Researchers have used deep learning to identify new antibiotic candidates against drug-resistant gonorrhea, a pathogen that develops resistance to standard treatments within a decade. The findings, published in Science Translational Medicine and Cell, demonstrate how machine learning can scan chemical libraries and generate novel molecules to replenish a shrinking pipeline of effective drugs.
The resistance challenge
Gonorrhea is the second most frequently reported sexually transmitted infection globally, causing tens of millions of cases annually. Left untreated, it leads to severe complications like infertility and pelvic inflammatory disease.
The primary clinical issue is the speed at which Neisseria gonorrhoeae adapts to new drugs. Melis Anahtar, assistant director of the clinical microbiology laboratory at Massachusetts General Hospital, pointed out the short lifespan of current treatments.
"We've seen this cycle of resistance development occur within just five to ten years after first-line roll-out, over and over again," Anahtar said. "To be able to prevail in this continuous arms race, we will need new antibiotics to fill the pipeline."
Screening chemical structures
A team led by James Collins at Harvard University and MIT, alongside Anahtar, pursued two parallel strategies. One study used deep learning to search existing chemical libraries, while a companion paper in Cell used generative AI to design entirely new molecules from minimal chemical fragments. This dual approach highlights how computational tools advance AI for Science & Research by accelerating early-stage drug discovery.
Instead of targeting specific bacterial proteins, which bacteria easily mutate to resist, the team trained their models on phenotypic data. They tested 38,650 small molecules in vitro to see if they could kill or suppress the pathogen in a cellular context.
To avoid selecting compounds that are simply toxic to human cells, the researchers applied strict computational and experimental filters. "We're not looking for the next bleach," Anahtar said. "We want candidates that are targeting something specific to the bacterial cell."
The Science Translational Medicine study identified two lead compounds: MP20, which increases bacterial membrane permeability, and A1, which binds to an enzyme essential for cell wall synthesis. The Cell study found two designed compounds, NG1 and DN1, which reduced bacterial burdens in animal models.
Testing in human tissue models
The researchers moved beyond standard mouse models to test the compounds in vaginal organ-on-a-chip systems developed at the Wyss Institute. These chips use human cells to mimic tissue-level interactions, offering a more accurate prediction of how drugs will behave in the human body. This integration of tissue models and machine learning represents a significant step forward for Research in drug development.
Using this system, the team observed that MP20 successfully crossed the epithelial barrier and killed N. gonorrhoeae in the vaginal lumen. The organ-on-a-chip platform also allows scientists to troubleshoot and optimize compounds by observing exactly how they interact with human tissue.
This approach helps avoid the translational pitfalls common when moving from animal models to human clinical trials. By relying on human cells rather than animal physiology, the systems provide a clearer picture of compound permeability and efficacy before entering expensive clinical phases.
Why this matters for science and research professionals
This work provides a concrete blueprint for integrating machine learning into early-stage antibiotic discovery. For scientists and researchers, it demonstrates that shifting from target-based screening to phenotypic deep learning models can uncover compounds with novel mechanisms of action.
It also validates the use of organ-on-a-chip systems for early efficacy testing. This offers a practical pathway to reduce the time and cost of bringing new antimicrobials to the clinic while addressing one of the most persistent bacterial threats.
Your membership also unlocks:
