BR Biomedical Reports 2049-9434 2049-9442 D.A. Spandidos BR-25-1-02154 10.3892/br.2026.2154 Articles Establishing a competence model in discriminating Mycobacterium species and Mycobacterium abscessus subspecies by using surface-enhanced Raman spectroscopy PichitphornApisith 1 FaksriKiatichai 1 2 KaewseekhaoBenjawan 1 PotisapChotima 1 Duong ThaiTran 1 EisiriJukgarin 1 JuntagranChadatan 1 NamwatWises 1 2 KhwansirikulPriyakorn 1 SalaoKanin 3 ChaisriSuwit 4 5 PhoksawatWisitsak 1 2 Research and Diagnostic Center for Emerging Infectious Diseases, Khon Kaen University, Khon Kaen 40002, Thailand Department of Microbiology, Faculty of Medicine, Khon Kaen University, Khon Kaen 40002, Thailand Department of Physiology, Yong Loo Lin School of Medicine, National University of Singapore, Singapore 117597, Singapore Chulabhorn International College of Medicine (CICM), Thammasat University, Pathum Thani 12120, Thailand Thammasat University Research Unit in Biomedical Science, Thammasat University, Pathum Thani 12121, Thailand Correspondence to: Dr Wisitsak Phoksawat, Department of Microbiology, Faculty of Medicine, Khon Kaen University, 123 Mittraphap Road, Muang Khon Kaen, Khon Kaen 40002, Thailand wisiph@kku.ac.th 072026 05052026 25 1 81 16 01 2026 09 04 2026 Copyright: © 2026 Pichitphorn et al. 2026 This is an open access article distributed under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.

Surface-enhanced Raman spectroscopy (SERS) is a powerful laser-based technique with high sensitivity and rapid detection. The present study aimed to establish a model for distinguishing Mycobacterium abscessus (nontuberculous mycobacteria; NTM) from M. tuberculosis H37Ra (MTB-H37Ra; a nonvirulent strain) based on genomic DNA detection, as well as distinguishing M. abscessus subspecies abscessus (MAB) from M. abscessus subspecies massiliense (MMAS). OnSpec-Prime SERS chips and a portable Raman spectrometer device were used. Machine learning approaches, including linear discriminant analysis (LDA), random forest, extreme gradient boosting (XGB), and logistic regression (LR), as well as the receiver operating characteristic (ROC) curve and area under the curve (AUC) were analyzed. A competence model to distinguish NTM from MTB-H37Ra was established; MAB and MMAS were potentially differentiated. At 15 and 50 ng/µl of genomic DNA, respectively, LR demonstrated 99.74 and 99.73% accuracy in differentiating NTM from MTB-H37Ra; XGB displayed 96.25 and 92.97% accuracy in differentiating between MAB and MMAS. LDA revealed clear clustering in each group. The ROC curves showed strong performance of the XGB model across various DNA concentrations. All models achieved an excellent to perfect AUC of 0.96 to 1.00. The present study established a competence model using SERS which may represent a rapid and high-accuracy detection approach, especially in M. abscessus subspecies-level discrimination in clinical specimens.

 surface-enhanced Raman spectroscopy Mycobacterium subspecies genomic DNA machine learning Funding: The present study was supported by a research grant for supporting lecturer to admit high potential student to study and research on his expert program year 2022 (grant no. 651JH114), joint funding of Graduate school and Faculty of Medicine, Khon Kaen University and the Invitation Research Funding (grant no. IN 67080), Faculty of Medicine, Khon Kaen University. Khon Kaen, Thailand. Introduction

Rapid infectious disease identification is critical in clinical decision making, enhancing epidemiological surveillance and outbreak control and reducing healthcare costs. The vibrational spectroscopic method, a recent technique which facilitates the identification of pathogens by their unique whole organism fingerprint, meets the requirements for speed and accuracy (1). Raman spectrometry (RS) is gaining popularity in numerous fields, particularly medical and clinical research (2). RS identifies pathogens through the inelastic scattering of photons from biological molecules following a primary radiation strike (3). A powerful technique, surface-enhanced Raman spectroscopy (SERS) can enhance the ability to identify biological molecules up to 1015 times that of traditional RS. SERS can detect very low concentrations of biological molecules such as proteins, DNA and metabolites) using minimal sample volumes (~1-2 µl) within 10 min; sample processing and analysis can be completed within 3 h. Therefore, SERS offers rapid detection, single-molecule detection with clearer signals, lower fluorescence backgrounds, and greater accuracy in complex matrices than traditional RS (4,5). It is also simpler, more time efficient, cost-effective and sensitive than other techniques (2), holding significant potential for development as a powerful tool for point-of-care testing (POCT).

Mycobacterium species include M. tuberculosis, M. leprae and nontuberculous mycobacteria (NTM), causative agents of various human diseases. NTM are responsible for pulmonary, cutaneous, and lymphatic infections as well as systemic disseminated disease, markedly affecting global morbidity and mortality rates (6,7). Globally, the incidence and prevalence of NTM diseases are on the rise (8,9). M. abcsessus is a significant, rapidly growing species comprising three primary subspecies: M. abscessus subsp. abscessus (MAB), M. abscessus subsp. massiliense (MMAS), and M. abscessus subsp. bolletii (10).

Multiple techniques have been developed for the diagnosis of Mycobacterium infections, including microbiological culture, biochemical assays and the polymerase chain reaction, matrix-assisted laser desorption/ionization-time of flight mass spectrometry (MALDI-TOF MS), and DNA sequencing (11-15). MALDI-TOF MS is a modern, reliable, accurate, high-throughput technique used to identify proteins via peptide mass fingerprinting and is gaining popularity for use in bacterial species-level identification (16,17). However, it remains limited in the discrimination of closely related M. abscessus subspecies. Thus, identifying these NTM at the subspecies level requires the implementation of novel approaches for simple, rapid and highly accurate discrimination (18).

The present study aimed to establish a SERS competence model for distinguishing Mycobacterium species and M. abscessus subspecies by detecting genomic DNA and using machine learning (ML) algorithms. SERS holds potential as a simple, rapid, accurate, and cost-effective method in routine microbiology testing based on genomic DNA analysis.

