AI Antibiotics MIT Built: How a Model Invents a Molecule

The AI antibiotics MIT researchers produced in 2025 came out of a pipeline with a shape every kid should see: 36 million molecules in, two useful drugs out. James Collins’s lab at MIT used generative algorithms to design more than 36 million candidate compounds, screened them computationally for antibacterial activity and toxicity, narrowed to roughly a thousand promising ones, and then hit the wall that stops most AI chemistry: of 80 compounds picked for synthesis in one branch of the project, chemists could only actually make two. Those two, named NG1 and DN1, went on to clear multidrug-resistant gonorrhea and MRSA infections in mice. The paper appeared in Cell on August 14, 2025.

Key Takeaways

• Two separate generative strategies were used: a fragment-based one (start from a chemical piece the model liked) and an unconstrained one (generate freely, only requiring chemical plausibility).
• Scale: 36 million-plus compounds generated. The fragment branch screened ~45 million fragments down to 4 million, then to 1 million, then generated ~7 million full molecules, then predicted ~1,000 as active, selected 80, and synthesized 2.
• NG1 kills drug-resistant Neisseria gonorrhoeae by targeting a protein called LptA that builds the bacterial outer membrane. DN1 kills MRSA by disrupting bacterial membranes more broadly.
• Both worked in mouse infection models. Neither has been tested in humans. The nonprofit Phare Bio is doing the follow-up medicinal chemistry.
• The synthesis bottleneck is the real story: a model can imagine a molecule far more easily than a chemist can build it.

Why anyone needed new antibiotics badly enough to try this

Almost no structurally new class of antibiotic has reached the clinic in decades, while resistance keeps spreading. The CDC counts more than 2.8 million antimicrobial-resistant infections and over 35,000 deaths per year in the United States alone (CDC).

Gonorrhea is a specific alarm. Neisseria gonorrhoeae has developed resistance to nearly every drug class thrown at it, and the World Health Organization lists it as a high-priority pathogen. MRSA, methicillin-resistant Staphylococcus aureus, is the other target: common in hospitals and in skin infections, including in kids.

Traditional antibiotic discovery meant screening soil bacteria and natural products, hoping something new turned up. That well has been largely tapped. Generative chemistry asks a different question: instead of searching what exists, can we design what doesn’t?

The AI antibiotics MIT pipeline, stage by stage

The work was led by James Collins, the Termeer Professor of Medical Engineering and Science at MIT, with first authors Aarti Krishnan, Melis Anahtar, and Jacqueline Valeri (MIT News).

What the models actually computed. Two algorithms did the generating. CReM (Chemically Reasonable Mutations) takes an existing molecule and makes small legal edits: add an atom, swap an atom, delete a group. Every edit respects real chemistry rules, so the output is at least buildable in principle. F-VAE (Fragment-based Variational Autoencoder) works differently: it learned patterns from over a million molecules in the ChEMBL database and assembles complete molecules out of fragments, the way you’d build a sentence from phrases you’ve seen before.

What data they used. ChEMBL, a public database of bioactive molecules, for learning what drug-like chemistry looks like. And their own screening data on which chemical fragments show antimicrobial activity, for training the activity predictors.

The two strategies. In the fragment-based branch, aimed at gonorrhea, the team first screened about 45 million chemical fragments and picked one promising piece, called F1, then asked the generators to build full molecules containing it. In the unconstrained branch, aimed at S. aureus, they let the algorithms produce molecules with no structural requirement beyond being chemically sensible, generating over 29 million.

The filters. Predicted antimicrobial activity. Predicted cytotoxicity, so the compound doesn’t poison human cells. And dissimilarity to existing antibiotics, because the whole point was a new structural class that resistance mechanisms haven’t seen.

What it got right. NG1 targets LptA, a protein involved in assembling the lipopolysaccharide outer membrane of Gram-negative bacteria. Hitting LptA is an unusual mechanism, which matters because existing resistance doesn’t apply. It cleared drug-resistant gonorrhea in lab cultures and in a mouse infection model. DN1 disrupted bacterial membranes more broadly and cleared MRSA skin infections in mice.

