Exploring Streptococcus pneumoniae capsular typing through MALDI-TOF mass spectrometry and machine-learning algorithms in Argentina: Identifying prevalent NON PCV13 serotypes alongside PCV13 serotypes

Laboratory surveillance of serotypes plays a crucial role in effectively implementing vaccines to prevent invasive pneumococcal diseases. The conventional method of serotyping, known as the Quellung reaction, is both time-consuming and expensive. However, the emergence of MALDI-TOF MS technology has revolutionized microbiology laboratories by enabling rapid and cost-effective serotyping based on protein profiles. In this study, we aimed to investigate the viability of utilizing MALDI-TOF MS technology as an adjunctive and screening method for capsular typing of . Our approach involved developing classification models based on MALDI-TOF MS to discern between strains originating from PCV13 (13-valent pneumococcal conjugate vaccine) and NON PCV13 isolates. Firstly, we established a comprehensive spectral database comprising isolates of serotypes present in the PCV13 vaccine, along with the top 10 most prevalent NON PCV13 serotypes based on local epidemiological data. This database served as a foundation for developing unsupervised models utilizing MALDI-TOF MS spectra, which enabled us to identify inherent patterns and relationships within the data. Our analysis involved a dataset comprising 215 new isolates collected from nationwide surveillance in Argentina. Our approach involved developing classification models based on MALDI-TOF MS to discern between strains originating from PCV13 (13-valent pneumococcal conjugate vaccine) and NON PCV13 isolates. Although our findings revealed suboptimal performance in serotype classification, they provide valuable insights into the potential of machine learning algorithms in this context. The sensitivity of the models ranged from 0.41 to 0.46, indicating their ability to detect certain serotypes. The observed specificity consistently remained at 0.60, suggesting a moderate level of accuracy in identifying non-vaccine serotypes. These results highlight the need for further refinement and optimization of the algorithms to enhance their discriminative power and predictive accuracy in serotype identification.By addressing the limitations identified in this study, such as exploring alternative feature selection techniques or optimizing algorithm parameters, we can unlock the full potential of machine learning in robust and reliable serotype classification of . Our work not only provides a comprehensive evaluation of multiple machine learning models but also emphasizes the importance of considering their strengths and limitati