Next Article in Journal
Simultaneous LC/MS Analysis of Carotenoids and Fat-Soluble Vitamins in Costa Rican Avocados (Persea americana Mill.)
Previous Article in Journal
Biological Role of Gellan Gum in Improving Scaffold Drug Delivery, Cell Adhesion Properties for Tissue Engineering Applications
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Molecules
Volume 24
Issue 24
10.3390/molecules24244516
molecules-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

A Machine Learning-Based Raman Spectroscopic Assay for the Identification of Burkholderia mallei and Related Species

by
Amira A. Moawad
1,2,†,
Anja Silge
3,4,†,
Thomas Bocklitz
3,5,
Katja Fischer
1,
Petra Rösch
3,
Uwe Roesler
6,
Mandy C. Elschner
1,*,
Jürgen Popp
3,4,5,* and
Heinrich Neubauer
1
1
Friedrich-Loeffler-Institut, Institute of Bacterial Infections and Zoonoses, Naumburger Str. 96a, 07743 Jena, Germany
2
Animal Health Research Institute, Agricultural Research Center, 12618 Dokki-Giza, Egypt
3
Institute of Physical Chemistry and Abbe Center of Photonics, Friedrich Schiller University, Helmholtzweg 4, 07743 Jena, Germany
4
InfectoGnostics Research Campus Jena, Center of Applied Research, Philosophenweg 7, 07743 Jena, Germany
5
Leibniz-Institute of Photonic Technology, Member of the Leibniz Research Alliance – Leibniz Health Technologies, Albert-Einstein-Str. 9, 07745 Jena, Germany
6
Institute for Animal Hygiene and Environmental Health, Free University Berlin, Robert-von Ostertag-Str. 7–13, 14163 Berlin, Germany
*
Authors to whom correspondence should be addressed.
†
These authors contributed equally to this work.
Molecules 2019, 24(24), 4516; https://doi.org/10.3390/molecules24244516
Submission received: 5 November 2019 / Revised: 3 December 2019 / Accepted: 4 December 2019 / Published: 10 December 2019
(This article belongs to the Section Analytical Chemistry)
Browse Figures
Versions Notes

Abstract

:
Burkholderia (B.) mallei, the causative agent of glanders, and B. pseudomallei, the causative agent of melioidosis in humans and animals, are genetically closely related. The high infectious potential of both organisms, their serological cross-reactivity, and similar clinical symptoms in human and animals make the differentiation from each other and other Burkholderia species challenging. The increased resistance against many antibiotics implies the need for fast and robust identification methods. The use of Raman microspectroscopy in microbial diagnostic has the potential for rapid and reliable identification. Single bacterial cells are directly probed and a broad range of phenotypic information is recorded, which is subsequently analyzed by machine learning methods. Burkholderia were handled under biosafety level 1 (BSL 1) conditions after heat inactivation. The clusters of the spectral phenotypes and the diagnostic relevance of the Burkholderia spp. were considered for an advanced hierarchical machine learning approach. The strain panel for training involved 12 B. mallei, 13 B. pseudomallei and 11 other Burkholderia spp. type strains. The combination of top- and sub-level classifier identified the mallei-complex with high sensitivities (>95%). The reliable identification of unknown B. mallei and B. pseudomallei strains highlighted the robustness of the machine learning-based Raman spectroscopic assay.

