Early antifungal resistance prediction based on MALDI-TOF mass spectrometry and machine learning

Abstract

Antimicrobial resistance (AMR) is a significant global health threat. Recent studies have shown that combining MALDI-TOF mass spectrometry with machine learning (ML) algorithms can accelerate AMR determination. However, these efforts have predominantly focused on bacterial pathogens. Given the significant morbidity, mortality and healthcare costs associated with fungal infections and their evolving antifungal resistance to treatment, there is a need for precision medicine approaches that enable early and accurate detection of antifungal resistance. We developed a machine-learning pipeline that integrates MALDI-TOF mass spectrometry data and drug features to predict antifungal resistance and identify spectral biomarkers for antifungal resistance. By leveraging the DRIAMS dataset, we included 658 pathogen spectra linked to 3,046 phenotypic antifungal resistance results. This dataset covered three drug classes and seven yeast species. The model was trained using categorical, phenotypic antifungal susceptibility testing results as ground truth. Our study systematically investigated the influence of various dimensionality reduction methods on MALDI-TOF mass spectra, antifungal encodings, and machine learning models using nested cross-validation to evaluate predictive performance. We identified that applying principal component analysis to MALDI-TOF mass spectra for dimensionality reduction, and training a multi-layer perceptron yielded the highest and most stable performance for the prediction of antifungal resistance. Our method achieved an AUPRC of 0.77 across the 10 highest-performing species-drug pairs and corresponding Matthews correlation coefficient of 0.69. The model demonstrated the best performance for the species-drug combinations of Candida albicans with micafungin, Candida parapsilosis with fluconazole, and Saccharomyces cerevisiae with itraconazole and fluconazole. By comparing established species-based guidelines, susceptibility test results, and machine learning predictions, we estimated that integrating our algorithm into antifungal selection could help avoid prescriptions to likely resistant pathogens in approximately 3 out of 10 patients for whom standard guidelines recommend such treatments.