Optical Photothermal Infrared (OPTIR) spectroscopy assisted by machine learning for lactic acid bacteria identification at strain level

Abstract

Lactic acid bacteria (LAB) are widely used in food, health, and biotechnology sectors, where accurate strain level identification is critical. Conventional methods, such as 16S rRNA sequencing, PCR-based fingerprinting (RAPD, AFLP), and MALDI-TOF mass spectrometry are powerful tools to identify bacteria at species level but often fail to resolve closely related strains due to limited taxonomic resolution, protocol sensitivity, or database dependence. In this study, we evaluated the capacity of Optical Photothermal Infrared (OPTIR) spectroscopy, a single-cell vibrational imaging technique, combined with supervised neural networks, to classify LAB at both species and strain levels. A total of 13 strains were analysed, including five Lactiplantibacillus plantarum, one Lactiplantibacillus pentosus, one Limosilactobacillus fermentum, three Lacticaseibacillus casei/paracasei, and three Streptococcus thermophilus, covering both intra- and inter-species diversity. Spectral data from LAB were acquired using a mIRage LS OPTIR system, preprocessed, and used to train a fully connected neural network for each level. The models achieved macro F1-scores of 97% for species level and 91% for strain level classification. These results demonstrate the potential of OPTIR, when integrated with machine learning, as a robust tool for high-resolution bacterial classification, with promising applications in microbiological quality control, probiotic selection, and microbial ecology.