Detection of Mycobacterium tuberculosis complex field infections in cattle using fecal volatile organic compound analysis through gas chromatography-ion mobility spectrometry combined with chemometrics

| Open Peer Review | Mycobacteriology | Research Article Detection of Mycobacterium tuberculosis complex field infections in cattle using fecal volatile organic compound analysis through gas chromatography-ion mobility spectrometry combined withchemometrics Pablo Rodríguez-Hernández,1 María José Cardador,2 Rocío Ríos-Reina,3 José María Sánchez-Carvajal,4 Ángela Galán-Relaño,5 Francisco Jurado-Martos,6 Inmaculada Luque,5 Lourdes Arce,2 Jaime Gómez-Laguna,4 Vicente Rodríguez-Estévez1 AUTHOR AFFILIATIONS See affiliation list on p. 14. ABSTRACT Bovine tuberculosis is considered a re-emerging disease caused by different species from the Mycobacterium tuberculosis complex (MTC), important not only for the livestock sector but also for public health due to its zoonotic character. Despite the numerous efforts that have been carried out to improve the performance of the current antemortem diagnostic procedures, nowadays, they still pose several drawbacks, such as moderate to low sensitivity, highlighting the necessity to develop alternative and innovative tools to complement control and surveillance frameworks. Volatilome analysis is considered an innovative approach which has been widely employed in animal science, including animal health field and diagnosis, due to the useful and interesting information provided by volatile metabolites. Therefore, this study assesses the potential of gas chromatography coupled to ion mobility spectrometry (GC-IMS) to discriminate cattle naturally infected (field infections) by MTC from non-infected animals. Volatile organic compounds (VOCs) produced from feces were analyzed, employing the subsequent information through chemometrics. After the evaluation of variable importance for the projection of compounds, the final discriminant models achieved a robust performance in cross-validation, as well as high percentages of correct classification (>90%) and optimal data of sensitivity (91.66%) and specificity (99.99%) in external validation. The tentative identification of some VOCs revealed some coincidences with previous studies, although potential new compounds associated with the discrimination of infected and non-infected subjects were also addressed. These results provide strong evidence that a volatilome analysis of feces through GC-IMS coupled to chemometrics could become a valuable methodology to discriminate the infection by MTC in cattle. IMPORTANCE Bovine tuberculosis is endemic in many countries worldwide and poses important concerns for public health because of their zoonotic condition. However, current diagnostic techniques present several hurdles, such as low sensitivity and complexity, among others. In this regard, the development of new approaches to improve the diagnosis and control of this disease is considered crucial. Volatile organic compounds are small molecular mass metabolites which compose volatilome, whose analysis has been widely employed with success in different areas of animal science including animal health. The present study seeks to evaluate the combination of fecal volatilome analysis with chemometrics to detect field infections by bovine tuberculo sis (Mycobacterium tuberculosis complex) in cattle. The good robust performance of discriminant models as well as the optimal data of sensitivity and specificity achieved highlight volatilome analysis as an innovative approach with huge potential. September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 1 Editor Florence Claude Doucet-Populaire, University Paris-Saclay, Clamart, France Ad Hoc Peer Reviewers Christian Gortázar, Instituto de Recursos Cinegéticos, CSIC-IREC, Ciudad Real, Spain; La'Tonzia L. Adams, Johns Hopkins University School of Medicine, Baltimore, Maryland, USA Address correspondence to Jaime Gómez-Laguna, [email protected]. The authors declare no conflict of interest. See the funding table on p. 15. Received 2 May 2023 Accepted 30 June 2023 Published 13 September 2023 Copyright © 2023 Rodríguez-Hernández et al. This is an open-access article distributed under the terms of the Creative Commons Attribution 4.0 International license. Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. KEYWORDS bovine tuberculosis, feces, gas chromatrography-ion mobility spectrome try, mycobacteria, volatile metabolites, chemometrics Mycobacterium tuberculosis complex (MTC) constitutes a wide group of slow-grow ing mycobacteria species where mainly Mycobacterium bovis (and others such as Mycobacterium caprae) stands out as the etiological agent for bovine tuberculosis (bTB) (1, 2), as well as for animal tuberculosis due to its multi-host nature (3). bTB is considered an infectious and chronic disease which affects not only different mammal species but also human beings, highlighting its zoonotic character (4, 5). Although human tubercu losis is mostly caused by Mycobacterium tuberculosis, about 30% of these cases have been related to M. bovis (zoonotic tuberculosis), especially in developing countries with significant livestock bTB prevalence (6–8). Apart from the above-mentioned danger to public health, animal tuberculosis has also a great economic impact in areas such as agriculture, wildlife, and trade (9). This disease reduces the productivity and sustainabil ity of cattle farms due to the compulsory slaughter of infected animals, especially in extensive livestock systems where wildlife reservoirs give rise to complex epidemiologi cal situations (10). Although major efforts have been made to develop control and surveillance programs in order to eradicate bTB, these are frequently considered inadequate or insufficient. In this regard, some drawbacks associated with current diagnostic techni ques and their poor performance, together with other epidemiological and geographical aspects, have been pointed out as responsible for the difficulty in achieving the objective of disease control, especially in endemic countries (10, 11). This context places bTB diagnosis under the spotlight: its time-consuming and tedious features, the lack of enough sensitivity (SE), the specific facilities and high expertise required by the current methodologies, as well as the influence of different variables on the final performance are limitations considered to be of crucial importance (12, 13). Therefore, the optimiza tion of alternative and reliable diagnostic methodologies to avoid such shortcomings has been considered by several authors as the key for bTB control on livestock farms (14). Volatilome is defined as the group of volatile organic compounds (VOCs) produced by a specific sample or organism, a term framed under volatilomics (15). These substances are small molecular mass metabolites which are generated directly to the environment, offering different advantages such as minimum and easy sample handling and sampling, and non-invasive and direct monitoring (16, 17). Therefore, volatilome analysis has been widely employed for different purposes and applications in animal science (18–21), including animal health and infectious diseases diagnosis (22, 23). In this regard, the potential of volatilome analysis as an innovative diagnostic tool against mycobacteria infection in livestock has recently been highlighted, as well as its several strengths such as the possible development of antemortem, portable, non-invasive methods able to improve the SE and specificity (SP) of the diagnosis (24). VOCs from different biological samples such as serum, feces, or breath have been successfully used to study and diagnose mycobacterial infections in various animal species (22, 23, 25, 26). However, some difficulties due to its initial phase and the scarce number of in vivo assays available have been identified. Regarding the infection of M. bovis in cattle, there are only a scarce number of studies published in which experimental infections and controlled conditions prevail, and the exhaled breath is the main sample evaluated (24). Thus, considering the benefits of this approach as well as the advantages that it would provide to the current diagnostic panorama of mycobacteria infection in livestock, the improvement of the state of knowledge and the number of in vivo assays of this VOC analysis application is considered important. In recent years, the combination of gas chromatography coupled to ion mobility spectrometry (GC-IMS) has made a breakthrough in the field of metabolomics due to its advantages of high SE, convenient operation, and low cost. In this regard, the goal of the present study was to evaluate and validate the combination of the GC-IMS technique and chemometrics to discriminate naturally infected cattle by MTC from non-infected animals. For this purpose, feces have Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 2 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. been selected as the sample of choice because of the interesting information which can be conveyed from the individual physiology (20) and their benefits compared with other biological matrices for the present goal (15, 24). MATERIALS AND METHODS Animals and classification criteria This study is part of a large project on cattle subjected to the surveillance and monitor ing for bTB in the framework of the Spanish national eradication program (27). A total of 31 adult dairy cows were included in the present study, belonging to 16 different dairy herds located in northern Andalusia. Two groups of animals were differentiated in terms of MTC infection: 19 cattle were diagnosed as infected (positive), and the remaining 12 individuals were diagnosed as