Application of a MALDI-TOF analysis platform (ClinProTools) for rapid and preliminary report of MRSA sequence types in Taiwan

Introduction

Since their emergence in the 1960s, methicillin-resistant Staphylococcus aureus (MRSA) infections have been a major health care concern worldwide (Chen & Huang, 2014; Walter et al., 2015; Wang et al., 2010). Many epidemiological studies have revealed that different multilocus sequence types (MLST) present specific characteristics such as virulence gene profiles (Recker et al., 2017; Schuenck et al., 2012; Wang et al., 2009, 2010, 2012). Understanding the evolution of MRSA lineages and the origin of infection is crucial in outbreak investigation. Molecular typing methods, such as pulsed-field gel electrophoresis and MLST, are highly expensive and labor intensive for epidemiological studies. Hence, the application of these methods in clinical practice is limited (Struelens et al., 2009). Sequence based typing method has been widely used since the past decade, and it provides adequately high resolution for confirming transmission. However, in regions with few medical resources and financial constraints, it remains relatively impracticable (Harris et al., 2013; Koser et al., 2012; Schwarze et al., 2018).

Recently, matrix-assisted laser desorption ionization time-of-flight mass spectrometry (MALDI-TOF MS) has been used in many clinical microbiology laboratories. This method can be used to identify bacterial species effectively and rapidly (Ge et al., 2016). The peptide or protein MS fingerprint of each bacterium can be generated and stored in a bacterial library for species identification. In addition, MALDI-TOF MS also provides an alternative solution for molecular typing methods (e.g., MLST) (Lartigue, 2013; Lu et al., 2012) and has the potential to offer lineage typing up to the subspecies level.

Studies that have adopted the MALDI-TOF MS approach (Lasch et al., 2014; Sauget et al., 2017; Ueda et al., 2015) have reported varying results; it may be due to several reasons. First, the predominant MRSA lineages in different areas are different and the discriminatory power of MALDI-TOF might therefore differ when compared between different MRSA lineages. Second, the bacterial MS fingerprints of isolates from one area may not match those of isolates from other areas. Third, in many of the published works, the MALDI-TOF mass spectra have been assessed manually. An objective, standardized, and automated protocol has not been widely applied thus far (Camoez et al., 2016). Finally, the data obtained from MALDI-TOF mass spectra are relatively complicated and may be analyzed by a wide variety of bioinformatics tools. A manual approach cannot ensure a consistently high-quality output, because of the potential for large interindividual or intraindividual variation in the interpretation of the data. Therefore, a reliable analysis platform is necessary to ensure comparability between reports in clinical practice. Although Staphylococcus protein A typing provides comparable performance as MLST typing with less cost and time, our study has adopted MLST typing because we aim to demonstrate the possibility to implement our method beyond S. aureus (Crisostomo et al., 2001; O’Hara et al., 2016). To validate the use of MALDI-TOF mass spectra in classifying MLST types of MRSA in Taiwan, we used the ClinProTools software for analyzing MALDI-TOF mass spectra to generate classification models of MRSA lineages. Accordingly, clinical microbiology laboratories may rapidly provide preliminary typing reports of MRSA, which may be further confirmed using a sequence-based method, thus enabling clinical practitioners to exclude an outbreak or transmission in time.

Materials and Methods

Results

Six MLST types of MRSA were identified in the isolates, namely ST5 (n = 40), ST45 (n = 8), ST59 (n = 62), ST239 (n = 179), ST241 (n = 12), and ST573 (n = 5). The isolates ST5, ST59, and ST239 were considered the major MLST types of MRSA because they exhibited numerous lineages. The isolates ST45, ST241, and ST573 were categorized together as other ST types.