1. Introduction

Most of the species belonging to the genus Burkholderia are known as plants’ associated pathogens with a soil reservoir. Two important exceptions are B. mallei and B. pseudomallei, which are implicated in life-threatening infections in human and animals. Genetic similarities and serological cross reactions between both pathovars make the identification and differentiation from each other difficult and challenging. The high infectious potential of both agents, increased resistance against many antibiotics and their small infectious dose, imply a need for fast and robust identification methods.
B. mallei is a Gram-negative non-motile bacterium belonging to the family Burkholderiaceae that mainly affects equines, causing the notifiable zoonotic disease glanders [1]. Glanders in equids is endemic in North Africa, South America, Middle East, and Asia [2]. Infection can be transmitted through direct contact with infected animals, skin cuts and abrasions, aerosol inhalation, and ingestion of contaminated drinking water and meat. The predominant generalized clinical signs are fever, drooping of the head, labored breathing, emaciation, swelling of limbs and joints. The cutaneous manifestation by multiple popular or pustular nodules and sometimes the typical yellowish-green nasal discharge with or without ulcerous nodules on the nasal mucosa can be observed [3]. The meat of infected equids can act as a reservoir for carnivorous animal infection [4]. The disease in human is occupational, affecting mainly veterinarians and laboratory and slaughterhouse workers in addition to horse owners. B. mallei is a host-adapted pathogen and has no environmental reservoir [5]. In contrast, its closely related species B. pseudomallei is a saprophyte with a reservoir in soil [6]. B. pseudomallei is a Gram-negative, motile, aerobic, nonspore-forming, and intracellular pathogen with a high resistance to environmental conditions [7]. The organism causes serious invasive infections in humans (including septicemia and pneumonia) and is the causative agent of melioidosis, an endemic disease affecting humans and many animal species in tropical areas with a high fatality rate [8,9]. It has frequently been reported in recent decades that melioidosis is an endemic disease of public health importance in Southeast Asia and Australia [10,11].
B. mallei and B. pseudomallei cause similar clinical symptoms in human and animals. The conventional microbiological identification for both have the disadvantages of being time and labor consuming and have to be performed in biosafety level 3 (BSL 3) conditions [12]. The soil bacterium B. thailandensis is phenotypically and genetically related to B. mallei and B. pseudomallei. Nonetheless, it shows less pathogenicity [6,13]. Since B. thailandensis and B. pseudomallei share the same reservoir, the appearance of B. thailandensis frequently decreases the assay specificity for the hazardous agents [13,14].
The B. cepacia complex (comprising the B. cepacia, B. multivorans, B. stabilis, B. ambifaria, B. dolosa, and B. cenocepacia spp.) is occupying ecological niches ranging from soil to hospital environments. Those species are considered as opportunistic pathogens to humans. They are suspected to cause cystic fibrosis [15]. B. glathei and B. phytofirmans represent the non-pathogen species of the genus Burkholderia [16,17,18].
The classification of B. mallei and B. pseudomallei as category B biothreat agents on one side and the increasing spread of pathogens due to international animal transport on the other side provoked researches to evaluate innovative diagnostic strategies to differentiate between both pathovars at one time. Beside the serodiagnosis, DNA microarray-based detection methods and matrix-assisted laser desorption/ionization mass spectrometry (MALDI-TOF MS) are evaluated by various research groups for the above-mentioned task [14,19,20].
The concept of Raman spectroscopic differentiation of bacteria combines the physical recording of bacterial Raman spectra with machine learning on validated reference spectra [21,22,23,24]. Stöckel et al. reported the successful differentiation between B. mallei and B. pseudomallei by Raman spectroscopy after inactivation of the bacteria with formaldehyde to handle them under BSL 1 condition [25]. Raman spectroscopy-named after Indian physicist C.V Raman-is a modern analytical tool, which uses monochromatic light sources in the visible, infrared, or ultraviolet range to investigate the biochemical composition of specimen. By simply irradiating a sample with laser light, the molecular composition of the probed sample volume can be analyzed. The photons, which are inelastically scattered by molecular bonds, are analyzed spectroscopically and the intensity of the inelastic scattering is plotted as a Raman spectrum [26]. Since Raman scattering can be observed through a microscope to measure very small sample volumes such as single bacterial cells, it became a promising tool for a wide range of microbiological applications. Within the cell’s structure, a phenotypic specific mixture of biochemical components is present. Probing a single bacterium results in a complicated Raman spectrum exhibiting overlapping Raman peaks originating from the cell’s typical components for example lipids, proteins, DNA/RNA, pigments and storage materials. Such a Raman spectrum acts as a molecular pattern, which consists of multiple features and can hardly be interpreted by comparing with a single reference spectrum. To utilize these Raman spectra for microbial diagnostic and bacterial identification, the spectral information is analyzed by multivariate statistics and machine learning. After many replicates, the class-specific Raman spectral pattern will be learned, and the algorithm can model the differences between the bacterial classes of interest [24].
Depending on the extent of distinguishable spectral phenotypes, the identification of bacteria based on Raman spectroscopic data is either successful, providing imprecise prediction or fail. It has been shown that the learning performance improves considerably if the expected biological or biochemical variances of a certain spectral phenotype are included in the training [27]. Phenotypical variations that typically mirrored in the Raman spectrum are contributed to: (i) The ecological setting from which the bacteria originate, (ii) the isolation procedures for the bacterial cells from their habitat, and (iii) the inactivation techniques to handle potential pathogen germs [28,29,30,31,32]. Furthermore, it was shown that the data pre-processing and a proper compilation of training collectives for the supervised machine learning in combination with hierarchical classification approaches improve the identification outcome significantly [27,33,34,35,36,37,38,39]. Once a reliable statistical model is established, the microbial diagnostic based on Raman spectroscopy is not dependent on time consuming cultivation, molecular or biochemical reactions. Only a small number of single cells (<100 isolated cells) or a minimum of biomass can uncover the identity of a specimen with a high level of accuracy. The only pre-requisite processing is the nondestructive isolation of the bacterial cells from the sample matrix to probe single intact cells [23]. The present study evaluates the reproducibility of a Raman based differentiation of Burkholderia spp. previously reported by Stöckel et al. carried out by an independent research laboratory and with an independent measurement set-up. In contrast to the study from Stöckel et al. bacteria in the current study are inactivated by heat instead of formaldehyde-inactivation to perform the analysis under biosafety level 1 conditions [25]. This study aims mainly to differentiate between the hazardous agents B. mallei and B. pseudomallei on a single cell level. It is investigated to which extent the spectral phenotypes form clusters analogous to the taxonomic pre-determined Burkholderia species. The potential to additionally detect and differentiate further relevant Burkholderia species is discussed. A representative panel of strains compromising Burkholderia from cell culture selections, round-robin tests, and well-characterized isolates (see Table 1) are measured to find out which spectral phenotypes are interfering the performance of the classification. According to the observed spectral phenotypes, classification tasks are defined to train predictive models for the stepwise differentiation of the most relevant Burkholderia classes. The performance of the classification is validated by independent batch cultures of the test strains. Finally, the statistic models are evaluated by the identification of Burkholderia strains, which are not included in the training database.

2. Material and Methods

2.1. Micobiology

The strains listed in Table 1 were stored in cryovials containing a cryopreservative (MICROBANK, Mast Diagnostica, Reinfeld, Germany) at −80 °C until further investigation. Four batches of each strain were cultivated onto culture plates (NCBAGL) produced from nutrient agar (OXOID, Wesel, Germany) supplemented with 7.5% calf blood (Fiebig-Nährstofftechnik GbR, Idstein, Germany) and 10% glycerol (Merck, Darmstadt, Germany) for 24 h at 37 °C. Using a 1 μL inoculation loop, a part of the cultivated strains was scraped from plates, suspended in 1 mL of 0.9% NaCl solution and heated in thermomixer at 99 °C for 15 minutes under shaking at 400 rpm (MHL23, HLC BioTech, Pforzheim, Germany. The viability of the bacteria after inactivation was tested by growth control. For this 100 µL of each heat inactivated sample was dispensed on NCBAGL plates and incubated for 7 days at 37 °C. The successful inactivation could thus be proven in all samples. The prepared batches were washed 3 times using 500 μL distilled water and centrifuged at 8000× g (Eppendorf 5418, Eppendorf, Hamburg, Germany) for 5 minutes. Finally, 50 μL of the suspension were applied to a nickel foil, air-dried and provided to the Raman spectroscopic investigation.

2.2. Raman Spectroscopy

Raman spectra were collected with a Raman microscope (BioParticle Explorer, rap.ID Particle Systems GmbH, Berlin, Germany). A solid-state frequency doubled Nd:Yag module (Cobolt Samba, 25 mW, Cobolt AB, Solna, Sweden) with an excitation wavelength of 532 nm was used. The laser light was focused through an 100x objective (Olympus MPlanFLN 100xBD) onto the sample. This result in a spot size <1 µm laterally so that approximately 7 mW hit the sample.The bacteria were measured from different regions of the specimens. The Rayleigh scattering was removed by two edge filters after collecting the 180°-backscattered light, while a thermoelectrically cooled CCD camera registered the light (Andor DV401-BV). A single-stage monochromator, consisting of a 920-line/mm grating, diffracted the backscattered light so that the spectral resolution accounted for about 8 cm−1. The integration time per Raman spectrum (15 to 3275 cm−1) was 5 seconds. Approximately 50 single-bacteria Raman spectra were measured per batch (biological replicate) and collected from four separately prepared batches for further analysis. The evaluation of the classification models was performed by using of about 20 spectra from other isolates (Table 2).

