Clostridioides difficile, transmitted rod, is the leading cause of infectious, antibiotic-related, nosocomial diarrhoea (Czepiel, Drozdz et al. 2019; Markovska, Dimitrov et al. 2023) . Certain strains of C. difficile can produce toxins that are responsible for intestinal damage and play an important role in the pathogenesis and evolution of the disease (Leffler and Lamont 2015; Krutova, Kinross et al. 2018) . Several C. difficile ribotypes (RTs) have been described, including those that show higher pathogenicity and/or involved in nosocomial outbreaks (Davies, Ashwin et al. 2016; ECDC 2022) . Ribotype B1/NAP1/027 (RT027) is of great clinical interest and has been reportedly associated with hospital outbreaks (He, Miyajima et al. 2013; Viprey, Davis et al. 2022) . Strains belonging to

RT027 are considered as <hypervirulent= due to the hyperproduction of toxins (Markovska, Dimitrov et al. 2023) . Other RTs, such as the hypervirulent RT078 and RT181, have recently been reported as highly prevalent (9.0%) and emerging in European countries, respectively (Kachrimanidou, Metallidis et al. 2022; Viprey, Davis et al. 2022) .

Rapid and reliable characterization of C. difficile is required to tackle transmission and avoid fatal outcomes (Krutova, Kinross et al. 2018) . PCR ribotyping, Multiple Locus Variant Analysis and Pulsed-Field Gel Electrophoresis (PFGE) are Europe and North America's most widely implemented genotyping methods, respectively (Calderaro, Buttrini et al. 2022) . These methods are considered the gold standard to differentiate C. difficile RTs. They are highly informative, but also time-consuming (48-96h respectively from bacterial culture) and laborintensive, requesting highly trained personnel and specific instruments, both factors lacking in most diagnostic laboratories (Bidet, Lalande, et al. 2000; Abad-Fau, Sevilla et al. 2023) . The Xpert® C. difficile BT assay has been implemented worldwide for the rapid differentiation of toxigenic C. difficile strains (Bai, Hao et al. 2021) . This test is fast (50 min from a direct sample) and simple to perform (Whang and Joo 2014) . It allows the detection of toxin B (tcdB), binary toxin (cdt) and the deletion in position 117 of the regulatory gene tcdC associated to RT027; however, the targeted genes do not allow the differentiation of RTs closely related to RT027, such as RT181 or RT176 (Kachrimanidou, Baktash, et al. 2020; Novakova, Kotlebova et al. 2020) . Furthermore, the cost per sample is still high (≥ 54 $, approximately 50€) (Chapin, Dickenson et al. 2011) . Spectrometry (MS), coupled with Machine Learning (ML) algorithms for spectra analysis, holds promise as an efficient method for distinguishing various C. difficile RTs rapidly (1h from bacterial culture) (Calderaro, Buttrini et al. 2022) , with a cost per isolate less than 1$ (Clark, Kaleta et al. 2013) . However, the current state-of-the-art of this approach has notable limitations, lacking a standard procedure for spectra analysis, an accessible code or software, and remaining exploratory in nature.

The objective of this study was to develop an efficient methodology for specific typing of the hypervirulent strains RT027 and RT181 utilizing MALDI-TOF MS and ML algorithms. For this purpose, we analysed two independent tools: Clover Mass Spectrometry Data Analysis Software (MSDAS), a commercial ad hoc software for MALDI-TOF MS spectra analysis (https://cloverbiosoft.com) considered state-of-the-art, and