Characteristic peaks for discrimination among various MLST types

The characteristic peaks were sorted using the corresponding P-values obtained in the (W/KW) test because the AD test revealed P < 0.05 (Table 1). The top 10 characteristic peaks, sorted using the W/KW test results, were selected for further statistical evaluation. The 10 peaks ranged from m/z 2,082 to 6,594 (Table 1). All the 10 selected peaks revealed P < 0.000001 in the W/KW test, thus indicating that they were informative and discriminative peaks in the MLST type classification.

Table 1:

Characteristic MALDI-TOF MS peaks in different MLST types of MRSA.

Mass	DAve	PW/KW	PAD	Average of peak intensity
				ST5	ST59	ST239	Other ST
2082.13	4.28	<0.000001	<0.000001	6.13	3.84	1.85	3.24
2415.79	38.19	<0.000001	<0.000001	24.57	6.13	44.32	19.72
2430.49	18.13	<0.000001	<0.000001	20.12	3.86	21.99	6.23
2880.03	11.2	<0.000001	<0.000001	1.15	1.88	12.35	5.38
2980.45	6.85	<0.000001	<0.000001	7.4	2.47	9.32	7.88
3276.8	2.94	<0.000001	<0.000001	2.61	3.73	0.79	2.08
3876.96	6.24	<0.000001	<0.000001	0.75	5.5	6.33	7
3892.91	4.91	<0.000001	<0.000001	5.61	0.69	0.87	2.77
6553.03	6.23	<0.000001	<0.000001	6.28	9.32	3.1	6.15
6593.54	4.33	<0.000001	<0.000001	1.98	3	6.3	5.48

DOI: 10.7717/peerj.5784/table-1

Notes:

DAve, difference between the maximal and the minimal average peak intensity of all classes; PW/KW, P-value obtained through Wilcoxon/Kruskal–Wallis test; PAD, P-value obtained through Anderson–Darling test; MALDI-TOF MS, matrix-assisted laser desorption ionization time-of-flight mass spectrometry.

In Table 1, the distribution of these peaks over various MLST types was further evaluated. High expression levels of m/z 2,430 and 3,893 and a low expression level of m/z 3,877 indicated the fingerprint of ST5; low expression levels of m/z 2,416, 2,980, and 3,893 indicated the fingerprint of ST59; high expression levels of m/z 2,416 and 2,880 and low expression levels of m/z 3,277 and 3,893 indicated the fingerprint of ST239. Based on 10 informative peaks, the odds ratio of different ST-pairs was calculated to determine the association between the peaks and respective ST-pairs (Table S1). Furthermore, the distribution of the isolates was plotted according to peak intensity of m/z 3,277 (x-axis) and m/z 6,594 (y-axis) (Fig. 1). The peaks at m/z 3,277 and 6,594 were the top two characteristic peaks among the ST types (Fig. 1). The scatter plot figures may be seen as a projection of a high-dimensional scatter space with various dimensions of the m/z peaks.

The average intensities of the 10 peaks of these lineages are further illustrated in Fig. 2. These results demonstrate the specific characteristics and patterns of peaks in the different MLST types. Specifically, the distribution of the ST59 lineages and ST239 lineages showed satisfactory separation (Fig. 1); ST5 lineages could also be distinguished from ST239 lineages based on distribution (Fig. 1). By contrast, ST5 lineages and ST59 lineages could not be satisfactorily discriminated using information from the peaks m/z 3,277 and 6,594 only (Fig. 1). Similarly, the lineages of other minor ST types mixed with other major ST types (i.e., ST5, ST59, and ST239) on the scatter plot (Fig. 1).

ML models for classification of various MLST types of MRSA

Various MLST types showed specific patterns of peaks expression, as presented in Table 1 and Figs. 3–5. To generate a comprehensive and objective classification, ML algorithms were used. In the ST5 binary classification models, the peaks at m/z 3,877 and m/z 3,893 showed satisfactory discriminative power (Fig. 3). The QC algorithm attained the highest cross-validation values (94.40%, Table 2).

Table 2:

Test performance of various lineage typing ML models.

