FT-IR-based strategy for Streptococcus pneumoniae serotyping using machine-learning classifiers

Rapid serotype identification of circulating Streptococcus pneumoniae is essential for both effective epidemiological surveillance and clinical management. The gold standard serotyping method (Quellung reaction) is time-consuming, labor-intensive and expensive. Fourier transform infrared (FT-IR) spectroscopy using the IR-Biotyper^® (Bruker Daltonics GmbH, Bremen, Germany) system has recently been proposed as a rapid and low-cost technique for serotype identification, based on spectral analysis of capsular polysaccharide components. The integration of machine learning algorithms within the IR-Biotyper^® enables automated analysis and classification of FT-IR spectra, allowing rapid prediction of pneumococcal serotypes. The present study evaluated the application of machine learning algorithms for serotype identification in clinical S. pneumoniae isolates using the IR-Biotyper^® system.

A database with 128 isolates, representing 30 different serotypes, was used to develop four classifiers. A global classifier was designed to predict twelve serotypes (3, 4, 6A/6C, 7C/7F, 12F, 15A/15B/15C, 19A/19F, 22F, 23A/23B/23F, 24B/24F, 35B and 38) achieving 99% accuracy. Additionally, three sequential sub-classifiers were developed to differentiate; 6A from 6C, 15A from 15B and 15C, and 23A from 23B and 23F, with validation dataset accuracies of 100%, 89% and 100% respectively. The evaluation of unknown serotypes showed 92% accuracy.

FT-IR spectroscopy showed high concordance with Quellung reaction, supporting its use as a rapid and cost-effective method for serotype identification of S. pneumoniae. Sequential sub-classification represents a practical strategy that could be used to classify additional serotypes of epidemiological relevance.