A total of 7 potential biomarker peaks were identified (Table 1) using Clover univariate analysis. Specific peaks for RT027 differentiation were found at 2463 m/z (Fig. 1a and 1b) and 4933 m/z (Fig. 1c and 1d). A peak at 4993 m/z (Fig. 1c and 1e) was absent in RT027 but present in RT181 and other RTs. Finally, peaks at 3353 m/z (Fig. 1f and 1g), 6187 m/z (Fig. 1h and 1i), 6651 m/z (Fig. 1j and 1k) and 6710 m/z (Fig. 1j and 1l) were shown to differentiate both RT027 and RT181 from Other RTs. Two of these peaks, at 6651 and 6710 m/z, were correlated with two isoforms of the methionine-cleaved ribosomal protein L28 using the Ribopeaks database (https://lcad.deinfo.uepg.br/~ppgca/ribopeaks/). This was also suggested by

the information about protein L28 for C. difficile available in the Uniprot database (https://www.uniprot.org/). The difference between these two isoforms is the aminoacidic substitution G9D (Table 2), according to Uniprot database. Overall, these biomarker peaks were used as input for all Clover MSDAS algorithms and as a prior Expert Knowledge EK in the DBLR-FS algorithm of AutoCdiff.

Differentiation of C. difficile ribotypes

In this study, it was demonstrated that the implementation of ML methods for the differentiation of C. difficile RTs based on MALDI-TOF MS spectra is a cost-effective, easy-toapply, and reliable tool with great potential for rapid screening of these isolates, offering a real alternative to complex typing methods such as PCR ribotyping and Pulsed-Field Gel Electrophoresis (PFGE). Its implementation could reduce the turn-around time for definitive characterization of the isolates to 24-48 hours from the reception of the clinical sample in the laboratory. The hands-on time of the method is approximately of 1 hour once the C. difficile isolates are grown on solid agar media.

ML-based classification methods were applied to the analysis of protein spectra obtained by MALDI-TOF MS for 1) the rapid differentiation of RT027 from other C. difficile isolates classified as <presumptive RT027= by the rapid Xpert® C. difficile BT assay and, 2) the rapid discrimination of RT027 and RT181 from RT027-like RTs and other RTs prevalent in Europe, both on independent test sets and real-time outbreak scenarios. Different ML algorithms and pre-processing pipelines consistently achieved this demonstration of versatility. Firstly, using a variety of algorithms trained and tested within two different packages: Clover MSDAS, a state-of-the art, proprietary software; and secondly, through AutoCdiff, an open-source package specifically developed for differentiation of C. difficile RTs within this study.

A total of 4 biomarker peaks (2463, 4933, 4993 and 6651 m/z) were found with Clover MSDAS software to differentiate the categories "RT027" and "Other RTs" in Experiment 1, and a total of 7 peaks (2463, 3353, 4933, 4993, 6187, 6651, 6710 m/z) when RT027 was differentiated from RT181 and Other RTs in Experiments 2 and 3. Three of these biomarkers correlated with those automatically found by DBLR-FS (4993, 6651, and 6710 m/z) for the same purpose. Some of these biomarkers have already been described in previous studies: protein peaks between 6647-6654 m/z and between 6707-6712 m/z (6651 m/z and 6710 m/z in this study, respectively) have been reportedly found in RT027 and RT176 (Emele, Joppe et al. 2019; Flores-Trevino, Garza-Gonzalez et al. 2019) . These peaks could correspond to two isoforms of the ribosomal protein L28. In this study, the isoform 6710 m/z was found in RT027 and RT181, suggesting the great similarity between these two RTs, while the isoform 6651 m/z was specifically found in the rest of the analysed RTs, including those classified as <presumptive RT027=. Therefore, the only RT176 isolate included in this study showed the 6651 m/z peak, contrary to the results showed by Emele et al., 2019 (Emele, Joppe et al. 2019) . Further, C. difficile RT176 isolates should be analysed to confirm this discrepancy. Furthermore, the peak at 3353 m/z was found to be specific for RT027 and RT181, as well as the peak at 6710 m/z, suggesting the possibility that it is the same isoform of the L28 with double charge ions. However, no biomarker peak with double charge was found for the other isoform of the L28 protein in addition to peak 6651 m/z.

