How physics and deep learning are helping in identifying different bacteria in seconds

Bacterial identification can take hours and often longer, precious time when diagnosing infections and selecting appropriate treatments. 

By teaching a deep learning algorithm to identify the "fingerprint" spectra of the molecular components of various bacteria,  researchers could now classify various bacteria in different media with accuracies of up to 98%.

Their results were made available online on Jan. 18 in Biosensors and Bioelectronics, ahead of publication in the journal's April (2022) issue.

Bacteria-induced illnesses, those caused by direct bacterial infection or by exposure to bacterial toxins, can induce painful symptoms and even lead to death, so the rapid detection of bacteria is crucial to prevent the intake of contaminated foods and to diagnose infections from clinical samples, such as urine.

By using surface-enhanced Raman spectroscopy (SERS) analysis boosted with a newly proposed deep learning model, researchers demonstrated a markedly simple, fast, and effective route to classify the signals of two common bacteria and their resident media without any separation procedures.

Raman spectroscopy sends light through a sample to see how it scatters. The results reveal structural information about the sample—the spectral fingerprint—allowing researchers to identify its molecules. The surface-enhanced version places sample cells on noble metal nanostructures that help amplify the sample's signals. However, it is challenging to obtain consistent and clear spectra of bacteria due to numerous overlapping peak sources, such as proteins in cell walls. "Moreover, strong signals of surrounding media are also enhanced to overwhelm target signals, requiring time-consuming and tedious bacterial separation steps.

To parse through the noisy signals, the researchers implemented an artificial intelligence method called deep learning that can hierarchically extract certain features of the spectral information to classify data. They specifically designed their model, named the dual-branch wide-kernel network (DualWKNet), to efficiently learn the correlation between spectral features. Such an ability is critical for analyzing one-dimensional spectral data. Despite having interfering signals or noise from the media, which make the general shapes of different bacterial spectra and their residing media signals look similar, high classification accuracies of bacterial types and their media were achieved with this method.

The most important thing is this method dramatically reduces analysis time.

Eojin Rho et al, Separation-free bacterial identification in arbitrary media via deep neural network-based SERS analysis, Biosensors and Bioelectronics (2022). DOI: 10.1016/j.bios.2022.113991