ST	Model	Selected peaks	Acc	Sen	Spe
ST5	QC	8	0.94 ± 0.04	0.94 ± 0.05	0.95 ± 0.05
	SNN	24	0.63 ± 0.05	0.91 ± 0.05	0.34 ± 0.04
	GA_KNN1	30	0.83 ± 0.03	0.68 ± 0.03	0.98 ± 0.05
	GA_KNN3	30	0.84 ± 0.03	0.71 ± 0.03	0.97 ± 0.02
	GA_KNN5	19	0.83 ± 0.03	0.68 ± 0.03	0.98 ± 0.02
	GA_KNN7	14	0.88 ± 0.03	0.79 ± 0.03	0.97 ± 0.02
ST59	QC	22	0.76 ± 0.03	0.85 ± 0.04	0.67 ± 0.04
	SNN	5	0.68 ± 0.04	0.57 ± 0.05	0.79 ± 0.04
	GA_KNN1	30	0.78 ± 0.03	0.62 ± 0.04	0.94 ± 0.03
	GA_KNN3	29	0.82 ± 0.02	0.70 ± 0.02	0.94 ± 0.03
	GA_KNN5	29	0.85 ± 0.03	0.74 ± 0.03	0.96 ± 0.02
	GA_KNN7	22	0.82 ± 0.02	0.67 ± 0.02	0.97 ± 0.02
ST239	QC	6	0.81 ± 0.02	0.91 ± 0.01	0.72 ± 0.02
	SNN	1	0.58 ± 0.03	0.72 ± 0.03	0.44 ± 0.04
	GA_KNN1	30	0.80 ± 0.02	0.88 ± 0.03	0.72 ± 0.02
	GA_KNN3	30	0.80 ± 0.02	0.85 ± 0.02	0.74 ± 0.02
	GA_KNN5	29	0.81 ± 0.02	0.89 ± 0.01	0.72 ± 0.02
	GA_KNN7	28	0.83 ± 0.02	0.90 ± 0.02	0.76 ± 0.02

DOI: 10.7717/peerj.5784/table-2

Notes:

QC, QuickClassifier; SNN, supervised neural network; GA, genetic algorithm; KNN, K-Nearest Neighbor.

The number following “GA-KNN” indicates the number of the nearest neighbors used in models.

Selected peaks: number of peaks selected by the models; Acc, accuracy; Sen, sensitivity; Spe, specificity. Performance metrics are expressed as mean ± standard error.

In ST59 binary classification models, peaks at m/z 2,980 and m/z 2,416 also showed satisfactory discriminative power (Fig. 4). The GA-KNN algorithm attained the highest cross-validation values among all the algorithms. The GA-KNN algorithm showed highest performance when the number of the nearest neighbor was set as 5 (85.00%, Table 2).

For the binary classification of ST239 vs non-ST239, the peaks at m/z 3,277 and m/z 6,553 showed moderate discriminative power (Fig. 5). The GA-KNN algorithm showed higher performance over other algorithms when the number of the nearest neighbor was set as 7 (82.80%, Table 2).

In this study, multiclass models for classifying an isolate into one of four lineages were designed. Generally, the GA-KNN algorithm outperformed the other algorithms; more specifically, this algorithm was suitable for detecting ST239 (Table 3). The GA-KNN algorithm could successfully detect the ST5, ST59, and ST239 lineages with an accuracy of >81.00%. However, none of the multiclass models showed reliable accuracy in detecting isolates of the other ST lineages (i.e., ST45, ST241, and ST573).

Table 3:

Accuracy of various multiclass models in classifying different ST lineages.