What it got wrong, or couldn’t do. The synthesis numbers are brutal and worth repeating. In the fragment branch, 80 compounds were selected for synthesis and 2 were successfully made. In the unconstrained branch, about 90 candidates yielded 22 synthesized. A model can propose a molecule that is chemically legal on paper and still practically unmakeable, because no known reaction sequence gets you there at reasonable cost. Nothing has been tested in humans. Mouse efficacy is a real milestone and also a stage where most antibiotic candidates eventually fail.

The analogy: a machine that designs LEGO ships

Picture a program that has studied a million LEGO instruction booklets. It has never touched a brick. Ask it for new spaceship designs and it produces 36 million of them, all obeying the rule that studs connect to sockets.

Now try to build them. Most designs call for a brick in a color and shape that doesn’t exist in any set you can buy. Some need a piece to be inserted in the middle after the hull is sealed, which is physically impossible. Out of 80 designs you picked, you manage to build two. Both fly better than anything in the booklets.

That is generative chemistry with today’s tools. “Chemically plausible” and “synthesizable this decade” are different standards, and the gap between them is where 78 out of 80 designs died.

The pipeline in numbers: where 36 million became 2

Stage	Fragment-based branch (gonorrhea)	Unconstrained branch (MRSA)
Starting pool	~45 million chemical fragments screened	No starting constraint
After antimicrobial prediction	~4 million fragments	—
After toxicity + novelty filters	~1 million fragments	—
Full molecules generated	~7 million (via CReM and F-VAE)	29+ million
Predicted active	~1,000 compounds	~90 candidates
Selected for synthesis	80	~90
Actually synthesized	2	22
Cleared infection in mice	1 (NG1)	1 (DN1)

Read this table left to right and the lesson is obvious: the computer stages shed orders of magnitude cheaply, and the laboratory stages are where the project nearly ends.

How to Teach Your Kid About AI-Designed Medicines

Ages 5–8: Build-a-Molecule With Toothpicks

Use marshmallows or clay balls as atoms and toothpicks as bonds. Give your kid a rule: every white ball gets exactly one stick, every black ball gets exactly four. Now ask them to invent three new shapes that follow the rules. Then ask them to build one they drew but didn’t check against the rules first. Usually it can’t be built. That’s the synthesis bottleneck, felt in the hands, with no chemistry vocabulary required.

Ages 9–12: The Filter Funnel Game

Write 40 nonsense “molecule names” on slips of paper. Apply three filters in order: keep only names with a vowel in position two (activity filter), then only those shorter than eight letters (toxicity filter), then only those that don’t rhyme with an existing name (novelty filter). Count survivors at each step. Then compare with MIT’s real numbers: 45 million to 4 million to 1 million. Ask which filter cut the most and why order matters.

Ages 13+: Read the Funnel and Find the Bottleneck

Give your teen the table above and ask one question: if you had unlimited computing power but the same chemistry lab, how many more drugs would you get? (Answer: almost none. The bottleneck is synthesis, not generation.) Then have them look up “retrosynthesis prediction,” the AI subfield trying to fix exactly that, and write three sentences on why it might matter more than better molecule generators. If they want a hands-on version, MolView lets them draw molecules in a browser and see 3D structures immediately.

The question to ask: “The AI designed 36 million molecules and chemists made 24 of them. Which number tells you more about how close we are to a new medicine?”

What to actually do at home

Separate “designed” from “delivered”

A headline saying AI created a new antibiotic is true and incomplete. The accurate version: AI proposed structures, humans built a couple dozen, two worked in mice, and human trials haven’t started. Teaching kids to append “and then what happened?” to every AI-science headline is a portable skill.

Point out that the mechanism matters, not just the result

NG1 works by hitting LptA. That specificity is why the compound matters: it’s a mechanism existing resistance genes weren’t selected against. Kids who learn to ask “how does it work?” instead of “does it work?” end up understanding medicine, not just consuming it.

Use the numbers, not adjectives

“36 million designed, 24 synthesized, 2 effective in mice” is more memorable and more honest than “breakthrough.” Our companion piece on the AI materials discovery funnel uses the same framing for materials science, and the parallel is exact.

