Predictive proteomics: Binary classification of Streptococcus pneumoniae vaccine types via MALDI-TOF MS and...

Laboratory surveillance of Streptococcus pneumoniae serotypes is crucial for the effective implementation of vaccines against invasive pneumococcal disease (IPD). The conventional serotyping method, the Quellung reaction, is time-consuming and expensive. Matrix-assisted laser desorption/ionization time-of-flight mass spectrometry (MALDI-TOF MS) has transformed clinical microbiology and offers a potential platform for rapid, cost-effective capsular typing based on protein profiles.

To evaluate the feasibility of using MALDI-TOF MS coupled with machine learning as an adjunctive or screening method for pneumococcal capsular typing. Specifically, we aimed to develop classification models to discriminate between isolates of PCV13 vaccine serotypes (VT) and non-PCV13 serotypes (NVT).

A custom spectral library was constructed from 133 clinical isolates of Streptococcus pneumoniae. The dataset, designed to reflect local epidemiology, comprised 147 spectra from all PCV13 serotypes and 210 spectra from the ten most prevalent non-PCV13 serotypes in Argentina. The non-uniform distribution of spectra across serotypes mirrored clinical availability and recent local IPD case data (approximate VT:NVT ratio of 35-40:60-65%), grounding the model in real-world epidemiology. The Clover MS Data Analysis Software was used to build a serotyping pipeline. After an initial unsupervised exploration using Principal Component Analysis (PCA), the dataset was split into training (70%, n = 249) and validation (30%, n = 108) sets. Multiple supervised classifiers-Partial Least Squares Discriminant Analysis (PLS-DA), Support Vector Machine (SVM), Random Forest (RF), K-Nearest Neighbors (KNN) and Light Gradient-Boosting Machine (LightGBM)-were trained and evaluated.

Unsupervised analysis (PCA, HCA, K-Means) showed limited discriminatory power for specific serotypes but highlighted the potential of machine learning for this task. In the supervised analysis, all classifiers demonstrated predictive potential. The Random Forest and LightGBM algorithms performed best, with Random Forest achieving a peak accuracy of 95.37% in distinguishing VT from NVT isolates using a 10-fold cross-validation strategy.

This study provides a proof-of-concept for an AI-powered, MALDI-TOF MS-based serotyping platform. The synergy of this technology with machine learning presents a promising, scalable alternative to traditional methods. Future work will focus on expanding the spectral database and refining the algorithms to improve accuracy and robustness, potentially establishing a new standard for pneumococcal surveillance.