Model	Selected peaks	ST5	ST59	ST239	Others
QC	24	0.65 ± 0.05	0.85 ± 0.03	0.79 ± 0.03	0.09 ± 0.07
SNN	25	0.65 ± 0.05	0.34 ± 0.05	0.25 ± 0.06	0.13 ± 0.09
GA_KNN1	30	0.58 ± 0.05	0.63 ± 0.04	0.84 ± 0.02	0.09 ± 0.07
GA_KNN3	26	0.73 ± 0.03	0.78 ± 0.03	0.90 ± 0.02	0.13 ± 0.08
GA_KNN5	27	0.73 ± 0.03	0.78 ± 0.03	0.94 ± 0.02	0.04 ± 0.07
GA_KNN7	22	0.92 ± 0.03	0.81 ± 0.03	0.94 ± 0.02	0.04 ± 0.07

DOI: 10.7717/peerj.5784/table-3

Notes:

QC, QuickClassifier; SNN, supervised neural network; GA, genetic algorithm; KNN, K-Nearest Neighbor.

The number following “GA-KNN” indicated the number of the nearest neighbor used in models; Selected peaks: number of peaks selected by the models. Accuracies were expressed as mean ± standard error.

Discussion

In this study, the major ST types (i.e., ST5, ST59, and ST239) of blood stream MRSA in Taiwan could be classified through ML-based MALDI-TOF mass spectra analysis with high accuracy. The analysis was conducted using a standardized system (i.e., ClinProTools) to obtain objective and consistent results. Moreover, the ST types of MRSA could be predicted accurately through MALDI-TOF mass spectra analysis before additional molecular typing methods (e.g., MLST). Biomarker peaks of various MRSA ST types were discovered through MALDI-TOF mass spectra analysis by using the ClinProTools software. Some of the biomarker peaks have been previously reported (Camoez et al., 2016; Josten et al., 2013; Sauget et al., 2017; Zhang et al., 2015) and validated in the study (namely m/z 3,277, 3,877, 3,893, 6,553, and 6,594), whereas some of them have not yet been widely validated (namely m/z 2,082, 2,416, 2,430, 2,880, and 2,980).

Nevertheless, reports on the MS characteristics of MRSA are inconsistent (Lasch et al., 2014; Sauget et al., 2017; Ueda et al., 2015). Although the accuracy of typing is sufficient in individual studies, the general patterns of specific lineages are not available (Sauget et al., 2017). Currently, bacterial lineages obtained from geographically diverse areas cannot be clearly discriminated using MALDI-TOF MS. However, more robust characteristic patterns of various types may be available in regions with geographic barriers (e.g., Taiwan Island) than in regions without these barriers. The characteristic patterns of specific lineages may be the result of the disseminated lineages in local areas. By contrast, establishing a localized solution by using an appropriate method of interpreting mass spectra may be more practical and crucial than establishing a generalized pattern. The isolates in this study were obtained from multiple reference hospitals in Taiwan. Consequently, the isolates used in the study can represent the molecular characteristics of MRSA in the local region. The localized molecular characteristics of MRSA may be sufficiently useful for clinical practice in regions with geographic barriers (e.g., Taiwan Island).

Currently, clinical microbiology laboratories generally use MALDI-TOF MS for bacterial identification because of its advantages in accuracy and effectiveness over traditional biochemical methods (Lartigue, 2013). Before analytical measurement by using MALDI-TOF MS, a protein extraction process is necessary. In-tube extraction or the direct deposit method are two common extraction methods used in clinical microbiology laboratories. In-tube extraction provides more purified intracellular components than the direct deposit method does, which results in high quality MALDI-TOF mass spectra and low noise. By contrast, the direct deposition of bacteria onto a steel plate is considered a less labor-intensive and more rapid preanalytical process than in-tube extraction. The time for the entire process and the turnaround time of MALDI-TOF MS can be considerably reduced by using the direct deposit method. Consequently, considering the relevance in routine practice, the direct deposit method was evaluated in this study.