 Materials and methods <sec> <title>Study design and workflow

The present study comprised two phases: Bacterial cultivation and data acquisition (Phase I) and data processing and computational analyses (Phase II; Fig. 1). In Phase I, collected samples of all bacteria maintained in the biobank stock were re-cultured. Previously, isolated M. abscessus was identified by whole-genome sequencing (WGS) (19). In the present study, the bacterial DNA were extracted, dropped on SERS substrates and air-dried. The DNA signals were measured using a portable Raman spectrometer (National Electronics and Computer Technology Center). In Phase II, the data were processed using the Python programming languages with the leave-one-out cross-validation (LOOCV) function to distinguish the NTM subspecies (MAB and MMAS) and MTB-H37Ra. The Raman mapping approach (20) was used to increase sensitivity. In the SERS spectra validation, spectra were evaluated based on the peak assignments. The ML approaches, including principal component analysis (PCA), uniform manifold approximation and projection (UMAP), linear discriminant analysis (LDA), random forest (RF), extreme gradient boosting (XGB), and logistic regression (LR), were analyzed. The performance of the constructed classification models was evaluated using the receiver operating characteristic (ROC) curve and area under the curve (AUC) to quantify discriminative ability.

 Bacterial isolation and cultivation

The bacteria used in the present study included NTM (M. abscessus subspecies: MAB and MMAS) and MTB-H37Ra (ATCC 25177; nonvirulent strain). The NTM were isolated from patients between 2012 and 2017 and retrieved from archived stock cultures at the Clinical Microbiology Laboratory Unit, Srinagarind Hospital, Khon Kaen University Khon Kaen, Thailand. MTB-H37Ra was kindly given by Professor Prasit Palittapongarnpim, Department of Microbiology, Faculty of Science, Mahidol University, Thailand (21). In the present study, the bacteria were re-subcultured on sheep blood agar (Quality Media and Supply Co. Ltd.) at 37˚C for 5 days. Middlebrook 7H9 broth liquid medium (Becton, Dickinson and Company) was used for MTB-H37Ra recultivation. The present study was approved and all protocols regulated by the Khon Kaen University Ethics Committee for Human Research (approval no. HE 591454).

In our previous study, genomic DNA samples were sequenced using the Illumina high-throughput platform at NovogeneAIT Genomics (Novogene Co., Ltd.). Based on this WGS analysis, NTM were divided into MAB and MMAS, corresponding to subspecies of M. abscessus identified by rpoB gene sequences (13). The accession numbers were shown as Table SI (19,22), however, one isolate of MMAS was identified by MALDI-TOF.

 Genomic DNA extraction

Bacterial genomic DNA was extracted from three full loops of pure colonies cultured on blood agar (Clinag Co. Ltd.). The bacterial pellet was washed twice with 1 ml Tris-EDTA buffer, then centrifuged at 13,416 x g and room temperature for 10 min. All genomic DNA was extracted using the cetyltrimethylammonium bromide (CTAB) method (19,23). Bacterial cell lysis began with incubation in 10 mg/ml lysozyme, followed by the sequential addition of 10% sodium dodecyl sulfate, 10 mg/ml proteinase K and CTAB solutions to disrupt the cell membranes. Cold absolute ethanol and 5 M sodium chloride were added to precipitate the DNA. Genomic DNA samples were resuspended in sterile nuclease-free water (Invitrogen; Thermo Fisher Scientific, Inc.) and measured using Nanodrop (Thermo Fisher Scientific, Inc.). A range of DNA concentrations (5, 10, 15, 20, 25 and 50 ng/µl) was prepared and stored at -20˚C until SERS analysis.

 Raman spectroscopy acquisition

The SERS measurements were acquired using a portable RS device, with 2 µl of 5, 10, 15, 20, 25 and 50 ng/µl DNA samples individually dropped onto the center of a silver-based SERS sensor and left to air dry for 10 min. The Raman spectrum was collected from the central region of the drop. Spectral acquisition was performed using a wavelength of 785 nm with an exposure time of 20 sec. Spectral data were collected from 25 points arranged in a 5x5 mapping grid with 0.1 mm spacing. Each sample was measured in triplicate. The optimal protocol in a previous study was used in this experiment (24). The Raman spectrum fingerprint region (600-1,800 cm-¹) represents the biological components of bacterial cells, including proteins, lipids, and nucleic acids (25). The key peak locations, primarily attributed to nucleic acid components, are summarized in Table SII. The observed Raman peaks correspond with those reported in previous studies (26-37).

 Raman spectral data processing

The Raman spectra from triplicate mapping (75 data points) of each isolate were collected. A total of 825 datasets were initially preprocessed by removing cosmic rays and outliers, followed by signal averaging for each Raman-mapped dataset. The analysis was performed by supervised ML using the Python programming language based on the LOOCV approach. The data included three classes (MAB, MMAS and MTB-H37Ra). The ML models were trained on PCA, UMAP, LDA, RF, XGB and LR algorithms from the Scikit-learn library (38). In this procedure, one Raman-mapped data point from each isolate was used as the test set; the remaining 824 data points from other isolated samples constituted the training set. These processes were repeated until all datasets were tested, and the average accuracy was recorded.

The performance of the classification models (LDA, RF, XGB and LR) was evaluated using ROC analysis. The AUC was used to quantify discriminative ability, calculated for each model to assess the trade-off between sensitivity (true positive rate) and specificity (1-false positive rate) across various classification thresholds. The ROC curves and AUC values were used to compare the overall classification efficacy of the ML approaches.

 Statistical analysis

Multiple comparisons for non-parametric independent samples were analyzed using the Kruskal-Wallis test. Adjusted P-values using Dunn-Bonferroni post-hoc analysis were later used to identify pairwise differences between groups. Pearson's and Spearman's correlation analyses were performed to assess the reproducibility of SERS spectra. The coefficient values converging toward 1.00 indicate high reproducibility and excellent consistency. Sensitivity, specificity, and accuracy were recorded using a confusion matrix in the Python programming language. The formulas are as follows: sensitivity=true positive/(true positive + false negative); specificity=true negative/(true negative + false positive); accuracy=average of sensitivity and specificity. P<0.05 was considered to indicate a statistically significant difference.