2.3. Data Pre-Processing

The open source software Gnu R (version R-3.6.0 Vienna, Austria) were used for all computations [40]. The pre-processing was carried out according to principles investigated in Bocklitz et al. [41]. The cosmic spike removal was performed according to the reference [42] and the threshold was set to 10. The used wavenumber positions of the wavenumber standard (4-acetamidophenol) were 329.2, 390.9, 465.1, 504, 651.6, 710.8, 797.2, 857.9, 968.7, 1105.5, 1236.8, 1278.5, 1323.9, 1371.5, 1561.5, 1648.4, 2931.1, 3064.6, and 3102.4 cm−1. A polynomial of degree 3 was utilized for wavenumber calibration [43] and the background was subsequently corrected by the SNIP algorithm [44]. The used wavenumber area was 300 cm−1 to 3100 cm−1 and the wavenumber area between 1800 cm−1 and 2600 cm−1 was excluded (Figure 1). As normalization a vector normalization was applied. Spectra of burned bacteria or material artefacts were excluded from the respective data set.
The number of spectra of each single batch (overview Figure 2) was reduced from the original number of recorded spectra to 20 representative spectra. A random sampling without replacement was performed (user-independent), which selects 3 spectra and merge them to a mean spectrum. This is done until no single spectrum of the data set is left (depending on the total number of spectra per batch, the last group of selected spectra may contain only two single cell spectra). After preprocessing the data were introduced to a Principal Component Analysis (PCA). A PCA score plot visualize the data’s inherent clusters without prior knowledge about the categorical (species) label. A PCA score plot of the first principal components was plotted ones for the full data set (including all Burkholderia species) and in the following for each determined subset (p–ma–thai-complex and c-gla-phy-complex). For visualization the mean score for each species was calculated and the standard deviation ellipse was plotted for an overview of the group extend.

2.4. Statistical Learning

A support vector machine (SVM was applied for machine learning in combination PCA. For each classification task an appropriate number of PCs (derived from the PCA objects of the particular classification levels) were introduced. The first 45 PCs were utilized to train the SVM of model 1, which separates the data of the B. mallei complex from the other Burkholderia species. The unbalanced class sizes were considered by a class weighted approach. The SVM model 2.1 was established by utilizing the first 40 PCs and SVM model 2.2 by using the first 25 PCs. Each SVM model was validated by a leave-one-batch-out cross-validation (LOBOCV) [27]. For this purpose, the Raman data except the data of the batches with the serial number x were used to construct a SVM model. This model was utilized to predict the classes of the left-out batch series. For example, at the top level, all 36 Burkholderia strains were included (see Figure 2). Each single batch included 20 spectra. Therefore, each hold out batch series included 1 batch × 36 strains × 20 spectra = 720 spectra and the training data 3 batches × 36 strains × 20 spectra = 2160 spectra. Accordingly model 2.1 included 26 strains and model 2.2 10 strains. This method is repeated so that spectra of each batch series are predicted once. The number of spectra per batch series used for validation purpose should not be confused with the class size. Raman data of the Burkholderia strains from the evaluation set (Table 2) were preprocessed separately and rotated into the respective PCA space of model 1, 2.1, and 2.2 before they were predicted by the respective SVM model.

3. Results

3.1. Data Management

The results of the Raman measurements are summarized in Figure 1. For each Burkholderia species a mean spectrum is shown. The standard deviation is drawn as the grey zone around each spectrum to visualize the within species variation. Noticeable are the spectra of B. mallei, B. pseudomallei, B. phytofirmans, and B. thailandensis. They exhibit the combined appearance of signals around 843, 1050, 1450, and 1735 cm−1 which are typical for Polyhydroxybutyrat (PHB), an intracellular storage material [45]. The signals of PHB were observed with different extents in the single cell Raman spectra of nearly every Burkholderia species (Figure S1). Overall irregularities like the appearance of PHB were equalized to some extend by summarizing the data of single cells within one culture batch to subsets of merged spectra by a randomized procedure described in the section material and methods (also visualized in the supporting information Figure S2). The reduced data matrices were used for an optimized training of the classifier. Such a data management makes the introduced approach reproducible and more robust against biological variations.

3.2. Classification Workflow for Burkholderia’s Raman Data

A prerequisite for machine learning and statistic modeling on Raman spectral data is a database containing validated reference spectra of bacterial cells exhibiting phenotypic variations. Different number of strains was available during the study for the representation of the Burkholderia species. The target species B. mallei and B. pseudomallei were available with a representative number of strains. For the remaining species one to two strains were included into the training (Table 1). For each strain, four independently cultivated batches were measured to include the variances from different culture plates and from day to day. The workflow for the data collection is visualized in Figure 2.
A task-oriented and hierarchically organized classification workflow for Burkholderia’s Raman data previously described by Stöckel et.al [25] was developed further in the present study. As a first step a PCA was applied to examine if the data contains inherent clusters. The PCA as an unsupervised data analysis tool finds the main axes of variance within a data set without prior knowledge about the categorical label. The PCA reduces the dimensionality of the Raman data by calculating a new set of principal components (PCs) to minimize redundant information without loss of spectral information.
The PCA score plot in Figure 3A shows the two directions of largest variance in the data and provides a valuable insight into the nature of Burkholderia’s Raman data. Literally, data with similar spectral phenotypes cluster together. The spread of the data points of each Burkholderia species was visualized by one times the standard deviation ellipse as an overview. The colors code the different complexes of the genus Burkholderia [6,15,46]. The obligate pathogen species of the B. mallei-complex are highlighted in red. The facultative pathogen species of the B. cepacia-complex are shown in blue and the non-pathogen B. phytofirmans, B. glathei, and B. thailandensis are represented by green. The PCA score plot shows that each species overlaps with other species, which was expected for bacteria of one genus.
Due to the clusters shown in the PCA plot, classes with shared spectral characteristics are grouped together under consideration of their diagnostic relevance. B. pseudomallei, B. mallei, and B. thailandensis were pooled and labelled as p–ma–thai-complex [15]. The species of the B. cepacia-complex pooled together with B. glathei and B. phytofirmans and joined to the ce–gla–phy-complex. Consequently, the first classification task was to differentiate between the major spectral classes of the dataset. In a next step the sub classes for the supervised machine learning were compiled. The PCA-score plot in Figure 3B show that the data of the p–ma–thai-complex fall into three clusters of the particular species. Therefore, the model 2.1 differentiates between the three species of the p–ma–thai-complex. In Figure 3C the PCA-score plots of the second data-subset are shown. The species of the B. cepacia complex cluster with B. glathei and the cluster of B. phytofirmans appears apart. Data of the B. cepacia-complex and B. glathei were pooled to a new joined class. Model 2.2 was trained to separate the joined cepacia complex from B. phytofirmans. Since the differentiation between the species of the B. cepacia-complex and B. glathei was not of interest, the separation of the species of the B. cepacia-complex was not further carried out. In summary, sub-classes are grouped following a narrowing path on classification. For each level a specific classification model was trained based on the data of the joined classes. A given decision on one level leads down to different classification paths.
The SVM is an efficient machine learning algorithm for Raman spectral data in combination with a dimensionality reduction technique like PCA. PCA reduces the number of features by choosing the most important ones that still represent a maximal part of the entire dataset. The cumulative explained variance of the principal components (PCs) for the top- and sub level data sets were examined and shown in the supporting information (Figures S3–S5). Less than 200 PCs of the 627 dimensions explain 90% of the data’s variance (the 627 dimensions derived from the 627 wavenumbers per spectrum). Within these 200 PCs the best number was screened for SVM input and adjusted depending on the complexity of the classification task. To prevent overfitting, the performance of each PCA-SVM model was optimized and tested by applying a leave one batch out cross validation (LOBOCV). This procedure provides biological and technical independent data sets for validation to get an idea of the model’s reliability. Model 1 gave the highest validation accuracy by introducing the first 45 PCs (the data set at the top level contained 2880 spectra). The complexity of the classification tasks decreased with each level in a hierarchically organized classification approach. The p–ma–thai-complex included 2000 spectra. Here, the classification result was best when 40 PCs were introduced. The ce–gla–phy-data set comprises the smallest number of strains and contained 880 spectra. Here 25 PCs were sufficient for classification.

