Tuberculosis has once again claimed the grim title of the world’s deadliest single-agent infectious disease, with the World Health Organization attributing 1.23 million deaths to Mycobacterium tuberculosis in 2024 alone. The bacterium’s extraordinary resilience stems largely from a unique double-membrane envelope, the outermost layer of which—the mycomembrane—is unlike any biological barrier known to science. This waxy, near-impenetrable fortress excludes most antibiotics, forcing clinicians to rely on months-long multidrug regimens that are increasingly compromised by resistance. Now, a multidisciplinary team led by the University of Massachusetts Amherst has unveiled a tandem of technologies that can dramatically accelerate the hunt for compounds capable of slipping through this defense, potentially slashing the time required to identify new tuberculosis drug candidates.
The challenge of crossing the mycomembrane is not merely one of size or charge. Its dense matrix of mycolic acids and other complex lipids creates a hydrophobic labyrinth that passively resists both hydrophilic and large hydrophobic molecules. Traditionally, researchers could only test potential permeators one molecule at a time, a bottleneck that rendered systematic exploration of chemical space impossibly slow. In 2023, microbiologist Sloan Siegrist and chemist Marcos Pires of the University of Virginia introduced a method called Peptidoglycan Accessibility Click-Mediated AssessmeNt, or PAC-MAN, which harnesses click chemistry to screen libraries of compounds in parallel. The technique uses a bioorthogonal probe that reacts with a tagged peptidoglycan layer only if the test compound has successfully breached the outer membrane, yielding a fluorescent signal proportional to permeability. While PAC-MAN marked a massive leap in throughput, Siegrist recognized that even pooled screening could not cover the vastness of potential drug-like molecules. “Marcos and I wanted to harness measurements of known chemicals to predict compound uptake for unknown chemicals,” she explained, bringing computational biologist Anna Green into the fold.
Green, an assistant professor in UMass Amherst’s Manning College of Information and Computer Sciences, specializes in decoding patterns in biological sequences and small molecules. Small molecules are notoriously difficult to analyze computationally because they lack the linear regularity of DNA or proteins; they vary wildly in topology, functional groups, and three-dimensional conformations. To tackle this, Green’s lab designed a custom deep learning architecture called the Mycobacterial Permeability neural Network, or MycoPermeNet. The model was trained on the PAC-MAN screening data—thousands of compounds with known permeation scores—and learned to associate structural motifs and physicochemical properties with the likelihood of crossing the mycomembrane. Unlike generic quantitative structure-activity relationship models, MycoPermeNet was built to capture the complex interplay between molecular flexibility, hydrogen-bonding capacity, and lipophilicity that governs transit through this atypical lipid barrier.
Once trained, the network could predict the permeability of entirely new structures based solely on their two-dimensional chemical graphs. Importantly, the model also pointed to specific molecular substructures and properties that favor permeation. The team discovered that moderately lipophilic compounds with a limited number of rotatable bonds and a restrained polar surface area were most adept at traversing the mycomembrane. These features align with the biophysical requirements of partitioning into a rigid, highly ordered lipid environment while still maintaining enough solubility to approach the membrane surface. The analysis also revealed that certain heterocyclic scaffolds repeatedly appeared among successful permeators, offering medicinal chemists concrete building blocks to prioritize in future drug design.
To validate the predictive power of their approach, the researchers interrogated large public and in-house datasets of compounds with known antimycobacterial activity. They found that the very same molecular features that MycoPermeNet associated with high permeability also correlated strongly with the ability to kill Mtb inside macrophages. This concordance suggests that passage through the outer membrane is not just a desirable attribute but a limiting step that dictates whole-cell efficacy. “The mycomembrane lets some molecules through and keeps others out,” Green said. “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.”
The dual PAC-MAN/MycoPermeNet platform now enables what the team calls “virtual permeation screening.” Rather than synthesizing and testing thousands of compounds, researchers can first filter massive virtual libraries—on the order of millions of molecules—through the machine learning model to identify candidates with a high predicted probability of membrane penetration. Only the most promising hits need to be advanced to the wet-lab PAC-MAN assay and subsequent minimum inhibitory concentration testing. In a field where a single round of traditional medicinal chemistry optimization can take months, such in silico triage could compress the early discovery timeline from years to weeks.
The study, published in Nature Microbiology, was a collaborative effort spanning microbiology, chemistry, computer science, and structural biology. Co-lead authors Irene Lepori, Nelson Evbarunegbe, Zichen Liu, and Shasha Feng were instrumental in executing the experimental and computational work, while senior authors Joel Freundlich and Wonpil Im contributed complementary expertise in antimycobacterial pharmacology and membrane simulations. The research was supported by the National Institutes of Health, UMass Amherst’s Institute for Applied Life Sciences, and the Gates Foundation, reflecting the global urgency of addressing tuberculosis.
Although the tools are not a drug themselves, they address the fundamental bottleneck that has historically stymied tuberculosis drug development: the inability to rationally design molecules that can reach their targets inside the bacterium. By revealing the chemical grammar of mycomembrane permeation, the UMass-led team has effectively handed the research community a decoder ring. The next steps will involve applying MycoPermeNet to diverse chemical libraries and coupling the permeability predictions with target-based screens, potentially unlocking entirely new classes of antitubercular agents. As drug-resistant strains continue to spread, such predictive platforms may prove indispensable in staying ahead of a pathogen that has co-evolved with humanity for millennia.
Subject of Research: Prediction of outer membrane permeation in mycobacteria using high-throughput screening and machine learning to accelerate tuberculosis drug discovery.
Article Title: Identification of chemical features for improved outer membrane permeation in mycobacteria using machine learning
Web References: https://doi.org/10.1038/s41564-026-02412-5
References: 10.1038/s41564-026-02412-5
Image Credits: Irene Lepori
Keywords: Tuberculosis, Mycobacterium tuberculosis, mycomembrane, drug permeability, machine learning, MycoPermeNet, PAC-MAN, click chemistry, antibiotic resistance, deep learning, drug discovery
