Fourier transform infrared (FT-IR) spectroscopy (IR Biotyper; Bruker) allows highly discriminatory fingerprinting of closely related bacterial strains. In this study, FT-IR spectroscopy-based capsular typing of Streptococcus pneumoniae was validated as a rapid, cost-effective, and medium-throughput alternative to the classical phenotypic techniques. A training set of 233 strains was defined, comprising 34 different serotypes and including all 24 vaccine types (VTs) and 10 non-vaccine types (NVTs). The acquired spectra were used to (i) create a dendrogram where strains clustered together according to their serotypes and (ii) train an artificial neural network (ANN) model to predict unknown pneumococcal serotypes. During validation using 153 additional strains, we reached 98.0% accuracy for determining serotypes represented in the training set. Next, the performance of the IR Biotyper was assessed using 124 strains representing 59 non-training set serotypes. In this setting, 42 of 59 serotypes (71.1%) could be accurately categorized as being non-training set serotypes. Furthermore, it was observed that comparability of spectra was affected by the source of the Columbia medium used to grow the pneumococci and that this complicated the robustness and standardization potential of FT-IR spectroscopy. A rigorous laboratory workflow in combination with specific ANN models that account for environmental noise parameters can be applied to overcome this issue in the near future. The IR Biotyper has the potential to be used as a fast, cost-effective, and accurate phenotypic serotyping tool for S. pneumoniae.
Keywords: FT-IR spectroscopy; Streptococcus pneumoniae; machine learning; pneumococcus; serotyping.