Discrimination between hypervirulent and non-hypervirulent ribotypes of Clostridioides difficile by MALDI-TOF mass spectrometry and machine learning

Two hundred forty clinical C. difficile isolates (157 training set and 83 validation set) from the German National Reference Center’s strain collection were tested (Table 1) [18]. Strains were pre-characterized by PCR-ribotyping with their selection based on their epidemiologic importance in Europe (Supplementary File S1).

For analysis, cryopreserved clinical isolates were thawed, sub-cultured on trypticase soy agar plates with 5% sheep blood (BD Biosciences, USA), and incubated at 37 °C for 48 h using an anaerobic chamber (Whitley, UK). Prior to further processing, fresh colonies underwent MALDI-TOF analysis for purity check (Bruker Daltonics, USA).

Off-plate ethanol/formic acid protein extraction protocol was used as described previously [19]. Briefly, 2–3 colonies were suspended in 300-μL liquid chromatography (LC-MS) grade water (Merck, Germany). Next, 900-μL absolute ethanol (Merck) were added followed by vortexing, then centrifuged (18,000 × g for 2 min). The supernatant was discarded and the bacterial pellet was completely dried. Cells were resuspended in 10 μL of 70% (v/v) formic acid and 10 μL of acetonitrile and thoroughly mixed and centrifuged (see above). One μL of the cleared supernatant was spotted four times (technical replicates) on the target plate. After air-drying, each spot was covered with 1 μL of saturated α-cyano-4-hydroxy-cinnamic acid (HCCA) matrix solution (Bruker). Measurements were performed with the Microflex LT smart mass spectrometer using the AutoXecute algorithm implemented in the Flexcontrol software (v.3.4, Bruker). To ensure biological reproducibility, this procedure was repeated with a new subculture of each isolate. Bacterial test standard (BTS, Bruker) was used for calibration. For species confirmation, acquired spectra were compared to the Bruker BDAL database (10,184 species-specific main spectra profiles) using the MALDI Biotyper compass explorer software (v.3.0).

Raw spectra were visualized using the FlexAnalysis software (Bruker), then exported to the Clover MS Data Analysis Software [20].

All spectra were preprocessed using default parameters: Smoothing (Savitzky–Golay filter: window length 11, polynomial order: 3); baseline removal (method: top-hat filter, factor 0.02); replicates alignment (constant tolerance: 0.2, linear tolerance: 2000 ppm) [21]. Obtained spectra from technical and biological replicates were combined to create one average spectrum per isolate that were used as input for generating peak matrices.

The Clover Biosoft platform was used for ML analyses utilizing pre-processed spectra. Firstly, spectra of 157 training set samples (Table 1) were used to distinguish between HVRTs and non-HVRTs. Three peak matrices were generated using different methods as previously described [21]. The “full spectrum method” uses each mass every 0.5 Da, regardless of its intensity, followed by a total ion current (TIC) normalization of the peak intensities. The “threshold method” (factor 0.01) excluded all peaks with an intensity <1% of the maximum intensity seen in each spectral profile and was coupled with a TIC normalization either before (TICp) or after (pTIC) removal of the minor peaks. For the individual peak identification in spectral profiles, a constant tolerance of 0.5 Da and linear tolerance of 500 ppm was applied [21]. All generated peak matrices were used as input for ML analyses utilizing unsupervised and supervised algorithms [22]. As an unsupervised algorithm, principal component analysis (PCA) was tested. For supervised algorithms, support vector machine (SVM), partial least square discriminant analysis (PLS-DA), k-nearest neighbor (KNN), and random forest (RF) were utilized. For internal validation, a 10-fold cross-validation was applied. Based on cross-validation results, confusion matrix, area under receive operating characteristic (AUROC) curve, and area under precision recall (AUPR) curve were used to estimate the prediction models’ performance. Secondly, HVRTs pre-processed spectra only were used for MS/ML subtyping.

The two best performing models in the cross validation (Table 2) were externally validated using pre-processed spectra of 83 new clinical isolates (validation set, Table 1) to evaluate their reliability and robustness.

Representative spectral profiles from different RTs are visualized in Fig. 1. Spectra of all isolates were correctly identified as C. difficile (Supplementary File S2).

Average spectra of 157 isolates (training set) were used to create three different peak matrices being tested by PCA (Fig. 2). When using the “full spectrum method” for peak matrix generation, PCA failed to separate HVRT from non-HVRT isolates (Fig. 2A).

Better separation was achieved, when either of the two “threshold methods” (pTIC and TICp) was applied combined with PCA (Fig. 2B, C). However, these test procedures were still insufficient to reliably separate HVRTs from non-HVRTs due to a subset of HVRTs belonging to RT027/176 merging with non-HVRTs (Fig. 2).

The TICp method showed the best separation between both groups and was thus used for downstream supervised ML analyses. SVM classification results displayed again only partial discrimination between HVRT and non-HVRT strains, as RT027/176 isolates clustered mostly together with non-HVRTs (Fig. 3A). In contrast, RF, PLS-DA, and KNN prediction models allowed for a much better discrimination (Fig. 3B–D).

After 10-fold cross validation of the supervised ML models, an overall accuracy of 99.4% was observed for the RF model, 98.7% for the PLS-DA model, 93.0% for the KNN model, and 78.3% for the SVM model (Table 2). The superior performances of the RF and PLS-DA models to reliably discriminate between HVRTs and non-HVRTs were confirmed by the ROC and PR curves with respective mean values of AUROC and AUPRC of 0.98 and 0.99 for RF, 0.99 and 1 for PLS-DA, 0.94 and 0.96 for KNN, and 0.74 and 0.79 for SVM (Supplementary File S3).

The two most discriminative algorithms (RF and PLS-DA) were next used for models’ external validation. When tested with the MALDI-TOF spectra of 83 new clinical C. difficile isolates (validation set) that were added blinded to the models. Both prediction models produced promising classification results with total accuracies of 98.8% (RF) and 97.6% (PLS-DA) (Table 3).

The respective mean values for AUROC and AUPRC confirmed the high performance of both models, with 0.98 and 0.92 (RF), and 0.96 and 0.97 (PLS-DA) (Supplementary File S4).

Given the promising separation of HVRTs and non-HVRTs by the RF and PLS-DA models, we wondered whether these two models could further discriminate between different HVRTs used in this study. However, when spectra of all isolates of the training set were included, no clear separation between specific HVRTs was attainable (Supplementary File S5). Thus, we next tested, if a better separation of certain HVRTs can be achieved by a two-step procedure, in which HVRTs were identified in a first step as described above. Next, we created a second peak matrix based on the average MALDI-TOF spectra of the training set HVRTs using the TICp method. With HVRTs’ peak matrix being used as input for PCA, three different clusters were observed (Fig. 4).

One cluster encompassed RT023 isolates, another cluster comprised RT027/176 isolates, while isolates of RT045, RT078, RT126, and RT127 grouped together in a third cluster. RF and PLS-DA algorithms confirmed the initial PCA findings (Fig. 5).

10-fold cross-validation resulted in 100% accuracy for both models (Table 4 and Supplementary File S6).

External validation of the two prediction models was next performed using average spectra of all 39 HVRT isolates from the validation set (Table 1). Overall accuracies of 92.3% (RF) and 97.4% (PLS-DA) were achieved (Table 5). However, three RT023 isolates were misclassified as RT045/078/126/127 (RF), while only one RT078 isolate was misclassified as RT023 (PLS-DA) (Table 5 and Supplementary File S7).