non-infected (negative). Although the majority of non-infected animals were reared in herds with an officially tuberculosis-free status, some of them came from positive farms. The diagnosis of the above-mentioned animals was performed through the bacteriological culture of retropharyngeal, tracheobronchial, and mesenteric lymph node tissue samples and a real-time polymerase chain reaction (real-time PCR) targeting IS6110 (27). The animals included in the study were considered infected when at least one lymph node tested positive to both procedures. Briefly, the microbiological culture consisted of a decontamination process of the tissue homogenates with an equal volume of 0.75% (wt/vol: 1/1) hexadecyl pyridinium chloride solution in agitation for 30 min (28). This was followed by a sample centrifuga tion process of 30 min at 1,500 × g, collecting afterwards the pellets with swabs and cultured in liquid media (MGIT 960, Becton Dickinson, Madrid, Spain) using an automat ized BD Bacter MGIT System (Becton Dickinson). A real-time PCR was run after the microbiological culture. In short, specific primers (IS6110-forward: 5′-GGTAGCAGACCTCACCTATGTGT-3′; IS6110-reverse: 5′-AGGCGTCGG TGACAAAGG-3′) and probe (IS6110-probe: 5′-FAM-CACGTAGGCGAACCC-MGBNFQ-3′) were used targeting IS6110 (3), a 68-bp region of the transposon IS6110, specific for MTC pathogens. A QuantiFast Pathogen PCR + IC Kit (Qiagen, Hilden, Germany) was employed to evaluate each sample in duplicate in the MyiQ2 Two-Color real-time PCR Detection System (Bio-Rad, Hercules, CA, USA). A detection limit ranging from 10 to 100 genomic equivalents as well as a cut-off set at Cq <38 were previously described (27). Following the manufacturer’s guidelines, an exogenous inhibition heterologous control (internal control assay) was included to evaluate the presence of certain inhibitors on the samples. Feces samples For the aim of the present study, individual fresh feces samples from 31 cattle were collected at the slaughterhouse. All samples were collected directly from the distal portion of the rectum after the official routine veterinary examination and in accordance with national and European regulations. Samples were directly collected into sterile containers and transported in refrigera tion to the laboratory, where they were stored at −18°C until analysis. Sample processing before volatilome evaluation consisted of overnight thawing and weighting of 1.5 g of fresh feces in a precision balance on the day of analysis. This amount of sample was placed in a 20-mL glass headspace vial which was subsequently closed with a magnetic screw cap and a polytetrafluoroethylene (PTFE) silicone septum. Volatilome analysis and instrumentation A FlavourSpec instrument (G.A.S. Gesellschaft für Analytics Sensorsysteme mbH, Dortmund, Germany) equipped with an autosampler (CTC Analytics AG, Zwingen, Switzerland) was employed for volatilome evaluation. Briefly, the samples inside glass headspace vials were incubated at 60°C for 15 min with agitation at 500 r.p.m. After Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 3 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. incubation, 100 µL of the headspace gas phase of the vial was automatically injected into the injector under splitless mode with the syringe of the autosampler, both at 80°C. Chromatographic separation was performed at 45°C on a HP-5 capillary column 30 m × 0.32 mm ID × 0.5 µm film thickness (Agilent, Santa Clara, CA, USA). Nitrogen (99.999% purity) was used as the carrier gas under the following programmed flow: 1 mL/min for 13 min, 15 mL/min for 2 min, 20 mL/min for 13 min, after which the flow stopped. The analytes were ionized in the ion mobility spectrometry (IMS) ionization chamber, where a radiation energy of 6.5 KeV and 9.8 cm of drift tube were maintained at 75°C. The drift gas (nitrogen gas) was set at 150 mL/min. Every spectrum was recorded as the average scan of 12. The integrated LAV (version 2.2.1) and VOCal (version 0.2.9) software were used for viewing and processing measurement data acquired by the instrument. The n-ketones C4–C9 (Sigma-Aldrich, Madrid, Spain) were employed as external references to calculate the retention index (RI) of VOCs under the same chromatographic conditions as the samples. Then, VOCs were identified by comparing the RI and drift time of standards in the GC-IMS Library (G.A.S. Gesellschaft für Analytics Sensorsysteme mbH). Chemometrics: data analysis Considering the magnitude of GC-IMS data, the resulting information after feces analysis was treated by means of chemometrics. Specifically, a targeted approach was selected to process volatilome information. The preparation of data for its processing was as follows: the topographic maps generated by GC-IMS equipment were individually examined to select those available signals or markers. Afterwards, LAV software was