Identification of Mycobacterium abscessus Subspecies by MALDI-TOF Mass Spectrometry and Machine Learning

ABSTRACT

Mycobacterium abscessus is one of the most common and pathogenic nontuberculous mycobacteria (NTM) isolated in clinical laboratories. It consists of three subspecies: M. abscessus subsp. abscessus, M. abscessus subsp. bolletii, and M. abscessus subsp. massiliense. Due to their different antibiotic susceptibility pattern, a rapid and accurate identification method is necessary for their differentiation. Although matrix-assisted laser desorption/ionization–time of flight mass spectrometry (MALDI-TOF MS) has proven useful for NTM identification, the differentiation of M. abscessus subspecies is challenging. In this study, a collection of 325 clinical isolates of M. abscessus was used for MALDI-TOF MS analysis and for the development of machine learning predictive models based on MALDI-TOF MS protein spectra. Overall, using a random forest model with several confidence criteria (samples by triplicate and similarity values >60%), a total of 96.5% of isolates were correctly identified at the subspecies level. Moreover, an improved model with Spanish isolates was able to identify 88.9% of strains collected in other countries. In addition, differences in culture media, colony morphology, and geographic origin of the strains were evaluated, showing that the latter had an impact on the protein spectra. Finally, after studying all protein peaks previously reported for this species, two novel peaks with potential for subspecies differentiation were found. Therefore, machine learning methodology has proven to be a promising approach for rapid and accurate identification of subspecies of M. abscessus using MALDI-TOF MS.

INTRODUCTION

Nontuberculous mycobacteria (NTM) are a group of mycobacteria present in the environment that, in some cases, can cause different types of infections in humans, such as pulmonary infections, skin and soft tissue infections, and disseminated infections (1). Mycobacterium abscessus is one of the most common and pathogenic NTM species isolated in clinical laboratories, causing respiratory infections, particularly in patients with cystic fibrosis (2). M. abscessus contains three subspecies: M. abscessus subsp. abscessus, M. abscessus subsp. massiliense, and M. abscessus subsp. bolletii (3). Here, they will be referred to as M. abscessus, M. massiliense, and M. bolletii, respectively.

The three subspecies show different susceptibility to clarithromycin, a decisive antibiotic for the treatment of these infections. Thus, M. bolletii and most strains of M. abscessus show resistance to clarithromycin, whereas M. massiliense is susceptible (4), although some exceptions can occur on rare occasions (5). The different antibiotic susceptibility pattern, in addition to the recommendation of the American Thoracic Society and Infectious Diseases Society of America (ATS/IDSA) to identify NTMs at the species level, makes it necessary to implement novel approaches for rapid and accurate discrimination of these three subspecies (6).

Currently, M. abscessus subspecies can only be identified by molecular methods, such as a commercial kit based on PCR-reverse hybridization (7) or by multiple gene sequencing (hsp65, rpoB, erm(41), etc.) (8, 9). On the other hand, the use of matrix-assisted laser desorption/ionization–time of flight mass spectrometry (MALDI-TOF MS) allows the reliable identification of most NTMs and has become the main identification method in several clinical laboratories (10, 11). However, differentiation of closely related species (like M. abscessus subspecies) remains a challenge. Although some studies have attempted subspecies identification by protein peak analysis (12–17), there is no consensus on the best strategy to follow. In last years, new approaches for data analysis from MALDI-TOF mass spectra have been applied, such as machine learning methods, which have the potential to get additional information than simple species identification (18).

The aim of this study was to evaluate MALDI-TOF MS and machine learning algorithms for the differentiation of M. abscessus subspecies. This study represents the first proof of concept for the identification of these species by applying MALDI-TOF MS and machine learning.

MATERIALS AND METHODS

Mycobacterial isolates.

