A new study demonstrated how AI can help identify novel antibiotic candidates against one of the world’s most persistent and rapidly evolving infections: gonorrhea.
Published in Science Translational Medicine, the work shows that deep learning models can successfully navigate millions of chemical structures to pinpoint compounds with activity against Neisseria gonorrhoeae, including strains resistant to multiple existing antibiotics. The findings point to a potential new strategy for replenishing a shrinking antibiotic pipeline as resistance continues to erode standard treatments.
Gonorrhea is the second most frequently reported sexually transmitted infection globally, with tens of millions of cases each year. In the US alone, more than 600,000 infections are reported annually. While often treatable, the infection can cause serious complications if left unchecked, including infertility in both men and women, pelvic inflammatory disease, and increased susceptibility to HIV. In rare cases, an infection that reaches the blood can lead to life-threatening conditions such as meningitis, sepsis, and cardiac complications.
The growing clinical challenge is not the infection itself, but the speed at which N. gonorrhoeae develops resistance to antibiotics. Even newly introduced treatments, such as Zoliflodacin and gepotidacin, are already expected to face resistance over time.
“We’ve seen this cycle of resistance development occur within just five to ten years after first-line roll-out, over and over again,” Melis Anahtar, first author of the study and Assistant Director of the Clinical Microbiology Laboratory at Massachusetts General Hospital, said in the press release. “To be able to prevail in this continuous arms race, we will need new antibiotics to fill the pipeline.”
The new study, led by James Collins of the Wyss Institute at Harvard University, MIT, and the Broad Institute, alongside Anahtar, Jackie Valeri, Majed Modaresi, and colleagues, explores how machine learning might accelerate that pipeline.
Two AI studies in parallel
This work was conducted alongside a second, closely related effort led by Aarti Krishnan, Anahtar, Veleri and Collins published in 2025 in Cell. While the Science Translational Medicine study focused on using deep learning to search vast, existing chemical libraries for promising antibacterial compounds, the Cell paper explored whether generative AI could go a step further — designing entirely new antibiotic molecules from scratch or from minimal chemical fragments.
“The gonorrhea model described in the Science Translational Medicine paper is the same one we used to score chemical fragments and de novo-generated compounds for N. gonorrhoeae in the Cell study, so the two papers are closely linked,” Anahtar told DDN. “The Cell paper is really about applying that model to a de novo design effort. Together, they show how a single deep learning framework can be used in very different ways to identify new antibiotic candidates.”
Mapping chemical space with machine learning
Over the years, gonorrhea has been treated with antibiotics that act on a range of different bacterial targets. Penicillins weaken the cell wall, tetracyclines inhibit protein synthesis, quinolones target DNA replication, and azithromycin blocks ribosomal function. Even the two most recent drugs in development act on type 2 topoisomerases — the same enzyme family targeted by fluoroquinolones — albeit at different binding sites.
The problem, Anahtar explained, is that drugs designed to fit neatly into a specific binding pocket are often highly vulnerable to resistance. “Just one or two point mutations can be enough to prevent binding altogether,” she said.
To try to sidestep that pattern and find novel targets, the team trained their model on phenotypic data rather than using a traditional target-based approach. The team tested 38,650 small molecules for their ability to inhibit the growth of N. gonorrhoeae in vitro, generating a large, experimentally grounded dataset that captured whether compounds could actually kill or suppress the pathogen in its cellular context. These results were then used to train a neural network capable of predicting antibacterial activity from chemical structure alone.
We’re not looking for the next bleach. We want candidates that are targeting something specific to the bacterial cell.
—Melis Anahtar, Massachusetts General Hospital
Phenotypic screening, however, comes with its own risks. “You can end up selecting compounds that are just generally toxic,” Anahtar noted. To address this, both studies applied stringent filtering — first computationally and then experimentally — using human cytotoxicity counter-screens to eliminate compounds that kill cells non-specifically. “We’re not looking for the next bleach,” she said. “We want candidates that are targeting something specific to the bacterial cell.”
That strategy paid off in both papers. In the Cell study, the team identified two novel compounds — dubbed NG1 and DN1 — as particularly promising. NG1 showed narrow-spectrum activity against pathogenic N. gonorrhoeae, while DN1 was designed to target Staphylococcus aureus but displayed broader activity against N. gonorrhoeae.
Crucially, both compounds acted through mechanisms distinct from those of commonly used antibiotics and were able to reduce bacterial burdens in animal models of infection. Beyond their immediate antibacterial effects, the researchers emphasized that both NG1 and DN1 sit within chemical families that are amenable to further medicinal chemistry optimization.
Similarly, the Science Translational Medicine study discovered two lead compounds. One, MP20, does not appear to have a single, clear protein target but instead seems to increase bacterial membrane permeability through an as-yet-unknown mechanism. The second, known as A1, has been shown to bind to alanine racemase, an enzyme essential for bacterial cell wall synthesis but not previously targeted by small-molecule antibiotics in gonorrhea.
This shows that both approaches are viable for uncovering antibiotic candidates that might otherwise be missed by conventional discovery strategies.
From computation to biological systems
Anahtar said she was particularly excited to test the compounds in organ-on-a-chip models. “With mouse models — especially for vaginal infection — you often just apply the compound and observe the final outcome,” she said. “It’s very difficult to interrogate what’s actually happening at the tissue level.”
Although still in their early stages, the vaginal organ-on-a-chip models developed by Donald Ingber’s lab at the Wyss Institute proved encouraging. The researchers were initially pleased simply to see N. gonorrhoeae adhere to, invade, and behave appropriately within the human cellular environment. The platform consists of a porous polymer membrane seeded on one side with multilayered human vaginal epithelial cells and, on the other, layers of human fibroblasts. Compounds can be introduced either from above, mimicking topical administration, or from below, modeling systemic delivery.
Using this system, the team showed that MP20 was able to traverse the epithelial barrier, pass through tight junctions, and kill N. gonorrhoeae residing in the vaginal lumen. “Seeing that was really interesting,” Anahtar said, noting that compound permeability is often one of the biggest unknowns in antibiotic development. “If something doesn’t work, you don’t always know why. In this case, it did work, which was exciting.”
Beyond proof of efficacy, Anahtar sees organ-on-a-chip models as powerful tools for troubleshooting and optimization. While the technology is still specialized and available in only a handful of laboratories, commercial versions of vaginal chips are now emerging. Because they rely on human cells rather than animal physiology, these systems may help avoid some of the translational pitfalls that arise when compounds need to be optimized for mice and then re-engineered for humans.
“It’s still early days, but I hope these models will allow for much deeper mechanistic interrogation — helping us understand why certain compounds perform well, why others don’t, and how to choose between multiple analogs for a very specific clinical indication," she said.
Toward a new antibiotic pipeline
This work builds on a growing body of evidence that machine learning approaches can deliver real traction in antibiotic discovery. In earlier studies, Collins and colleagues used similar AI-driven strategies to identify several promising drug candidates, including halicin and abaucin. Halicin, originally investigated as a potential diabetes drug, was found to be highly effective against Clostridioides difficile and pan-resistant Acinetobacter baumannii infections in murine models. Abaucin, by contrast, showed narrow-spectrum activity against A. baumannii — a huge advantage for more selective antibiotic treatment.
By combining deep learning with large-scale chemical screening and physiologically relevant models, this approach offers a way to systematically explore antibiotic candidates at a scale and speed that would previously have been unfeasible.
As antibiotic resistance continues to rise globally, AI-enabled discovery strategies may become increasingly central to maintaining the effectiveness of future treatments.
