Researchers at Harvard’s Wyss Institute, Massachusetts Institute of Technology, and the Broad Institute have developed a deep learning-based antibiotic discovery approach to address the growing crisis of multidrug-resistant gonorrhea.
The study, published in Science Translational Medicine, successfully deployed artificial intelligence to screen millions of compounds and identified entirely new chemical structures that can kill pathogens through new cellular pathways.
The cycle of antimicrobial resistance
Gonorrhea is the second most commonly reported sexually transmitted infection (STI) in the world, with more than 600,000 cases reported annually in the United States alone. If left untreated, gonococcal infections can lead to pelvic inflammatory disease, infertility in both men and women, increased risk of HIV infection, and life-threatening systemic complications such as meningitis and sepsis.
Two new oral antibiotics, zoliflodacin and gepotidacin, were recently approved for the treatment of genitourinary gonorrhea, the first completely new classes of antibiotics for this infection in more than 30 years, but history shows that gonococcal bacteria adapt rapidly. Significant resistance typically emerges within 5 to 10 years of initial product deployment. To break this continuing arms race, scientists need to discover chemical structures that target unusual biological pathways and reduce the statistical frequency of pathogen resistance.
Building an AI detection pipeline
To uncover these “hidden gems” of antimicrobial activity, the research team designed a multi-step machine learning workflow.
- Train the model:
- The researchers manually tested 38,650 small molecules in a laboratory assay against Neisseria gonorrhoeae. They used this empirical dataset to train a predictive deep learning model that recognized patterns associated with anti-gonococcal activity.
- Virtual screening:
- A trained AI model was used to virtually screen a large compound library containing approximately 6 million small molecules.
- Filtering and separation:
- Computer screening yielded 213 potential candidates. Through continuous biological growth assays, resistance tracking, and toxicity filters to protect human cells, the team isolated two highly potent compounds with very low resistance frequencies.
New cellular targets
Researchers used proteomic analysis to identify the precise biological mechanism of the leading candidate aminothiazole compound, named A1.
A1 specifically binds to and inhibits an enzyme called alanine racemase. This enzyme is required by Neisseria gonorrhoeae to synthesize and repair the bacterial protective cell wall. Although a variety of existing antibiotics target cell wall biosynthesis, selectively neutralizing alanine racemase with small molecules constitutes an entirely new mode of action in the treatment of gonorrhea.
Verification using organ chips and animal models
To confirm that the compound could work outside of a simulated environment, the team tested their discovery in a complex physiological tissue environment.
The researchers, working with the Vaginal Chip team at the Wyss Institute, introduced the first compound, MP20, into a microfluidic model lined with living human vaginal epithelial cells. This treatment successfully reduced the pathogen titer within the device.
Additionally, they tested a second compound, A1, in a vaginal infection model in live mice. Application of five topical treatments of A1 in a 24-hour window significantly reduced the concentration of N. gonorrhoeae compared to untreated controls.
Although A1 compounds require further medicinal chemistry optimization and hit-to-lead development before moving to human clinical trials, the pipeline’s success demonstrates that combining high-quality biological data with artificial intelligence can quickly uncover therapeutic compounds that are beyond the reach of science.
