In this study, four different bacterial strains were cultivated: Escherichia coli (ATTC 10536), Pseudomonas putida (DMSZ 291), Staphylococcus arlettae (CVD059), Shewanella oneidensis (ATCC 700550). All bacteria were cultivated in Luria Bertani (LB) broth except Shewanella oneidensis, cultivated in Tryptic Soy (TS) media.

Cell suspensions were grown overnight (18 h) in the corresponding liquid culture media to prepare agar plate samples for UV hyperspectral imaging. The optical density of the overnight culture was measured and diluted to a cell concentration of approximately 10⁸ CFU·mL^-1 (OD 0.12). Afterward, four individual spots were created for each bacterial culture plate, inoculating 5 µL of cell suspension per spot. Plates corresponding to Escherichia coli (EC) and Staphylococcus arlettae (SA) were incubated at 37°C, while Pseudomonas putida (PP) and Shewanella oneidensis (SO) at 30°C.

The set-up used for UV hyperspectral imaging was adapted from the system described in a previous work (Al Ktash et al., 2023). Figure 1 illustrates the UV hyperspectral imaging system based on a back-illuminated CCD camera (Apogee Alta F47: Compact, inno-spec GmbH, Nürnberg, Germany) and a spectrograph (RS 50–1938, inno-spec GmbH, Nürnberg, Germany) with a slit width of 30 μm. The CCD camera has a resolution of 1,024 × 1,024 pixels (spatial × spectral) and a pixel size of 13 μm × 13 μm. The integration time was optimized and set at 300 m to measure bacterial colonies. The set-up also incorporates a Deuterium lamp (SL3 Deuterium Lamp, StellarNet Inc., Tampa, Florida) as the light source. The hyperspectral set-up contains a conveyor belt (700 mm × 215 mm × 60 mm, Dobot Magician, Shenzhen Yue-jiang Technology Co., Ltd., Shenzhen, China), where samples are placed and displaced until the detection area. The conveyor belt moves at a constant speed of 0.15 mm/s.

The UV hyperspectral imaging data were acquired with the SI-Cap-GB version V3.3.x.0 software (inno-spec GmbH, Nürnberg, Germany), which automatically calculated the values of the reflectance measurements. In order to correct the illumination variations during the scanning line and the scattering produced by the different topographies of the bacterial colonies, raw spectra corresponding to each pixel were converted to reflectance spectra after radiometric calibration using the following Equation 1 (Boldrini et al., 2012). R e f l e c t a n c e = R R 0 = I s a m p l e − I d a r k I r e f e r e n c e − I d a r k (1) where R is the intensity reflected from the sample, and R₀ is the intensity reflected from a specific reference material with high reflectivity, in this case, Polytetrafluoroethylene (PTFE) or Teflon). I_sample corresponds to the spectral intensity of the sample, I_dark is the intensity of the dark, and I_reference is the intensity of the reference sample, in this case PTFE. Absorption is calculated as the negative logarithm of reflectance (Abs = -log (R/R₀)). All the data presented in the graphs of this research article are represented using absorbance, calculated as the logarithm of the reflectance.

The acquired hyperspectral data were analyzed using the HydraPCA software, a custom-made program in Matlab (MATLAB 9.2.0, Mathworks, MA, United states), which was designed to extract spectra from individual pixels. Considering the sensor size (1,024 pixels × 1,024 pixels), the program analyzed 1,024 pixels in the x-direction (lateral resolution in x), and 1,024 pixels in the wavelength λ direction ranging from 225 nm to 410 nm (variables/columns in the PCA matrix). By moving the conveyor belt line by line, a lateral resolution in y is realized.

Multivariate data analysis was carried out with Aspen Unscrambler™ version 10.5.1″(Aspen Technology Inc., Bedford, MA, United states). First, UV spectra were preprocessed to calculate PCA. Linear baseline correction and Savitzky–Golay smoothing (5 points, symmetric, 3rd polynomial order) were carried out. PCA models were then calculated with mean centering, cross-validation, and the NIPALS algorithm. QDA was used with mean centering and segmented cross-validation for model building according to the sample type and the Kernel algorithm.

The prediction performance was analyzed by calculating the accuracy (Equation 2), specificity (Equation 3), and sensitivity (Equation 4) of the model (Dankers et al., 2019). A c c u r a c y = T N + T P T N + T P + F P + F N (2) S p e c i f i t y = T N T N + F P (3) S e n s i t i v i t y = T P T P + F N (4) where TP is true positive, FP is false positive, TN is true negative, and FN is false negative prediction.

The four bacterial species selected for this research—Escherichia coli, Pseudomonas, Staphylococcus, and Shewanella—exhibit distinct characteristics while sharing traits that make them suitable for UV-based detection (details in Table 2). Both Escherichia coli and Pseudomonas are gram-negative bacteria. E. coli is a significant pathogen in the healthcare field and a widely used model organism in research, while Pseudomonas is notable for its environmental importance and its role as an opportunistic pathogen, particularly in immunocompromised individuals. Pseudomonas, produces pigments like pyocyanin which offer unique spectral features under UV light, whereas E. coli lacks pigments but absorbs UV light through its structural components, such as proteins and nucleic acids. Staphylococcus, a gram-positive bacterium, is medically significant due to its involvement in various infections and produces pigments like staphyloxanthin. Shewanella, another gram-negative bacterium, has also been reported as an opportunistic pathogen which contains a wide variety of light absorbing c-cytochromes. Despite differences in morphology, pigmentation, and ecological roles, all four species demonstrate potential for differentiation using UV hyperspectral imaging.