A total of 325 clinical isolates of M. abscessus obtained from 232 different patients were included in this study. They encompassed 157 M. abscessus, 116 M. massiliense, and 52 M. bolletii isolates. The isolates were obtained from Hospital General Universitario Gregorio Marañón (HGM; Madrid, Spain), Hospital Universitario La Princesa (HLP; Madrid, Spain), Instituto de Salud Carlos III-Centro Nacional de Microbiología (ISCIII; Majadahonda, Spain), Hospital Universitari de Bellvitge (HUB; Hospitalet de Llobregat, Spain), Oslo University Hospital (Oslo, Norway), Brest University Hospital (Brest, France), Radboud University Medical Centre (Nijmegen, The Netherlands), and Sciensano (Brussels, Belgium). All isolates are described in Table S1 in the supplemental material.

Bacterial cultures and protein extraction procedure.

All isolates were previously identified by PCR-reverse hybridization (GenoType NTM-DR; Hain Lifescience, Nehren, Germany) and/or whole-genome sequencing by Kmer finder (19). All HGM, HLP, and ISCIII isolates were cultured from frozen stocks on 7H11 agar plates until growth was observed. Among HUB isolates, 38 were cultured on 7H11 agar plates and 48 on Löwenstein-Jensen (bioMérieux; Marcy l’Etoile, France) media. In all cases, the isolates were incubated at 37°C until growth was observed (4 to 7 days). The protein extraction procedure for MALDI-TOF MS analysis was performed as previously described (11). First, a 1 μL loopful of biomass was suspended in 300 μL of high-performance liquid chromatography quality water and then heat inactivated in a dry bath at 95°C for 30 min. After this, 900 μL of ethanol was added, the tubes were centrifuged at 13,000 rpm for 2 min, and the supernatant was discarded. After the supernatant was centrifuged and discarded again, the pellet was dried at room temperature. Then, 0.5-mm silica/zirconia beads were added together with 10 μL of acetonitrile. The tubes were vortexed briefly and sonicated for 15 min. After sonication, 10 μL of formic acid was added, and the tubes were vortexed for 10 s and centrifuged at 13,000 rpm for 2 min. One microliter of the supernatant was deposited onto the MALDI target plate (Bruker Daltonics; Bremen, Germany) in triplicates, allowed to dry, and covered with 1 μL of α-cyano-4-hydroxycinnamic acid.

Spectra acquisition by MALDI-TOF MS and data processing.

Acquisition of protein spectra was performed using the MBT Smart MALDI Biotyper (Bruker Daltonics) in the range of 2,000 to 20,000 Da. All spots were read three times, resulting in nine protein spectra per isolate. The spectra were exported and processed with Clover MS Data Analysis software (Clover Biosoft, Granada, Spain). The processing pipeline consisted of (i) smoothing by Savitzky-Golay filter (window length = 11; polynomial order = 3); (ii) baseline subtraction by Top-Hat filter (factor = 0.02); (iii) alignment of spectra with 2 Da of constant tolerance and 300 ppm of linear mass tolerance; and (iv) normalization by total ion current.

Predictive models and external validation.

Once the spectra were processed, unsupervised principal component analysis (PCA) and hierarchical cluster analysis and supervised partial least squares discriminant analysis (PLS-DA), support vector machine (SVM), random forest (RF), and K-nearest neighbors (KNN) algorithms were applied for the creation of predictive models using Clover MS Data Analysis software. A total of 43 isolates (20 M. abscessus, 15 M. massiliense, and 8 M. bolletii) collected in HGM, ISCIII, and HUB were included in the test set for the creation of the predictive models; they represented a total of 539 mass spectra. These isolates were randomly selected to represent all the variability observed previously (subspecies, morphology, culture media, and geographical origin). M. massiliense and M. bolletii spectra were balanced by oversampling to obtain the same number of spectra for each category. Internal validation was performed by 10-fold cross-validation. For external validation, 282 isolates collected in all centers were used (137 M. abscessus, 101 M. massiliense, and 44 M. bolletii), and the identification obtained in each of the three spots used was considered.

