹), and proteins (761, 1,007, 1,158, 1,282, and 1,439 cm⁻¹) likely contributed to enhanced biofilm formation (30, 35–38, 40). Interestingly, the spectral disparities between typical yeast cells and FC yeast cells of C. auris were more striking. For instance, in FC yeast cells of C. auris, elevated Raman intensity of lipids (860 cm⁻¹) aligns with the enrichment of genes related to fatty acid metabolism. Additionally, upregulated expression of genes encoding regulatory factors likely contributed to the heightened Raman peak intensity of proteins (1,002 and 1667 cm⁻¹) in FC yeast cells of C. auris. Other Raman peak discrepancies might be associated with sugars, nucleic acids, lipids, and proteins. Although we inferred the potential molecular foundations underlying the traits of C. auris based on the Ramanome of single cells, it remains a challenge to directly interpret the complex and high-dimensional spectral signatures of specific biomolecules. Thus, future efforts should focus on developing a comprehensive spectral database and robust artificial intelligence algorithms.

In conclusion, we devised a proof-of-concept method that combines Raman spectroscopy with artificial intelligence to precisely and promptly identify and determine the characteristics of C. auris at the single-cell level. This innovative approach holds promise for improved outcomes in combating C. auris infection and controlling its transmission. Future work will focus on transitioning this proof-of-concept technique into a fully validated clinical diagnosis protocol.

MATERIALS AND METHODS

Strains and cultivation

The Candida strains used in this study are listed in Table S4. A yeast extract peptone dextrose (YPD) medium (2% dextrose, 2% peptone, 1% yeast extract, and 2% agar) was used for routine growth of all Candida strains. The fungal cells were incubated at 37°C for 2.5 days for spectral analysis and at 25°C for 3 days for phenotype identification. An additional 0.5% Phloxine B was added to the YPD medium to facilitate the capture of colony and cellular images. To investigate the influence of culture conditions on the spectrum, C. auris clade I CBS12768 was cultured on YPD medium and Salmonella Shigella medium [1% mixture of peptic digest of animal tissue and pancreatic digest of casein (1:1), 2% dextrose, and 2% agar] with different culture times of 2, 2.5, and 3 days, and different culture temperatures of 30°C and 37°C.

Raman spectrum acquisition

Around 6–10 monoclonal colonies were randomly collected from the medium plates and washed triple with 0.9% sodium chloride (NaCl). Then fungal cells were resuspended and adjusted to 1 × 10⁵ cells/mL in 0.9% NaCl. Furthermore, 100–200 μL of fungal cell suspension was pipetted onto a quartz chip for spectral acquisition. Spectral acquisition was carried out as previously described (21, 45). The specific single-cell Raman spectrum acquisition quantity for each label is shown in Table S5. Raman spectra of all strains were collected from single cells in yeast form. For cellular phenotype evaluation, the single cell was dragged to a specific well on the chip by laser tweezers for subsequent cell culture right after acquiring its Raman spectrum.

Data processing and analysis of Raman spectra

The Raman spectra were converted to an ASCII file using Winspecs software. The “Ramanpro0.4.2” software package, developed in the R language, was employed for processing the Raman spectra. Briefly, cosmic rays were removed, and background was subtracted from all Raman spectra. The spectra were smoothed using the Savitzky-Golay filter, and baseline correction was achieved by Modpolyfit. Subsequently, the spectra were normalized using the Min-Max normalization method.

For data analysis, principal component analysis and t-distributed stochastic neighbor embedding were employed. PCA was executed using the R package “Ramanpro0.4.2” with the confidence interval parameter set at 95% to delineate different categories. t-SNE was performed using the R package “Rtsne” with the perplexity parameter set to 25.

Machine learning model construction and feature Raman spectral peak extraction

The machine learning model was constructed using the CNN algorithm as previously described (45). To establish the CNN model, labels were defined based on clades, species, or phenotypic characteristics. For each label, no fewer than 200 Raman spectra were used for training. The optimal CNN model was selected by performing a 10-fold cross-validation to assess diagnostic accuracy through the ROC curve, which plots sensitivity vs specificity. After determining the optimal, the accuracy rate of classification model was evaluated using an independent clinical isolate as the testing data set, comprising 100 Raman spectra per label, and sourced independently from the training data set. For the antifungal drug susceptibility model, we utilized the results of antifungal susceptibility using the broth dilution method as the true label to validate the model’s accuracy.

Both ORSFE method and PCA were employed for extracting feature Raman spectral peaks as previously described (45, 46).

Antifungal susceptibility testing

Antifungal susceptibility testing was performed in accordance with the guidelines outlined in the Clinical and Laboratory Standards Institute standard M27 (47). In brief, C. auris cells were initially cultured in YPD medium plates at 37°C for 24 h. The colonies were collected and washed twice with ddH₂O, and then fungal cells were adjusted to 5 × 10³ cells/mL in RPMI 1640 medium (wt/vol, 1.04% RPMI-1640, 3.45% 3-morpholinopropanesulfoinc acid (MOPS), NaOH used for pH adjustment to 7.0). Approximately 500 fungal cells in 0.2 mL RPMI 1640 medium were mixed with a serial twofold concentration of fluconazole or amphotericin B in a 96-well plate for minimum inhibitory concentration assay. The tentative breakpoints of fluconazole and amphotericin B for C. auris were set as ≥32 and ≥2 mg/L, respectively (48).

Clinical mixed sample identification details

C. albicans SC5314 and C. auris clade I CBS12768 were mixed in YPD medium. Before spectrum collection, the strains were added to artificial urine to prepare samples simulating clinical mixed infection. A total of 100 single cells were selected for spectrum collection. Then, the CNN model was used to identify the spectra, and the results were compared with internal transcribed spacer (ITS) sequencing. Sequencing was performed by Beijing Tsingke Biotech Co., Ltd using ITS primers (ITS1: TCCGTAGGTGAACCTGCGG, ITS4: TCCTCCGCTTATTGATATGC).

DATA AVAILABILITY

The authors confirm that the data supporting the findings of this study are available within the article and its supplemental material.

SUPPLEMENTAL MATERIAL

The following material is available online at https://doi.org/10.1128/aem.01025-24.

Supplemental material. aem.01025-24-s0001.docx.

Figures S1 to S7 and Tables S1 to S5.

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.