Sniffing out antimicrobial resistance

Bacterial breath tests could soon become a new tool in the arsenal against antimicrobial resistance, by enabling doctors to streamline infection diagnosis and identify the most appropriate treatment without the need for lengthy lab procedures. That idea comes from an international team of engineers, microbiologists, and machine learning specialists writing in this week’s Cell Biomaterials (2025, DOI:10.1016/j.celbio.2025.100125).

The researchers say that detecting signature volatile compounds emitted by microbes will enable doctors to make faster and more informed diagnoses and reduce the overuse of broad-spectrum antibiotics for common illnesses such as urinary tract infections.

Existing culture-based diagnostic tests often take days or even weeks—an unacceptable delay for patients that encourages doctors to prescribe treatment without knowing details of the specific pathogen responsible. This lack of rapid diagnostic methods is a key limitation driving the misuse of antibiotics. In 2023, the World Health Organization launched the Antimicrobial Resistance Diagnostic Initiative and called on researchers to develop new technologies to enhance infection-testing capabilities.

Andreas Güntner, an engineer at Swiss Federal Institute of Technology (ETH), Zurich, is part of the group sniffing out a solution. “Historically, we know already that different bacteria had, even by nose, recognizable and very distinct smells,” Güntner says. These signature scents—from the “sweet, grape-like” fragrance of Pseudomonas aeruginosa to the putrid stench of Clostridium infections—arise from volatile intermediates produced by the pathogen’s metabolism, he says. For example, Clostridium’s foul odor is largely caused by butyric acid and other fermentation products.

Particular combinations and concentrations of these volatile organic compounds (VOCs) are therefore characteristic of biosynthetic processes and provide a fingerprint of not just the species but the strain of bacteria. “These metabolic pathways are genetically determined, and on a genetic basis is also where you determine antimicrobial resistance,” Güntner says. “What we need is to identify the metabolic signatures that are expected to be diagnostic at the strain level.”

But the magnitude of the challenge should not be underestimated. The gaseous headspace of biological samples such as urine and blood contains a hugely complex mixture of compounds, only a small subset of which provide meaningful information. Machine learning will thus play a crucial part in the development and operation of this technology, both to identify promising biomarkers and to inform the design of appropriate sensor systems, according to Güntner and his colleagues.

“It is very interesting,” José Belizário, a microbiologist at the University of São Paulo, says in an email to C&EN. “The development of sensors for reliable and repeatable VOC detection would help to expand the volatilomic signatures of different strains [and] large-scale machine learning models could offer a powerful strategy to predict biological targets and their downstream effects.”

Although the proposal is still many years from becoming a standardized test, Güntner is optimistic; the team has already established a proof-of-concept system. “We have identified discriminatory compounds for Escherichia coli [indoles], and we have sensors capable of actually detecting some of these biomarkers,” he says. “The challenge for the upcoming 5-10 years is to really develop [this approach] to cover a significant amount of relevant bacterial types.”