Likelihood-based bacterial identification approach for bimicrobial mass spectrometry data

Abstract

Mass spectrometry is a potential diagnostic tool for rapid bacterial detection. However, in order to use this technology in clinical settings, it is important to develop sound statistical algorithms that can accurately analyze polymicrobial mass spectrometry data. Here, we propose a likelihood-based bacterial identification algorithm for bimicrobial mass spectrometry data. Specifically, we introduce a two-component mixture model with partially known labels. This method can model peaks with unknown origins. It also considers errors in mass-to-charge ratios and intensities of peaks between observed and reference mass spectra. Coupled with a decoy strategy, the likelihood is used to identify bacterial species and to measure uncertainty of such identifications. Using two real mass spectrometry datasets, we demonstrate the superior performance of our approach in accurate bacterial identifications, compared to model-free approaches. Example datasets and R codes for the proposed method are freely available under MIT license at https://github.com/soyoungryu/BacID.