Automated web-based typing of Clostridioides difficile ribotypes via MALDI-TOF MS

Abstract

Background: Clostridioides difficile is a major cause of hospital-acquired diarrhea and a driver of nosocomial outbreaks, yet rapid, accurate ribotype identification remains challenging. We sought to develop a MALDI-TOF MS-based workflow coupled with machine learning to distinguish epidemic toxigenic ribotypes (RT027 and RT181) from other strains in real time.

Results: We analyzed MALDI-TOF spectra from 379 clinical isolates collected across ten Spanish hospitals and identified seven discriminant biomarker peaks. Two peaks (2463 and 4993 m/z) were uniquely associated with RT027, while combinations of five additional peaks reliably identified RT181. Our classifiers-implemented both in the commercial Clover MSDAS platform and the open-access AutoCdiff web tool-achieved up to 100% balanced accuracy in ribotype assignment and proved robust in real-time outbreak simulations.

Conclusions: This study demonstrates that MALDI-TOF MS combined with tailored machine learning can deliver rapid, high-precision ribotype identification for C. difficile. The freely available AutoCdiff models ( https://bacteria.id ) offer an immediately deployable solution for clinical laboratories, with the potential to enhance outbreak surveillance and control.

Keywords: Biomarker peaks; Classification; Clostridioides difficile; Clostridium difficile; MALDI-TOF MS; Machine learning; Neural network; Outbreak; Random forest; Ribotyping.

Fig. 1

DBLR-FS graphical model. The training of this algorithm is divided into two phases: (1) Kernelization: Both the MALDI-TOF training spectra and a data subset (relevance vectors) are combined into a linear kernel to define the dual space with the same dimensions as number of samples. (2) Classification: The linear kernel along with the dual weights define a linear regression over outputs . Moreover, the prior distribution of (dependent on , see Supplementary Material 1) is initialized with higher values at peaks known to be relevant based on expert knowledge from microbiologists. Finally, output is introduced on a Bayesian LR to determine the most likely ribotype

Fig. 2

FA-VAE training and prediction phases. Training phase (left): Paired MALDI-TOF spectra and known ribotypes are fed into two modules. The VAE encoder–decoder compresses each spectrum () and reconstructs the original signal, while the Bayesian LR module embeds the ribotype label (). The two vectors are then merged into a single global encoding (G) that captures both spectral and ribotype information for each isolate. Prediction phase (right): A new spectrum is passed through the VAE encoder to produce , which is projected into the global space G. From G, the model extracts the ribotype vector and uses the Bayesian classifier to output the most likely ribotype (for example, RT181 with 80% probability)

Fig. 3

Screenshot of the upload page web interface

Fig. 4

Screenshot of the results page web interface

Fig. 5

Peaks used for creation of predictive models for Clover MSDAS. a Peak 2463 m/z. b Intensities of 2463 m/z according to model categories. c Peak 3353 m/z. d Intensities of 3353 m/z. e Peaks 4933 and 4993 m/z. f Intensities of 4933 m/z. g Intensities of 4993 m/z. h Peak 6187 m/z. i Intensities of 6187 m/z. j Peaks 6651 and 6710 m/z. k Intensities of 6651 m/z. l Intensities of 6710 m/z

Fig. 6

Proposed laboratory workflow with the implementation of C. difficile typing by MALDI-TOF MS methodology presented in this paper. The workflow shows an Xpert^®-dependent option and a direct-culture option for laboratories without a Xpert^® system available. PFGE: Pulsed-Field Gel Electrophoresis. WGS: Whole-Genome Sequencing