A machine learning approach to detect Clostridioides difficile toxigenic biomarkers by MALDI-TOF-MS

LEDESMA MARTÍN; SOFIA MELIAN; CRISTINA LEGARIA; BARBERIS CLAUDIA; CARLOS VAY

Puerto de Acapulco, Guerrero. México

Congreso; 8vo Simposio de la Sociedad Mexicana de Proteómica; 2019

Sociedad Mexicana de Proteómica

Clostridiodes difficile (CD) is the leading cause of nosocomial diarrhea associated with antibiotic treatment. The infection caused by this bacterium is raising and present a high rate of mortality. The current microbiological diagnosis includes the serological detection of both GDH enzyme (a marker of CD) and toxins A and B (the toxigenic feature), direct from fecal matter. However, toxins detection has been characterized by low sensitivity (~60%). The main goal of this work was to design a simple, trustworthy, and fast diagnosis algorithm to detect differential biomarkers between toxigenic and non-toxigenic CD by MALDI-TOF mass spectra analysis. Forty-three non-duplicated CD isolates (32 toxigenic and 11 non-toxigenic) had been studied during the period 2017-2019 from positive GDH fecal matter of hospitalized patients in Hospital de Clínicas ?José de San Martín.? Briefly, strains were isolated in blood agar with cefoxitin and grown in the anaerobic chamber after the ethanolic shock. MALDI-TOF-MS (Bruker BioTyper 3.1, score>2) was used to performed identification, and seven continuous mass spectra were acquired with a Microflex LT mass spectrometer for each isolate. Finally, an in-house PCR to detect tcdB gen (encoded for toxin B) was used to classify toxigenic status. The dataset contains 336 mass spectra from 43 different microbial isolates, including both technical (~3) and biological (~2) replicates for all strains. Initially, the dataset was randomly partitioned into a training set (60% of isolates) and a test set (40 % of isolates). Programmed feature (peaks) selection was performed in training set with a binary discriminant analysis ranking (BDA) seeking for biomarkers of each condition (toxigenic/non-toxigenic). The extracted features were then used to train a BDA predictive model. Metrics such as accuracy (A), sensitivity (S), and specificity (E), adding to the receiver operating characteristic (ROC) curves were used to evaluate the performance of the model in the test set. Peaks were bioinformatically assigned using databases such as expasy/TagIden, and Uniprot. The top fifteen selected peaks have a performance of A=95%, S=100%, E=80%, and AUC=0.9. Fourteen out of fifteen of the selected peaks were absent in the toxigenic strains, some of these peaks were assigned as phage proteins, which correlates with features of PaLoc chromosomal insertion site and the prophage modulation of toxin production. In summary, MALDI-TOF-MS could be used as a complementary tool to differentiate tcdB+/tcdB- strains of CD, moreover the ability to classify non-toxigenic strains (100% S) would be promising for confirm this status and avoid the antibiotic overtreatment of this patient.