Machine learning–enhanced spectroscopy detects toxins

When an ecological disaster releases toxins into the environment, extensive sampling and analysis is vital to ensure the health and safety of the people affected. Surface-enhanced vibrational spectroscopies, such as Raman and infrared spectroscopy, offer scientists a way to detect low concentrations of toxic compounds without chemical tagging. But the vibrational spectra can be difficult to analyze. Now researchers have demonstrated that machine learning can streamline the analysis of surface-enhanced Raman scattering spectra to readily detect toxins in human tissue (Proc. Natl. Acad. Sci. U.S.A. 2025, DOI: 10.1073/pnas.2422537122).

Surface-enhanced Raman scattering (SERS) is a phenomenon wherein the Raman signal of a molecule is boosted by the electronic properties of a substrate, which often consists of metal nanoparticles. “The dirty little secret of SERS is that everybody's substrate gives you slightly different-looking spectra of the same molecule,” says Naomi Halas, a professor at Rice University and coleader of the work. So individual groups must develop their own spectral databases, she says, making it challenging to streamline SERS to use for environmental sampling.

Instead of designing a new surface for SERS, Halas collaborated with her Rice University colleague Ankit Patel to develop a machine learning program to simplify the SERS data generated with a previously designed gold nanosphere substrate. The algorithm identifies peaks characteristic of specific molecules and improves the spectrum’s signal-to-noise ratio.

To demonstrate that machine learning–enhanced SERS would work with complex samples, Halas and her colleagues used the technique to identify polycyclic aromatic hydrocarbons (PAHs)—molecules associated with a variety of poor health outcomes—in human placental tissue from self-identified smokers and nonsmokers. With the algorithm-enhanced spectra, it was a simple task to find and characterize the PAHs with a traditional Raman spectral database.

“For SERS, the biggest technology challenge is the signal stability,” says Bo Tan of Toronto Metropolitan University, who is unaffiliated with the new research. Generally, scientists try to improve a SERS signal with hardware, she says. This is the first time she’s seen machine learning be used to improve the signal. If an algorithm were able to integrate the spectra from all the different sensor types into a single database, she says, “that could be very useful.”