MSACL 2022 Abstract

INTRODUCTION:
Over the last years, MALDI-TOF Mass Spectrometry has demonstrated its utility for nontuberculous mycobacteria identification. However, the accurate identification of close related species and subspecies is still challenging. One of the mycobacterial groups with more clinical interest is Mycobacterium abscessus complex, which is composed of three subspecies that show different antibiotic susceptibility patterns: M. abscessus, M. massiliense and M. bolletii. Their differentiation is currently performed by laborious molecular tests, like PCR-reverse hybridization or gene sequencing. The aim of this study was to apply MALDI-TOF MS and Machine Learning algorithms for accurate differentiation of these three subspecies.

METHODS:
A total of 82 clinical isolates of M. abscessus complex were analysed. Of these, 24 isolates previously identified at subspecies level by whole genome sequencing were included in the training set for the development of prediction models (15 M. abscessus, 7 M. massiliense and 2 M. bolletii) and 58 isolates identified by PCR-reverse hybridization (GenoType NTM-DR, Hain Lifescience, Nehren, Germany) were included in the validation set and used for external validation (37 M. abscessus, 16 M. massiliense and 5 M. bolletii). All isolates were analysed in triplicate by MALDI Biotyper (Bruker Daltonics, Bremen, Germany) and 3 spectra per spot were recorded. The spectra were processed using Clover MS Data Analysis software (Clover Biosoft, Granada, Spain). Three Machine Learning algorithms were applied: Partial Least Squares (PLS), Support Vector Machine (SVM) and Random Forest (RF).

RESULTS:
Internal validation of the training set (n=24) using k-fold cross-validation (k=10) yielded 100% of correct classification in the three algorithms used. For external validation (n=58), PLS algorithm correctly identified 58.6% of isolates and SVM identified 79.3% of isolates. RF showed to be the best algorithm, achieving 96.5% of correct identification at subspecies level. Using RF, only 2 strains of M. abscessus were misclassified as M. massiliense. Protein peak analysis found 3 specific peaks (m/z 3460, 4652 and 5672) for M. bolletii. Although no specific peaks were found for the other two subspecies, up to 4 protein peaks (m/z 2174, 3087, 3120 and 3864) were present in most M. abscessus or M. massiliense isolates.

CONCLUSION:
This study suggests that the use of Machine Learning analysis could be promising for accurate identification of the three M. abscessus subspecies. Therefore, the implementation of this methodology may represent a simple, cost-effective and rapid way to differentiate M. abscessus complex isolates in clinical microbiology laboratories.