Connect it to why doctors resist prescribing antibiotics

The resistance problem is the whole reason this research exists. If you’ve ever left a pediatrician’s office without the antibiotic you expected, that’s the front line. Our guide on antibiotic overuse in children explains the clinical side, and this research is one of the long-term answers.

What not to do

Don’t tell your kid that AI “discovered” NG1 the way Fleming discovered penicillin. It’s a different act. Fleming noticed something real on a plate. The model produced a structure that humans then had to synthesize and test. Both count as science, but the roles are not interchangeable, and conflating them makes the next decade of headlines harder to read.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can state the funnel out loud: millions designed, dozens built, two worked in mice. Precision on those three numbers is the goal.
• Month 2 red flags: They say “AI cured MRSA.” Pull the paper’s mouse-model language back up and add the missing words: in mice, not yet in people.
• Month 3 self-check: Ask them to find another AI-drug headline and identify its funnel. If the article doesn’t give one, that absence should bother them. That reflex is the outcome worth having.

Frequently Asked Questions

Are NG1 and DN1 available as medicines?

No. Both showed efficacy in mouse infection models, which is an early preclinical stage. Human trials require extensive safety testing, dose-finding, and manufacturing work first. The nonprofit Phare Bio is exploring structural modifications to both compounds, and the MIT team is extending the platform to Mycobacterium tuberculosis and Pseudomonas aeruginosa.

How is this different from AlphaFold or protein prediction?

Protein structure prediction figures out the shape of a molecule that already exists in nature. Generative chemistry invents molecules that don’t exist and proposes their structures. Different problem, different failure mode: a wrong structure prediction is a wrong answer, while a wrong generated molecule is often an unbuildable one.

Why could chemists only make 2 out of 80 compounds?

Because “chemically reasonable” on a screen is a lower bar than “synthesizable in a real lab at reasonable cost.” Building a molecule requires a sequence of reactions that each work, in an order that doesn’t destroy earlier parts. Many generated structures have no known route. This is why retrosynthesis prediction is an active AI research area.

Does this mean AI will replace chemists?

The MIT numbers argue the opposite. The generative stage was fast and cheap; the human chemistry stage was the constraint. If anything, this kind of work increases the value of synthetic chemists, because their capacity now limits how many AI ideas can be tested at all.

What’s LptA and why does it matter for kids to know?

LptA is a protein that helps Gram-negative bacteria assemble their outer membrane, the protective layer that also makes those bacteria hard to kill. Attacking it is unusual, which means bacteria haven’t evolved defenses against that specific attack. For a kid, it’s a clean illustration of why how a drug works determines whether resistance already exists.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Krishnan, A., Anahtar, M., Valeri, J., et al. (2025). “Generative AI-designed compounds active against drug-resistant bacteria.” Cell, August 14, 2025. Summary: https://news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814
2. MIT News. (2025, August 14). “Using generative AI, researchers design compounds that can kill drug-resistant bacteria.” https://news.mit.edu/2025/using-generative-ai-researchers-design-compounds-kill-drug-resistant-bacteria-0814
3. MIT Technology Review. (2025, October 21). “AI-designed compounds can kill drug-resistant bacteria.” https://www.technologyreview.com/2025/10/21/1124755/a-i-designed-compounds-can-kill-drug-resistant-bacteria/
4. Centers for Disease Control and Prevention. “About Antimicrobial Resistance.” https://www.cdc.gov/antimicrobial-resistance/about/index.html
5. Centers for Disease Control and Prevention. (2019). Antibiotic Resistance Threats in the United States. https://www.cdc.gov/antimicrobial-resistance/media/pdfs/2019-ar-threats-report-508.pdf
6. Euronews Health. (2025, August 15). “MIT scientists use AI to develop new antibiotics for stubborn gonorrhoea and MRSA.” https://www.euronews.com/health/2025/08/15/mit-scientists-use-ai-to-develop-new-antibiotics-for-stubborn-gonorrhoea-and-mrsa
7. MIT Institute for Medical Engineering and Science. James Collins lab profile. https://hst.mit.edu/node/2594