Precisely Identifying Growth Phases of Living Bacteria using Open-Set Deep Learning-Driven Single-Cell Raman Spectroscopy

Owing to population cellular heterogeneity, it is still challenging to identify and determine the growth phases of spore-forming bacterial cells from a complex culture environment with a single-cell resolution. Knowledge of the physiological state of spore-forming bacteria can help to yield optimal fermentation performance, identify the spore origin in food safety, and reveal the transmission dynamics of spore-associated pathogens in controlling infectious diseases. Therefore, being able to precisely identify and predict the growth phases of individual living bacterial cells is of considerable significance. Here, we proposed an advanced strategy named open-set deep learning-driven single-cell Raman spectroscopy to identify various growth phases of individual <i>Bacillus pumilus</i> cells/spores. Unlike a single network architecture, our proposed deep learning configuration was designed by integrating a convolutional neural network with a long short-term memory (LSTM) model. To further enhance the diversity of original Raman spectra data sets yielded by our home-built single-cell Raman spectroscopy platform, an interpolation algorithm-enhanced spectral shifting strategy was employed to augment the Raman spectra data sets. For time-dependent single-cell Raman spectra of <i>B. pumilus</i> cells/spores sampled at 13 various growth time points, our proposed deep learning configuration working in a closed-set environment could achieve a high prediction accuracy of 96.52 ± 0.88%. Moreover, the relative classification contribution remembered by LSTM showed that the significant dynamic changes in physiological states appeared from 12 to 20 h. Particularly, the two Raman bands located at 1017 and 1655 cm^-1 have shown high classification contributions in identifying various growth time points sampled at 12, 16, 24, and 36 h. More importantly, our proposed deep learning configuration can also work in an open-set environment via the creation of an enhanced Softmax module, which was developed by integrating an enhanced threshold strategy with an adjacency confidence margin mechanism. Amazingly, an open-set average prediction accuracy as high as 92.15 ± 1.67%, including 13 distinct growth time points of <i>B. pumilus</i> cells/spores, unseen growth time points (6 h, 72 h), and three unknown <i>Bacillus</i> species, was achieved, indicating that open-set deep learning model-driven single-cell Raman spectroscopy can effectively identify trained growth time points of individual <i>B. pumilus</i> cells/spores, unseen growth time points of individual <i>B. pumilus</i> cells/spores that were excluded from the training data sets, and unknown <i>Bacillus</i> species. It can be foreseen that open-set deep learning-driven single-cell Raman spectroscopy has shown great promise for identifying growth phases of live bacteria in complex practical environments.