Classification of bacterial biological warfare agent simulants through 2D Py-GC/MS data coupled with deep learning

Identification and detection of biological warfare agents (BWAs) constitute a critical task for both biodefence applications and public health protection. Gas Chromatography coupled with Mass Spectrometry (GC/MS) represents a robust analytical platform due to its sensitivity to Volatile Organic Compounds (VOCs) and Fatty Acid Methyl Esters (FAMEs). However, the high chemical complexity of microorganisms necessitates more comprehensive molecular profiling. In this context, Pyrolysis-GC/MS (Py-GC/MS) is employed to thermally decompose intact bacterial cells, generating high-dimensional chromatographic signatures comprising a broad spectrum of volatile degradation products. In this study, eight bacterial species, including BWA simulants, Bacillus atrophaeus, Francisella philomiragia, Escherichia coli, Streptococcus mitis, Yersinia enterocolitica subsp. enterocolitica, Acinetobacter baumannii, Agromyces mediolanus, and Staphylococcus epidermidis, were analyzed at three concentration levels using a Py-GC/MS device. Four complementary data representations were derived from each measurement: 2D GC×MS chromatograms, 1D Total Ion Count (TIC) signals, FAME feature profiles, and Principal Component (PC) projections. The resulting dataset is publicly released to support reproducibility and further research. Machine Learning (ML) and Deep Learning (DL) models are subsequently employed to extract discriminative features and perform classification. For Gram type classification, each data representation-model pair is trained and validated on four bacterial classes and evaluated on four unseen classes, enabling a direct assessment of out-of-distribution generalization. For species-level identification, models are trained and evaluated on a subset of four bacterial classes, including two BWA simulants. The experimental results demonstrate that 2D chromatographic representations, when coupled with advanced DL architectures, consistently achieve superior performance in terms of accuracy and F1-score. Furthermore, the Limit of Detection (LOD) is systematically evaluated across all experimental settings, providing additional insight into the sensitivity and robustness of each data representation-model pair.