3.3. Hierarchical Classification

The confusion table of model 1 (Table 3) shows that 95.5% of the p–ma–thai-complex and 83.4% of the ce–gla–phy-complex were predicted correctly. The result reveal that model 1 is sensitive for the recognition of the p–ma–thai-complex with a considerably low false negative rate of 4.5%. The false positive rate of the p–ma–thai-complex examined by the LOBOCV was 16.6%. That means that more spectra of the ce–gla–phy-complex were misclassified by the model and therefore the specificity for the targets of the p–ma–thai-complex diminish. B. phytofirmans could be identified as an important interfering species because its spectral phenotype shows a strong overlap with B. pseudomallei (Figure 2A).
The confusion table of model 2.1 (Table 4) shows the result for differentiation between the three species of the p–ma–thai-complex. Sensitivities of 91.4% and 91.9% for B. mallei and pseudomallei were reached respectively. For B. thailandensis 65% spectra were correctly classified. Misclassifications of B. thailandensis are mainly due to the confusion with B. mallei but not vice versa. In contrast, the confusion of B. thailandensis with B. pseudomallei was insignificant.
The confusion table of model 2.2 (Table 5) summarizes the differentiation between data of the joined B. cepacia-B. glathei class (99.2% sensitivity) and B. phytofirmans (91.2% sensitivity).

3.4. Identification of New Strains

After training and validating the classification performance of the models, their generalizability was evaluated. Raman data of new B. mallei and B. pseudomallei strains were recorded, which were not taken into account for the training. The pre-processing and the exclusion of artefacts were performed in the same way like for the training data. The only exception was the averaging procedure, which was only performed for an optimized training approach. For an overview the spectra of the test strains are plotted in Figure 4. The new data were predicted by model 1 and model 2.1 and the results presented in Table 6 and Table 7.
According to the predictions performed by model 1 the majority of all spectra were put correctly into the joined class of the p–ma–thai-complex (Table 6). Ten of the twelve strains were properly identified with sensitivities ranging between 94% and 100%. A higher number of spectra were recorded for the strain 14RR5392, an isolate originated from a Green Iguana in Prague [47]. Four independent biological replicates were tested here and 95% of the data were correctly predicted by model 1. The lowest sensitivities achieved by model 1 ranged between 80% and 90% (strain 06RR1054 with 88.9% and 11RR2812 with 83.3% correctly identified spectra). Raman data which were assigned to the p–ma–thai complex by model 1, were introduced to model 2.1 for species determination. The results are shown in Table 7. It is noteworthy that none of spectra was misclassified as B. thailandensis. Misclassifications occur only infrequently between B. mallei and B. pseudomallei. For ten of the twelve strains the sensitivities ranged between 90% and 100%. The lowest level of identification accuracy was found for the B. mallei strain 10RR1381 with 87% accuracy. However, one should note that the classification outcome refers to the classification of single cells representing only a small percentage of the biomass of one culture or batch. The results reveal the strength of Raman spectroscopic based identification that only a small number of representatives (or a minimum of biomass) already reveal the identity of the specimen with a high level of accuracy.

4. Discussion

The concept of Raman spectroscopic differentiation of bacteria combines the physical recording of bacterial Raman spectra with supervised machine learning on validated reference spectra. Stöckel et al. reported that a single SVM was not capable of discriminating the Raman data of B. mallei and B. pseudomallei alongside other Burkholderia and Pseudomonas species [25]. Therefore, the Raman spectra of B. mallei and B. pseudomallei were pooled together and treated as a joined class. A top-level SVM separated this joined class from the remaining species and a sub-level SVM exclusively performs on data of B. mallei and B. pseudomallei for definitive species separation. Identification accuracies of more than 90% could be achieved on the spectra level. The hierarchically organized classification workflow for Raman data of Burkholderia previously described by Stöckel et.al [25] was further developed in the present study.
For an optimal supervision of the learning process, data’s inherent clusters were considered for the compilation of joined classes. Valuable insights into the nature of the Burkholderia’s Raman data are provided and information about the interfering species are delivered. By applying a LOBOCV, a realistic estimation of the strength and limits of Burkholderia species identification based on Raman spectroscopic data can be elucidated. The strength of the method is exhibited by the high sensitivities for the identification of the target species B. mallei and B. pseudomallei which follows the results of Stöckel et al. This is supported by the identification of the new and not referenced B. mallei and B. pseudomallei strains. The sensitives for the identification of the new Burkholderia strains reached for ten of the twelve strains between 90% and 100%. The results of the present study provided independent evidence of reproducibility for the classification and identification of B. mallei and B. pseudomllei based on Raman data. However, the quality of a classifier is also dependent on the specificity for the identification of a target species. The in comparison more frequently occurring misclassification of the ce–gla–phy-complex as B. mallei related agents at the top-level limits the reliability of the model. It is suggested that the misclassified class might be not sufficiently represented in the training data set and therefore model 1 insufficiently captures the underlying pattern of the data. This also applies to B. thailandensis which was represented by only one strain at the level of model 2.1. An increased classification accuracy is expected by a proper representation of the species-specific spectral phenotype. Even it is essential to have B.thailandensis in the database, it is important to mention that the bacterium is not an expected contaminant in sample material for clinical diagnostics, and any misclassifications could be not over-interpreted.
In contrast, model 2.1 perfectly predicts the classes for B. mallei and B. pseudomallei in the training and generalizes to the new strains that hasn’t been utilized before.
To address a specific diagnostic, the problem of sample size planning and the compilation of training data for machine learning has to be optimized so that the differences between species with a similar spectral phenotype can be properly modelled. The present study highlighted the potential of a machine learning-based Raman spectroscopic assay as a microbial diagnostic tool. It was shown that the performance of the method for the identification of B. mallei and B. pseudomallei could be reproduced by an independent laboratory and with independent measurement equipment.
As a next step, the model transferability for bacterial Raman data has to be optimized. With a view to multicenter identification purposes established database can then be shared between different laboratories. Once the access to a comprehensive bacterial Raman data collection is provided, task-specific compilations of data subsets can be used to answer new upcoming diagnostic questions.

Supplementary Materials

The following are available online at https://www.mdpi.com/1420-3049/24/24/4516/s1.