Knowledge of the bacterial molecular type of MRSA is crucial while performing epidemiological studies on bacterial outbreak. In this study, specific peaks were identified for major clonal lineages of MRSA. The peaks m/z near 3,277, 3,877, 3,893, 6,553, and 6,594 identified in this study have been reported in previous studies as characteristic peaks in discriminating MRSA clonal complexes (Camoez et al., 2016; Josten et al., 2013; Wolters et al., 2011). Zhang et al. (2015) reported the highest expression of the peaks at m/z 3,277 and 6,554 in ST59. The peak at m/z 6,594 was recognized as the SA1452 protein and as a biomarker of the clonal complex 8 and USA-300 lineages; both these MRSA lineages are prevalent in communities and hospitals (Boggs, Cazares & Drake, 2012; Josten et al., 2013; Wolters et al., 2011). In this study, the peak at m/z 6,594 was characteristic for ST239, which were the most prominent ST type of bloodstream MRSA and prominent HA MRSA (Table 1). By contrast, some of the characteristic peaks have not previously been reported and validated as discriminative peaks (i.e., peaks at m/z 2,082, 2,416, 2,430, 2,880, and 2,980) (Sauget et al., 2017); however, they play a crucial role in MLST type classification. For example, the peak at m/z 2,980 was a distinguishing peak for the ST59 MRSA lineages (Fig. 4), which is one of the characteristic ST types of the CA MRSA in Taiwan (Huang & Chen, 2011). Moreover, the peaks at m/z 2,416 and 2,430 were noted as characteristic peaks in the ST5 and ST239. In the QC models, the peak at m/z 2,416 was proven to represent psm-mec, which is strongly associated with SCCmec III and VIII (Queck et al., 2009). Briefly, although the peak at m/z 2,416 is commonly expressed peak in MRSA, the expression level, in addition to its presence, may serve as an informative feature in classification of MRSA lineages.

Furthermore, for identifying subtle differences in the MALDI-TOF mass spectra for preliminary reporting of AST or subspecies results, not only single characteristic peaks but also specific combinations of characteristic peaks may be beneficial. Single MS peaks have provided some characteristics of each major ST type (Table 1; Figs. 1–5). Additional integration of these characteristic peaks by using ML models may generate a more comprehensive and robust result for ST typing of MRSA (Table 2) than that generated using single peaks. Consequently, the accuracy of the subspecies classification by using a combination of peaks may be higher and more resistant to variations in analysis than that using a single peak because of comprehensive interpretation. Zhang et al. (2015) reported the successful use of ML models (by ClinProTools) in analyzing MALDI-TOF mass spectra to classify various MLST types of MRSA. High performance of the binary models (including ST5, ST45, ST59, and ST239 binary classification models) were described. However, the high performance might have resulted from over fitting because only one specific combination of training sets and validation sets was used for performance evaluation.

This study had several limitations. Firstly, our study did not adopt nucA polymerase chain reaction or other sequencing methods to exclude S. argenteus, which were previously identified as S. aureus using molecular typing (Thaipadungpanit et al., 2015; Tong et al., 2015). The MRSA isolates were collected from an island with geographic barriers, thus resulting in a relative simple ST distribution. Possibly, our method might be unsuitable in other country with wide variety of bacterial lineage. We thus recommend researchers who intend to adopt our approach to train and validate their own model by using regional MRSA lineages. Another limitation of this study is that the MRSA isolates in this study are all from previously collected samples under the SMART program. This indicates the model may not exhibit sufficient accuracy in performance on new MRSA isolated in the future. Moreover, only six MLST types (namely ST5, ST45, ST59, ST239, ST241, and ST573) of MRSA were included for models training and validation. Based on the study design, ML learning models can be used for detecting the major lineages of MRSA in Taiwan only, but they cannot be used for detecting new clones, which were not included in model training. Consequently, the models are designed to be used for reporting preliminary lineage information in outbreak investigation. When a new clone emerges in the future, the ML models could be tuned and updated using the newly collected datasets. Our study has demonstrated that in regions with limited medical resources, an MALDI-TOF-based MLST typing model may serve as a valuable method for timely intervention in the transmission or outbreak of MRSA.

Supplemental Information