employed to convert raw IMS data to .csv format, obtaining each individual signal (or feature) intensity. This step allowed data use and processing by MATLAB software (The Math works Inc., Natick, MA, USA, 2007) and its PLS Toolbox extension (Eigenvector Research, Inc., Manson, WA, USA) . Subsequently, a principal component analysis (PCA) was initially performed in order to reduce data dimensionality and thus explore data, visualize potential trends and detect possible outliers. Data were preprocessed by autoscaling (mean centering and scaling to unit variance) previously to modeling. Classification models were based on the application of partial least squares-discriminant analysis (PLS-DA). For the calibration and cross-validation (CV; venetian blinds) of the models, the data set was split into a training set of 75% of the samples (calibration set) and a test set with the remaining 25% (validation set), which was used for the subsequent external validation (as blind samples). Considering the limited number of samples available for the present study, the sample set was randomly split three times, each with the validation and calibra tion percentages previously mentioned, to improve model robustness and accuracy. Moreover, the different splits between training and test sets were always done by keeping the ratio of samples as in the original set and balancing the representation of each group. The final number of latent variables selected in PLS-DA models was assessed by the minimum root mean square classification error in CV and prediction. The variable importance for the projection (VIP) of every feature initially selected was also obtained to reduce data complexity and refine the final discriminant model, attempting to improve discrimination success by eliminating those features with a limited influence on model performance (VIP value <1) and keeping those considered significant (VIP value >1). VIP evaluation and the subsequent feature reduction were successively performed several times. Thus, sequential PLS-DA classification models were developed by selecting in each one the features with VIP value of ≥1, due to the VIPs being scaled in such a way that all compounds that have a VIP of ≥1 are considered relevant (29). Therefore, after the step of feature selection, PLS-DA models were recalculated, repeating the whole process (VIP evaluation, feature reduction, and model repetition) until the obtention of optimal results. The main goal was to achieve the greatest performance but without affecting the discriminant strength of models. This strategy was performed with the three sets of split samples. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 4 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. The diagnostic performance of the present approach was evaluated by comparing with the microbiological culture. However, since the culture is considered an imperfect reference technique for MTC diagnosis (30), the statistical Epidat (version 3.1) software (Galician Health Service, Spain) was used to estimate the adjusted SE and SP. Lastly, MetaboAnalyst (version 5.0) software was used to perform a volcano plot statistical analysis. This analysis combined the results from fold change analysis and P value from t-test to visually identify those significant features based on their statistical significance as well as to select potential biomarkers of the two groups in comparison. This representation was performed with the autoscaled intensity from all the VOCs detected after GC-IMS analysis. As mentioned above, VOC signals in the topographic plots were tentatively identified according to their retention and drift time using VOCal software and the available GC-IMS library (G.A.S. Gesellschaft für Analytics Sensorsysteme mbH). RESULTS Evaluation of topographic maps and feature selection GC-IMS analysis of fecal samples resulted in two-dimensional-topographic plots where the different VOCs detected are characterized by the drift and retention times (x and y axes, respectively) and the intensity, attending to the signals’ color and size. Fig. 1 shows a comparison between the topographic map of two fecal samples as examples: the first from an MTC-infected cow (Fig. 1A) and the second belonging to a non-infec ted individual (Fig. 1B). Some differences between the two samples can be visualized regarding the two topographic plots presented in the figure: the number, as well as the intensity of the signals, seems to be lower in the case of the infected animal (Fig. 1A), compared with the topographic plot of the negative animal (Fig. 1B). However, there were also some VOCs which showed a greater intensity in the first case (Fig. 1A). After the visual examination of all the 31 topographic plots generated by GC-IMS analysis of feces, a total of 260 features were finally selected. Subsequently, the resulting data were evaluated