 Results <sec> <title>Minimal concentration of bacterial genomic DNA for SERS detection

The minimal detectable concentration for genomic DNA was determined using SERS. For bacterial genomic DNA samples, Raman maps were acquired in triplicate. Within each map, 25 individual Raman spectra were collected, resulting in a total of 75 spectra (25 spectra/map x3 maps) per sample. Raman mapping can obtain all signals from each specified point. To ensure signal responsiveness and usability prior to sample analysis, each SERS substrate was thoroughly tested; it was confirmed that the SERS chip exhibited specific background peaks at 1,003 and 1,600 cm-¹ (Fig. 2A). M. abscessus genomic DNA was then dropped at a concentration of 50 ng/µl onto the SERS chip. Prominent peaks were observed in the 600-1,800 cm-¹ range at 726, 755, 783, 927, 983, 1,053, 1,098, 1,165, 1,187, 1,250, 1,330, 1,486 and 1,578 cm-¹. Genomic DNA samples at concentrations of 5, 10, 15, 20 and 25 ng/µl were also tested. Clear peak visibility was maintained at 15-50 ng/µl (Fig. 2B). At lower concentrations, some peaks disappeared and interference from other peaks were more noticeable. Therefore, 15 ng/µl may be considered the minimum detectable concentration with clear DNA features, with 50 ng/µl demonstrating the most reliable spectral analysis.

 Raman spectra of genomic DNA

Raman spectra of genomic DNA at 15 and 50 ng/µl were obtained using SERS at 785 nm (Fig. 3). The characteristic bands of the two M. abscessus subspecies (MAB and MMAS) and MTB-H37Ra were analyzed. Peaks 726, 755, 781 and 1,317 cm-¹, corresponding to nucleic acids, were the most intense in MAB and MMAS; MTB-H37Ra exhibited the strongest peaks at 755, 781, and 1,317 cm-¹. The 1,000-1,250 cm-¹ region corresponds to the asymmetric stretching mode of the phosphate deoxy group, with the peak at 1,097 cm-¹ the most prominent for Mycobacterium. The 1,331 cm-¹ band corresponds to vibrations of bases coupled with sugar vibrations, specifically adenine and guanine bases. Vibrational bands in the 1,450-1,800 cm-¹ range, corresponding to double bond stretching of residual bases, were observed at 1,486 and 1,577 cm-¹ in MAB and MMAS; MTB-H37Ra showed a weaker band at 1,486 cm-¹.

 SERS spectral analysis compared with Mycobacterium species and subspecies

Raman spectra exhibiting a relative standard deviation <50% were selected as reliable peak markers (Table SII). Notably, all three bacteria exhibited abundant peak assignments primarily related to genomic DNA. SERS performance was subsequently evaluated using eight candidate peak markers at 15 and 50 ng/µl DNA concentrations (Figs. 4 and 5). These peak markers revealed significant spectral differences, indicating variations in biomolecule content and abundance across species (NTM: MAB/MMAS vs. MTB-H37Ra) and even subspecies (MAB and MMAS). These distinct biomolecular profiles allow for the identification of bacterial features using useful factors. The eight candidate peaks were selected and evaluated by the significance of their P-values, highlighting their discriminatory power.

 Discriminating Mycobacterium subspecies using ML approaches

SERS spectral data were analyzed using PCA (Fig. 6A), UMAP (Fig. 6B), and LDA (Fig. 6C). UMAP and PCA could not clearly distinguish the bacterial groups. However, LDA effectively separated the three bacteria into distinct clusters, demonstrating that SERS spectral data contain sufficient information for reliable differentiation when combined with appropriate dimensionality reduction techniques.

To further explore the utility of this approach, ML was employed for classification. This involved the LOOCV strategy, with four models evaluated: LDA, XGB, RF, and LR (Table I). The XGB model demonstrated the highest performance in NTM subgroup classification (MAB vs. MMAS), achieving accuracies of 96.25 and 92.97% (F1-scores 96.32 and 92.94%) at 15 and 50 ng/µl DNA, respectively. The LR model proved most effective at distinguishing between the NTM and MTB-H37Ra groups, delivering near-perfect discrimination with accuracies of 99.74 and 99.73% (F1-scores 99.86 and 99.85%) at both concentrations. ROC curves were generated to evaluate the performance of each model and determine sensitivity and specificity, and the best AUC values of each model were calculated (Fig. 7). Notably, the sum of the sensitivity (true positive rate) and specificity (1-false positive rate), as represented by the calculated AUC values, was consistently close to 1, highlighting the effectiveness of the models at distinguishing the different bacterial classes. The ROC curves confirmed these findings, with the XGB model showing strong performance in the NTM group (MAB and MMAS; Fig. 7A) and the LR model achieving a perfect AUC of 1 for distinguishing the NTM from the MTB-H37Ra group (Fig. 7B).

 Discussion

RS and SERS technologies offer several key advantages and hold potential utility in clinical diagnostics. These techniques enable high-throughput data generation, have rapid turnaround times and are cost-effective, rendering them efficient and accessible. Recently, RS has shown significant promise in the diagnosis of cancer (39,40), dengue fever (41), diabetes (1) and tuberculous meningitis (42). RS and SERS have also been applied in TB diagnosis (2). While MALDI-TOF MS is the main identification method for most NTM (43), M. abscessus and M. tuberculosis differentiation remains a challenge and M. abscessus identification at the subspecies level is limited. Some studies have attempted subspecies identification using protein peak analysis (44,45); however, this is not the best strategy to follow and requires validation with larger collections of clinical isolates to confirm utility in a microbiology laboratory setting.

SERS could be considered an improvement on MALDI-TOF MS with its ability to distinguish molecular biomolecules at subspecies resolution, higher sensitivity, lower sample volume requirement, lack of need for extensive sample pretreatment and special matrix reagents and shorter turnaround time (46). Compared with WGS (the gold standard for the M. abscessus complex), SERS also requires lower sample volumes and less preparation prior to analysis, has shorter processing times, and is considerably more cost-effective (47). Regarding long-term storage potential, nucleic acids (genomic SERS-based or WGS) exhibit far greater stability and longevity than proteins, which typically require immediate analysis following sample collection.

There is a need for novel approaches offering simple, rapid, cost-effective, and highly accurate discrimination of these bacteria at the species and subspecies levels. The present study was to establish a competence SERS model for differentiating genomic DNA of Mycobacterium species (NTM and MTB-H37Ra) and subspecies (MAB and MMAS), with the goal of developing a powerful tool for clinical diagnosis. However, the evaluations did not include clinical validation or testing on external datasets, which are critical points and should be addressed in future studies.