Bacterial agar plate colonies were produced for the four bacterial strains [Escherichia coli (ATTC 10536), Pseudomonas putida (DMSZ 291), Staphylococcus arlettae (CVD059), Shewanella oneidensis (ATCC 700550)]. The four colonies were grown until reaching a minimum size of 13 µm in diameter, corresponding to the hyperspectral imager resolution (13 μm × 13 µm). Then, the plates were analyzed using UV hyperspectral imaging in the wavelength range between 225 and 400 nm. Control agar plates with LB and TSA without bacteria were also prepared and used as control samples.

Figure 2 shows images of the different steps carried out for the UV-hyperspectral image analysis of the colonies on agar plates. First, agar plates under analysis were located on the conveyor belt. Once reaching the detection area (i.e., below the lamp, Figure 2A), the camera took hyperspectral images in reflectance mode (Figure 2B). As shown in Figure 2B, the bacterial colonies could be distinguished from the surrounding medium, being able to count the number of colonies in the sample, i.e., in this case four. The area corresponding to each colony contained multiple pixels (between 220 and 250 pixels in the case of overnight cultures of Escherichia coli), which could be easily selected for bacterial identification through spectral analysis. Even when each pixel provided an independent UV-spectrum, the average spectrum of the 220–250 individual pixels was considered representative of the specific bacterial colony and used to compare colonies between them (Figure 2C).

Representative agar plates of each sample type are presented in Figures 3A1–6. In addition to an agar plate containing four colonies for each bacterial strain (Shewanella oneidensis, Escherichia coli, Pseudomonas putida and Staphylococcus arlettae), a control agar plate was analyzed for each agar type (LB and TSA)—six plates in total. As shown, the colonies corresponding to each bacterial strain differed in morphology, topography, and color. This resulted in species-specific absorbance spectra for each bacterial type, i.e., bacterial fingerprints, which should permit the differentiation and identification of each bacterial strain (Figure 3B). It can be observed that the spectra from LB and TSA agar (Figures 3B1, 3B3) clearly differed from plates containing bacterial colonies. LB and TSA agar presented a small peak near 230 nm (peak 1), which was not observed in bacterial colonies. In contrast, the samples containing bacteria colonies show peaks between 260 and 280 nm (Figures 3B2–B6; peak 2) and a shoulder in the range between 290 nm and 320 nm (Figures 3B2–B6; peak 4). Variances from 300 nm to higher wavelengths are probably related to the different biochemical composition of bacteria, e.g., the presence of specific pigments.

In each agar plate, four discrete colonies were grown and used to acquire hyperspectral images. Three of the colonies were used to create a model and the fourth one was employed for future sample prediction.

Principal component analysis (PCA) was used to extract the essential spectral characteristics of all spectroscopic data obtained from hyperspectral imaging (Uddin et al., 2021), thus enabling straightforward data analysis and visualization. The 3D score plot of the PCA model created for the agar and bacterial colony samples at each sample pixel is shown in Figure 4A. This model explains 92% of the total variance of the spectral changes by the first three principal components. The first principal component PC1 (87%), captures the largest amount of data. The following components, PC2 (4%) and PC3 (1%), capture the remaining variation in descending order. The closer the samples are in the score plot, the more similar they are concerning the three components.

The PC1 variable influences 87% of the recorded data and, as indicated by the loadings, it has a constant positive correlation along the spectra without significant variations. On the other hand, PC2 and PC3 have more variations and, as can be observed in the 2D score plot for components PC2 and PC3 (Supplementary Figure S1), they are responsible for a better grouping of the different samples. All bacteria except SO have a negative correlation with PC2 and differ in the correlation to PC3, which simplifies their classification. The difference in PC2 will make SO better identified against the other bacterial species. On the other hand, when considering the samples that are positive in PC2, LB agar is negatively correlated to PC3, TSA has a close to null relation, and SO is positively related. Hence, the samples can be distinguished and grouped only considering three PCs. The loading plot (Figure 4B) also indicates that the most significant variation in the PCs starts around the 290 nm wavelength. Thus, in future experiments, the analyzed spectral range may be reduced, making it possible to decrease the amount of collected data.

In order to validate the quality of the PCA model, the scores of the first three PCs were used to make a quadratic discriminant analysis (PCA-QDA). For each sample, PCA-QDA gives a value that correlates with a group, and thus it is possible to determine the group to which the sample belongs to. The sample is assigned then to the corresponding group, either a bacterial species or an agar type. The model performance was tested with segmented cross-validation according to sample type (Avram et al., 2020).

