AI is now finding and even designing new antibiotic candidates, including compounds active against drug-resistant bacteria such as MRSA and gonorrhea. That is real scientific progress. But discovery was never the main reason the antibiotic pipeline is thin. New antibiotics earn little because doctors rightly hold them in reserve, and AI does nothing to fix that.
In 2020, a team led by James Collins at MIT published a paper in Cell describing halicin, a molecule that a deep learning model had flagged as a likely antibiotic from a library of existing compounds. It killed a wide range of bacteria in the lab, including some resistant to existing drugs. In 2023, a group from McMaster University and MIT reported abaucin in Nature Chemical Biology, a compound found with machine learning that targeted Acinetobacter baumannii, one of the most feared hospital pathogens. Later that year, Collins’s group reported in Nature a new structural class of compounds active against MRSA, found using explainable deep learning.
In August 2025, the same group went further. In a Cell paper, they described using generative AI to design antibiotics from scratch, exploring tens of millions of theoretical compounds. Two candidates stood out: NG1, active against drug-resistant gonorrhea, and DN1, which cleared MRSA skin infections in mice. Both appear to act on bacterial membranes through mechanisms not seen in existing antibiotics. Phare Bio, a nonprofit linked to the project, is advancing them through preclinical work.
This is exactly what the field has been asking for: new chemical structures, new mechanisms and a faster route to candidates. And yet the most important thing to know about the antibiotic pipeline has not changed. The companies that bring new antibiotics to market have repeatedly gone broke doing it.
That is the thesis of this piece. AI is solving the part of the antibiotic problem that was difficult but relatively cheap. The part that is killing the pipeline is economic, and no model can design its way around it.
How is AI being used to discover antibiotics?
Broadly, in three ways.
The first is screening existing libraries. A model is trained on compounds known to inhibit bacterial growth, then used to score millions of other molecules that have never been tested as antibiotics. Halicin and abaucin came from this approach. It is fast and cheap because the compounds already exist.
The second is explaining what makes a compound work. Collins’s 2023 Nature work used models that identified chemical substructures linked to activity, then searched for compounds sharing them. That helps chemists understand why a molecule works, which matters for improving it.
The third is generative design. Rather than searching existing libraries, models propose new molecules. The August 2025 Cell work generated tens of millions of theoretical candidates, filtered them computationally, and had a small number synthesized and tested. The difficulty is that many designed molecules cannot easily be made: in the gonorrhea arm of that work, only two of 80 selected compounds could be synthesized. Still, the ones that could be made and worked were structurally unlike existing antibiotics.
The common promise is novelty. Resistance spreads fastest to drugs that resemble ones bacteria have already seen. Genuinely new structures and mechanisms buy time.
Why is the antibiotic pipeline so thin?
The scientific problem is real. Finding compounds that kill bacteria, particularly Gram-negative bacteria with tough outer membranes, without harming human cells is hard. Decades of screening produced diminishing returns. AI directly helps here.
But the economic problem is bigger. A new antibiotic that works against resistant bacteria is exactly the drug doctors should use as little as possible. Good stewardship keeps it in reserve for patients who have no other option, to slow the development of resistance. Most infections are treated with cheap, older generic antibiotics. So a new antibiotic, however valuable to society as insurance, sells in small volumes.
Meanwhile, the cost of clinical development and post-approval obligations is the same as for any drug. The results have been stark. Achaogen won FDA approval for plazomicin, an antibiotic for complicated urinary tract infections, in 2018, and filed for bankruptcy in 2019. Other small antibiotic developers have struggled or failed after approval too. Large pharmaceutical companies have largely left the field, because the expected returns cannot compete with oncology or immunology.
The human cost of that failure is enormous. A major analysis in The Lancet estimated 1.27 million deaths directly attributable to bacterial antimicrobial resistance in 2019. The Global Research on Antimicrobial Resistance project’s 2024 forecast, also in The Lancet, projected more than 39 million deaths directly attributable to resistance between 2025 and 2050 if trends continue. As we argued in the slow pandemic has two faces, the world faces both too much antibiotic use and too few new antibiotics, and the two problems are linked.
Does AI make antibiotics cheaper to develop?
It makes early discovery cheaper and faster. That is a smaller share of total cost than headlines suggest.
The expensive part of bringing any drug to market is clinical development: phase 1 safety studies, and then larger trials to show the drug works in patients with specific infections. For antibiotics, trials have particular challenges, including recruiting patients with resistant infections, which are unpredictable and scattered. After approval, companies must fund manufacturing, surveillance and market access, often for years before sales grow.
AI can shorten the path to a candidate and may improve the odds that candidates are safe and effective. But even if discovery became almost free, the approved drug would still face the same small market. A company with a better AI-discovered antibiotic would still face Achaogen’s problem.
A worked example: why a better molecule doesn’t fix the business
Here is a hypothetical, with round numbers for illustration only.