First, the minimal genomic DNA concentration required for the SERS technique was determined. DNA was detected for all three bacteria at concentrations of 5-50 ng/µl, indicating that these concentrations allow for the observation of characteristic signals in the Raman spectral range of 600-1,800 cm-¹ (25). SERS analysis was performed at both 15 and 50 ng/µl DNA concentrations; while higher concentrations yielded clearer peaks, peak detection was possible at the lower concentration, confirming high sensitivity. This aligns with previous findings of a strong SERS signal for DNA at ≥1 nM, with decreasing signal intensity at lower concentrations (48). These results demonstrate that SERS is effective even with limited sample amounts and that selecting appropriate DNA concentrations is crucial to avoid background interference and ensure accurate analysis.

The SERS spectra of genomic DNA from Mycobacterium species revealed several distinct characteristic bands, providing a foundation for bacterial differentiation. This analysis confirmed key spectral features previously reported for nucleic acids: the prominent bands at 726 cm-¹ (adenine) and 781 cm-¹ (cytosine) served as reliable markers for the nucleotide bases (26), while the 1,097 cm-¹ peak, corresponding to the phosphodioxy (PO2-) backbone (49), confirmed the presence of DNA. Distinct purine and pyrimidine markers were also observed, including bands for guanine (1,317 cm-¹) and adenine (1,331 cm-¹), which fall within the ‘DNA fingerprint’ region (26). The 1,486 cm-¹ peak indicates the hydrogen bonding state at guanine’s N7 site (50) and the 1,577 cm-¹ band is a key purine marker (27). The analysis also revealed a strong band at 755 cm-¹ (the vibrations from supercoiled DNA structures), a finding consistent with previous research (51). This suggested that SERS may be sensitive to secondary and tertiary structures. This structural information, combined with the complete SERS spectral fingerprint, could serve as a unique and powerful signature for distinguishing bacterial species.

Despite the fact that Mycobacterium species such as MAB/MMAS and MTB-H37Ra belong to the same genus (with the close genetic relationship between MAB and MMAS accounting for the broad similarity of their SERS Raman peaks), their distinct genetic makeup results in clear differences in their SERS spectra. This analysis revealed a notable distinction between MAB, MMAS and the MTB-H37Ra strain. Specifically, the peaks at 1,486 and 1,577 cm-¹ were prominent in MAB and MMAS but notably weaker in MTB-H37Ra and a strong peak at 928 cm-¹ was detected exclusively in the MTB-H37Ra spectrum. These molecular-level differences provide a unique ‘spectral fingerprint’, suggesting that the genetic distance between these organisms directly influences their SERS spectra and enables the discrimination of even closely related species. This capacity for differentiation based on subtle genetic and biochemical variations highlights the potential utility of SERS in bacterial identification and classification.

The spectra analyses revealed significant differences in eight key Raman peaks across the MAB, MMAS, and MTB-H37Ra genomic DNA samples. The boxplot analysis provided critical insight into the effect of DNA concentration and genetic relationship on SERS discrimination. At the low concentration of 15 ng/µl, the SERS spectra of the genetically similar MAB and MMAS were largely indistinguishable, but both were distinct from the more genetically distant MTB-H37Ra strain. This indicated that SERS effectively differentiates between species but has limited resolution for very closely related subspecies at lower concentrations. At the higher DNA concentration of 50 ng/µl, the spectral differences of all three bacteria were more pronounced, suggesting that a higher DNA concentration enhances the ability to detect subtle variations. Consequently, increasing the DNA concentration may improve the discriminatory power of SERS analysis, positioning it as a precise method of distinguishing even very closely related bacterial species.

LDA revealed clear clustering among MAB, MMAS and MTB-H37Ra, highlighting the potential of combining SERS with ML for bacterial differentiation. The XGB model effectively distinguished between the NTM (MAB vs. MMAS); the LR model was effective at distinguishing the NTM group from MTB-H37Ra, achieving the highest accuracy of 99.74%. ROC analysis confirmed the accuracy of the models. These findings align with a previous study reporting that ML methods achieved an average identification accuracy of 90.73% for 12 common pathogenic bacteria and a 99.92% accuracy in distinguishing between antibiotic-sensitive and -resistant strains of Acinetobacter baumannii (52).

Compared with MALDI-TOF MS, SERS represents a high-efficiency technique for identifying Mycobacterium species and subspecies. SERS uses purified genomic DNA, reducing the risk of sample contamination. Conversely, background noise from proteins in the culture media can present a challenge with MALDI-TOF MS. In addition, SERS utilizes a machine small enough to be developed into a portable POCT device, allowing for faster and more cost-effective processing; MALDI-TOF MS requires a larger processing area and is more expensive. SERS may therefore be preferable in clinics or areas with limited resources, making it a more practical tool in treatment settings.

In addition to SERS-based detection in mycobacterial studies, Perumal et al (53) and our work both emphasize the development of SERS-based models for Mycobacterium detection. Perumal et al (53) established successfully SERS spectra to characterize three major mycolic acid (MA) forms (α-MA, methoxy-MA and keto-MA) of MTB combining PCA and functional PCA for dimensionality reduction and LR, as well as LDA. By contrast, our study used genomic DNA to discriminate species (NTM vs. MTB-H37Ra) and NTM subspecies (MAB vs. MMAS). Here, we employed both supervised methods (LDA) and unsupervised methods (PCA and UMAP) for spectral clustering, alongside advanced ML algorithms (XGB, LR and RF) with LOOCV, which is likely to yield more rigorous insights and higher resolution. Collectively, these complementary approaches advance SERS-based technology toward comprehensive mycobacterial diagnostics.

Although these findings are promising for bacterial differentiation, it is crucial to address the limitations of the present study. The portable Raman spectrometer, while excellent for POCT, simpler, faster and more compact, has a lower resolution and sensitivity than benchtop systems, restricting its use to preliminary screening; parallel validation with conventional Raman is recommended. However, prior studies have reported that portable RS devices are sufficiently effective for medical and environmental biomolecule measurement (54-56). There is also the issue of the long-term stability of SERS chips as prolonged storage may alter surface and plasmonic activity, affecting signal reproducibility. In the present study, all SERS chips were vacuum-sealed during storage and used within 30 days of opening per the manufacturer protocol in a single experimental session.