Ethics statement.

The Ethics Committee of the Gregorio Marañón Hospital (CEIm) evaluated this project and considered that all the conditions for waiving informed consent were met, since the study was conducted with microbiological samples and not with human products.

RESULTS

Analysis of isolates by unsupervised algorithms.

Through the analysis of all isolates by PCA (Fig. 1), different variables included in the study were examined: the M. abscessus subspecies, the morphology of the colonies, the type of culture media, and the geographical origin of the isolates. As can be observed in Fig. 1, the geographical origin of the isolates formed clusters, especially from strains collected in Madrid hospitals (HGM, HLP, and ISCII), Barcelona (HUB), Norway, and The Netherlands. On the other hand, isolates from different subspecies and different morphology overlapped in both clusters, as well as isolates from HUB, which were analyzed in two different culture media.

FIG 1.

Principal component (PC) analysis of all isolates included in the study, colored according to different characteristics. (A) Comparison of M. abscessus subspecies. (B) comparison of colony morphology. (C) Comparison of culture media. (D) Comparison of geographical zone of origin.

Analysis of isolates by supervised algorithms.

(i) Internal validation of predictive models. The results for each algorithm after applying a 10-fold cross-validation are shown in Table 1. The algorithms PLS-DA (Fig. S1), SVM, and RF showed the same accuracy (99.8%), with only 1 spectrum of the 539 misclassified as other category (Table S2), while KNN was the algorithm with lower accuracy (Fig. S1).

TABLE 1.

Accuracy results for internal 10-fold cross validation and external validation using Spanish isolates

Algorithm	M. abscessus (%)	M. bolletii (%)	M. massiliense (%)	Total (%)
10-Fold cross validation
PLS-DA	99.4	100	100	99.8
SVM	99.4	100	100	99.8
RF	100	100	99.4	99.8
KNN	79.9	97.2	92.2	89.8
External validation
PLS-DA	58.0	84.4	73.9	67.7
SVM	89.0	81.3	90.8	88.1
RF	93.0	78.1	91.8	90.1
KNN	44.0	25.0	61.8	47.8

(ii) External validation of predictive models using Spanish isolates. Blind analysis of the 201 Spanish isolates used for external validation showed that PLS-DA and KNN produced low identification rates, while the RF algorithm yielded 90.1% correct classification (Table 1). With this algorithm, M. bolletii obtained the lower identification rate, with 18 spectra misclassified (Table 2). Since RF yielded the highest identification rate (Fig. 2A), results obtained with this algorithm were further analyzed. Among the three identifications obtained in each spot for each isolate, the subspecies obtained in at least two spots were considered the final identification. A total of 184 (91.5%) of the isolates obtained uniform identification results (Table 3), and only 1 isolate obtained different subspecies identification for each spot. Among the isolates with identical identification, the accuracy rate of subspecies-level identification was higher than in those with only two matching identifications. Moreover, the probability of correct identification provided by RF was evaluated to establish a confidence cutoff. For all subspecies, 172 (85.6%) isolates obtained a probability higher than 60% (Fig. 2B), so this cutoff was proposed for a confident result. Considering the categorical result, 89.6% of isolates were correctly identified at the subspecies level, while establishing the confidence cutoff at 60% of probability, the accuracy rate increased to 95.9% (Fig. 2C). Moreover, when both parameters were considered (same identification in 3 spots and confidence higher than 60%), out of the 170 isolates that met these criteria, 164 (96.5%) were correctly identified (Table 3). Then, a total of 6 isolates were misidentified: 2 M. abscessus identified as M. massiliense, 2 M. massiliense as M. abscessus, and 2 M. massiliense identified as M. bolletii. The three subspecies performed similarly, with a similar area under the curve (AUC) between them (Fig. 2D). Finally, positive predictive values (PPV) were evaluated for each subspecies. Considering all identification results, the PPV obtained was 90.3% for M. abscessus, 89.3% for M. bolletii, and 90.0% for M. massiliense. When we considered only those isolates with a probability result higher than 60%, the PPV increased to 97.6%, 91.7%, and 95.3% for each subspecies, respectively (Fig. 2E).

