Researchers from the Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, along with collaborators at Jiangnan University and Wenzhou Medical University, have developed a rapid method to identify lactic acid bacteria directly on agar plates. Their study, published in Spectrochimica Acta Part A: Molecular and Biomolecular Spectroscopy, combines adaptive Raman spectroscopy with a novel deep learning model to address challenges in microbial analysis.

Lactic acid bacteria play vital roles in food production, pharmaceuticals, and agriculture, but traditional identification methods face limitations. Techniques like genetic testing or mass spectrometry require extensive sample preparation, destroy colonies, and cannot perform in situ detection. Existing Raman spectroscopy methods, which analyze molecular vibrations using laser light, struggle with biological heterogeneity within bacterial colonies, reducing accuracy for industrial applications.

The team created a two-part solution involving an adaptive colony Raman acquisition method based on signal-to-noise ratio screening. This technique automatically identifies optimal measurement points within colonies by analyzing spectral quality. It minimizes errors caused by uneven biological structures that previously hampered identification. They paired this with the Raman Swin Transformer, a deep learning model adapted from image-processing algorithms. This model efficiently captures subtle spectral patterns using a specialized attention mechanism to analyze local and global features in Raman data.

Using a near-infrared laser, the system scanned bacterial colonies in situ through agar plates. Each measurement took half a second. The researchers tested the system on fourteen lactic acid bacteria species, collecting over seven thousand spectral samples. The Raman Swin Transformer model achieved ninety-eight point two percent accuracy in species classification. When identifying new strains of previously learned species from different sources, the system maintained over seventy percent accuracy. The model outperformed existing alternatives in robustness tests.

This non-destructive approach enables real-time selection of bacterial colonies for industrial applications like dairy fermentation or probiotic development. By accelerating strain screening from days to minutes, the technology could streamline microbial resource banking and quality control in biotechnology. The researchers note the framework could extend to other industrially important microorganisms.