Finally, the limited biological sample size, although partially mitigated by performing technical triplicates (75 spectra per sample), constrains the generalizability and predictive accuracy of the developed model. The current dataset, while providing initial insights, risks bias and overfitting due to the small, imbalanced distribution of isolates (such as 10 NTM vs. a single strain of MTB-H37Ra). However, reproducibility assessments of SERS spectra, and Pearson's and Spearman's correlation analysis were displayed as Fig. S1 and Table SIII. Expanding the sample size may improve the model and enhance the predictive power for reliable differentiation.

In conclusion, the present study represented a proof-of-concept study: Novel competence SERS models were successfully established using technology to differentiate Mycobacterium at the species and subspecies levels. The present study used a portable RS device which could be developed as a promising POCT in the future. SERS analysis combined with ML was effectively employed to characterize MAB and MMAS genomic DNA, establishing a valuable Raman spectral database for M. abscessus subspecies classification, as well as differentiating these from MTB-H37Ra. Detection was based on genomic DNA, a stable sample source for long-term storage, and was possible at very low concentrations (15 ng/µl). ML models, particularly XGB, achieved high classification accuracies of 96.25% (94.71% sensitivity, 97.91% specificity, and an AUC of 99%) for distinguishing between MAB and MMAS, with LR exceeding 99.74% accuracy (99.86% sensitivity, 98.46% specificity, and an AUC of 100%) for differentiating NTM from MTB-H37Ra. These findings demonstrated the potential utility of SERS in genomic-level infectious disease diagnosis.

 Supplementary Material Reproducibility of SERS spectra across different genomic DNA concentrations. The dot plots illustrate the coefficient of determination (R<sup>2</sup>) values obtained from genomic DNA at concentrations of (A) 15 ng/<italic>μ</italic>l and (B) 50 ng/<italic>μ</italic>l. The y-axis indicates the R<sup>2</sup> values, reflecting the degree of the correlation and consistency between replicates. The x-axis denotes the specific bacterial isolate tested: MAB, MMAS, and MTB-H37Ra. Numbers (1, 2 and 3) under each isolate represent independent technical replicates. The high R<sup>2</sup> values (mostly >0.8) observed across both concentrations and all strains demonstrate the high precision and reliability of the method. SERS, surface-enhanced Raman spectroscopy; MAB, <italic>M</italic>. <italic>abscessus</italic> subspecies <italic>abscessus</italic>; MMAS, <italic>M. abscessus</italic> subspecies <italic>massiliense</italic>; MTB-H37Ra/H37Ra, <italic>M. tuberculosis</italic> H37Ra. Accession numbers of <italic>Mycobacterium</italic> species and subspecies used. Specific DNA-related Raman shifts and their corresponding peak assignments. Reproducibility assessment of SERS spectra via Pearson’s and Spearman’s correlation analysis. Acknowledgements

Not applicable.

 Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

 Authors' contributions

AP was responsible for conceptualization, review, sample administration, data curation, data analysis and writing the original draft. TDT, PK and CJ were responsible for sample administration and data curation. JE was responsible for data analysis and software consultancy. AP, TDT and JE confirm the authenticity of all the raw data. BK, CP, WN, KS and SC were responsible for conceptualization, review and editing. KF was responsible for conceptualization, funding, review and editing. WP was responsible for conceptualization, funding, data analysis, editing, finalized manuscript and corresponding. All authors read and approved the final manuscript.

 Ethics approval and consent to participate

The present study was approved by the Khon Kaen University Ethics Committee for Human Research in accordance with the 1964 Declaration of Helsinki (approval no. HE 591454). Leftover clinical specimens were provided for the completion of the present study.

 Patient consent for publication

Not applicable.

 Competing interests

The authors declare that they have no competing interests.

