Tufts University researchers are using AI and machine learning to more quickly identify potential narrow-spectrum antibiotics to prevent and treat Lyme disease, thanks to a generous gift from an alumnus that sped up their timeline by two to three years.

The work has already identified several hundred unique compounds that kill Borrelia burgdorferi, the bacteria that causes Lyme disease, while leaving other bacteria unaffected. Such narrow-spectrum antibiotic drug development is vital to prevent a drug that might kill Lyme bacteria from also killing helpful bacteria in the body’s microbiome or leading to drug resistance.

The gift has now enabled researchers to attract significant additional funding from the National Institutes of Health (NIH) and private foundations to expand their efforts.

“Without the anonymous gift, these efforts might never have gotten off the ground, or would have proceeded much more slowly,” says Linden Hu, the Paul and Elaine Chervinsky Professor in Immunology at Tufts University School of Medicine and co-director of the Tufts Lyme Disease Initiative. “The NIH doesn’t tend to fund pilot projects—they want ideas that already have data that suggest greater investment will yield success. This funding enabled us to get that proof of concept.”

Lyme disease, which is caused by a bite from a deer tick, affects approximately 475,000 people in the U.S. each year, predominantly in the Northeast and Mid-Atlantic states, the Upper Midwest, and on the West Coast in Northern California, Oregon, and Washington. While most cases can be treated successfully with a course of antibiotics, about 10-20% of people infected develop symptoms including fatigue, “brain fog,” and joint and muscle pains that persist for months to years after antibiotic treatment ends.

Researchers are hunting for ways to prevent Lyme disease and post-treatment Lyme disease syndrome (PTLDS). One such path is to identify new types of antibiotics that could be taken preventatively by people living in areas where the disease is endemic, in much the same way people take anti-malarial drugs when traveling to or living in areas where malaria is prevalent.

The goal is to find narrow-spectrum antibiotics that will kill B. burgdorferi but not kill other common harmful bacteria, such as E. coli or Staphylococcus aureus, and do not affect “good” bacteria that are part of our normal flora. Broad spectrum anti-microbials that kill multiple organisms risk creating drug resistance if taken by large numbers of people for long periods of time.

One effort, led by Hu and Maha Farhat, associate professor of bioinformatics at Harvard University, has screened 60,000 existing compounds to determine which can kill the Lyme disease bacteria. Those that looked promising were then counter-screened to make sure they killed only B. burgdorferi.

This traditional compound screening process is expensive and involves putting each molecule into a test tube or well with a sample of the bacteria, and seeing which, if any, kill the bacteria. Screening 60,000 compounds yielded several hundred that are active against B. burgdorferi, only a handful of which will likely prove worth pursuing.

Engineering New Drugs

Based on these initial screenings, the researchers then developed AI models which could comb through a large swath of the estimated 1 X 10⁶⁰ compounds (that’s 1 followed by 60 zeroes) in the drug-like chemical space much more quickly and efficiently—and at far lower cost—to identify additional potential compounds that have a high likelihood of being effective against the Lyme bacteria.

In addition, Farhat’s team is using the information to design chemical compounds that “should” be active against the Lyme bacteria (and only Lyme) based on their initial screenings. They do this by building a generative model that can ‘imagine’ any molecule type capable of killing the Lyme bacteria.

This allowed the team to develop engineered compounds that have additional favorable qualities. For example, they seem to kill the Lyme bacteria more effectively, are less toxic, or can be taken orally, for example.

The AI process is predictive and isn’t always right, but it helps narrow the field of compounds worthy of further testing much more quickly and less expensively, says Hu.

Along with Bree Aldridge, professor at the School of Medicine and School of Engineering, and Trever Smith II, research assistant professor at the School of Medicine, Hu and Farhat were recently funded by an NIH grant to further refine their efforts, screen more compounds, and design more potential compounds that could be used to prevent and treat Lyme disease. They will also see if they can determine how and why the drugs that kill the Lyme bacteria successfully attack Lyme—and only Lyme.

Aldridge’s research has previously focused on the tuberculosis bacterium and included development of an AI-powered tool called DECIPHAER that helps pinpoint how bacteria are being killed by antibiotics. DECIPHAER will play a key role in the new Lyme research.

“A compound can work to kill bacterial cells by targeting different major cellular functions like their ability to produce a cell wall, proteins, or energy,” explains Aldridge. “We want to understand how each compound actually kills bacteria, and we can do so using what we call morphological profiling, which is taking pictures of cells after they have been treated with a compound, and seeing how their cells fell apart.”

“It’s a ‘guilt by association’ algorithm,” she says. “So, if a new compound causes cells to disintegrate or die in the same physical way that they die with a cell wall-acting agent, then we assume that the new compound is also a cell wall-acting agent. We aren’t the first team to use this idea, but we are excited that with DECIPHAER, we can now bring a multi-omics approach to learn more specific mechanistic details.”

DECIPHAER links those details to reveal more exact insights into how the compounds are affecting the cells and exactly why the bacteria are dying, predicting a compound’s molecular impact from images alone and revealing how the compounds work in different conditions or in different combinations.

This information, in turn, can potentially be used to identify and design even more effective drugs to prevent and treat Lyme disease.

The donor’s support has also helped Peter Gwynne’s research to pinpoint weakness in the Lyme disease genome that can be exploited by identifying which portions control vital functions of the organism and might be disrupted by existing compounds.

“If you think of the genome as being like a subway map, most bacteria have relatively large genomes, with multiple pathways used to manage vital biological functions,” explains Gwynne, a research assistant professor in the Molecular Biology and Microbiology Department at the School of Medicine. “If a drug attacks and breaks one pathway, these bacteria have a backup pathway to use to perform that vital function, in much the same way that a large subway system has multiple ways to get from one destination to another if one route has a breakdown.”

The Lyme bacteria, in contrast, has a small genome and appears to have no backups or, at best, a limited number of backup pathways for critical functions. “This means that if we can identify compounds that break a critical pathway, the bacteria have no backup, and we can kill the Lyme bacteria,” Gwynne adds.

The donor’s gift paid for screening thousands of compounds to identify a handful of promising drug candidates that can attack some of these choke points. New funding from the Bay Area Lyme Foundation (BALF) is now furthering efforts by Gwynne and colleagues, and he also will be using AI computational modeling to speed the screening process.

“Significant funding from individuals is incredibly important to early-stage research like this,” says Hu. “It enables us to prove an idea is worth funding from the federal government and other foundations. It helps us attract talented people like Bree Aldridge and Maha Farhat, who may not be working in an area like Lyme disease, but can be brought together to help us because there is funding available.”

This gift, in particular, enabled the team to do in six months what otherwise would have taken two to three years, Hu estimates. Because the researchers were able to move quickly in an era when changes in AI and machine learning are moving so rapidly, their proposals for additional funding were still cutting edge when submitted, which enhanced the ability to get additional financial support.

“As researchers, we risk falling behind in exploiting the potential of AI unless we can move as swiftly as the artificial intelligence field is moving,” Hu adds. “Thanks to the donor, this research team has already identified several potential drug candidates that the University hopes to patent and that we hope will one day will move to testing in humans. And we believe the additional funding we have been able to garner will help us design even more effective drugs that can one day be used to prevent and treat Lyme disease.”