through chemometrics to confirm the preliminary visual differences previously mentioned. The initial PCA, which was performed using the 31 samples analyzed, showed a slight trend of grouping according to infected and non-infected samples by means of PC1, although with some sample overlapping. It did not reveal any outliers within the entire set of samples (Fig. 2); therefore, volatilome information from all the 31 fecal samples evaluated by GC-IMS analysis was used to develop discriminant models. Discriminant models The PLS-DA models developed in the present study with three calibration sets (a total of 72 samples) were used to classify samples of the corresponding three validation sets (a total of 21 samples) as blind samples (external validation). The confusion matrix as well as results related to the classification and diagnostic performance of the external validation and CV obtained in models are available in Table 1. Concerning external validation results obtained with the initial PLS-DA models, only one sample from MTC-infected cattle (belonging to Set 2) was incorrectly classified as non-infected. Hence, an average percentage of classification error of 8.33% was obtained for this group. As with the infected group, only one sample from the group of negative animals (belonging to Set 3) was incorrectly classified inside the group of infected cattle, reaching an 11.11% mean classification error in this group. The correct category classification percentage of each set of samples is also collected in Table 1. Finally, the global percentage of correct classification of the model was calculated consider ing the above-mentioned mean correct classification percentage of each category, achieving 90.28%. Regarding the model diagnostic performance, SE and SP parameters are summarized in Table 1. Results of 91.66% and 99.99% were obtained for mean SE and SP, respectively. Individual SE and SP data of each split of samples are displayed in Table 1. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 5 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. Despite the good classification and diagnostic performance obtained for external validation of discriminant models, the results of CV were quite different. In general, lower percentages of correct classification, SE and SP, were reached, ranging from 55% to 77%. Against this background, which reflects a low model robustness, VIP evaluation was subsequently performed in order to refine PLS-DA models and improve CV and external validation. The score plots of the three PLS-DA models performed with Sets 1, 2, and 3 are available in Fig. 3. The resulting representation shows a clear separation between the groups of infected and non-infected animals. Nevertheless, a slight overlapping between groups is present in the score plot of Sets 2 and 3 (explaining the low classification rates obtained for CV; i.e., the low robustness of the models of these two sets): the spider representation design of score plots in Set 2 shows three infected samples close to the center of the non-infected group and one sample from the non-infected group close to the center of positive category. This is also in line with the classification results of external validation, since misclassified samples in confusion matrices belonged to these two sets (Sets 2 and 3). Model improvement by means of the evaluation of variable importance for the projection and selection of relevant compounds VIP values of the 260 features employed for the PLS-DA model development were obtained for the three sets of samples. Subsequently, several reductions of features were applied considering the criteria explained in the methodology subsection (features whose VIP value was below 1 were discarded): for Set 1, four reductions of features were applied, obtaining seven features (VOCs numbered as 125, 140, 158, 175, 180, 191, and FIG 1 Topographic plots obtained after feces analysis by GC-IMS for MTC-infected (A) and non-infected (B) cattle. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 6 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. FIG 2 PCA score plot obtained with volatilome information after GC-IMS analysis of feces. TABLE 1 Results of external and cross-validation (confusion matrices, classification performance, and diagnostic performance) obtained by the initial PLS-DA models developed with three different sets of samples (no VIP evaluation performed) Cross-validation External validation Real classification Real classification Infected (Set 1/Set 2/Set 3) Non-infected (Set 1/Set 2/Set 3) Infected (Set 1/Set 2/Set 3) Non-infected (Set 1/Set 2/Set 3) Confusion matrices Model prediction Infected (Set 1/Set 2/Set 3) 9/10/10 4/4/2 4/3/4 0/0/1 Non-infected (Set 1/Set 2/Set 3) 6/5/5 5/5/7 0/1/0 3/3/2 Total number of samples (Set 1/Set 2/Set 3) 15/15/15 9/9/9 4/4/4 3/3/3 Classification performance Correct category classification percentage (Set 1/Set 2/Set 3) 60.0/66.67/66.67 55.56/55.56/77.78 100.0/75.0/100.0 100.0/100.0/66.67 Mean classification error (%) ± SD (%) 35.55 ± 3.85 37.03 ± 12.83 8.33 ± 14.43 11.11 ± 19.24 Total correct classification (%) ± SDa (%) 63.71 ± 8.51 90.28 ± 15.29 Diagnostic performance SEa (%) (Set 1/Set 2/Set 3) 60.2/66.89/67.01 99.99/75.0/99.99 SPa (%) (Set 1/Set 2/Set 3) 69.35/75.26/99.1 99.99/99.99/99.99 Mean SEa (%) ± SDa (%) 64.70 ± 3.90 91.66 ± 14.48 Mean SPa (%) ± SDa (%) 81.24 ± 15.75 99.99 ± 0.0 aSD, standard deviation; SE, sensitivity; SP, specificity. