Predicting Staphylococcus aureus spa Types Using MALDI-TOF Mass Spectrometry and Machine Learning

Staphylococcus aureus is a leading cause of intramammary infections (IMI) in Canadian dairy herds and is frequently isolated from both clinical and subclinical mastitis cases. Persistent IMI caused by S. aureus are of particular concern, as they are often associated with economic losses due to reduced milk quality and yield, and its contagious nature allows cow-to-cow transmission during milking. Due to its reproducibility and discriminatory power, spa typing is employed in research setting and could possibly be employed in epidemiological investigations to track the transmission of S. aureus strains within and between herds, and to explore associations with virulence factors. For instance, a given spa type could be associated with IMI persistency and clinical outcome. Matrix-Assisted Laser Desorption/Ionization Time-of-Flight Mass Spectrometry (MALDI-TOF MS) is widely used for rapid and accurate bacterial species identification and has recently shown potential for characterizing genotypic traits of S. aureus. The objective of this study was to evaluate the performance of machine learning algorithms in predicting the 6 most common S. aureus spa types found in Canadian dairy herds, using only MALDI-TOF MS spectra. A total of 373 S. aureus isolates from the Mastitis Pathogens Culture Collection that had previously undergone spa typing were included in the study. These isolates underwent MALDI-TOF MS, and the resulting spectra were pre-processed, including smoothing, baseline correction, intensity calibration, and peak alignment. From the training data, 132 reference peaks were identified, and feature matrices were generated for training and validation sets. Principal Component Analysis (PCA) was used to investigate patterns within the data, while Random Forest (RF) models were trained to predict the 6 main S. aureus spa types. The PCA reveal that certain spa types were clearly distinguishable. The final RF models achieved a good accuracy for 3 out of the 6 spa types. For spa type t529, the accuracy was 89.0% (95% CI: 83.3, 93.2%), resulting in an AUC of 0.88. For spa type t605, the accuracy was 99.4% (95% CI: 96.8% - 99.9%), resulting in an AUC of 0.94. Finally, for spa type t13401, the accuracy was 97.1% (95% CI: 93.4% - 99.1%), resulting in an AUC of 0.89. For the other 3 spa types, model performance was poor, with AUC values of 0.53, 0.68, and 0.62 for spa types t359, t267, and t2445, respectively. These findings suggest that MALDI-TOF MS can reasonably predict, and at no extra costs, certain spa types of S. aureus, highlighting its potential as a cost-effective and accessible alternative to more resource-intensive molecular typing methods in epidemiological studies.