The peak at 4928 m/z (4933 m/z in this study) has been described before as an uncharacterized protein present in RT027 (Corver, Sen et al. 2019) . Our results correlated with this finding and showed that the 4933 m/z peak is specific to RT027 and the 2463 m/z peak, which we suggest may be the same protein with double charge ions. However, the peak at 4993 m/z, absent in RT027 and present in the other two remaining categories ("RT181" and "Other RTs") had not been previously described, probably due to the limited literature about the RT181 available so far.

A recent paper using Clover MSDAS reported a balanced accuracy >95.0% when differentiating hypervirulent C. difficile RTs (Abdrabou, Sy et al. 2023) . However, RT027 and RT176 could not be differentiated, although two separation steps were implemented. Despite not having established a specific category for RT176, this study validates a methodology to differentiate RTs with characteristics like RT176 ("presumptive RT027" by Xpert® C. difficile BT and cause of nosocomial outbreaks), as is the case of RT181. Our web application tool serves a dual purpose in this context. It acts as a practical platform for researchers to apply this methodology for their own data analysis and, importantly, functions as a data collection hub. By aggregating user-contributed data, the application becomes a vital resource for continually enhancing and refining our ML models. This evolving database is expected to significantly improve our models' capability to distinguish an expanding array of RTs with similar characteristics, paving the way for more precise and comprehensive typing in the near future.

In this study, Light-GBM and RF, operating within Clover MSDAS, exhibited superior performance in distinguishing RT027 from other "presumptive RT027" isolates. This pattern was consistent across all algorithms, excluding PLS-DA, when differentiating RT027 and RT181 from the rest of the analyzed RTs in Experiment 2. Furthermore, all algorithms achieved 100% accuracy in classifying the outbreak C. difficile isolates.

In AutoCdiff, the DBLR-FS with EK and RF algorithms showed the highest accuracy in differentiating RT027 from other "presumptive RT027" isolates. The Bayesian methods in AutoCdiff, like DBLR-FS, do not need cross-validation and provide valuable uncertainty measurements, enhancing their practical utility. In our second experiment, only one isolate, RT173 from the <Other RTs= group, was misclassified as RT027, likely due to the presence of the 4933 m/z peak, commonly associated with RT027. This highlights the need for further analysis with a larger collection of this RT. Significantly, all algorithms in AutoCdiff correctly classified the isolates in the C. difficile outbreak scenarios, underscoring their effectiveness and ease of use in real-world applications.

Therefore, our results highlight the efficacy of the described methodology for rapid screening of C. difficile isolates. The implementation of the developed methodology would allow efficient control of the detected cases and early treatment of the affected patients, optimizing hospital resources and representing a substantial improvement in the workflow of clinical microbiology laboratories. In our commitment with the wide diffusion of MALDI-TOF MS-based bacterial typing, we have released the developed classification models for both Clover MSDAS and AutoCdiff, alongside the dataset used to train all algorithms. While access to Clover MSDAS offers a free plan with limited functionality and a paid version for full access, the owner generously offers access to the free full version for three months, and its use demands minimal knowledge about ML. In contrast, all classification algorithms crafted within AutoCdiff are freely accessible at www.bacteria.id. Users simply need to register and upload their MALDI-TOF MS spectra to the website for the AutoCdiff algorithms to classify them into RT027, RT181, or another RT. Besides, the dataset containing the MALDI-TOF spectra from the isolates analyzed in this study has been made accessible

Although this study represents a step forward in automating MALDI-TOF MS data for bacterial typing, it still has limitations. The first one is the number of isolates analyzed, sourced exclusively from Spain but sourcing from different hospitals of the country. Through the use of the web-based classification models at www.bacteria.id, we expect to contact with other research groups interested in the validation and expansion of the collection of C. difficile isolates. Due to the importance of RT027 and RT181, in this study we focussed mainly in these two RTs but we plan to differentiate other important RTs in future collaborative studies.

