Skip to main content
Premium Trial:

Request an Annual Quote

Machine Learning May Extend MALDI Mass Spec-Based Antibiotic Resistance Testing


NEW YORK – New research by scientists at ETH Zurich and the University of Basel could extend MALDI mass spectrometry-based antibiotic susceptibility testing used in clinical microbiology.

The development could prove a boon for Bruker and BioMérieux, currently the dominant players in the field, who have both made efforts to push their MALDI platforms into the AST space.

In a study published in Nature Medicine last month, the Swiss researchers used machine learning to predict antibiotic susceptibility directly from MALDI-TOF mass spectra. In the study, they used a training dataset containing 300,000 mass spectra comprising more than 750,000 antibiotic susceptibility phenotypes in microorganisms collected at four medical institutions to build classifiers for predicting antibiotic resistance. They also trained resistance classifiers specific to three pathogen-antibiotic pairings of strong clinical interest — Staphylococcus aureus and oxacillin, Escherichia coli and ceftriaxone, and Klebsiella pneumoniae and ceftriaxone, finding that they performed with an area under the curve of .80, .74, and .74, respectively.

Additionally, the researchers looked at records from a set of 63 cases of E.coli, K. pneumoniae, or S. aureus infections in which physicians had made decisions regarding antibiotic use after the pathogen had been identified but before conventional AST information was available. They found that had their classifier been used, it would have recommended a different course of therapy in nine of the 63 cases. It would have correctly recommended de-escalation of treatment in seven cases, correctly recommended maintenance of the then-current treatment in one case, and incorrectly recommended escalation of therapy in one case.

Over the last decade, Bruker's MALDI Biotyper and BioMérieux's Vitek MS clinical microbiology platforms have seen broad uptake as tools for microbial identification as they are able to make IDs more quickly than conventional biochemical approaches and more cheaply than molecular testing. Both companies have also worked to extend their platforms to antibiotic resistance testing, though only in a limited way thus far.

Bruker, for instance, currently offers CE-IVD-marked Biotyper tests for cephalosporin and carbapenem resistance. It sells both tests in the US as research-use-only versions. The Biotyper tests for resistance to these antibiotics work by detecting the hydrolysis products generated by resistant bacteria. The company also offers a CE-IVD-marked module for automated early detection of carbapenem resistance in K. pneumoniae. This module uses the Biotyper to look for peaks associated with the blaKPC protein, which can confer resistance to carbapenems.

Bruker also several years ago introduced a new version of the Biotyper, the Biotyper Sirius system, which is able to run in negative-ion mode, enabling more effective analysis of bacterial lipids, which are helpful for assessing antibiotic resistance.

The machine learning method presented in the recent study has the advantage of using the same mass spec measurements used for identifying pathogens to assess antibiotic susceptibility, potentially making for a highly streamlined process.

"You are simply reusing the data that the MALDI-TOF machine is producing anyway," said Adrian Egli, head of clinical microbiology at University Hospital Basel and a senior author on the study. "So it is a very straightforward approach."

This would tie in well with suggestions from industry observers that a simple, highly automated approach will be necessary for MALDI to make significant inroads into the AST space.

In addition to conventional AST testing, MALDI-based methods are competing with an emerging group of other rapid AST technologies from companies like Gradientech, Specific Diagnostics, Accelerate Diagnostics, and Q-linea, which use technologies like advance image analyses to speed phenotypic testing.

Bruker was not a part of the Swiss study but is "very interested" in the approach presented by the team, said Markus Kostrzewa, senior VP of innovations at the firm. He said the company is working internally with its own machine learning and MALDI mass spec experts and is in discussions with outside researchers about the approach's potential.

"We are exploring the potential to translate this idea into broadly available applications and products," Kostrzewa said.

He noted that in the early days of MALDI-based microbial identification, there was hope that the spectra used for identification could also be used for AST applications but that this proved "too optimistic" at the time. The new work could lead to a revival of this idea, he said.

While a MALDI-based method is not likely to replace conventional AST, it could provide initial early guidance for doctors, he suggested. "Such early information might be of particular value for appropriate bloodstream infections where time really matters," he said, noting that this application would complement and expand Bruker's existing MALDI Biotyper Sepsityper kit.

Kostrzewa suggested that while doctors were unlikely to de-escalate therapy based on such an approach, it could lead them to escalate treatment. This could limit the approach's ability to reduce utilization of broad-spectrum antibiotics, he noted.

Egli said that his team's machine learning method does not appear to rely on measuring signal primarily from proteins directly linked to antibiotic resistance mechanisms but rather is picking up broader similarities between bacteria that exhibit resistance. Based on sensitivity analyses he and his colleagues conducted, he also believes that the approach is picking up differences in metabolic profiles between resistant and nonresistant organisms.

A number of questions remain to be worked out about the method, he noted. For instance, it is unclear how well the models trained on data from the four hospital facilities will work on spectra from centers in other locations.

"The four centers we had were all in a relatively close area," he said. "So I think the next step for the study would be to take it to somewhere like, say, Berlin or London where you could expect that the strains circulating there are slightly different."

Another question is how frequently clinicians would need to update the data they use to train their classifiers. Given the speed of bacterial evolution, it is plausible that a classifier trained on data from several years before would no longer be effective.

"If you take a dataset from five years ago, I think it would be very difficult to predict resistance today based on that particular machine learning algorithm," Egli said. "So you need to have a relatively close dataset in terms of time. We would need to train and update the algorithm once a year or something like that, maybe."

He said that he and his team were currently in discussions with colleagues in infectious disease and intensive care to plan a multi-center prospective clinical trial to follow up on their initial results.

"We're trying to find out now, really, where this is suitable to use," he said. "We'll do a prospective validation of the study to really check what the impact is in a randomized controlled trial, where one group of patients will get the algorithm and the other will not, and then we can measure, for instance, time to optimal antibiotic therapy, and we can say if the algorithm has an impact on it or not."

Ultimately, Egli said, "our dream would be to basically build a web-based platform that could be deployed by interested individuals, so that they could use it as a tool in their own local centers with their own data."