Author Contributions

For research articles with several authors, a short paragraph specifying their individual contributions must be provided. The following statements should be used “Conceptualization, A.A.M., A.S., P.R. and M.C.E.; Methodology, A.A.M., A.S., P.R. And M.C.E.; Software, T.B.; Validation, A.A.M., A.S.; Formal Analysis, A.S. and T.B.; Investigation, A.A.M., K.F., M.C.E.; Resources, M.C.E., H.N.; Data Curation, A.S. and T.B.; Writing-Original Draft Preparation, A.A.M. and A.S.; Writing-Review & Editing, M.C.E., J.P., H.N.; Visualization, A.S. and T.B.; Supervision, J.P., U.R. and H.N.; Project Administration, M.C.E.; Funding Acquisition, Y.Y.

Acknowledgments

The authors would like to thank the German Federal Foreign Office in the frame of the German Biosecurity Program.

Conflicts of Interest

None to declare.

References

  1. Saxena, A.; Pal, V.; Tripathi, N.K.; Goel, A.K. Development of a rapid and sensitive recombinase polymerase amplification-lateral flow assay for detection of Burkholderia mallei. Transbound. Emerg. Dis. 2019, 66, 1016–1022. [Google Scholar] [CrossRef] [PubMed]
  2. Pal, V.; Saxena, A.; Singh, S.; Goel, A.K.; Kumar, J.S.; Parida, M.M.; Rai, G.P. Development of a real-time loop-mediated isothermal amplification assay for detection of Burkholderia mallei. Transbound. Emerg. Dis. 2018, 65, e32–e39. [Google Scholar] [CrossRef] [PubMed]
  3. Malik, P.; Singha, H.; Goyal, S.K.; Khurana, S.K.; Tripathi, B.N.; Dutt, A.; Singh, D.; Sharma, N.; Jain, S. Incidence of Burkholderia mallei infection among indigenous equines in India. Vet. Rec. Open 2015, 2, e000129. [Google Scholar] [CrossRef] [PubMed] [Green Version]
  4. Khaki, P.; Mosavari, N.; Khajeh, N.S.; Emam, M.; Ahouran, M.; Hashemi, S.; Taheri, M.M.; Jahanpeyma, D.; Nikkhah, S. Glanders outbreak at Tehran Zoo, Iran. Iran. J. Microbiol. 2012, 4, 3–7. [Google Scholar]
  5. Dedie’, K.; Bockemqhl, J.; Kqhn, H.; Volkmer, K.J.; Weinke, T. Rotz und Melioidose. In Bakterielle Zoonosen bei Tier und Menschen. Stuttgart7Ferdinand Enke Verlag; 1993; pp. 159–168. [Google Scholar]
  6. Dance, D.A. Ecology of Burkholderia pseudomallei and the interactions between environmental Burkholderia spp. and human-animal hosts. Acta Trop. 2000, 74, 159–168. [Google Scholar] [CrossRef]
  7. Foong, Y.C.; Tan, M.; Bradbury, R.S. Melioidosis: A review. Rural Remote Health 2014, 14, 2763. [Google Scholar]
  8. Dance, D.A. Melioidosis. Curr. Opin. Infect. Dis. 2002, 15, 127–132. [Google Scholar] [CrossRef]
  9. Estes, D.M.; Dow, S.W.; Schweizer, H.P.; Torres, A.G. Present and future therapeutic strategies for melioidosis and glanders. Expert Rev. Anti Infect. Ther. 2010, 8, 325–338. [Google Scholar] [CrossRef] [Green Version]
  10. Alwarthan, S.M.; Aldajani, A.A.; Al Zahrani, I.M.; Bukhari, H.A. Melioidosis: Can Tropical Infections Present in Nonendemic Areas? A Case Report and Review of the Literature. Saudi J. Med. Med. Sci. 2018, 6, 108–111. [Google Scholar] [CrossRef]
  11. Limmathurotsakul, D.; Golding, N.; Dance, D.A.; Messina, J.P.; Pigott, D.M.; Moyes, C.L.; Rolim, D.B.; Bertherat, E.; Day, N.P.; Peacock, S.J.; et al. Predicted global distribution of Burkholderia pseudomallei and burden of melioidosis. Nat. Microbiol. 2016, 1, 15008. [Google Scholar] [CrossRef] [Green Version]
  12. Cheng, A.C. Melioidosis: Advances in diagnosis and treatment. Curr. Opin. Infect. Dis. 2010, 23, 554–559. [Google Scholar] [CrossRef]
  13. Schmoock, G.; Ehricht, R.; Melzer, F.; Rassbach, A.; Scholz, H.C.; Neubauer, H.; Sachse, K.; Mota, R.A.; Saqib, M.; Elschner, M. DNA microarray-based detection and identification of Burkholderia mallei, Burkholderia pseudomallei and Burkholderia spp. Mol. Cell. Probes 2009, 23, 178–187. [Google Scholar] [CrossRef]
  14. Karger, A.; Stock, R.; Ziller, M.; Elschner, M.C.; Bettin, B.; Melzer, F.; Maier, T.; Kostrzewa, M.; Scholz, H.C.; Neubauer, H.; et al. Rapid identification of Burkholderia mallei and Burkholderia pseudomallei by intact cell Matrix-assisted Laser Desorption/Ionisation mass spectrometric typing. BMC Microbiol. 2012, 12, 229. [Google Scholar] [CrossRef] [Green Version]
  15. Coenye, T.; Vandamme, P. Diversity and significance of Burkholderia species occupying diverse ecological niches. Environ. Microbiol. 2003, 5, 719–729. [Google Scholar] [CrossRef] [Green Version]
  16. Mitter, B.; Petric, A.; Shin, M.W.; Chain, P.S.; Hauberg-Lotte, L.; Reinhold-Hurek, B.; Nowak, J.; Sessitsch, A. Comparative genome analysis of Burkholderia phytofirmans PsJN reveals a wide spectrum of endophytic lifestyles based on interaction strategies with host plants. Front. Plant Sci. 2013, 4, 120. [Google Scholar] [CrossRef] [Green Version]
  17. Sessitsch, A.; Coenye, T.; Sturz, A.V.; Vandamme, P.; Barka, E.A.; Salles, J.F.; Van Elsas, J.D.; Faure, D.; Reiter, B.; Glick, B.R.; et al. Burkholderia phytofirmans sp. nov., a novel plant-associated bacterium with plant-beneficial properties. Int. J. Syst. Evol. Microbiol. 2005, 55 Pt 3, 1187–1192. [Google Scholar] [CrossRef]
  18. Stopnisek, N.; Zühlke, D.; Carlier, A.; Barberán, A.; Fierer, N.; Becher, D.; Riedel, K.; Eberl, L.; Weisskopf, L. Molecular mechanisms underlying the close association between soil Burkholderia and fungi. ISME J. 2016, 10, 253–264. [Google Scholar] [CrossRef] [Green Version]
  19. Fehlberg, L.C.; Andrade, L.H.; Assis, D.M.; Pereira, R.H.; Gales, A.C.; Marques, E.A. Performance of MALDI-ToF MS for species identification of Burkholderia cepacia complex clinical isolates. Diagn. Microbiol. Infect. Dis. 2013, 77, 126–128. [Google Scholar] [CrossRef]
  20. Gassiep, I.; Armstrong, M.; Norton, R.E. Identification of Burkholderia pseudomallei by Use of the Vitek Mass Spectrometer. J. Clin. Microbiol. 2019, 57, e00081-19. [Google Scholar] [CrossRef] [Green Version]
