The bacterium has a harder time evolving a defense against what it has never encountered.
For decades, Neisseria gonorrhoeae has outpaced every antibiotic humanity has deployed against it, rendering each new treatment a temporary truce rather than a lasting peace. Now, a team spanning Harvard, MIT, and the Broad Institute has turned deep learning loose on a library of six million compounds, seeking molecular structures so unfamiliar to the bacterium that resistance itself becomes difficult. Two candidates have emerged from that search — one of them targeting an enzyme the pathogen needs to build its own cell wall through a mechanism never before exploited in this infection. The work does not yet offer a cure, but it offers something rarer: a method that might finally let medicine search faster than a pathogen can evolve.
- Gonorrhea infects hundreds of thousands of Americans each year, and the two newest antibiotics approved to treat it are already shadowed by the near-certainty that resistance will follow within a decade.
- The bacterium's speed of adaptation has made conventional drug discovery feel like a foot race against a sprinter — each new weapon risks obsolescence before it reaches widespread use.
- Researchers trained a deep learning model on nearly 39,000 laboratory-tested molecules, then unleashed it on six million virtual compounds, collapsing years of screening into a targeted shortlist of 213 candidates.
- Two compounds survived rigorous potency, resistance, and toxicity testing — one of them, A1, disables an enzyme critical to the bacterium's cell wall construction through a previously unknown mechanism.
- Animal models and a microfluidic human tissue chip both showed meaningful reductions in bacterial load, moving the discovery from computational promise toward biological reality.
- The compounds still require medicinal chemistry refinement before clinical use, but the pipeline itself may matter more than any single drug — offering a scalable way to stay ahead of a pathogen that has never stopped adapting.
Gonorrhea infects tens of millions globally each year, and in the United States alone more than 600,000 cases are reported annually. Left untreated, the disease causes infertility in both sexes, pelvic inflammatory disease, heightened HIV transmission risk, and — when bacteria spread systemically — meningitis or sepsis. The infection's true danger, however, is not what it does to the body today but how quickly it dismantles the medicines designed to stop it.
Neisseria gonorrhoeae has developed resistance to first-line antibiotics within five to ten years of their introduction, repeatedly. Two new oral drugs — zoliflodacin and gepotidacin — were recently approved, representing the first entirely new antibiotic classes for this infection in over three decades. History offers little comfort: if deployed widely, resistance will almost certainly follow. Medicine needs a continuous supply of new weapons, and it needs them faster than conventional discovery can produce them.
A team led by James Collins at Harvard's Wyss Institute, alongside Melis Anahtar at Massachusetts General Hospital and collaborators at MIT and the Broad Institute, built a machine learning pipeline to address exactly this problem. They first tested 38,650 small molecules in the laboratory to determine which could inhibit bacterial growth, using that data to train a deep learning model. The model was then deployed against a virtual library of roughly six million compounds, flagging 213 candidates for closer examination. After growth, resistance, and toxicity screening, two compounds stood out for their potency against multi-drug resistant strains and their remarkably low rates of resistance generation.
The lead candidate, A1, is an aminothiazole compound that works by inhibiting alanine racemase — an enzyme the bacterium requires to construct its cell wall. While cell wall biosynthesis has been targeted before, this specific mechanism was novel. The team validated the finding using proteomic and genetic tools and continues to investigate precisely how A1 disables the enzyme.
To test real-world relevance, the researchers moved beyond standard laboratory conditions. Using a microfluidic vaginal chip developed by Donald Ingber's group — a human-relevant tissue model — they showed that one compound, MP20, significantly reduced pathogen levels alongside vaginal epithelial cells. In a mouse model of vaginal infection, five doses of A1 over 24 hours substantially lowered bacterial concentration compared to untreated animals.
The results are promising but preliminary. A1 still requires optimization through medicinal chemistry before it could become a viable drug. The deeper significance, though, lies in the pipeline: a deep learning approach capable of screening vast chemical libraries — including compounds synthesized on demand — to surface unexpected structures as starting points for future antibiotics. Published in Science Translational Medicine, the study suggests that when artificial intelligence is paired with high-quality biological data and human-relevant models, the search space expands and the timeline compresses. In an arms race where the enemy has always adapted faster, that may be the most important finding of all.
Gonorrhea infects tens of millions of people worldwide each year. In the United States alone, more than 600,000 cases are reported annually, making it the second most common sexually transmitted infection. Left untreated, the disease ravages the body in ways that extend far beyond the initial infection. Women and men both face infertility. Pelvic inflammatory disease can develop. The virus increases the likelihood of HIV transmission. If the bacteria escape the genitals or throat and spread systemically, they can inflame the heart, trigger meningitis, or cause sepsis.