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 7 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. 220) as the definitive number for discriminant model development (the remaining 253 features were removed). In the case of Sets 2 and 3, this variable reduction process was repeated three times, deleting 243 and 250 features, respectively. The final numbers of features employed for PLS-DA model development in these sets were 17 for Set 2 (VOCs 16, 22, 29, 47, 56, 112, 125, 135, 152, 169, 180, 190, 191, 197, 198, 226, and 234) and 10 for Set 3 (VOCs 112, 119, 152, 158, 175, 179, 180, 191, 197, and 234). Figure 4 presents the whole process of VIP evaluation and feature reduction performed for Set 2 as an FIG 3 Score plots of PLS-DA models performed using three different sets of samples. FIG 4 Process of VIP evaluation and feature reduction performed for Set 2. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 8 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. example. Figure 4A shows the initial 260 features, while Figure 4D presents the final 17 features employed for the definitive and refined model. In general, the reduction of features after VIP evaluation resulted in a greater performance of classification models developed with Sets 1, 2, and 3 of samples. The results obtained for CV and external validation of the final PLS-DA models are summar ized in Table 2. As can be seen, the general percentage of correct classification for CV and external validation was significantly higher (90.37% and 95.83%, respectively) than the data obtained for the initial models (63.71% and 90.28%, respectively) (Table 1). This was achieved since the classification error was significantly reduced. SE and SP parameters were also improved for CV results: 95.66% and 99.99%, respectively, which highlight a better robustness of models (Table 2). SE and SP achieved for external validation in model refinement was the same as those obtained before VIP evaluation: 91.66% and 99.99%, respectively, which reveal a good and constant performance of the final models after data reduction through VIP evaluation. In general, the final number of VOCs selected after VIP evaluation for the three sets of samples was 23 (some of them were common to the three sets). Table 3 shows informa tion about the intensity, standard deviation, P value and fold change result of this group of features. Furthermore, some of these above-mentioned 23 compounds were identified based on the searching results of the GC-IMS Library: dimer of acetic acid methyl ester (VOC 22), propanoic acid (VOC 47), 3-methylbutanal (VOC 119), 3-methyl-1-pentanol (VOC 158), pentenal (VOC 175), methyltrimethoxysilane (VOC 180), ethyl 2-methylpropa noate (VOC 190), thiophene (VOC 191), and hexanal (VOC 220). The information about this identification is also included in Table 3. Relevant fecal volatile organic compounds suggested by volcano plot in infected and non-infected cattle In addition to the VIP evaluation, the importance of some of these VOCs was confirmed by another statistic test, the volcano plot, which highlighted the relevance of 26 VOCs detected after GC-IMS analysis of feces (Fig. 5). It was designed using results of fold change analysis and t-test of all the 260 VOCs initially obtained. This type of represen tation enabled discerning, after quick observation, those VOCs which obtained the greatest changes and were also significant. Three out of the 260 VOCs represented in the volcano plot (numbered as 220, 237, and 251) were significant for the group of TABLE 2 Results of external and cross-validation (confusion matrices, classification performance, and diagnostic performance) obtained by the final PLS-DA models developed with three different sets of samples (with VIP evaluation performed) Cross-validation External validation Real classification Real classification Infected (Set 1/Set 2/Set 3) Non-infected (Set 1/Set 2/Set 3) Infected (Set 1/Set 2 /Set 3) Non-infected (Set 1/Set 2/Set 3) Confusion matrices Model prediction Infected (Set 1/Set 2/Set 3) 14/15/14 1/2/1 3/4/4 0/0/0 Non-infected (Set 1/Set 2/Set 3) 1/0/1 8/7/8 1/0/0 3/3/3 Total number of samples (Set 1/Set 2/Set 3) 15/15/15 9/9/9 4/4/4 3/3/3 Classification performance Correct category classification percentage (Set 1/Set 2/Set 3) 93.33/100.0/93.33 88.89/77.78/88.89 75.0/100.0/100.0 100.0/100.0/100.0 Mean classification error (%) ± SDa (%) 4.45 ± 3.85 14.81 ± 6.41 8.33 ± 14.43 0.0 ± 0.0 Total correct classification (%) ± SDa (%) 90.37 ± 7.39 95.83 ± 10.21 Diagnostic performance SEa (%) (Set 1/Set 2/Set 3) 93.50/99.99/93.50 75.0/99.99/99.99 SPa (%) (Set 1/Set 2/Set 3) 99.99/99.99/99.99 99.99/99.99/99.99 Mean SEa (%) ± SDa (%) 95.66 ± 3.75 91.66 ± 14.48 Mean SPa (%) ± SDa (%) 99.99 ± 0.0 99.99 ± 0.0 aSD, standard deviation; SE, sensitivity; SP, specificity. Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 9 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235. tuberculosis: a review of current and emerging diagnostic techniques in view of their relevance for disease control and eradication. Transbound Emerg Dis 57:205–220. https://doi.org/10.1111/j.1865-1682.2010.01148. x 10. Cowie CE, Beck BB, Gortazar C, Vicente J, Hutchings MR, Moran D, White PCL. 2014. Risk factors for the detected presence of Mycobacterium bovis in cattle in South central Spain. Eur J Wildl Res 60:113–123. https://doi. org/10.1007/s10344-013-0757-0 11. Skuce RA, Allen AR, McDowell SWJ. 2012. Herd-level risk factors for bovine tuberculosis: a literature review. Vet Med Int 2012:621210. https:/ /doi.org/10.1155/2012/621210 12. Maurer DL, Ellis CK, Thacker TC, Rice S, Koziel JA, Nol P, VerCauteren KC. 2019. Screening of microbial volatile organic compounds for detection of disease in cattle: development of lab-scale method. Sci Rep 9:12103. https://doi.org/10.1038/s41598-019-47907-w 13. Nienhaus A, Schablon A, Costa JT, Diel R. 2011. Systematic review of cost and cost-effectiveness of different TB-screening strategies. BMC Health Serv Res 11:247. https://doi.org/10.1186/1472-6963-11-247 14. Fend R, Geddes R, Lesellier S, Vordermeier H-M, Corner LAL, Gormley E, Costello E, Hewinson RG, Marlin DJ, Woodman AC, Chambers MA. 2005. Use of an electronic nose to diagnose Mycobacterium bovis infection in badgers and cattle. J Clin Microbiol 43:1745–1751. https://doi.org/10. 1128/JCM.43.4.1745-1751.2005 15. Rodríguez-Hernández P, Cardador MJ, Arce L, Rodríguez-Estévez V. 2022. Analytical tools for disease diagnosis in animals via fecal volatilome. Crit Rev Anal Chem 52:917–932. https://doi.org/10.1080/10408347.2020. 1843130 16. Sinha R, Khot LR, Schroeder BK, Si Y. 2017. Rapid and non–destructive detection of Pectobacterium carotovorum causing soft rot in stored potatoes through volatile biomarkers sensing. Crop Protection 93:122– 131. https://doi.org/10.1016/j.cropro.2016.11.028 17. Singh KD, Del Miguel GV, Gaugg MT, Ibañez AJ, Zenobi R, Kohler M, Frey U, Sinues PM-L. 2018. Translating secondary electrospray Ionizationhigh-resolution mass spectrometry to the clinical environment. J Breath Res 12:027113. https://doi.org/10.1088/1752-7163/aa9ee3 18. Recharla N, Kim K, Park J, Jeong J, Jeong Y, Lee H, Hwang O, Ryu J, Baek Y, Oh Y, Park S. 2017. Effects of amino acid composition in pig diet on odorous compounds and microbial characteristics of swine excreta. J Anim Sci Technol 59:28. https://doi.org/10.1186/s40781-017-0153-5 19. Pérez-Calvo E, Wicaksono AN, Canet E, Daulton E, Ens W, Hoeller U, Verlhac V, Celi P, Covington JA. 2019. The measurement of volatile organic compounds in faeces of piglets as a tool to assess gastrointesti nal functionality. Biosystems Engineering 184:122–129. https://doi.org/ 10.1016/j.biosystemseng.2019.06.005 20. Karthikeyan K, Muniasamy S, SankarGanesh D, Achiraman S, Ramesh Saravanakumar V, Archunan G. 2013. Faecal chemical cues in water buffalo that facilitate estrus detection. Anim Reprod Sci 138:163–167. https://doi.org/10.1016/j.anireprosci.2013.02.017 21. Sever A, Abd Elkadir A, Matana Y, Gopas J, Zeiri Y. 2015. Biomarkers for detection and monitoring of B16 melanoma in mouse urine and feces. J Biomark 2015:841245. https://doi.org/10.1155/2015/841245 22. Spooner AD, Bessant C, Turner C, Knobloch H, Chambers M. 2009. Evaluation of a combination of SIFT-MS and multivariate data analysis for the diagnosis of Mycobacterium bovis in wild badgers. Analyst 134:1922–1927. https://doi.org/10.1039/b905627k 23. Nol P, Ionescu R, Geremariam Welearegay T, Barasona JA, Vicente J, de Jesus Beleño-Sáenz K, Barrenetxea I, Jose Torres M, Ionescu F, Rhyan J. 2020. Evaluation of volatile organic compounds obtained from breath and feces to detect Mycobacterium tuberculosis complex in wild boar (Sus scrofa) in Doñana National Park, Spain. Pathogens 9:346. https://doi. org/10.3390/pathogens9050346 24. Rodríguez-Hernández P, Rodríguez-Estévez V, Arce L, Gómez-Laguna J. 2021. Application of volatilome analysis to the diagnosis of mycobacte ria infection in livestock. Front Vet Sci 8:635155. https://doi.org/10.3389/ fvets.2021.635155 25. Bergmann A, Trefz P, Fischer S, Klepik K, Walter G, Steffens M, Ziller M, Schubert JK, Reinhold P, Köhler