From Code to Cure: Inside the Wyss Institute’s AI-Driven Fight Against Superbug *N. gonorrhoeae*

The growing threat of multidrug-resistant Neisseria gonorrhoeae has pushed researchers to abandon traditional, slow-paced drug discovery in favor of artificial intelligence. At Harvard’s Wyss Institute for Biologically Inspired Engineering, a team led by Core Faculty member Jim Collins—working with colleagues at MIT and the Broad Institute—has achieved a series of breakthroughs that don’t just screen existing drug libraries, but invent entirely new antibiotics from scratch using generative deep learning [8, 9, 52].

How the Wyss Institute used generative AI to create new antibiotics

The Collins lab’s recent work, published in the journal Cell, describes a two-pronged generative AI framework for designing antibiotics against drug-resistant N. gonorrhoeae and Staphylococcus aureus (MRSA) [7, 8]. The team used graph neural networks to systematically evaluate more than 100 million chemical fragments in silico, predicting core scaffolds with selective antibacterial activity against each pathogen . From there, they deployed variational autoencoders and genetic algorithms to expand these promising fragments into larger, fully formed molecules with the desired drug-like properties [7, 8].

In total, the models designed over 36 million candidate compounds, which the researchers computationally filtered for predicted antibiotic activity, low toxicity, and synthesizability [8, 16]. Ultimately, the team synthesized 24 of the most promising AI-designed molecules and tested them in the lab. Seven of those compounds showed antibacterial activity, and two lead candidates—designated NG1 (targeting gonorrhea) and DN1 (targeting MRSA)—demonstrated potent bactericidal effectiveness against multidrug-resistant strains in both laboratory and animal studies [8, 7, 55]. These molecules are structurally distinct from any existing antibiotics and appear to work by novel mechanisms that disrupt bacterial cell membranes .

A critical detail is that the Wyss/MIT team has not stopped at in vitro and animal testing. Collins has disclosed that he has worked directly with Wyss Founding Director Donald Ingber to leverage the Institute’s “organ-on-chip” microfluidic cell-culture devices to test the efficacy of AI-discovered antibiotics in human tissue-like environments . These platforms allow researchers to study how drugs behave in living human tissues, complementing traditional animal experiments and providing a more nuanced view of therapeutic potential before a compound ever enters human trials .

The broader AI revolution in antimicrobial discovery

The Wyss/MIT work is not an isolated incident. It reflects a fundamental shift in how the scientific community approaches antimicrobial resistance. AI is no longer just speeding up the screening of existing compound libraries; it is being used to design “new-to-nature” molecules, mine the proteomes of extinct organisms for antimicrobial peptides, and predict resistance patterns in real time from genomic data [17, 18, 20, 26].

• Generative design: The Cell paper from Collins’ lab is part of a wave that includes other groups using large language models (LLMs) to generate novel antimicrobial peptides effective against clinical superbugs [7, 22].
• Molecular de-extinction: Researchers have trained deep learning models to mine extinct organism proteomes for hidden antibiotic peptides, uncovering molecules humanity has never seen before .
• Resistance prediction: Separate machine learning studies have shown that CatBoost models can predict gonococcal resistance to key antibiotics with cross-validated AUROC scores as high as 0.97, outperforming previous best methods by 4-7 percentage points . This means clinicians could soon receive real-time, laboratory-grade forecasts of resistance using only routinely collected patient data.
• A paradigm shift: A comprehensive NIH review frames AI not just as a helpful tool but as a “catalyst” for antimicrobial discovery, addressing the core problem that traditional pipelines are too slow and too expensive to keep pace with evolving multidrug-resistant pathogens .

The Wyss Institute’s foundational role in this shift is hard to overstate. Collins’ earlier deep learning work, also done with MIT collaborators, was responsible for discovering halicin in 2019—the first new class of antibiotics identified in decades, and the first discovered leveraging an AI-powered platform [9, 47]. The newer generative-AI work for gonorrhea is a direct evolution of that same research program, moving from "AI as a screen" to "AI as a designer" [7, 50].

FDA approvals catch up: new oral drugs for gonorrhea

While the Wyss Institute’s generative AI candidates (like NG1) remain preclinical, the antibiotic discovery field received a major validation in December 2025. On December 11 and 12, the U.S. Food and Drug Administration approved two new oral medications to treat uncomplicated urogenital gonorrhea—the first entirely new treatment options in decades [33, 40, 35].

• Zoliflodacin (brand name Nuzolvence) is a first-in-class, single-dose oral spiropyrimidinetrione topoisomerase inhibitor developed by Innoviva Specialty Therapeutics in partnership with the Global Antibiotic Research and Development Partnership (GARDP) and the U.S. National Institute of Allergy and Infectious Diseases [31, 33, 39]. The Phase 3 trial of 930 patients found the drug to be non-inferior to the standard-of-care injectable regimen .
• Gepotidacin (brand name Blujepa) is a first-in-class oral triazaacenaphthylene antibacterial developed by GSK. It was initially approved in March 2025 for urinary tract infections in women, and then received a supplemental FDA approval in December 2025 for gonorrhea treatment in patients with limited or no alternative options [34, 37, 40].

Both drugs are structurally novel oral antibiotics, a critical feature because the previous standard of care—an injectable ceftriaxone-based regimen—posed logistical barriers and was increasingly challenged by rising resistance [36, 44]. However, the approvals come with important caveats. Both zoliflodacin and gepotidacin showed limited success against pharyngeal gonorrhea in earlier Phase 2 trials, meaning their use will need to be carefully managed . And neither was discovered using AI. Instead, they reflect the continued importance of traditional, non-AI small-molecule drug development, even as AI accelerates the pipeline of preclinical candidates [7, 8].

Where it all leads

The Wyss Institute’s work, and the broader AI-driven antibiotic movement it represents, sits at a pivotal intersection. On one side, generative AI models are now capable of designing structurally novel compounds that kill multidrug-resistant “superbugs” in the lab and in animal models [7, 48]. On the other, the December 2025 FDA approvals of zoliflodacin and gepotidacin prove that new chemical entities can win regulatory approval and reach patients in urgent need of alternatives to failing frontline antibiotics [33, 35]. The next step—the marriage of AI-designed candidates with human organ-on-chip testing—has already begun inside Collins’ lab .

If this integrated approach succeeds, the future of antibiotic discovery could look radically different: deep learning models propose entirely novel molecules, organ-on-chips validate their safety and efficacy in human tissue environments, and the most promising candidates move rapidly toward clinical trials. For a pathogen like N. gonorrhoeae, which the WHO and CDC have placed on their highest-priority watchlists due to its alarming resistance trajectory, the stakes could not be higher [41, 5]. The Wyss Institute’s AI-designed antibiotics may still be preclinical, but they represent a proof of concept that we can now teach machines to invent the medicines we desperately need.