An AI-first startup discovers a novel antibiotic active against a resistant Gram-negative pathogen. AI cuts its discovery costs sharply, perhaps saving a few years and tens of millions of dollars compared with traditional screening. It raises money, runs trials and wins approval.
Now it has to sell. Hospital stewardship committees, doing their job, restrict the drug to confirmed or strongly suspected resistant infections. In the first years, only a few thousand patients a year receive it in its launch market. Even at a high price per course, revenue reaches tens of millions of dollars a year, while the company pays for manufacturing, a sales and medical team, post-marketing studies and regulatory obligations in multiple countries. It burns cash. Investors, looking at Achaogen, decline to fund the long wait. The drug survives only if a larger company buys it cheaply, or it does not survive at all.
Notice that AI changed only the first line of that story. Every line after approval is the same. The better the stewardship, the worse the business, which is exactly the right outcome for patients and exactly the wrong one for investors.
The strongest case that AI changes the economics
The counterargument has some force.
If AI can reduce discovery costs substantially, and nonprofits such as Phare Bio can carry candidates through early development with philanthropic and public money, the private capital needed per approved antibiotic falls. The pipeline might then be sustained by a mix of public, philanthropic and small private funding, rather than needing large pharma returns.
AI might also improve the odds in trials, by finding compounds with fewer safety problems or clearer mechanisms, reducing expensive failures. And more novel candidates mean more shots on goal for the same money.
These points are right, and they make AI-enabled discovery valuable. But they do not solve the post-approval problem, which is where antibiotic companies have died. A cheaper route to approval still ends in the same market.
What would actually fix the antibiotic market?
The fix most experts support is to pay for antibiotics in a way that is not tied to volume of sales, often called pull incentives. The idea is to pay developers for the value of having a drug available, like an insurance premium, rather than for each prescription.
The United Kingdom has gone furthest. After a pilot run by NICE and NHS England with two antibiotics, cefiderocol and ceftazidime-avibactam, the NHS adopted a subscription-style model in 2024, paying fixed annual fees for selected antimicrobials based on assessed value rather than the number of doses used. In the United States, the PASTEUR Act, which would create a similar system, has been introduced in Congress several times without being enacted. Other countries have explored their own versions, and many experts argue that pull incentives need to be coordinated across wealthy countries to be large enough to matter.
This mirrors a pattern we saw in the partnerships that worked: in global health, public money often achieves most when it creates a market, not when it funds a single product.
What this means for funders, policymakers and founders
For AI and biotech founders working on antibiotics, the business model question should come before the platform. A plan that ends at approval is not a plan. Know which pull incentives exist, how you would qualify and who might acquire the asset.
For philanthropic and public funders, AI-enabled discovery is a good investment only if something downstream will carry the resulting candidates. Funding discovery without funding development and market access produces papers, not medicines.
For policymakers, the case for pull incentives becomes stronger, not weaker, as AI improves discovery. More candidates will reach the point where the market fails. Wealthy countries that want new antibiotics to exist when they need them must be willing to pay for them in a way that does not depend on overusing them.
For clinicians and hospital leaders, stewardship remains essential. Holding new antibiotics in reserve is correct, and the right answer to the economic problem is a payment reform, not looser prescribing.
Frequently asked questions
Has AI discovered a new antibiotic?
AI has identified and designed several promising antibiotic candidates, including halicin in 2020, abaucin in 2023, a new structural class active against MRSA in 2023, and the generatively designed NG1 and DN1 in 2025. These have shown activity in laboratory and animal studies. None had completed human clinical trials or been approved as of late 2025.
Why don’t pharmaceutical companies make new antibiotics?
Because new antibiotics typically earn little revenue. To slow resistance, doctors reserve new drugs for patients with no other options, so sales volumes are low, while development and post-approval costs are high. Several companies, including Achaogen after plazomicin’s approval, have gone bankrupt after bringing antibiotics to market. Most large pharmaceutical companies have left the field.
What is the UK antibiotic subscription model?
It is a payment system in which the NHS pays drug companies a fixed annual fee for access to selected antimicrobials, based on their assessed value, rather than paying per dose. It began as a pilot with two drugs, cefiderocol and ceftazidime-avibactam, and was expanded in 2024. The aim is to reward developers without encouraging overuse.
How many people die from antibiotic resistance?
A major Lancet analysis estimated that 1.27 million deaths were directly attributable to bacterial antimicrobial resistance in 2019, with many more deaths associated with it. A 2024 forecast from the Global Research on Antimicrobial Resistance project, also in The Lancet, projected more than 39 million attributable deaths between 2025 and 2050 if current trends continue.
The takeaway
AI has earned its place in antibiotic discovery. It is producing new structures, new mechanisms and a faster path to candidates, which the field badly needed. But the pipeline is not empty because scientists ran out of ideas. It is thin because the world asks developers to build insurance and then pays them per claim.
Until that changes, the best AI-designed antibiotic will face the same fate as the best human-designed one: approval, praise and a business that cannot survive its own success.
If your health system could pay for one antibiotic to exist whether or not it is used, would it? Tell us what stands in the way.
