By Deena Beasley
Oct 7 (Reuters) – Novartis’ experimental del-desiran for a rare type of muscular dystrophy was expected to be a winner, with the Swiss company’s CEO predicting annual peak sales of $5 billion. When a late-stage trial missed its goal last month, shares fell 11%, erasing $30 billion in market value.
AI start-up BioinvestGPT was not surprised. In July, it ran a simulated trial to measure how virtual patients would respond to del-desiran and predicted an insignificant clinical benefit.
It and other artificial intelligence companies say they are working with some drugmakers on virtual trials to boost the odds an experimental medicine will prove safe and effective when used by actual patients in an expensive traditional clinical trial.
The global biopharmaceutical industry spends some $140 billion a year on human clinical testing, but only around 12% of drug candidates are approved by regulators — a rate that has changed little in decades.
AI firms say simulations are being used to identify which pharmaceutical programs are worth moving forward and to assess the value of acquisition targets. As AI tools get better at flagging clinical risks, they say unexpected trial failures may become less frequent.
Human clinical trials typically start with small Phase 1 studies primarily to gauge safety, followed by mid-stage Phase 2 and the large Phase 3 trials required by regulators that also assess efficacy. The process takes years, compared with a month or less for some AI simulations.
“We shouldn’t only ask how to run trials faster. We should ask how to run fewer trials that are going to fail,” said Francisco Beca, chief medical officer at QuantHealth, an AI clinical trial simulation platform headquartered in Tel Aviv.
Investment in AI drug discovery more than doubled to $8.4 billion in 2025 compared to 2023, according to a recent report from McKinsey. It said spending is currently focused where the technology works best today, particularly molecule design, rather than the industry’s major bottlenecks, such as proving a drug will work as intended through lengthy trials.
Pharma companies are “dipping their toes” into AI trial modeling, said McKinsey partner Alex Devereson. They are using different types of the technology, internally or with partners, to do things like assess drug candidates before committing capital to a program, he said.
US health regulators last week announced a set of initiatives aimed at speeding drug trials. If successful, the program could help create a path for predictive AI to be used in clinical development, a federal health official said.
FIVE OUT OF SIX CORRECT
Copenhagen-based BioinvestGPT in July shared with Reuters its detailed analyses of the likely outcome of several high-profile drug trials before the results were known and has so far been correct in five of six.
They include the negative result for del-desiran, success for Moderna and Merck’s melanoma vaccine, a weak clinical benefit for AstraZeneca and Ionis’ heart drug Wainua, the first failed trial for Novo Nordisk’s heart drug ziltivekimab, and success for Vaxcyte’s pneumococcal vaccine.
The fundamental goal of the pharma industry is proof of a superior clinical benefit compared to the standard of care, said Bragi Lovetrue, who co-founded BioinvestGPT with his wife Idonae Lovetrue in 2024.
The AI platform uses DNA sequencing to simulate a human body matching the eligibility criteria for a specific clinical trial. A virtual trial is then done using a model of the test drug.
“We can pinpoint the reason why a drug is effective and safe, and in many cases, why not,” Lovetrue said.
The simulations are not always right. The predictive AI forecast positive results for Novartis’ pelacarsen, which lowers blood levels of a cholesterol-carrier called lipoprotein(a). Novartis in September said the drug did not cut the risk of a major heart attack or stroke in patients with a genetic risk factor in a late-stage trial.
A subsequent analysis showed “we got the mechanism wrong,” by not accounting for genetically set variations in the size of the lipoprotein, Lovetrue said.
QuantHealth, which uses real-world data and AI to simulate patient-level responses to therapies, has published its simulations of ulcerative colitis and cholesterol drug trials.
BIOGEN, TAKEDA TRIALS PREDICTED TO FAIL
BioinvestGPT has done predictions for a wide range of trials.
For anticipated trial results expected before year-end, the AI company predicts failure for two Phase 3 trials of Biogen’s litifilimab in the most common type of lupus, as well as for Phase 2 studies of Japanese drugmaker Takeda’s zasocitinib in Crohn’s disease and ulcerative colitis.
The AI model shows that both drugs are “suboptimal” for those specific trial populations.
Biogen’s drug, which targets an immune cell receptor called BDCA2, is also being studied in a different type of lupus. Takeda filed recently for US approval of zasocitinib, which blocks an enzyme called tyrosine kinase 2, in plaque psoriasis and has a Phase 3 trial in psoriatic arthritis.
Takeda research chief Andy Plump said TYK2 was identified as a target by analyzing human genetics, while machine learning was used to optimize and “polish” the structure of the daily pill.
“I have immense confidence in this mechanism,” Plump said. “I don’t think we are near being able to use tools like AI to make definitive predictions.”
Diana Gallagher, Biogen’s head of clinical development for multiple sclerosis, immunology and Alzheimer’s, said the company uses “every tool available to us,” including AI.
She said only two biologic drugs have been approved for lupus and AI algorithms that rely on historical data could be prone to predicting a negative result.
QuantHealth’s Beca and other experts said human trials will always be necessary, but AI simulations can be valuable when assessing the potential of an experimental drug.
“Is it still ethical in 2026 to expose patients to a trial that most likely will fail?” he said. “With the advancement of this (AI) technology, come 2027, 2028, or 2029, probably the answer is going to be that it no longer is.”
(Reporting By Deena Beasley; Editing by Caroline Humer and Bill Berkrot)
