AI Discovers Two New Antibiotics for Drug-Resistant Gonorrhea

Scientists have developed a deep learning model that could significantly accelerate the discovery of novel antibiotic compounds. In a study published today in Science Translational Medicine, the AI algorithm was able to identify two compounds effective against gonorrhea infections, each operating through mechanisms unlike those of any antibiotics in use today.

“Faced with the threat of untreatable gonorrhea, new antibiotics are urgently needed,” writes James J. Collins, PhD, professor of medical engineering and science at the Broad Institute of MIT and Harvard. “Our work establishes a much-needed hit discovery tool to address the growing crisis of antimicrobial resistance for this pathogen.”

According to the CDC and WHO, gonorrhea is one of the most urgent antibiotic-resistant threats worldwide. When left untreated, this sexually transmitted infection can lead to serious health complications. Yet treatment is becoming increasingly challenging as strains of Neisseria gonorrhoeae continue to develop resistance to existing antibiotics.

“Previous first-line therapies for gonorrhea—including penicillin, tetracycline, ciprofloxacin, cefixime, and azithromycin among others—are no longer recommended because of excessively high resistance rates in circulating strains,” says Collins. “The last remaining first-line monotherapy, ceftriaxone, is at risk of becoming obsolete, with resistance rates of over 10% in parts of the world.”

Traditional antibiotic discovery relies on high-throughput screening methods that search large chemical libraries for promising compounds. However, this approach cannot keep pace with the fast expansion of chemical space, which today englobes over 75 billion compounds.

To overcome this limitation, Collins and colleagues turned to deep learning algorithms capable of making large-scale screening significantly more time- and cost-efficient. They selected a predictive graph neural network (GNN), a model that represents molecules as graphs and can virtually screen thousands of molecules per second. Compared to large language models, the GNN more accurately identified molecules with activity against gonorrhea that were structurally different from both the training data and existing antibiotics.

To train the model, the team screened 38,650 small molecules with diverse structures and identified those capable of inhibiting the growth of N. gonorrhoeae. The model was then able to virtually screen through nearly six million compounds, which led to identification of 83 antibiotic candidates with confirmed activity against N. gonorrhoeae.

Among them, two compounds stood out for having distinct structures from existing antibiotics, known as MP20 and A1. In mice and an organ-on-a-chip model of vaginal gonorrhea infection, both molecules rapidly and selectively killed the bacteria without inducing drug resistance.

Proteomics analyses revealed the mechanism of action of each antibiotic candidate. MP20 was found to disrupt the bacterial membrane and damage DNA, while A1 targeted an enzyme essential for the synthesis of the bacterial cell wall.

“There is a critical need to fill the trickling antimicrobial development pipeline with promising antibiotic candidates, especially ones with distinct mechanisms of action,” says Collins. “We have shown that these compounds can retain potency against even the most highly resistant strains of N. gonorrhoeae, exhibit selective killing toward bacteria, and act through orthogonal mechanisms from those of existing first-line antibiotics.”