H, Miekisch W. 2015. In vivo volatile organic compound signatures of Mycobacterium avium subsp. paratuberculosis. PLoS One 10:e0123980. https://doi.org/10.1371/ journal.pone.0123980 26. Stahl RS, Ellis CK, Nol P, Waters WR, Palmer M, VerCauteren KC. 2015. Fecal volatile organic Ccompound profiles from white-tailed deer (Odocoileus virginianus) as indicators of Mycobacterium bovis exposure or Mycobacterium bovis bacille calmette-guerin (BCG) vaccination. PLoS One 10:e0129740. https://doi.org/10.1371/journal.pone.0129740 27. Sánchez-Carvajal JM, Galán-Relaño Á, Ruedas-Torres I, Jurado-Martos F, Larenas-Muñoz F, Vera E, Gómez-Gascón L, Cardoso-Toset F, Rodríguez- Gómez IM, Maldonado A, Carrasco L, Tarradas C, Gómez-Laguna J, Luque I. 2021. Real-time PCR validation for Mycobacterium tuberculosis complex detection targeting IS6110 directly from bovine lymph nodes. Front Vet Sci 8:643111. https://doi.org/10.3389/fvets.2021.643111 28. Corner LA, Trajstman AC. 1988. An evaluation of 1-hexadecylpyridinium chloride as a decontaminant in the primary isolation of Mycobacterium bovis from bovine lesions. Vet Microbiol 18:127–134. https://doi.org/10. 1016/0378-1135(88)90058-2 29. Mehmood T, Liland KH, Snipen L, Sæbø S. 2012. A review of variable selection methods in partial least squares regression. Chemom Intell Lab Syst 118:62–69. https://doi.org/10.1016/j.chemolab.2012.07.010 30. Courcoul A, Moyen J-L, Brugère L, Faye S, Hénault S, Gares H, Boschiroli M-L. 2014. Estimation of sensitivity and specificity of bacteriology, histopathology and PCR for the confirmatory diagnosis of bovine tuberculosis using latent class analysis. PLoS One 9:e90334. https://doi. org/10.1371/journal.pone.0090334 31. Romero KI, Fernandez-Maestre R. 2018. Ion mobility spectrometry: the diagnostic tool of third millennium medicine. Rev Assoc Med Bras (1992) 64:861–868. https://doi.org/10.1590/1806-9282.64.09.861 32. Hernández-Mesa M, Ropartz D, García-Campaña AM, Rogniaux H, Dervilly-Pinel G, Le Bizec B. 2019. Ion mobility spectrometry in food analysis: principles, current applications and future trends. Molecules 24:2706. https://doi.org/10.3390/molecules24152706 33. Tian H, Li S, Wen H, Zhang X, Li J. 2020. Volatile organic compounds fingerprinting in faeces and urine of Alzheimer's disease model SAMP8 mice by headspace-gas chromatography-ion mobility spectrometry and headspace-solid phase microextraction-gas chromatography-mass spectrometry. J Chromatogr A 1614:460717. https://doi.org/10.1016/j. chroma.2019.460717 34. Rodríguez-Hernández P, Cardador MJ, Ríos-Reina R, Simões J, Arce L, Rodríguez-Estévez V. 2023. Feed supplementation detection during the last productive stage of the acorn-Fed Iberian pig through a faecal volatilome analysis. Animals (Basel) 13:226. https://doi.org/10.3390/ ani13020226 35. Purkhart R, Köhler H, Liebler-Tenorio E, Meyer M, Becher G, Kikowatz A, Reinhold P. 2011. Chronic intestinal mycobacteria infection: discrimina tion via VOC analysis in exhaled breath and headspace of feces using differential ion mobility spectrometry. J Breath Res 5:027103. https://doi. org/10.1088/1752-7155/5/2/027103 36. Wang S, Chen H, Sun B. 2020. Recent progress in food flavor analysis using gas chromatography–ion mobility spectrometry (GC–IMS). Food Chem 315:126158. https://doi.org/10.1016/j.foodchem.2019.126158 37. Ellis CK, Stahl RS, Nol P, Waters WR, Palmer MV, Rhyan JC, VerCauteren KC, McCollum M, Salman MD. 2014. A pilot study exploring the use of breath analysis to differentiate healthy cattle from cattle experimentally infected with Mycobacterium bovis. PLoS One 9:e89280. https://doi.org/ 10.1371/journal.pone.0089280 38. Turner C, Knobloch H, Richards J, Richards P, Mottram TTF, Marlin D, Chambers MA. 2012. Development of a device for sampling cattle breath. Biosyst Eng 112:75–81. https://doi.org/10.1016/j.biosystemseng. 2012.03.001 39. Peled N, Ionescu R, Nol P, Barash O, McCollum M, VerCauteren K, Koslow M, Stahl R, Rhyan J, Haick H. 2012. Detection of volatile organic compounds in cattle naturally infected with Mycobacterium bovis. Sens Actuators B: Chem 171–172:588–594. https://doi.org/10.1016/j.snb.2012. 05.038 40. Cho YS, Jung SC, Oh S. 2015. Diagnosis of bovine tuberculosis using a metal oxide-based electronic nose. Lett Appl Microbiol 60:513–516. https://doi.org/10.1111/lam.12410 41. Ellis CK, Rice S, Maurer D, Stahl R, Waters WR, Palmer MV, Nol P, Rhyan JC, VerCauteren KC, Koziel JA. 2017. Use of fecal volatile organic compound analysis to discriminate between non-vaccinated and BCG-vaccinated cattle prior to and after Mycobacterium bovis challenge. PLoS One 12:e0179914. https://doi.org/10.1371/journal.pone.0179914 Research Article Microbiology Spectrum September/October 2023 Volume 11 Issue 5 10.1128/spectrum.01743-23 16 Downloaded from https://journals.asm.org/journal/spectrum on 17 November 2023 by 150.214.182.235.