  21. Lorenz, B.; Wichmann, C.; Stockel, S.; Rosch, P.; Popp, J. Cultivation-Free Raman Spectroscopic Investigations of Bacteria. Trends Microbiol. 2017, 25, 413–424. [Google Scholar] [CrossRef]
  22. Orelio, C.C.; Beiboer, S.H.; Morsink, M.C.; Tektas, S.; Dekter, H.E.; van Leeuwen, W.B. Comparison of Raman spectroscopy and two molecular diagnostic methods for Burkholderia cepacia complex species identification. J. Microbiol. Methods 2014, 107, 126–132. [Google Scholar] [CrossRef] [PubMed]
  23. Pahlow, S.; Meisel, S.; Cialla-May, D.; Weber, K.; Rosch, P.; Popp, J. Isolation and identification of bacteria by means of Raman spectroscopy. Adv. Drug Deliv. Rev. 2015, 89, 105–120. [Google Scholar] [CrossRef] [PubMed]
  24. Stöckel, S.; Kirchhoff, J.; Neugebauer, U.; Rosch, P.; Popp, J. The application of Raman spectroscopy for the detection and identification of microorganisms. J. Raman Spectrosc. 2015, 47, 89–109. [Google Scholar] [CrossRef]
  25. Stöckel, S.; Meisel, S.; Elschner, M.; Melzer, F.; Rosch, P.; Popp, J. Raman spectroscopic detection and identification of Burkholderia mallei and Burkholderia pseudomallei in feedstuff. Anal. Bioanal. Chem. 2015, 407, 787–794. [Google Scholar] [CrossRef]
  26. Petry, R.; Schmitt, M.; Popp, J. Raman spectroscopy—A prospective tool in the life sciences. ChemPhysChem Eur. J. Chem. Phys. Phys. Chem. 2003, 4, 14–30. [Google Scholar] [CrossRef]
  27. Guo, S.; Bocklitz, T.; Neugebauer, U.; Popp, J. Common mistakes in cross-validating classification models. Anal. Methods 2017, 9, 4410–4417. [Google Scholar] [CrossRef]
  28. Harz, M.; Rosch, P.; Peschke, K.D.; Ronneberger, O.; Burkhardt, H.; Popp, J. Micro-Raman spectroscopic identification of bacterial cells of the genus Staphylococcus and dependence on their cultivation conditions. Analyst 2005, 130, 1543–1550. [Google Scholar] [CrossRef]
  29. Kloß, S.; Lorenz, B.; Dees, S.; Labugger, I.; Rösch, P.; Popp, J. Destruction-free procedure for the isolation of bacteria from sputum samples for Raman spectroscopic analysis. Anal. Bioanal. Chem. 2015, 407, 8333–8341. [Google Scholar] [CrossRef]
  30. Meisel, S.; Stockel, S.; Elschner, M.; Rosch, P.; Popp, J. Assessment of two isolation techniques for bacteria in milk towards their compatibility with Raman spectroscopy. Analyst 2011, 136, 4997–5005. [Google Scholar] [CrossRef]
  31. Neugebauer, U.; Schmid, U.; Baumann, K.; Holzgrabe, U.; Ziebuhr, W.; Kozitskaya, S.; Kiefer, W.; Schmitt, M.; Popp, J. Characterization of bacterial growth and the influence of antibiotics by means of UV resonance Raman spectroscopy. Biopolymers 2006, 82, 306–311. [Google Scholar] [CrossRef]
  32. Neugebauer, U.; Schmid, U.; Baumann, K.; Ziebuhr, W.; Kozitskaya, S.; Deckert, V.; Schmitt, M.; Popp, J. Towards a detailed understanding of bacterial metabolism—Spectroscopic characterization of Staphylococcus epidermidis. ChemPhysChem Eur. J. Chem. Phys. Phys. Chem. 2007, 8, 124–137. [Google Scholar] [CrossRef]
  33. Kloß, S.; Kampe, B.; Sachse, S.; Rösch, P.; Straube, E.; Pfister, W.; Kiehntopf, M.; Popp, J. Culture independent Raman spectroscopic identification of urinary tract infection pathogens: A proof of principle study. Anal. Chem. 2013, 85, 9610–9616. [Google Scholar] [CrossRef]
  34. Munchberg, U.; Kampe, B.; Rosch, P.; Bauer, M.; Popp, J. Micro-Raman Spectroscopy: A novel method for identification of Sepsis pathogens. Infection 2011, 39, S97. [Google Scholar]
  35. Silge, A.; Heinke, R.; Bocklitz, T.; Wiegand, C.; Hipler, U.C.; Rosch, P.; Popp, J. The application of UV resonance Raman spectroscopy for the differentiation of clinically relevant Candida species. Anal. Bioanal. Chem. 2018, 410, 5839–5847. [Google Scholar] [CrossRef]
  36. Silge, A.; Schumacher, W.; Rosch, P.; Da Costa, P.A.; Gerard, C.; Popp, J. Identification of water-conditioned Pseudomonas aeruginosa by Raman microspectroscopy on a single cell level. Syst. Appl. Microbiol. 2014, 37, 360–367. [Google Scholar] [CrossRef]
  37. Stöckel, S.; Meisel, S.; Elschner, M.; Rosch, P.; Popp, J. Identification of Bacillus anthracis via Raman Spectroscopy and Chemometric Approaches. Anal. Chem. 2012, 84, 9873–9880. [Google Scholar] [CrossRef]
  38. Stöckel, S.; Meisel, S.; Elschner, M.; Rösch, P.; Popp, J. Raman Spectroscopic Detection of Anthrax Endospores in Powder Samples. Angew. Chem. Int. Ed. 2012, 51, 5339–5342. [Google Scholar] [CrossRef] [Green Version]
  39. Stöckel, S.; Schumacher, W.; Meisel, S.; Elschner, M.; Rösch, P.; Popp, J. Raman Spectroscopy-Compatible Inactivation Method for Pathogenic Endospores. Appl. Environ. Microb. 2010, 76, 2895–2907. [Google Scholar] [CrossRef] [Green Version]
  40. Team, R.C. R: A Language and Environment for Statistical Computing Vienna, Austria: R Foundation for Statistical Computing. Available online: http://www.R-project.org/ (accessed on 10 November 2019).
  41. Bocklitz, T.; Walter, A.; Hartmann, K.; Rosch, P.; Popp, J. How to pre-process Raman spectra for reliable and stable models? Anal. Chim. Acta 2011, 704, 47–56. [Google Scholar] [CrossRef]
  42. Ryabchykov, O.; Bocklitz, T.; Ramoji, A.; Neugebauer, U.; Foerster, M.; Kroegel, C.; Bauer, M.; Kiehntopf, M.; Popp, J. Automatization of spike correction in Raman spectra of biological samples. Chemom. Intell. Lab. Syst. 2016, 155, 1–6. [Google Scholar] [CrossRef]
  43. Dorfer, T.; Bocklitz, T.; Tarcea, N.; Schmitt, M.; Popp, J. Checking and Improving Calibration of Raman Spectra using Chemometric Approaches. Z. Phys. Chem. 2011, 225, 753–764. [Google Scholar] [CrossRef]