The real crisis, though, is not the infection itself but the pathogen's ability to outrun medicine. Neisseria gonorrhoeae evolves resistance to antibiotics with stunning speed. Two new oral drugs—zoliflodacin and gepotidacin—were recently approved to treat uncomplicated urogenital gonorrhea. They represent the first entirely new antibiotic classes developed for this infection in more than three decades. Yet history suggests a grim timeline. The bacterium has developed significant resistance to first-line treatments within five to ten years of their introduction, again and again. If these two new drugs are deployed widely, resistance will almost certainly follow. The cycle is nearly inevitable. To stay ahead, medicine needs a constant supply of new weapons.
A team led by James Collins at Harvard's Wyss Institute, working with Melis Anahtar at Massachusetts General Hospital and colleagues across MIT and the Broad Institute, has proposed a different strategy. Rather than waiting for resistance to emerge and then scrambling to develop new drugs, they used artificial intelligence to search for antimicrobial compounds that might target pathways the bacterium has never encountered. The logic is elegant: if a drug attacks an uncommon cellular mechanism, the pathogen will have a harder time evolving a defense. To test this hypothesis, the researchers built a machine learning pipeline.
They began by testing 38,650 small molecules in the laboratory to see which ones could inhibit N. gonorrhoeae growth. This data became the training set for a deep learning model. Once the model proved it could identify promising candidates with chemical structures unlike conventional antibiotics, the team deployed it against a virtual library of approximately 6 million compounds. The AI flagged 213 candidates for further investigation. After running growth inhibitory assays, resistance tests, and toxicity screens, two compounds emerged as particularly promising. Both showed strong potency against multi-drug resistant strains and generated resistance at remarkably low frequencies.
The lead candidate, called A1, is an aminothiazole compound with a previously unknown mechanism of action against gonorrhea. Using proteomic analysis, the researchers discovered that A1 binds to and inhibits alanine racemase, an enzyme the bacterium needs to construct its cell wall. While other antibiotics target cell wall biosynthesis, directly inhibiting alanine racemase with a small molecule was novel. The team validated this finding using genetic tools and is now investigating the precise mechanics of how A1 disables the enzyme.
But laboratory success does not guarantee clinical relevance. The researchers took the next step by testing their compounds in environments that mimic human infection. Working with Donald Ingber's group at the Wyss Institute, they used a microfluidic vaginal chip—a human-relevant tissue model—to demonstrate that one compound, MP20, significantly reduced pathogen levels when introduced into the device alongside vaginal epithelial cells. In a mouse model of vaginal infection, five doses of A1 administered over 24 hours substantially lowered bacterial concentration compared to untreated controls.
These results are encouraging but preliminary. Anahtar emphasized that A1 requires further validation and optimization through medicinal chemistry before it could become a clinically viable drug. The deeper promise, though, lies in the pipeline itself. The deep learning approach has potential to screen vastly larger chemical libraries—including compounds synthesized on demand—to uncover unexpected structures as starting points for future antibiotic development. In an arms race where the enemy evolves faster than conventional drug discovery can respond, this method offers a way to expand the search space and compress the timeline. The study, published in Science Translational Medicine, demonstrates that when artificial intelligence meets high-quality biological data and human-relevant models, compounds that would otherwise remain hidden can be brought into focus. The question now is whether this discovery process can accelerate fast enough to keep pace with a pathogen that has proven itself a master of adaptation.
Citas Notables
We've seen the cycle of resistance development occur within just five to 10 years after first-line roll-out, it has happened over and again. To be able to prevail in this continuous arms race, we will need new antibiotics to fill the pipeline.— Melis Anahtar, M.D., Ph.D., Assistant Director of the Clinical Microbiology Laboratory at Massachusetts General Hospital
This study showcases the enormous power of AI combined with high quality biological data sets in the discovery of potentially therapeutic compounds that otherwise would be entirely out of reach.— Donald Ingber, M.D., Ph.D., Wyss Founding Director
La Conversación del Hearth Otra perspectiva de la historia
Why does gonorrhea resistance develop so quickly compared to other bacterial infections?
The bacterium has a short generation time and enormous population size. When you expose billions of cells to an antibiotic, even rare mutations that confer resistance get amplified rapidly. It's a numbers game—and the numbers favor the pathogen.
So the AI approach is trying to find compounds the bacterium hasn't "seen" before?
Exactly. If you hit it with a drug that targets an uncommon pathway, the existing resistance mechanisms don't help. The bacterium has to evolve something entirely new, which takes longer and happens less frequently.
How confident are the researchers that A1 will actually work in patients?
They're cautious. The mouse model and the vaginal chip are promising, but they're not human bodies. There's a long road from "this works in the lab" to "this is safe and effective in people."
What happens if resistance develops to these new compounds too?
That's the real question. The hope is that by using AI to find structurally novel compounds, you buy time—maybe years instead of months. But eventually, yes, resistance will probably emerge. The goal is to stay ahead of it.
Does this change how we think about antibiotic discovery?
It does. Instead of screening thousands of compounds by hand, you can screen millions computationally and focus human effort on the most promising candidates. It's not a cure for resistance, but it's a way to accelerate the entire cycle.