TABLE 2.

Accuracy of external validation for all algorithms tested over 201 isolates, representing 603 mass spectra

Actual/predicted	M. abscessus	M. bolletii	M. massiliense	Percent correct
PLS-DA
M. abscessus	174	122	4	58.0%
M. bolletii	3	78	15	81.2%
M. massiliense	0	54	153	73.9%
Total PLS-DA				67.2%
SVM
M. abscessus	264	7	29	88.0%
M. bolletii	15	77	4	80.2%
M. massiliense	13	6	188	90.8%
Total SVM				87.7%
RF
M. abscessus	277	5	18	92.3%
M. bolletii	15	78	3	81.2%
M. massiliense	9	8	190	91.8%
Total RF				90.4%
KNN
M. abscessus	134	58	108	44.7%
M. bolletii	29	23	44	24.0%
M. massiliense	47	32	128	61.8%
Total KNN				47.3%

FIG 2.

Analysis of mass spectra by random forest (RF) algorithm. (A) RF plot of the model. (B) Percentages of identification probably obtained by RF on validation isolates. (C) Number of correctly identified isolates according to probability cutoff obtained by RF. (D) Receiver operating characteristics (ROC) and precision recall curves for validation isolates by RF. AP, average precision. Precision-recall curve shows the trade-off between the precision (low false-positive rate) and recall (low false-negative rate). (E) Total positive predictive value (PPV) for RF results and PPV using a 60% probability cutoff. (F) Feature importances of mass peaks for RF model.

TABLE 3.

Random forest accuracy according to the identification obtained in each spot^a

RF identification	No. isolates (%)	No. correct (%)
Same ID in 3 spots	184 (91.5)	174 (94.6)
Same ID in 2 spots	16 (8.0)	7 (43.8)
Three different ID	1 (0.5)	0 (0)
Isolates with confidence >60%	172 (85.6)	165 (95.9)
Same ID (3 spots) and confidence >60%	170 (84.6)	164 (96.5)

^aRF, random forest; ID, indentification.

(iii) External validation with isolates from other countries. To evaluate the usefulness of the model for its application in other laboratories, a larger classification model was developed using the 120 isolates (40 of each subspecies) sourced from Spanish hospitals and it was applied to the 81 strains obtained from 4 European countries. This new RF model increased the correct subspecies-level identification rate to 88.9%. Using the RF algorithm, the previous model obtained 69.1% of correct classification at the subspecies level (Table S3).

Specific peak analysis for subspecies discrimination.

All protein peaks reported in previous studies were searched among the analyzed isolates (Table 4). No unique subspecies specific peaks were found, although some of them were present in most strains of certain subspecies. Thus, almost all M. abscessus isolates showed peaks at 2,081, 3,378, and 7,637 m/z; M. bolletii showed at 2,081, 3,123, 3,463, and 7,637 m/z; and M. massiliense showed at 3,378, 4,385, and 6,711 m/z. In addition, two novel potential peaks were found in this study: 2,673 m/z (Fig. S3), which was present in 88.9% of M. abscessus isolates, 17.1% of M. bolletii, and 7.3% of M. massiliense; and 6,960 m/z (Fig. S3), which was present in 90.5% of M. abscessus isolates, 9.8% of M. bolletii, and 26.0% of M. massiliense (Table 4).

TABLE 4.