  44. Ryan, C.; Clayton, E.; Griffin, W.; Sie, S.; Cousens, D. SNIP, a statistics-sensitive background treatment for the quantitative analysis of PIXE spectra in geoscience applications. Nucl. Instrum. Methods Phys. Res. Sect. B Beam Interact. Mater. Atoms 1988, 34, 396–402. [Google Scholar] [CrossRef]
  45. Ciobotă, V.; Burkhardt, E.-M.; Schumacher, W.; Rösch, P.; Küsel, K.; Popp, J. The influence of intracellular storage material on bacterial identification by means of Raman spectroscopy. Anal. Bioanal. Chem. 2010, 397, 2929–2937. [Google Scholar] [CrossRef] [PubMed]
  46. Gilad, J.; Harary, I.; Dushnitsky, T.; Schwartz, D.; Amsalem, Y. Burkholderia mallei and Burkholderia pseudomallei as bloterrorism agents: National aspects of emergency preparedness. IMAJ RAMAT GAN 2007, 9, 499. [Google Scholar]
  47. Elschner, M.C.; Thomas, P.; El-Adawy, H.; Mertens, K.; Melzer, F.; Hnizdo, J.; Stamm, I. Complete Genome Sequence of a Burkholderia pseudomallei Strain Isolated from a Pet Green Iguana in Prague, Czech Republic. Genome Announc. 2017, 5. [Google Scholar] [CrossRef] [PubMed] [Green Version]
Sample Availability: Samples of the compounds are available or not from the authors.
Molecules 24 04516 g001 550
Figure 1. The mean spectra of the Burkholderia species were shown in panel A and B The grey zone around the solid lines visualizes the standard deviation of the specie’s mean. Representative Raman signals are marked.
Figure 1. The mean spectra of the Burkholderia species were shown in panel A and B The grey zone around the solid lines visualizes the standard deviation of the specie’s mean. Representative Raman signals are marked.
Molecules 24 04516 g001
Molecules 24 04516 g002 550
Figure 2. Hierarchically organized classification workflow for Burkholderia’s Raman data. Model 1 includes all Raman data and separate the p–ma–thai-complex (includes B. mallei, B. pseudomallei, and B. thailandensis) from other Burkholderia species. The dataset is split at the following level. Model 2.1 includes the Raman data of the p–ma–thai-complex and differentiates between the three species. Model 2.2 classifies the data of c-gla-phy-complex and differentiates between the cluster of B. cepacia-complex and non-pathogen Burkholderia species B. phytofirmans. Each Burkholderia species was represented by at least one strain and from each strain 4 batches were measured to provide biological and technical independent data sets for validation.
Figure 2. Hierarchically organized classification workflow for Burkholderia’s Raman data. Model 1 includes all Raman data and separate the p–ma–thai-complex (includes B. mallei, B. pseudomallei, and B. thailandensis) from other Burkholderia species. The dataset is split at the following level. Model 2.1 includes the Raman data of the p–ma–thai-complex and differentiates between the three species. Model 2.2 classifies the data of c-gla-phy-complex and differentiates between the cluster of B. cepacia-complex and non-pathogen Burkholderia species B. phytofirmans. Each Burkholderia species was represented by at least one strain and from each strain 4 batches were measured to provide biological and technical independent data sets for validation.
Molecules 24 04516 g002
Molecules 24 04516 g003 550
Figure 3. The PCA score plot shows the two directions of largest variance in the data and provides a valuable insight into the nature of Burkholderia’s Raman data. The spread of the data points of each Burkholderia species was visualized by the standard deviation ellipse for an overview. The colors codes the Burkholderia complexes. The obligate pathogen species of the B. mallei-complex are highlighted in red. The facultative pathogen species of the B. cepacia-complex are shown in blue and the non-pathogen B. phytofirmans, B. glathei, and B. thailandensis are visualized in green. A: The panel shows the whole Burkholderia data set (Model 1). B: Score plot of the p–ma–thai complex show three clusters (Model 2.1). C: Score plot of the remaining Burkholderia species (Model 2.2).
Figure 3. The PCA score plot shows the two directions of largest variance in the data and provides a valuable insight into the nature of Burkholderia’s Raman data. The spread of the data points of each Burkholderia species was visualized by the standard deviation ellipse for an overview. The colors codes the Burkholderia complexes. The obligate pathogen species of the B. mallei-complex are highlighted in red. The facultative pathogen species of the B. cepacia-complex are shown in blue and the non-pathogen B. phytofirmans, B. glathei, and B. thailandensis are visualized in green. A: The panel shows the whole Burkholderia data set (Model 1). B: Score plot of the p–ma–thai complex show three clusters (Model 2.1). C: Score plot of the remaining Burkholderia species (Model 2.2).
Molecules 24 04516 g003
Molecules 24 04516 g004 550
Figure 4. Mean spectra of the Burkholderia strains introduced to the hierarchical classification model for identification. The grey zone around the solid lines visualizes the standard deviation of the strain’s mean spectra.
Figure 4. Mean spectra of the Burkholderia strains introduced to the hierarchical classification model for identification. The grey zone around the solid lines visualizes the standard deviation of the strain’s mean spectra.
Molecules 24 04516 g004
Table
Table 1. Burkholderia strains used for the training data set, their origin and number of Raman spectra.
Table 1. Burkholderia strains used for the training data set, their origin and number of Raman spectra.
SpeciesLaboratory NumberName of StrainSourceno. Spectra
B. mallei211101RR0419BogorBfR261
300102RR0118ATCC 23344 Typst. BfR235
061102RR0551Bfr 237BfR229
290103RR0041MukteswarBfR217
080304RR0090ZagrebBfR250
041206RR1051NCTC 10260IMB236
041206RR1052NCTC 10230IMB260