 References GuevaraETorres-GalvánJCRamírez-ElíasMGLuevano-ContrerasCGonzálezFJUse of Raman spectroscopy to screen diabetes mellitus with machine learning toolsBiomed Opt Express94998501020183031991710.1364/BOE.9.004998 KhanSUllahRShahzadSAnbreenNBilalMKhanAAnalysis of tuberculosis disease through Raman spectroscopy and machine learningPhotodiagnosis Photodyn Ther2428629120183035975710.1016/j.pdpdt.2018.10.014 RavanshadRKarimi ZadehAAmaniAMMousaviSMHashemiSASavar DashtakiAMirzaeiEZareBApplication of nanoparticles in cancer detection by Raman scattering based techniquesNano Rev Exp9137355120173041071010.1080/20022727.2017.1373551 ZhangCWinnardPT JrDasariSKominskySLDoucetMJayaramanSRamanVBarmanILabel-free Raman spectroscopy provides early determination and precise localization of breast cancer-colonized bone alterationsChem Sci974375320172962914410.1039/c7sc02905e LiZLiZChenQZhangJDunhamMEMcWhorterAJFengJMLiYYaoSXuJMachine-learning-assisted spontaneous Raman spectroscopy classification and feature extraction for the diagnosis of human laryngeal cancerComput Biol Med14610561720223560548610.1016/j.compbiomed.2022.105617 WuUIHollandSMHost susceptibility to non-tuberculous mycobacterial infectionsLancet Infect Dis1596898020152604996710.1016/S1473-3099(15)00089-4 CassidyPMHedbergKSaulsonAMcNellyEWinthropKLNontuberculous mycobacterial disease prevalence and risk factors: A changing epidemiologyClin Infect Dis49e124e12920091991194210.1086/648443 IdeSNakamuraSYamamotoYKohnoYFukudaYIkedaHSasakiEYanagiharaKHigashiyamaYHashiguchiKEpidemiology and clinical features of pulmonary nontuberculous mycobacteriosis in Nagasaki, JapanPLoS One10e012830420152602094810.1371/journal.pone.0128304 ShahSKMcAnallyKJSeoaneLLombardGALaPlaceSGLickSDhillonGSValentineVGAnalysis of pulmonary non-tuberculous mycobacterial infections after lung transplantationTranspl Infect Dis1858559120162736898910.1111/tid.12546 LeeMRYangCYShuCCLinCKWenYFLeeSWKoJCWangJYLeeLNYuCJFactors associated with subsequent nontuberculous mycobacterial lung disease in patients with a single sputum isolate on initial examinationClin Microbiol Infect21250.e1e720152565854510.1016/j.cmi.2014.08.025 JeongJKimSRChangCLLeeSHIdentification of mycobacteria species by HPLC and species distribution during five years at Ulsan university hospitalKorean J Lab Med28243320081830925210.3343/kjlm.2008.28.1.24(In Korean) KehrmannJKurtNRuegerKBangeFCBuerJGenoType NTM-DR for identifying Mycobacterium abscessus subspecies and determining molecular resistanceJ Clin Microbiol541653165520162703048710.1128/JCM.00147-16 Kham-NgamIChetchotisakdPAnantaPChaimaneePReechaipichitkulWLulitanondVNamwatWFaksriKDifferentiation between persistent infection/colonization and re-infection/re-colonization of Mycobacterium abscessus isolated from patients in Northeast ThailandInfect Genet Evol68354220193052186510.1016/j.meegid.2018.12.001 KwonYSKohWJDiagnosis of pulmonary tuberculosis and nontuberculous mycobacterial lung disease in KoreaTuberc Respir Dis771520142511469610.4046/trd.2014.77.1.1 PfyfferGEMycobacterium: General characteristics, laboratory detection, and staining procedures. In Manual of Clinical Microbiology. Jorgensen JH, Carroll KC, Funke G, Pfaller MA, Landry ML, Richter SS, Warnock DW, Carroll KC, Bernard KA, Dumler JS, et al (eds). John Wiley & Sons, Ltd., pp536-569, 2015. DingleTCButler-WuSMMaldi-tof mass spectrometry for microorganism identificationClin Lab Med3358960920132393184010.1016/j.cll.2013.03.001 NithimongkolchaiNHinwanYKaewseekhaoBChareonsudjaiPReungsangPKraiklangRChareonsudjaiSWonglakornLChetchotisakdPSirichoatAMALDI-TOF MS analysis of Burkholderia pseudomallei and closely related species isolated from soils and water in Khon Kaen, ThailandInfect Genet Evol11610553220233799588510.1016/j.meegid.2023.105532 GriffithDEAksamitTBrown-ElliottBACatanzaroADaleyCGordinFHollandSMHorsburghRHuittGIademarcoMFAn official ATS/IDSA statement: diagnosis, treatment, and prevention of nontuberculous mycobacterial diseasesAm J Respir Crit Care Med17536741620071727729010.1164/rccm.200604-571ST KaewprasertONonghanphithakDChetchotisakdPNamwatWOngRTHFaksriKWhole-genome sequencing and drug-susceptibility analysis of serial Mycobacterium abscessus isolates from Thai patientsBiology (Basel)11131920223613879810.3390/biology11091319 EspositoABonifacioASergoVFornasaroSLabel-free surface enhanced Raman scattering (SERS) on centrifugal silver plasmonic paper (CSPP): A novel methodology for unprocessed biofluids sampling and analysisBiosensors (Basel)1146720213482168310.3390/bios11110467 ChangsenCFranzblauSGPalittapongarnpimPImproved green fluorescent protein reporter gene-based microplate screening for antituberculosis compounds by utilizing an acetamidase promoterAntimicrob Agents Chemother473682368720031463846510.1128/AAC.47.12.3682-3687.2003 ZhengHLuLWangBPuSZhangXZhuGShiWZhangLWangHWangSGenetic basis of virulence attenuation revealed by comparative genomic analysis of Mycobacterium tuberculosis strain H37Ra versus H37RvPLoS One3e237520081858405410.1371/journal.pone.0002375 De AlmeidaINDa Silva CarvalhoWRossettiMLCostaERDe MirandaSSEvaluation of six different DNA extraction methods for detection of Mycobacterium tuberculosis by means of PCR-IS6110: Preliminary studyBMC Res Notes656120132437346110.1186/1756-0500-6-561 BottaRChindaudomPEiamchaiPHorprathumMLimwicheanSChananonnawathornCPatthanasettakulVKaewseekhaoBFaksriKNuntawongNTuberculosis determination using SERS and chemometric methodsTuberculosis (Edinb)10819520020182952332310.1016/j.tube.2017.12.008 Alunni CardinaliMCasagrande PierantoniDCaponiSCorteLFiorettoDCardinaliGMeso-Raman approach for rapid yeast cells identificationBiophys Chem25410624920193145461210.1016/j.bpc.2019.106249 BoikoVVRomanyukVRGnatyukOPIlchenkoOOKarakhimSOKorovinAVDovbeshkoGIVibrational spectra of DNA in the confined interglobular volume of photonic crystalJ Biol Phys4410111620182946443410.1007/s10867-018-9480-0 PrescottBSteinmetzWThomasGJ JrCharacterization of DNA structures by laser Raman spectroscopyBiopolymers232352561984670448710.1002/bip.360230206 SerbanDBenevidesJMThomasGJ JrDNA secondary structure and Raman markers of supercoiling in Escherichia coli plasmid pUC19Biochemistry4184785320021179010610.1021/bi011004z NeugebauerUClementJHBocklitzTKrafftCPoppJIdentification and differentiation of single cells from peripheral blood by Raman spectroscopic imagingJ Biophotonics357958720102044983110.1002/jbio.201000020 MatthewsQJirasekALumJDuanXBroloAGVariability in Raman spectra of single human tumor cells cultured in vitro: Correlation with cell cycle and culture confluencyAppl Spectrosc6487188720102071905010.1366/000370210792080966 KimJParkHJKimJHChangBParkHKLabel-free detection for a DNA methylation assay using Raman spectroscopyChin Med J (Engl)1301961196720172877654910.4103/0366-6999.211874 ZanganaSVeresMBonyárASurface-enhanced raman spectroscopy (SERS)-Based sensors for deoxyribonucleic acid (DNA) detectionMolecules29333820243906491510.3390/molecules29143338 MaquelinKChoo-SmithLPEndtzHPBruiningHAPuppelsGJRapid identification of Candida species by confocal Raman microspectroscopyJ Clin Microbiol4059460020021182597610.1128/JCM.40.2.594-600.2002 MovasaghiZRehmanSRehmanIURaman spectroscopy of biological tissuesAppl Spectrosc Rev424935412007 KendallCHutchingsJBarrHShepherdNStoneNExploiting the diagnostic potential of biomolecular fingerprinting with vibrational spectroscopyFaraday Discuss149279290333-35620112141318610.1039/c005379a LiuZDavisCCaiWHeLChenXDaiHCirculation and long-term fate of functionalized, biocompatible single-walled carbon nanotubes in mice probed by Raman spectroscopyProc Natl Acad Sci USA1051410141520081823073710.1073/pnas.0707654105 HuangHShiHFengSChenWYuYLinDChenRConfocal Raman spectroscopic analysis of the cytotoxic response to cisplatin in nasopharyngeal carcinoma cellsAnal Methods52602662013 PedregosaFVaroquauxGGramfortAMichelVThirionBGriselOBlondelMPrettenhoferPWeissRDubourgVScikit-learn: Machine learning in pythonJ Mach Learn Res12282528302011 FengXFoxMCReichenbergJSLopesFCPSSebastianKRDunnAKMarkeyMKTunnellJWSuperpixel Raman spectroscopy for rapid skin cancer margin assessmentJ Biophotonics13e20196010920203186787810.1002/jbio.201960109 TraynorDDuraipandianSBhatiaRCuschieriKMartinCMO'LearyJJLyngFMThe potential of biobanked liquid based cytology samples for cervical cancer screening using Raman spectroscopyJ Biophotonics12e20180037720193065381910.1002/jbio.201800377 NaseerKAminASaleemMQaziJRaman spectroscopy based differentiation of typhoid and dengue fever in infected human seraSpectrochim Acta A Mol Biomol Spectrosc20619720120193011457910.1016/j.saa.2018.08.008 SathyavathiRDingariNCBarmanIPrasadPSRPrabhakarSNarayana RaoDDasariRRUndamatlaJRaman spectroscopy provides a powerful, rapid diagnostic tool for the detection of tuberculous meningitis in ex vivo cerebrospinal fluid samplesJ Biophotonics656767220132288777310.1002/jbio.201200110 RodriguezLZhangZWangDRecent advances of Raman spectroscopy for the analysis of bacteriaAnal Sci Adv4819520233871592310.1002/ansa.202200066 FangousMSMougariFGouriouSCalvezERaskineLCambauEPayanCHéry-ArnaudGClassification algorithm for subspecies identification within the Mycobacterium abscessus species, based on matrix-assisted laser desorption ionization-time of flight mass spectrometryJ Clin Microbiol523362336920142500904810.1128/JCM.00788-14 PanageaTPincusDHGrogonoDJonesMBryantJParkhillJFlotoRAGilliganPMycobacterium abscessus complex identification with matrix-assisted laser desorption ionization-time of flight mass spectrometryJ Clin Microbiol532355235820152594860710.1128/JCM.00494-15 TakeiSTeramotoKSekiguchiYIharaHTohyaMIwamotoSTanakaKKhasawnehAHoriuchiYMisawaSIdentification of Mycobacterium abscessus using the peaks of ribosomal protein L29, L30 and hemophore-related protein by MALDI-MS proteotypingSci Rep141118720243875526710.1038/s41598-024-61549-7 WangYZhangZSunYWuHLuoLSongYRecent advances in surface-enhanced raman scattering for pathogenic bacteria detection: A reviewSensors (Basel)25137020254009611710.3390/s25051370 Van LieropDFauldsKGrahamDSeparation free DNA detection using surface enhanced Raman scatteringAnal Chem835817582120112151709510.1021/ac200514e MovileanuLBenevidesJMThomasGJ JrTemperature dependence of the Raman spectrum of DNA. II. Raman signatures of premelting and melting transitions of poly(dA).poly(dT) and comparison with poly(dA-dT).poly(dA-dT)Biopolymers6318119420021178700610.1002/bip.10022 BenevidesJMLiTLuXJSrinivasanAROlsonWKWeissMAThomasGJ JrProtein-directed DNA structure II. Raman spectroscopy of a leucine zipper bZIP complexBiochemistry3954855620001064217910.1021/bi990053x Christens-BarryWAMartinJCLebowitzJRaman spectroscopy of supercoiled and nicked ColE1 plasmidBiopolymers28151515261989267342110.1002/bip.360280904 LuWLiHQiuHWangLFengJFuYVIdentification of pathogens and detection of antibiotic susceptibility at single-cell resolution by Raman spectroscopy combined with machine learningFront Microbiol13107696520233668764110.3389/fmicb.2022.1076965 PerumalJDinishUBendtAKKazakeviciuteAFuCYOngILHOlivoMIdentification of mycolic acid forms using surface-enhanced Raman scattering as a fast detection method for tuberculosisInt J Nanomedicine136029603820183032359010.2147/IJN.S171400 KhongkaewPCruzJBertottoJPCárdenasVAlcalàMNuchtavornNPhechkrajangCA comparative study of benchtop and portable NIR and Raman spectroscopic methods for the quantitative determination of curcuminoids in turmeric powderFoods11218720223589277210.3390/foods11152187 Hernández-JiménezMRevillaIVivar-QuintanaAMGrabskaJBećKBHuckCWPerformance of benchtop and portable spectroscopy equipment for discriminating Iberian ham according to breedCurr Res Food Sci810067520243829234410.1016/j.crfs.2024.100675 BrawleyHNJuárezIDKurouskiDZwartSRSmithSMFeasibility of portable Raman SERS for blood biomarker monitoring in spaceflight conditionsLife Sci Space Res (Amst)4821622420264156540310.1016/j.lssr.2025.11.013

