Why 90% of Drugs Fail in Clinical Trials, and What AI and the FDA Are Changing
Most medicines that reach human volunteers never reach a pharmacy. The reasons are less mysterious than they sound, and two forces arriving at the same time, AI drug design and the FDA’s retreat from animal testing as the default, are aimed squarely at them.
The short answer: About 9 in 10 drug candidates that enter human trials fail. Across 9,704 development programs started from 2011 to 2020, only 7.9% of drugs entering Phase 1 went on to FDA approval, and the trip took about 10.5 years on average. Most failures come from a drug not working well enough in people (roughly 40 to 50%) or causing harm that cannot be managed (about 30%). Part of the problem is that animal tests are a mixed predictor of human side effects: good at clearing drugs that are safe, weaker at flagging the specific harms that later appear.
- Odds from Phase 1 to approval: 7.9% (BIO, Informa and QLS, 2011 to 2020)
- Median cost per approved drug: $985 million, counting failures (JAMA, 2020)
- Biggest single reason for failure: lack of efficacy in patients
- What is changing: AI models that predict toxicity before a molecule is made, and an FDA rule, effective February 4, 2027, that stops treating animal tests as the legal default
Picture the funnel a medicine travels. Thousands of molecules are made and screened. A handful look promising enough to test in mice, rats, dogs or monkeys. One is chosen, filed with the FDA, and given to a few dozen healthy volunteers or patients at a tiny starting dose. That moment, the first human dose, is where most people assume the hard part is over. It is where the hard part starts.
Max Jaderberg, president of the Alphabet company Isomorphic Labs, put the root cause bluntly in a November 2026 interview: the field is simply not very good at predicting how a molecule will behave in a human body, and animal models help less than people think because “a rat isn’t a human.” His company is betting billions that AI can close that gap. To judge whether it can, it helps to know exactly where the gap is.
How Many Drugs Fail in Clinical Trials?
The most widely cited benchmark comes from a 2021 report by the Biotechnology Innovation Organization (BIO), Informa Pharma Intelligence and QLS Advisors, which followed 9,704 development programs from 2011 to 2020. Its headline number: a drug entering Phase 1 had a 7.9% chance of eventually winning FDA approval. That is down from 9.6% in the group’s previous ten-year look.
The odds are not the same for every kind of medicine. Biologics, such as antibodies and other large protein drugs, reached approval 9.1% of the time, compared with 5.7% for conventional small molecules, the pills most people picture. Cancer drugs had the lowest odds of any disease area, which matters because cancer is where much of the money and much of the AI work is aimed.
Each phase is its own filter. Phase 1 asks whether people can tolerate the drug and at what dose. Phase 2 asks whether it does anything useful in patients with the disease. Phase 3 asks whether that benefit holds up in hundreds or thousands of people against the current standard of care. Phase 2 is the steepest drop in the funnel: it is where a drug that looked elegant in a mouse first has to move a real disease in a real person, and most do not.
Those failures are expensive. Using public filings for 63 drugs approved from 2009 to 2018, researchers writing in JAMA estimated a median research and development cost of $985.3 million per approved drug and a mean of $1.34 billion, once the cost of the failures is spread across the winners (Wouters et al., 2020). For cancer and immune drugs, the median estimate was $2.77 billion. Every failed program is baked into the price of the ones that succeed, and eventually into what patients and insurers pay.
The Four Reasons Most Drugs Fail
A 2022 review from University of Michigan pharmacologists, titled plainly “Why 90% of clinical drug development fails and how to improve it?”, sorted failures into four buckets (Sun et al., 2022):
- Lack of clinical efficacy, about 40 to 50%. The drug hits its target, but hitting the target does not change the disease enough. Often the target was the wrong one, or the disease works differently in people than in the animal model.
- Unmanageable toxicity, about 30%. The dose needed to help is too close to the dose that harms.
- Poor drug-like properties, about 10 to 15%. The molecule is not absorbed, does not reach the right tissue, or is cleared too fast.
- Commercial and strategic reasons, about 10%. The market changed, a competitor got there first, or the company ran out of money.