Figure 5 illustrates the confusion matrix (Supplementary Table S1) for the agar media and bacterial colony samples, which evaluates the reliability of the classification model. It summarizes the number of correct and incorrect predictions made by the model on a data set by comparing the predicted and actual labels. The overall accuracy of the model was 90.3%. The high diagonal bars represent the accordance between the predicted and actual values, corresponding to dark-grey highlighted diagonal elements in Supplementary Table S1. The correctly predicted pixel number was much more significant than the incorrect predictions, corresponding to off-diagonal elements light-highlighted in Supplementary Table S1.

The sensitivity and specificity of the model were determined for all samples. Spectra corresponding to the TSA and LB agar were predicted with 96.6% and 93.8% of sensitivity and 98.9% and 98.5% of specificity. Most of the mispredicted ones corresponded to the other agar types. So, it could be concluded that the model can distinguish bacterial colonies from the agar. Regarding the bacterial colonies, SO was the best-predicted bacteria with 98.2% sensitivity and 99.6% specificity. The bacteria identified with the lowest sensitivity (78.2%) and specificity (95.7%) was PP.

Once the model was developed, it was used to predict the four colonies and certify its ability to identify different bacterial colonies. After analyzing the pixels corresponding to each colony, the probability of belonging to one of the four bacterial strains or the two agar types was determined with the current model (Table 1). The model indicated that there is a 98% chance for colony 1 to correspond to SO, 92% for colony 2 to be EC, 78% for colony 3 to be PP, and 83% for colony 4 to pertain to SA. Thus, all colonies could be correctly predicted after classifying them according to the highest probability.

UV-Hyperspectral images of colonies containing agar plates allowed for easy detection and selection of the areas of interest for the spectral analysis. One of the first differences to highlight was the peak at 230 nm, which was only present in the LB and TSA agar controls. This peak has been related to unfolded (denaturalized) proteins (Liu et al., 2009). The process of preparing the agar culture plates includes a sterilization step of 20 min at 121°C of the medium prior to plating. This high-temperature step denaturalizes the proteins present in the medium, being responsible for the appearance of this peak. In contrast, bacteria are seeded and grown at physiological conditions, so proteins and nucleotides associated with them and present on the surface of the plate after cell culturing are not denaturalized. This aspect was confirmed by the presence of absorption peaks between 260 and 280 nm (peak 2), which are associated with non-denaturalized proteins and nucleotides.

Spectral differences above 300 nm may be attributed to variations in the biochemical composition of bacteria, concretely to the presence of natural pigments, like carotenoids, flavins, phenazines, quinones, monascines, violaceins, indigoidines or melanins, able to absorb in the UV-VIS range (Celedón and Díaz, 2021). Pseudomonas species, for example, contain pyocyanin, pyochelin, pyoverdine, and pyomelanine (Srivastava et al., 2022; Hoegy et al., 2014; Hamad et al., 2020; Mould et al., 2020), Staphylococcus produce staphyloxantin and zeaxanthin (Pelz et al., 2005; Ram et al., 2020), and Shewallena, characterized to contain high concentration of a wide variety of c-cytochromes as part of its electron transport chain (Kaila and Wikström, 2021; Gennis, 1987), also presents pyomelanine as electron acceptor in the iron reduction process (Kuttan et al., 2023) (Table 2).

The multivariate data analysis used to identify patterns and relationships between the spectra extracted from the samples showed that 92% of the total variance can be explained by using three PCs. PC1 has a constant positive correlation along the spectra without significant variations. On the other hand, PC2 and PC3 are responsible for the sample grouping, as evidenced in the 2D score plot (Supplementary Figure S1).

The model for bacterial colony identification by discriminant analysis generated based on the principal components showed an overall accuracy of 90.26%. Moreover, sensitivity and selectivity values for each bacterial strain and agar type were high and confirmed by the validation test made with the colonies used for prediction. These results suggest UV-hyperspectral can effectively discriminate between bacterial colonies based on their unique spectral signatures.

As can be observed in Table 3, previous approaches using hyperspectral imaging for colony identification in agar plates have been realized in the VIS-NIR spectral range. To the best of our knowledge, this study is the first to use UV-hyperspectral imaging for purposes and reach similar accuracy yields. Moreover, in this research, the agars used were similar, and for generic growth, there were no chromogenic and specific agars as in some of the research. It is well known that these types of agars generate changes in the appearance of some bacterial colonies, affecting their spectral properties and making their identification easier. However, they are more expensive than generic growth agars. The model developed in this research was not only able to detect colony growth in both culture media but also was able to distinguish between them, indicating that, if necessary, prediction models could be developed depending on the agar used for the growth.

These results demonstrate that UV hyperspectral imaging has great potential in bacteria identification, offering a reliable alternative to the methods based on human interpretation. However, future approaches should be centered on increasing the number of bacterial species used to build a prediction model and try to cover the wide variety of bacterial species that can be found at the different scenes where plate culturing is used for identification purposes (i.e., environmental, clinical, food and water analysis). Yeast and molds of interest could also be included, as well as different agar types. Additionally, recent computer science and programming advances could allow for the automation of all the identification and quantification processes (Yoon et al., 2015). This automation could enhance efficiency and ensure the scalability and reproducibility of results.