041206RR1055NCTC 120-ListerIMB266
041206RR1056NCTC 10247IMB220
041206RR1057BfR M2IMB279
240609RR5318Dubai7IMB243
010411RR2811Bahrain1FLI257
B. pseudomallei041206RR1058HollandIMB304
041206RR1059PITT 521IMB315
041206RR1060PITT 225AIMB315
041206RR1061HeckeshornIMB313
041206RR1062NCTC 1688IMB297
041206RR1063EF15660IMB314
041206RR1064PITT 5691IMB272
060406RR074003-04450IMB317
060406RR074503-04448IMB324
290103RR0046ATCC 23343BfR315
120107RR0019BozenMSB316
250413RR3267A101-10RKI314
081210RR1369Bp 9/H05410-0490RKI333
B. thailandensis090804RR0288DSM 13276DSMZ250
B. cepacia130303RR0117ATCC 25608DSMZ271
120707RR0672DSM 7288DSMZ326
B. cenocepacia180507RR0377ATCC BAA-245DSMZ295
B. dolosa180507RR0376DSM 16088DSMZ331
030718RR17093DSM 26124DSMZ365
B. multivorans180507RR0375DSM 13243DSMZ337
B. stabilis180507RR0378DSM 16586 DSMZ318
B. ambifaria150408RR2192DSM 16087DSMZ331
B. glathei150408RR2194ATCC 29195DSMZ294
B. phytofirmans111109RR8565 DSM 17436DSMZ332
DSMZ: German Collection of Microorganisms and Cell Cultures, Braunschweig; BfR: Federal Institute for Risk Assessment, Berlin; IMB: Sanitary Academy of the Armed Forces, Munich; RKI: Project QUANDHIP, Robert Koch Institute, Berlin; FLI: Friedrich-Loeffler-Institut, Jena, MSB: Medical Service Bozen.
Table
Table 2. Burkholderia strains used for validation, their origin and number of Raman spectra.
Table 2. Burkholderia strains used for validation, their origin and number of Raman spectra.
SpeciesLaboratory NumberName of StrainSourceno. Spectra
B. mallei040203RR0053M1BfR28
041206RR1054ATCC 23344Typst. IMB36
251109RR8925ATCC 23344Typst. RKI21
081210RR1381ATCC 23344 Typst.RKI23
010411RR2812010411RR2812FLI42
010411RR2813010411RR2813FLI26
040411RR2899040411RR2899FLI30
040411RR2900040411RR2900FLI40
100713RR4351NCTC 10245RKI25
150614RR6088M3BfR68
150614RR6089U5BfR24
B. pseudomallei120214RR5392VB976100VML302
DSMZ: German Collection of Microorganisms and Cell Cultures, Braunschweig; BfR: Federal Institute for Risk Assessment, Berlin; IMB: Sanitary Academy of the Armed Forces, Munich; RKI: Project QUANDHIP, Robert Koch Institute, Berlin; FLI: Friedrich-Loeffler-Institut, Jena VML: Vet. Med. Laboratory GmbH, Ludwigsburg; MSB: Medical Service Bozen.
Table
Table 3. Results of the Leave one Batch Out cross-validation of Model 1.
Table 3. Results of the Leave one Batch Out cross-validation of Model 1.
True
Model 1p–ma–thai-complex 1c-gla-phy-complex 2
Identified asp–ma–thai-complex 11986133
c-gla-phy-complex 294667
Sensitivities in % 95.5 83.4
1 Burkholderia p–ma–thai-complex (includes B. mallei, B. pseudomallei and B. thailandensis). 2 Burkholderia c-gla-phy-complex (includes B. cepacia-complex, B. glathei, B. phytofirmans).
Table
Table 4. Results of the Leave one Batch Out cross-validation of Model 2.1.
Table 4. Results of the Leave one Batch Out cross-validation of Model 2.1.
True
Model 2.1B. malleiB. pseudomalleiB. thailandensis
Identified asB. mallei8778023
B. pseudomallei549565
B.thailandensis29452
Sensitivities in % 91.4 91.9 65
Table
Table 5. Results of the Leave one Batch Out cross-validation of Model 2.2.
Table 5. Results of the Leave one Batch Out cross-validation of Model 2.2.
True
Model 2.2joined cepacia complex 3B. phytofirmans
Identified asjoined cepacia complex 37147
B. phytofirmans673
Sensitivities in % 99.2 91.3
3Burkholderia joined cepacia complex (includes species of the B. cepacian complex and B. glathei).
Table
Table 6. Results of the identification of unknown Burkholderia strains. Results of model 1 summarized as confusion-table of the strain label versus the predicted classes.
Table 6. Results of the identification of unknown Burkholderia strains. Results of model 1 summarized as confusion-table of the strain label versus the predicted classes.
True
Model 103RR005306RR105409RR892510RR138111RR281211RR281311RR289911RR290013RR435114RR539214RR608814RR6089
Identified asp–ma–thai-complex 12832202335252940252876424
c-gla-phy-complex 20410711001540
Sensitivities in % 10088.995.210083.395.296.71001009594.1100
1 Burkholderia p–ma–thai-complex (includes B. mallei, B. pseudomallei, and B. thailandensis). 2 Burkholderia c-gla-phy-complex (includes B. cepacia-complex, B. glathei, B. phytofirmans).
Table
Table 7. Data identified by model 1 as p–ma–thai-complex were introduced to model 2.1 for species identification.
Table 7. Data identified by model 1 as p–ma–thai-complex were introduced to model 2.1 for species identification.
True
Model 2.103RR005306RR105409RR892510RR138111RR281211RR281311RR289911RR290013RR435114RR539214RR608814RR6089
Identified asB.mallei283220203123294024246324
B. pseudomallei00034200126310
B. thailandensis000000000000
Sensitivities in % 1001001008788.6921001009691.698.4100

Share and Cite

MDPI and ACS Style

Moawad, A.A.; Silge, A.; Bocklitz, T.; Fischer, K.; Rösch, P.; Roesler, U.; Elschner, M.C.; Popp, J.; Neubauer, H. A Machine Learning-Based Raman Spectroscopic Assay for the Identification of Burkholderia mallei and Related Species. Molecules 2019, 24, 4516. https://doi.org/10.3390/molecules24244516

AMA Style

Moawad AA, Silge A, Bocklitz T, Fischer K, Rösch P, Roesler U, Elschner MC, Popp J, Neubauer H. A Machine Learning-Based Raman Spectroscopic Assay for the Identification of Burkholderia mallei and Related Species. Molecules. 2019; 24(24):4516. https://doi.org/10.3390/molecules24244516

Chicago/Turabian Style

Moawad, Amira A., Anja Silge, Thomas Bocklitz, Katja Fischer, Petra Rösch, Uwe Roesler, Mandy C. Elschner, Jürgen Popp, and Heinrich Neubauer. 2019. "A Machine Learning-Based Raman Spectroscopic Assay for the Identification of Burkholderia mallei and Related Species" Molecules 24, no. 24: 4516. https://doi.org/10.3390/molecules24244516

Article Metrics

Back to TopTop