Besides, the need for a bacterial culture for MALDI-TOF analysis prolongs the detection period for 24-48h until colony growth, whereas the Xpert® C. difficile BT can be performed from direct samples in a shorter time. However, the reduction in costs is significant compared to this technique, and while many microbiology laboratories do not have PCR ribotyping or Xpert® C. difficile BT available, MALDI-TOF is ubiquitous to almost all of them. In conclusion, the integration of MALDI-TOF MS with ML methods presents a promising approach for effectively distinguishing clinically significant C. difficile RTs, offering a streamlined and time and effort-effective alternative to traditional molecular methods. A proposal for implementation of this methodology in the clinical laboratory is showed in Figure 2. Notably, the ML methods described in this study showed a high accuracy for the differentiation of RT027 and RT181 from other C. difficile RTs and their implementation significantly reduces the time to results from 120 hours (PCR-ribotyping) to 24-48h, providing a rapid screening solution, especially in the case of a suspected outbreak caused by hypervirulent RTs. The availability of the developed methods could empower laboratories with limited molecular resources to enhance their diagnostic capabilities. Finally, the findings underscore the transformative potential of ML applied to MALDI-TOF MS data, showcasing its capacity to revolutionize the landscape of C. difficile strain typing in clinical microbiology.

A total of 363 clinical isolates (Table 4) belonging to the most prevalent C. difficile RTs in Spain and Europe were included in the study (ECDC 2022) . A subgroup of 348 isolates were sampled in 10 different hospitals in Spain from faeces of patients with diarrheal symptoms between 2010-2022. The remaining 15 C. difficile isolates sourced from two nosocomial outbreaks that took place in Madrid during May-June 2023 in the Hospital Central de la Defensa Gómez Ulla (HCDGU; n=12) and Hospital General Universitario Gregorio Marañón (HGUGM; n=3). All isolates were analysed by Xpert® C. difficile BT assay -Cepheid, USA- (Bai, Hao et al. 2021) . Subsequently, they were also characterized by PCR-ribotyping and analysis of the amplicons in capillary electrophoresis (Stubbs, Brazier et al. 1999; Indra, Huhulescu et al. 2008; Marin, Martin et al. 2015) . Briefly, the intergenic spacer region located between the 16S rRNA and 23S rRNA encoding regions was amplified by PCR. The DNA fragments were subsequently analysed by capillary electrophoresis. Phylogenetic analysis of the ribotyping profiles was performed using the Bionumerics 5.0 software (bioMérieux, Marcy l9Etoile, France), as described before (Reigadas, Alcala et al. 2018) .

Prior to MALDI-TOF MS analysis, the isolates were plated on Brucella Blood Agar (Beckton Dickinson®, Franklin Lakes, NJ, US) and incubated under anaerobic conditions at 37ºC. Incubation lasted 48h to allow sufficient time for C. difficile growth. Before analysis, all isolates were re-identified using MALDI-TOF MS technology in an MBT Smart MALDI Biotyper (Bruker Daltonics, Bremen, Germany) with the updated database containing 11,096 mass spectra profiles, applying the standard protocol for rapid, on-plate protein extraction with 100% formic acid followed by HCCA (α-Cyano-4-hydroxycinnamic acid) matrix solution.

For the acquisition of MALDI-TOF MS spectra, a small number of colonies of each C. difficile isolate was transferred to the MALDI plate in duplicate, overlaid with 1μl of 100% formic acid, allowed to dry and covered with 1μl of organic HCCA matrix. Two spectra were acquired from each spot on the MALDI plate in the positive ion mode within the 2,000 to 20,000 Da range. Consequently, for each bacterial isolate, a minimum of four spectra were available, consistent with the conditions applied for spectra analysis (Candela, Arroyo et al. 2022) . With a subset of 275 isolates, spectra from three consecutive days were collected to test the repeatability of the method. Excluding the outliers and flat lines, each isolate was represented by an average of 5.26 spectra, with a standard deviation of 1.32. This cumulative approach involved a total of 1,879 spectra in the context of this study.