Study workflow (experimental design and analysis). The design of the present study comprised two phases. Phase I (blue background; left hand): Bacterial cultivation and data acquisition; collected bacterial samples maintained in the biobank stock were re-cultured. Bacterial DNA were then extracted, dropped on SERS substrates, and air-dried. DNA signals were measured using a portable Raman spectrometer. Phase II (peach-pink background; right hand): Data processing and computational analyses; data were processed using the Python programming languages using the LOOCV function to distinguish the NTM subspecies (MAB and MMAS) and MTB-H37Ra. For validation, SERS spectra were evaluated based on the peak assignments. The ML approaches (PCA, UMAP, LDA, RF, XGB, and LR) were analyzed. Performance of the constructed classification models was evaluated using ROC curves and the AUC to quantify discriminative ability. LOOCV, leave-one-out cross-validation; NTM, nontuberculous mycobacteria; MAB, M. abscessus subspecies abscessus; MMAS, M. abscessus subspecies massiliense; MTB-H37Ra, M. tuberculosis H37Ra; SERS, surface-enhanced Raman spectroscopy; ML, machine learning; PCA, principal component analysis; UMAP, uniform manifold approximation and projection; LDA, linear discriminant analysis; RF, random forest; XGB, extreme gradient boosting; LR, logistic regression; ROC, receiver operating characteristic; AUC, area under the curve.