The authors’ central argument is worth knowing because it is the same argument AI companies make. Drug makers, they write, have spent decades optimizing for potency and specificity, how tightly a molecule grips its target, while paying too little attention to where the drug actually goes in the body. A molecule that grips perfectly but piles up in the liver instead of the tumor will need a high dose, and high doses are where toxicity lives. Their proposed fix is a framework that scores candidates on tissue exposure alongside potency before anyone commits to a trial.
How Well Do Animal Tests Predict Human Side Effects?
Before any drug reaches a person, regulators have historically required safety studies in at least two animal species, usually one rodent and one non-rodent. The question of how well those studies predict human harm has a more nuanced answer than either side of the animal testing debate tends to admit.
The classic study pooled data from 12 drug companies on 150 compounds that caused 221 human toxic events during development (Olson et al., 2000). Animal studies in a rodent plus a non-rodent had flagged the same organ toxicity 71% of the time. Non-rodents alone caught 63%; rodents alone caught 43%. Animals were best at predicting blood, gut and heart problems and worst at predicting skin reactions. Notably, this study only looked at drugs that did cause human toxicity, so it measures how often animals caught a real problem, not how often they raised false alarms.
A larger, blinded industry database from the IQ Consortium filled in that missing half. Across 182 molecules with both animal and Phase 1 human data, animal studies had a sensitivity of 48% and a positive predictive value of 43%, but a specificity of 84% and a negative predictive value of 86% (Monticello et al., 2017). In plain terms: when animals showed no toxicity in an organ, people usually did not either. When animals did show toxicity, people developed it less than half the time. The authors, who work for Amgen, Lilly, Pfizer and other drug makers, concluded that this supports animal testing as a gate for safe entry into human trials.
Cancer drugs look worse. An MD Anderson analysis of 108 cancer drugs found that animal toxicity findings did not correlate strongly with what patients experienced in Phase 1, with a median positive predictive value of 0.65 and negative predictive value of 0.50 (Atkins et al., 2020). The authors called for an urgent look at new approaches.
Featured Partner
Invest in the Infrastructure Behind Modern Medicine
As healthcare expands beyond hospital walls, the buildings and campuses supporting that shift are generating compelling returns for investors who move early. The Healthcare Real Estate Fund offers qualified investors direct access to a curated portfolio of medical office, outpatient, and specialty care facilities.
Learn More →So when an AI executive or an FDA commissioner says animal models are poor predictors, the fair translation is this: they are reasonably good at telling you a drug is probably safe to try, and much less reliable at telling you which specific harm a drug will cause in a person. For a field trying to raise its odds above 8%, the second weakness is the expensive one.
When Animal Tests Miss: The TGN1412 Trial
The case that still shapes how Phase 1 trials are run happened in London in March 2006. Six healthy young men received a single infusion of TGN1412, an experimental antibody designed to stimulate immune T cells. Within 90 minutes, all six developed a violent systemic inflammatory reaction. Within 12 to 16 hours they were critically ill with lung injury, kidney failure and clotting abnormalities, and two needed intensive organ support for more than a week (Suntharalingam et al., 2006). All six survived.
The dose they received was reported to be hundreds of times lower than one monkeys had tolerated in safety testing. The drug’s target behaved differently in human immune cells than in the animals used to test it, a difference no species study could have revealed. The episode led regulators to change how first-in-human doses are chosen for drugs that act on the immune system, and it remains the clearest example of why “safe in animals” is a starting point, not a guarantee.
What the FDA Is Changing About Animal Testing
US policy has moved quickly. The FDA Modernization Act 2.0, signed in December 2022, removed the legal assumption that animal tests are required and allowed human biology-based methods instead. On April 10, 2025, the FDA announced a stepwise plan to reduce and in some cases replace animal studies, starting with monoclonal antibodies.
The pace picked up from there. In December 2025 the agency issued draft guidance aimed at phasing out the use of monkeys in testing antibody drugs, and on March 18, 2026 it released draft guidance on validating new approach methodologies, known as NAMs. These include organoids, organ-on-a-chip systems, human cell assays and computer models. FDA Commissioner Marty Makary said animal testing has a “poor track record of predicting safety and efficacy in humans.”
Then, on September 21, 2026, the FDA announced a final rule replacing animal testing language throughout its regulations with broader terms such as “nonclinical tests.” It takes effect on February 4, 2027. The agency was careful to say animal studies may still be needed when they are the best tool for a specific question. The rule does not ban animal testing; it ends its status as the default.