Each software tool (Clover MSDAS and AutoCdiff) uses a specific preprocessing. Clover MSDAS applied the following pre-processing pipeline: 1) Square root variance stabilization; 2) Smoothing via Savitzky-Golay filter (window length: 11; polynomial order: 3); 3) Baseline subtraction using Top-Hat filter (factor 0.02); 4) Peak alignment with constant mass tolerance of 2 Da and linear mass tolerance of 600 ppm; and 5) TIC normalization. On the other hand, AutoCdiff applied this pre-processing pipeline (Weis, Horn et al. 2020; Weis, Cuénod et al. 2022; Guerrero-López, Sevilla-Salcedo et al. 2023) : 1) Square root variance stabilization; 2) Smoothing via Savitzky-Golay filter (window length: 11; polynomial order: 3); 3) Baseline subtraction using Top-Hat filter (factor 0.02); and 4) TIC normalization. Replicates were not merged but treated as unique spectrum to mimic a real-time scenario in both

Different ML algorithms were employed to process MALDI-TOF MS spectra. Two development approaches were adopted: (i) utilizing commercial ad-hoc software tailored for bioanalysis, and (ii) employing open-source algorithms, developed in Python and integrated into a free, user-friendly web application tool created for this study.

The first was Clover MSDAS, a commercial ad hoc software developed by Clover Bioanalytical Software (Granada, Spain, https://cloverbiosoft.com) for pre-processing of the protein spectra, biomarker detection and automatic typing with well-known ML algorithms: Random Forest (RF), K-Nearest Neighbor (KNN), Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Machine (SVM), and Light Gradient Boosting Machine (Light-GBM). In Clover MSDAS, all algorithms underwent training using biomarkers identified by the software, as described in previous studies (Hamidi, Bagheri Nejad et al. 2022; Busby, Doyle et al. 2023; Candela, Arroyo et al. 2023) . The use of these biomarkers for differentiation of specific categories is referred to as Expert-Knowledge (EK). The extraction of EK involved conducting a univariate analysis in Clover MSDAS, utilizing peak intensities to assign binary labels for Area Under the Curve (AUC) calculations. Peaks yielding an AUC greater than 0.7 were subsequently selected as potential biomarkers.

On the other hand, AutoCdiff is a novel, open-source tool developed in our study, designed to streamline the process of automatic bacteria typing. This tool has been allocated in a webpage (https://bacteria.id) to allow MALDI users to efficiently analyse MALDI-TOF MS spectra by simply uploading raw data. The tool features a comprehensive array of pre-built Bayesian ML models, including sophisticated algorithms like the Factor Analysis-Variational AutoEncoder (FA-VAE) (Guerrero-López, Sevilla-Salcedo et al. 2022) and Dual Bayesian Logistic Regression with Feature Selection (DBLR-FS) (Belenguer-Llorens, Sevilla-Salcedo et al. 2022) . In addition, AutoCdiff includes foundational models such as Random Forest (RF) and Decision Trees (DT), to meet a variety of analytical needs. In AutoCdiff, algorithms employed the complete spectrum for input, unrestricted to specific biomarkers. Specifically, for DBLR-FS, an additional approach allows the incorporation of EK as input to augment the data to speed up the model convergence. Consequently, biomarkers extracted by Clover MSDAS were utilized as Bayesian priors for the model's weights, thereby enriching the model with EK.

Regarding cross-validation Clover MSDAS utilized fixed parameters: RF with a minimum of 2 isolates per split, 1 leaf, 100 estimators, and sqrt maximum features; SVM with a linear kernel; Light-GBM with 32 leaves, a 0.1 learning rate, and 100 estimators; KNN with 5 neighbors; and PLS with 2 components. In AutoCdiff, a 10-fold cross-validation was applied for DT and RF algorithms, experimenting with various parameters like maximum depth (2, 4, 6, 8), minimum isolates per split (2, 4, 6), minimum leaves (1, 2, 4), and maximum features (auto, sqrt, log2). The FA-VAE and DBLR-FS models, due to their probabilistic nature, bypassed the need for cross-validation, enabling direct estimation of model hyperparameters and providing as output probabilities and uncertainties.