Presence of all protein peaks reported in previous studies and in the present one among isolates of each subspecies^a

Peaks	M. abscessus		M. bolletii		M. massiliense
	Present study (%)	Previous studies (%)	Present study (%)	Previous studies (%)	Present study (%)	Previous studies (%)
Previously reported  peaks (m/z)
2,081	96.0	31.7–96.5	100	100	53.1	0
3,108	10.3	0	17.1	0	79.2	100
3,123	84.1	100	95.1	100	9.4	0
3,354	4.0	1.4	12.2	NA	3.1	100
3,378	99.2	100	36.6	0	97.9	100
3,463	23.8	0	90.2	66.7	31.2	0
4,385	31.7	0–1.4	29.3	0	89.6	89.5–100
4,391	53.2	98.6–100	41.5	100	9.4	0–5.2
6,711	32.5	0	73.2	NA	90.6	100
7,637	92.6	93.2–100	90.2	100	62.5	0
7,667	11.9	3.4	2.4	0	41.7	88.1–100
8,508	7.1	4.9	14.6	NA	9.4	84.6
8,768	2.4	0–0.7	7.3	0	70.8	38.4–100
8,782	72.2	89.2–100	61.0	100	6.2	0–2.8
9,475	78.6	17–100	73.2	100	16.7	0–9.5
Novel potential  peaks (m/z)
2,673	88.9	NA	17.1	NA	7.3	NA
6,960	90.5	NA	9.8	NA	26.0	NA

^aNA, not analyzed.

DISCUSSION

Differentiation of M. abscessus subspecies by MALDI-TOF MS has been attempted in previous studies using conventional peak analysis (12–17). However, many variables can hinder this objective, such as the culture media used, the morphology of the colonies, or the geographic origin of the strains (15, 20). Moreover, due to the large number of protein peaks that are usually found in mass spectra, accurate identification based on only a few peaks may not be entirely reliable. Therefore, it is necessary to apply novel strategies capable of analyzing a large amount of data, such as machine learning methodologies (18, 21).

In the present study, we applied machine learning using both unsupervised and supervised algorithms. The first approach by unsupervised methodology (PCA) did not provide subspecies differentiation (Fig. 1). In the case of morphology, no important spectral differences between smooth and rough variants were observed. The main difference of these two variants is the expression of glycopeptidolipids on the surface (22, 23), and due to MALDI-TOF MS analyzing mainly ribosomal proteins, these differences were not detected. On the other hand, because this is a proof of concept and identification of mycobacteria from liquid media could be more complex (24), only solid culture media were evaluated, and differences between 7H11 and Löwenstein-Jensen were not observed. Interestingly, several clusters obtained by PCA corresponded to the geographic origin of the isolates. All strains obtained from the three Madrid hospitals (HGM, HLP, and ISCIII) were grouped together, while those from Barcelona (HUB) were separated. All isolates from Madrid were analyzed in the same hospital (HGM) and Barcelona isolates in HUB by the same operator, so differences in experience and preparation of the protein extracts were discarded. Thus, although the MALDI-TOF MS model in both centers was the same (MBT Smart Biotyper) and the acquisition of spectra was performed with the same technical parameters, some differences were still present in the mass spectra of these isolates, as observed in Fig. 1D. Moreover, strains collected in Norway and The Netherlands clustered together. Therefore, these results may suggest that there could be differences in the mass spectra of M. abscessus from different origins, suggestive of ecotypes, and highlight the importance of including strains from diverse origins in this type of study.

The application of supervised machine learning algorithms was targeted to the differentiation of the three subspecies. Among the four algorithms tested, the lower results were obtained by PLS-DA (Fig. S2A), SVM (Fig. S2B), and KNN (Fig. S2C), while Random Forest was able to identify a greater number of isolates at the subspecies level (Table 1). As recommended by other studies, the identification of NTM by MALDI-TOF MS should be performed in two or three replicates (11), so for more accurate identifications, we used three spots for each isolate. When the identification of the three spots was considered, the accuracy was higher in those cases where the same subspecies was obtained in all spots (Table 3). The categorical result of RF is accompanied by a probability result, so we aimed to establish a confidence cutoff to reach higher accuracy of identification. Without applying probability cutoff, RF correctly identified 90.1% of the isolates. Most of the isolates (172; 85.6%) obtained probability results above 60% (Fig. 2B), so when the cutoff was established at 60%, the accuracy rate increased to 95.9% (Fig. 2C). Moreover, by applying this cutoff, the PPV for all subspecies increased to higher than 90% (Fig. 2E).

