AI-Driven Drug Discovery Accelerates Search for Tuberculosis Drugs Against World's Deadliest Infection

Fighting the world's deadliest infection with PAC-MAN and AI

Tuberculosis, caused by the bacterium Mycobacterium tuberculosis (Mtb), is the world's deadliest single-agent infection, responsible for 1.23 million deaths in 2024, according to the World Health Organization. The bacterium's unique outer cell membrane is notoriously hard to penetrate, making few drugs, including antibiotics, effective in treating the disease. However, a research team led by the University of Massachusetts Amherst has developed a pair of techniques that can vastly speed the search for better tuberculosis drugs. Published in the journal Nature Microbiology, the team's approaches first measure which chemical compounds are able to slip across the outer membrane and then use those measurements to predict other compounds that can get into the Mtb cell. "Mtb is unique," says Sloan Siegrist, associate professor of microbiology at UMass Amherst and, along with Anna Green, assistant professor in UMass Amherst's Manning College of Information and Computer Sciences, one of the paper's senior authors. "Not only does it have two membranes that protect the cell from antimicrobial chemical compounds that we might use to kill it, its outer membrane is unlike any other biological barrier out there." It's largely thanks to this outer membrane, called the mycomembrane, that Mtb is so resilient to both the body's immune system and antibiotics. Siegrist's lab specializes in finding chinks in the mycomembrane, which are crucial for developing drugs that can quickly and effectively treat tuberculosis. The only problem is that there are uncountable numbers of chemical compounds, and until recently, researchers had to test them one at a time to see which ones might get into Mtb cells. Then, in 2023, Siegrist co-authored a paper with Marcos Pires, professor of chemistry at the University of Virginia, announcing a technique called Peptidoglycan Accessibility Click-Mediated AssessmeNt, or PAC-MAN, which could test many compounds in parallel rather than one at a time. Yet, despite PAC-MAN's huge advance in efficiency, it wasn't enough. "Marcos and I wanted to harness measurements of known chemicals to predict compound uptake for unknown chemicals, so we brought in computational biologists and chemists, including my colleague Anna Green from UMass Amherst's Manning College of Information and Computer Sciences." Green's specialty is using computation to understand patterns in biological compounds. "Small molecules can be particularly difficult to analyze computationally," she says. "Because they come in all different sizes with a wide range of molecular connections, you can't describe them with a single measurement -- by weight, say, or size." This is where artificial intelligence comes in. Green and her lab designed a machine learning model, the Mycobacterial Permeability neural Network (MycoPermeNet), trained on the PAC-MAN screening data. Once trained, the model can predict how readily a compound permeates the mycomembrane from its chemical structure alone and points to the physical properties and molecular substructures that help a compound slip past Mtb's defenses. Using PAC-MAN and MycoPermeNet, the team identified a host of attributes that predict how well a compound is able to sneak its way past the mycomembrane and found in large data sets that these same features also correlate with a compound's ability to kill Mtb. "The mycomembrane lets some molecules through and keeps others out," says Green. "There must be something about this membrane, and about the chemistry of each molecule, that decides which ones get in -- and our combined tools help us figure out which ones can get through, and why."

New pair of techniques can speed up the search for better tuberculosis drugs

University of Massachusetts AmherstJul 6 2026Reviewed Tuberculosis, caused by the bacterium Mycobacterium tuberculosis (Mtb) is the world's deadliest single-agent caused infection, responsible for 1.23 million deaths in 2024, according to the World Health Organization. The bacterium's unique outer cell membrane is notoriously hard to penetrate, making few drugs, including antibiotics, effective in treating the disease. However, a research team led by the University of Massachusetts Amherst has developed a pair of techniques that can vastly speed up the search for better tuberculosis drugs. Published in the journal Nature Microbiology, the team's approaches first measure which chemical compounds are able to slip across the outer membrane and then use those measurements to predict other compounds that can get into the Mtb cell. "Mtb is unique," says Sloan Siegrist, associate professor of microbiology at UMass Amherst and, along with Anna Green, assistant professor in UMass Amherst's Manning College of Information and Computer Sciences, one of the paper's senior authors. "Not only does it have two membranes that protect the cell from antimicrobial chemical compounds that we might use to kill it, its outer membrane is unlike any other biological barrier out there." It's largely thanks to this outer membrane, called the mycomembrane, that Mtb is so resilient to both the body's immune system and antibiotics. Siegrist's lab specializes in finding chinks in the mycomembrane, which are crucial for developing drugs that can quickly and effectively treat tuberculosis. The only problem is that there are uncountable numbers of chemical compounds, and, until recently, researchers had to test them one-at-a-time to see which ones might get into Mtb cells. Then, in 2023, Siegrist coauthored a paper with Marcos Pires, professor of chemistry at the University of Virginia, announcing a technique called Peptidoglycan Accessibility Click-Mediated AssessmeNt, or PAC-MAN, which could test many compounds in parallel rather than one at a time. Yet, despite PAC-MAN's huge advance in efficiency, it wasn't enough. "Marcos and I wanted to harness measurements of known chemicals to predict compound uptake for unknown chemicals, so we brought in computational biologists and chemists, including my colleague Anna Green from UMass Amherst's Manning College of Information and Computer Sciences." Green's specialty is using computation to understand patterns in biological compounds. "Small molecules can be particularly difficult to analyze computationally," she says. "Because they come in all different sizes with a wide range of molecular connections, you can't describe them with a single measurement-by weight, say, or size." This is where Artificial Intelligence comes in. Green and her lab designed a machine learning model, the Mycobacterial Permeability neural Network (MycoPermeNet), trained on the PAC-MAN screening data. Once trained, the model can predict how readily a compound permeates the mycomembrane from its chemical structure alone and points to the physical properties and molecular substructures that help a compound to slip past Mtb's defenses. Using PAC-MAN and MycoPermeNet, the team identified a host of attributes that predict how well a compound is able to sneak its way past the mycomembrane and found in large datasets that these same features also correlate with a compound's ability to kill Mtb. "The mycomembrane lets some molecules through and keeps others out," says Green. "There must be something about this membrane, and about the chemistry of each molecule, that decides which ones get in-and our combined tools help us figure out which ones can get through, and why." This work was supported by the National Institutes of Health, UMass Amherst's Institute for Applied Life Sciences and the Gates Foundation. Irene Lepori and Nelson Evbarunegbe (both UMass Amherst), Zichen Liu (University of Virginia) and Shasha Feng (Lehigh University) were the co-lead authors, while Joel Freundlich (Rutgers University-New Jersey Medical School) and Wonpil Im (Lehigh University) were senior authors alongside Siegrist, Green and Pires.