Deep learning-based analysis and identification of single-particle mass spectra of bacteria.

Abstract

Single-particle mass spectrometry (SPMS) has the potential to identify bacterial species. However, this crucial topic has received limited attention in research. This investigation aims to fill this gap by combining SPMS with supervised learning algorithms to distinguish six bacterial species. The study begins by collecting particle size and mass spectra data for six bacteria and four biomass combustion products (BCPs) using SPMS. These data are used to compare particle sizes and create a comprehensive dataset containing mass spectra for all ten subjects. The mass spectra peak ratio method is then employed to differentiate between bacteria and BCPs, highlighting their distinct distributions of PO₃^-/PO₂^- and CNO^-/CN^- in scatter plots. In addition to this, the study compares the mass spectrometry ion features of bacteria and BCPs and evaluates the classification performance of support vector machines (SVM), multi-layer perceptrons (MLP), and convolutional neural networks (CNN) using five criteria. The Score-Weighted Class Activation Mapping (Score-CAM) method is used to visualize and analyze the CNN models, extracting and analyzing the key ionic features that the CNN models relied on for classification. The results demonstrate that the mass spectra peak ratio method effectively distinguishes bacteria from BCPs. The CNN and MLP algorithms can not only accurately distinguish between bacteria and BCPs but also precisely identify different types of bacteria. The overall classification accuracy of the CNN and MLP models exceeds 96%. The key ions obtained using the Score-CAM method exhibit varying degrees of signal intensity differences among different bacteria, which helps to understand the compositional differences between various bacterial species. This study provides an effective methodology for the in-depth analysis of SPMS data.