A Copenhagen startup says it simulated a drug trial in July and got the answer the real trial gave in September. The smart reading of that is not “AI replaces the lab.”
In July, a small Copenhagen company called BioinvestGPT sent Reuters its analyses of several high-profile drug trials whose results were not yet public. One of them covered del-desiran, a Novartis drug for a rare form of muscular dystrophy. Novartis’s chief executive had talked about peak annual sales of as much as $5 billion. The startup’s simulation said the clinical benefit would be insignificant.
In September, the real Phase 3 trial missed its main goal, days after Novartis had already taken a separate hit when pelacarsen failed. Reuters ties the 11 percent share drop and roughly $30 billion in lost market value to the del-desiran result.
Reuters reported the whole record on October 7. BioinvestGPT has now called five of six trials correctly. I have been turning that number over for a day, and I think most of the commentary is reading it the wrong way around.
What the company actually claims
According to Reuters, the platform uses DNA sequencing data to build a virtual patient group that matches a specific trial’s eligibility rules. It then runs a model of the drug against that group. Some of these simulations finish in a month or less. A traditional trial takes years.
The hits Reuters listed:
- Moderna and Merck’s melanoma vaccine: predicted to succeed, and it did.
- AstraZeneca and Ionis’s heart drug Wainua: predicted to show weak clinical benefit, and it did.
- A Novo Nordisk cardiovascular study: predicted to fail, and it did.
- Vaxcyte’s pneumococcal vaccine: predicted positive, and it was.
- Novartis’s del-desiran: predicted insignificant benefit, and the trial missed.
The miss was Novartis’s cholesterol drug pelacarsen. BioinvestGPT predicted positive results. The drug failed to reduce the risk of major heart attacks or strokes in Phase 3. The company owned the error. Co-founder Bragi Lovetrue told Reuters, “we got the mechanism wrong,” because the model did not account for genetically set variation in lipoprotein size.
Two things about that record matter before anyone gets excited. First, six is a small number. TipRanks, which covered the same story, notes that the five-of-six record is based on a small sample. Second, the predictions Reuters received were ones the company chose to share in advance. I have not found a complete public, timestamped list of every call the company has made. Until there is one, treat the hit rate as interesting, not proven.
The wrong reading
The headline version of this story is “AI predicts drug failures, Wall Street should panic.” That is the reading that moves stock tickers and gets clicks. It also misses what the numbers say about the money.
Pharma spends roughly $140 billion a year on human clinical testing, and about 12 percent of drug candidates end up approved, according to the figures in the Reuters and TipRanks reports. Look at those two numbers together. A large share of that spend goes to drugs that do not make it.
Nobody needs a model to replace the trial. A trial is how you find out what a drug does in people, and the pelacarsen miss is a clean example of why. The model got the biology partly wrong, and only the real trial could settle it. Takeda’s head of research, Andy Plump, made the same basic point when asked about the startup’s prediction that Takeda’s Crohn’s and colitis drug zasocitinib will fail. He said he remains highly confident in the drug’s biological mechanism, and he cautioned that AI cannot yet make definitive clinical predictions.
I think he is right about the “definitive” part. And I think it does not matter nearly as much as it sounds.
What a rough forecast is worth
Suppose a model is right two thirds of the time, or three quarters, or even only more often than a coin flip. It still changes what a smart organization does, as long as nobody treats it as a verdict.
Think about what a decision at a drug company looks like. You have twelve candidates and money for four Phase 3 trials. Each trial costs hundreds of millions. Today the choice is made by committee, with real data and a lot of human judgment, and it is wrong a lot. Now add a cheap, fast signal that says “this one looks poorly matched to the patient group you are about to test it on.” You do not cancel the drug because of that. You ask why. You look at the patient selection. You change the enrollment criteria, or you test a different population, or you move that candidate to the back of the line and spend the time on something else.
That is the use I find convincing. Not skipping the research, but steering where the money and the months go. The model is a compass. It is not a map, and it is definitely not the territory.
The BioinvestGPT story, read that way, has a very specific lesson. On Biogen’s lupus drug litifilimab, the model’s conclusion was that the drug is suboptimal for the patient populations in the trials. That is a statement about trial design as much as about the drug. If it is right, the useful response from Biogen is not to panic. It is to check whether the enrollment is wrong.
The part where I stay skeptical
I want to be fair to the doubters, because I am one of them on several points.
A virtual patient built from sequencing data is only as good as what we currently know about disease biology. The pelacarsen miss shows what happens when something in that biology is not captured. There is also the incentive problem. A startup that publishes predictions has every reason to publicize the hits, and a track record chosen after the fact looks better than one committed to in advance. The only fix for that is a public log, written before the results, that anyone can check later.
Biogen’s Diana Gallagher gave Reuters the strongest counterargument. She said the company uses every tool available to it, AI included, and that algorithms trained on historical data may lean toward predicting failure in lupus, because only two biologics have ever been approved for it. A model that has mostly seen failure will tend to forecast failure. I cannot rule that out from the outside, and neither can the startup yet.
And markets can overreact. If a few high-profile calls turn out right, traders may start treating the model’s output as if it were a trial result. That is how a useful forecasting tool turns into a rumor mill with a nicer interface.
What to watch next
The company has put out two more predictions, and both say failure. Biogen’s litifilimab is in two Phase 3 studies for systemic lupus. Takeda’s zasocitinib is in Phase 2 trials for Crohn’s disease and ulcerative colitis. Real data will arrive on both. If the model is right again, that will be seven of eight, still a small sample but harder to wave away. If it is wrong on both, we will have learned something just as useful about where these simulations break.
Either way, the thing I would watch is not the stock move. It is whether any company announces that it changed a trial because of a simulation. That would be the first sign that the compass is being used the way I think it should be.
The takeaway
I do not think a Copenhagen startup just made clinical trials obsolete. I think it showed that a cheap, fast, imperfect signal can point at where an expensive, slow, necessary process is likely to go wrong. For a $140 billion a year industry where nearly nine in ten candidates fail, that is worth a great deal even if the signal is wrong now and then.
Use the forecast to decide where to look. Then go look.
Sources: Reuters (Deena Beasley, October 7, 2026) and TipRanks (Ran Melamed, October 7, 2026), both reporting on BioinvestGPT’s predictions. Figures are as reported by those outlets.
Chris Meredith writes about AI, technology, and what it actually means for real people. Follow along on Substack: monkeyattack.substack.com