The current model was evaluated on strains from other European countries to evaluate its application in other laboratories. Due to the identification rate being lower than that obtained with Spanish isolates, the largest model with more isolates increased this rate to 88.9% of correct identification. This fact shows that the development of bigger models can increase their accuracy and it could be easily exported to more laboratories in the future.

To analyze the protein peaks found in this study, all previously reported peaks were searched and compared with our isolates. In most cases, the detection rate of the peaks was similar to previous reports (Table 4) and, in addition, the most important peaks were found in the range of 2,000 to 10,000 Da (Fig. 2F). However, remarkable discrepancies in a few cases were found. Some peaks were found with a lower presence than reported previously: this is the case of the peak at 4,391 m/z (Fig. S4A) in M. abscessus and M. bolletii, with only half of our isolates presenting it; the peak around 8,782 m/z (Fig. S4B) that was present in 61% of M. bolletii isolates in comparison with 100% reported by Suzuki et al. (13) and Kehrmann et al. (16); and the peak around 7,667 m/z (Fig. S4C) in M. massiliense that was found in 47.1% of our isolates. Strikingly, peaks at 3,354 and 8,508 m/z were found only in a few M. massiliense isolates, while they were previously reported in most isolates of this subspecies (14, 15). On the other hand, peaks that were reported as absent in some subspecies, were found in some of our isolates. That was the case of 3,108 (Fig. S4D) and 4,385 m/z (Fig. S4A) in M. abscessus and M. bolletii; 3,123 m/z (Fig. S4D) in M. massiliense; 3,378 m/z in M. bolletii (Fig. S3A); 3,463 m/z (Fig. S4E) in M. abscessus and M. massiliense; and 6,711 m/z (Fig. S4F) in M. abscessus. The greater differences were in peaks at 2,081 (Fig. S3B) and 7,637 m/z, which have never been reported in M. massiliense (13, 17), and we found them in more than 50% of M. massiliense isolates. All these differences could have been influenced by two factors. First, it is important to include a high number of strains, representing the three subspecies to confirm that the peaks found are specific to them. The second factor is the geographic origin of the isolates. There have been reported differences in peak patterns according to the origin of the strains (15), so multicentric studies are needed to search common peaks worldwide and create accurate identification algorithms. On the other hand, two novel potential peaks have been found: 2,673 (Fig. S3C) and 6,960 m/z (Fig. S3D), both of them present in most M. abscessus isolates and in the low number of isolates from the other subspecies.

Due to the variability in the detection of peaks observed previously, the present study showed that the application of novel methodologies for data analysis, such as machine learning, could be an innovative way to improve the accuracy of MALDI-TOF MS in the identification of M. abscessus subspecies. Recently, other novel strategies have been evaluated for the same purpose. Jia Khor et al. (25) used the MALDI Biotyper Sirius system (Bruker Daltonics) for the detection of subspecies-specific lipids and were able to differentiate a few M. abscessus isolates. On the other hand, Bajaj et al. (26) evaluated for the first time the liquid chromatography-mass spectrometry for the identification of M. abscessus subspecies. However, these novel methods need to be validated with larger collections of clinical isolates to confirm their utility in a microbiology laboratory setting.

In conclusion, the high correct identification rate of M. abscessus subspecies obtained in this study, states the utility of machine learning strategy for identification purposes. The validation of the developed method could allow in the near future the automatic discrimination of M. abscessus subspecies in the routine of the clinical microbiology laboratory by the addition of a greater number and diversity of isolates.