Machine Learning Approach for Candida albicans Fluconazole Resistance Detection Using Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry

• Published in: Frontiers in microbiology Vol. 10; p. 3000
• Main Authors: Delavy, Margot, Cerutti, Lorenzo, Croxatto, Antony, Prod'hom, Guy, Sanglard, Dominique, Greub, Gilbert, Coste, Alix T
• Format: Journal Article
• Language: English
• Published: Switzerland Frontiers Media S.A 14.01.2020
• Subjects: Candida albicans; diagnostic; fluconazole resistance; machine learning; MALDI-TOF MS; Microbiology; Candida albicans; MALDI-TOF MS; diagnostic; machine learning; fluconazole resistance
• ISSN: 1664-302X; 1664-302X
• DOI: 10.3389/fmicb.2019.03000

causes life-threatening systemic infections in immunosuppressed patients. These infections are commonly treated with fluconazole, an antifungal agent targeting the ergosterol biosynthesis pathway. Current Antifungal Susceptibility Testing (AFST) methods are time-consuming and are often subjective. Moreover, they cannot reliably detect the tolerance phenomenon, a breeding ground for the resistance. An alternative to the classical AFST methods could use Matrix-Assisted Laser Desorption/Ionization Time-of-Flight (MALDI-TOF) Mass spectrometry (MS). This tool, already used in clinical microbiology for microbial species identification, has already offered promising results to detect antifungal resistance on non-azole tolerant yeasts. Here, we propose a machine-learning approach, adapted to MALDI-TOF MS data, to qualitatively detect fluconazole resistance in the azole tolerant species . MALDI-TOF MS spectra were acquired from 33 clinical strains isolated from 15 patients. Those strains were exposed for 3 h to 3 fluconazole concentrations (256, 16, 0 μg/mL) and with (5 μg/mL) or without cyclosporin A, an azole tolerance inhibitor, leading to six different experimental conditions. We then optimized a protein extraction protocol allowing the acquisition of high-quality spectra, which were further filtered through two quality controls. The first one consisted of discarding not identified spectra and the second one selected only the most similar spectra among replicates. Quality-controlled spectra were divided into six sets, following the sample preparation's protocols. Each set was then processed through an R based script using pre-defined housekeeping peaks allowing peak spectra positioning. Finally, 32 machine-learning algorithms applied on the six sets of spectra were compared, leading to 192 different pipelines of analysis. We selected the most robust pipeline with the best accuracy. This LDA model applied to the samples prepared in presence of tolerance inhibitor but in absence of fluconazole reached a specificity of 88.89% and a sensitivity of 83.33%, leading to an overall accuracy of 85.71%. Overall, this work demonstrated that combining MALDI-TOF MS and machine-learning could represent an innovative mycology diagnostic tool.