Raman spectrum patterns obtained from SERS analysis. (A) Raman fingerprinting of each SERS chip, analyzed at 785 nm. Pre-analysis was conducted before the chips were used to investigate sample specimens. (B) Raman fingerprinting of Mycobacterium abscessus at specific concentrations (15 and 50 ng/µl) quantified via Nanodrop. Prominent peak positions are indicated above each peak, allowing for a clear comparison related to DNA content. Raman spectra are displayed as plots of Raman intensity vs. Raman shift (cm-¹). SERS, surface-enhanced Raman spectroscopy; a.u., arbitrary unit.

SERS spectrum results for three bacteria: MAB, MMAS and MTB-H37Ra. (A) Raman spectra from DNA at 15 ng/µl. (B) Raman spectra from DNA at 50 ng/µl. The solid lines indicate the mean spectra; the shaded regions represent the standard deviation. Numbers above the peaks correspond to the strong peak positions, with dotted lines marking the precise location of each peak. All Raman spectra are plotted as Raman intensity vs. Raman shift (cm-¹). SERS, surface-enhanced Raman spectroscopy; MAB, M. abscessus subspecies abscessus; MMAS, M. abscessus subspecies massiliense; MTB-H37Ra, M. tuberculosis H37Ra; a.u., arbitrary unit.

Box plot comparisons of peak intensities for eight candidate peaks for bacterial DNA content at 15 ng/µl. The specific Raman intensities examined were at 726, 755, 781, 1,097, 1,317, 1,331, 1,486 and 1,577 cm-¹. The y-axis represents the relative intensity of the peaks; the x-axis indicates the bacterial DNA species. Multiple comparisons were analyzed by Kruskal-Wallis test prior to examining pairwise differences between groups by using Dunn-Bonferroni post-hoc analysis. *P<0.05, **P<0.01, and ****P<0.0001; n.s., non-significant. H37Ra, M. tuberculosis H37Ra; MMAS, M. abscessus subspecies massiliense; MAB, M. abscessus subspecies abscessus; a.u., arbitrary unit.

Box plot comparisons of peak intensities for eight candidate peaks for bacterial DNA content at 50 ng/µl. The specific Raman intensities examined were at 726, 755, 781, 1,097, 1,317, 1,331, 1,486, and 1,577 cm-¹. The y-axis represents the relative intensity of the peaks; the x-axis indicates the bacterial DNA species. Multiple comparisons were analyzed by Kruskal-Wallis test prior to examining pairwise differences between groups by using Dunn-Bonferroni post-hoc analysis. **P<0.01, ***P<0.001 and ****P<0.0001; n.s., non-significant. H37Ra, M. tuberculosis H37Ra; MMAS, M. abscessus subspecies massiliense; MAB, M. abscessus subspecies abscessus; a.u., arbitrary unit.

Plot analysis of PCA, UMAP and LDA. Plots compare three bacteria at genomic DNA concentrations of 15 (left hand) and 50 ng/µl (right hand). (A) PCA, (B) UMAP, and (C) LDA were analysed and represented in dot plot. Blue dots represent MAB, green dots indicate MMAS, and red dots represent MTB-H37Ra. PCA, principal component analysis; UMAP, uniform manifold approximation and projection; LDA, linear discriminant analysis; MAB, M. abscessus subspecies abscessus; MMAS, M. abscessus subspecies massiliense; MTB-H37Ra, M. tuberculosis H37Ra.

The LDA, XGB, RF and LR analyses of the ROC curve. (A) MAB vs MMAS. (B) NTM group vs. MTB-H37Ra. LDA, linear discriminant analysis; XGB, extreme gradient boosting; RF, random forest; LR, logistic regression; ROC, receiver operating characteristic; MAB, M. abscessus subspecies abscessus; MMAS, M. abscessus subspecies massiliense; NTM, nontuberculous mycobacteria; MTB-H37Ra, M. tuberculosis H37Ra.

Sensitivity, specificity, and accuracy of the machine learning approach on the Raman spectral datasets with surface-enhanced Raman spectroscopy (SERS) sensors.

  SERS at 15 ng/µl (785 nm) SERS at 50 ng/µl (785 nm)
Classification model Sensitivity (%) Specificity (%) Accuracy (%) F1-score (%) Sensitivity (%) Specificity (%) Accuracy (%) F1-score (%)
A, NTM group: MAB (n=5) and MMAS (n=5)
XGB 94.71 97.91 96.25 96.32 93.22 92.73 92.97 92.94
LDA 77.52 77.59 77.67 77.63 81.05 81.47 81.26 81.29
RF 94.71 97.91 94.81 96.32 84.70 93.42 88.58 89.17
LR 92.29 93.02 92.65 92.68 92.10 89.27 90.63 90.45
B, NTM group: MAB (n=5) and MMAS (n=5); MTB-H37Ra (n=1)
XGB 97.19 97.83 97.23 98.51 99.13 96.97 98.94 99.42
LDA 98.14 85.25 97.10 98.42 99.70 87.18 98.41 99.12
RF 93.78 100 93.94 96.79 96.20 100 96.41 98.06
LR 99.86 98.46 99.74 99.86 99.85 98.57 99.73 99.85

The values were calculated as sensitivity=true positive/(true positive + false negative); specificity=true negative/(true negative + false positive); and accuracy=average sensitivity and specificity. SERS, surface-enhanced Raman spectroscopy; MAB, M. abscessus subspecies abscessus; MMAS, M. abscessus subspecies massiliense; MTB-H37Ra/H37Ra, M. tuberculosis H37Ra; NTM: nontuberculous mycobacterium; XGB, extreme gradient boosting; RF, random forest; LDA, linear discriminant analysis; LR, logistic regression.