To evaluate the accuracy of the available ML algorithms, three experiments were designed: i) an initial experiment was created to specifically differentiate RT027 among the isolates labelled as "Presumptive RT027" by the Xpert® C. difficile BT assay (Figure 3). To facilitate this evaluation, a subset of isolates (n=100) that were categorized as "presumptive RT027" was analysed. These isolates belonged to RT027 (n=42) and other toxigenic RTs (n=58). The classification models were tested with a set of 31 independent isolates belonging to the same categories (Table 4). ii) The second experiment aims to differentiate the main toxigenic ribotypes of our setting, namely RT027 and RT181, directly from the MALDI-TOF MS spectra obtained. To carry out this evaluation, a new dataset (n=269) was built where the most prevalent RTs in Spain and Europe (ECDC 2022; Viprey, Davis et al. 2022) , such as RT001, RT014, RT106, RT207, RT023, RT002, and RT017 were represented (Table 4). Isolates from this experiment were classified into three distinct categories: i) RT027 (n=44), ii) RT181 (n=43) and iii) "Other RTs" (n=182). To validate this model, an independent subset of strains (n=79) from the same 3 categories was used. The distribution of C. difficile strains in the training and testing sets is represented in Table 4.iii) Finally, a third experiment was carried out in real-time with isolates from two nosocomial outbreaks to evaluate the efficacy of ML algorithms in a real-world scenario. All algorithms underwent training with the three different categories of the previous experiment using the whole bacterial collection, encompassing 348 C. difficile isolates (Table 4). Subsequently, these algorithms were put to the test with isolates sourcing from two outbreak scenarios belonging to HCDGU (n=12) and HGUGM (n=3) in real time. (Fig. 3).

The authors declare no conflict of interests. MJA and LM are employees of Clover This work is partially supported by grants PID2020-115363RB-I00, PID2021-123182OB-I00 and TED2021-132366B-I00 funded by MCIN/AEI/10.13039/501100011033, by the project PI18/00997 from the Health Research Fund (FIS. Instituto de Salud Carlos III. Plan Nacional de I+D+I 2013-2016) of the Carlos III Health Institute (ISCIII, Madrid, Spain) partially financed by the European Regional Development Fund (FEDER) 8A way of making Europe9. The funders had no role in the study design, data collection, analysis, decision to publish, or preparation/content of the manuscript. AC (Rio Hortega CM21-00165), DRT (Sara Borrell CD22-00014) and BRS (Miguel Servet CPII19/00002) are funded by ISCIII. AGL and MBS are the recipients of the Intramural predoctoral contracts 2020 and 2022, respectively, from the Health Research Center of the Hospital Gregorio Marañón -IISGM-.

Aligned with our dedication to the FAIR (Findable, Accessible, Interoperable, and Reusable) principles, we have made the dataset from this study publicly accessible on Zenodo. It can be found at https://doi.org/10.5281/zenodo.10370872.

m/z RT027

RT181

RT: Ribotype.

The code for all models implemented in AutoCdiff, which are open-source, is freely accessible at our GitHub repository: https://github.com/aguerrerolopez/Clostridium. This repository includes comprehensive documentation and usage notes to facilitate replication and further exploration of our models. For additional information or specific inquiries regarding the code, the authors are available to respond to reasonable requests. the Curve values are represented in this table.

2463 0.93 3353 0.85 0.90 4933 0.98 4993 0.98 0.77 6187 0.88 0.89 6651 0.97 6710 0.96 ribotypes (RTs) included in this study. L28-M indicates the mass of the protein excluding the N-terminal methionine.

6837

6710

6779 6651