Identification of bacterial spores by Raman and OPTIR spectroscopy coupled with machine learning.

Accurate identification of bacterial endospores is essential for contamination tracing in the food industry; however, rapid species-level discrimination remains difficult with conventional methods. In this study, co-localised optical photothermal infrared (OPTIR) and Raman microspectroscopy were evaluated for the single-spore classification of six food-relevant spore-forming species, each represented by a single strain or isolate, spanning the Bacillus subtilis group, the Bacillus cereus group, and Heyndrickxia sporothermodurans. Co-localised OPTIR and Raman spectra were acquired from the same individual spores and analysed separately or after multimodal concatenation. After preprocessing, unsupervised exploration by principal component analysis showed biologically meaningful inter-species structure but incomplete class separation, indicating the need for supervised modelling. A one-dimensional convolutional neural network was therefore developed and evaluated under a leave-one-day-out cross-validation design using independent acquisition days, with a radial basis function support vector machine used as baseline. Among the tested representations, the concatenated Raman + OPTIR model achieved the best external performance, with a mean overall accuracy of 87.26%, balanced accuracy of 86.82%, and macro-F1 score of 86.73%, outperforming Raman alone, OPTIR alone, and the support vector machine baseline. The gain from multimodal fusion was class-dependent. These results demonstrate that co-localised OPTIR and Raman spectroscopy provide complementary biochemical information at the individual spore level and support a promising culture-free strategy for species-level identification of food-relevant bacterial spores.