For patients, the practical meaning is subtle but real. A drug company can now build its safety case partly on human-derived tissue models and computational predictions, and the FDA has written down how it will judge whether those methods are trustworthy. That opens a door for exactly the kind of prediction AI companies are selling.
Can AI Raise the Odds?
AI drug design attacks the failure problem from two directions. The first is choosing better targets, the proteins or pathways a drug should hit, which goes after the biggest failure bucket, lack of efficacy. The second is designing molecules that are predicted to be safe and well-behaved before anyone makes them.
Isomorphic Labs describes its approach this way: generative models and design agents propose thousands of candidate molecules, then separate “virtual lab” models score them for properties such as binding and toxicity, so chemists only make the most promising few. Jaderberg said the company now makes 20 to 50 molecules where the old approach screened hundreds of thousands, with a high share of hits. Those figures are company-reported, and none of Isomorphic’s molecules had entered a human trial as of November 16, 2026; our Isomorphic Labs profile tracks its timeline and its target of a first human study by the end of 2026.
The best early evidence comes from companies that are further along. A 2024 analysis by Boston Consulting Group of AI-native biotech pipelines found that AI-discovered molecules passed Phase 1 80 to 90% of the time, well above historical averages, but passed Phase 2 only about 40% of the time, roughly the industry norm, on a small sample (Jayatunga et al., 2024). That split is telling. AI appears to be good at designing molecules with drug-like properties that people tolerate. It has not yet shown it can pick targets that work better against disease, which is the larger of the two failure buckets.
The strongest single data point so far is Insilico Medicine’s rentosertib, whose target and molecule were both found with AI and which improved lung function in a randomized trial in idiopathic pulmonary fibrosis. We cover that result, and its limits, in the clinical verdict on AI drug discovery and our status report on AI-designed drugs in human trials. The longer-range bet, simulating whole cells well enough to predict efficacy before a trial, is the subject of our piece on the race to build a virtual cell.
What Would Prove AI Is Working
Funding rounds and design-speed claims are not the scoreboard. The numbers worth watching are the ones in this article:
- Phase 2 success rates for AI-designed drugs climbing clearly above the industry’s historical level, across many programs, not one.
- Fewer Phase 1 surprises, meaning toxicities that AI and human tissue models predicted before dosing, published in peer-reviewed form.
- Regulatory acceptance, such as an FDA filing that relies on a validated NAM or computational model in place of an animal study.
- Cost per approved drug falling in studies like the JAMA estimate, which would be the clearest sign that fewer failures are being paid for.
Until those appear, “AI-designed” describes how a molecule was made, not whether it works. That distinction is the one most headlines skip.
What This Means for You
If you or someone you love is considering a clinical trial, especially an early-phase cancer trial, the statistics above are not a reason to say no. Phase 1 trials exist to find the right dose safely, start low, and are closely monitored. They are worth asking hard questions about: what animal or lab data supports this drug, what side effects the team is watching for, and what happens if you need to stop. You can look up any trial by name or number on ClinicalTrials.gov.
When you see a headline about an AI-designed drug, ask three questions. Has a person received it yet? What phase is it in? And is the claim from a peer-reviewed paper or a company press release? Those answers tell you more than the size of the funding round.
Finally, the honest timeline matters for your own health. Even if AI raises the odds, a drug entering Phase 1 today is still years from your pharmacy. The levers you control now, the foundations of nutrition, sleep, movement, breath and mindset, remain the best-proven defense against the Four Villains of Health and Longevity: heart disease, cancer, neurodegeneration and metabolic disease. The goal is to stay healthy long enough for better medicines to reach you.
Related: our AI healthcare breakthroughs tracker and our look at why tomorrow’s medicines start as experiments few would fund today.
What percentage of drugs fail in clinical trials?
About 90%. In a study of 9,704 programs from 2011 to 2020, only 7.9% of drugs that entered Phase 1 trials went on to FDA approval. Cancer drugs had the lowest odds and biologics did better than conventional small molecules.
Why do most drugs fail in clinical trials?
Most fail because they do not work well enough in patients (roughly 40 to 50% of failures) or because they cause harm that cannot be managed at an effective dose (about 30%). Smaller shares fail because of poor absorption or distribution in the body (10 to 15%) or for commercial and strategic reasons (about 10%).
How well do animal tests predict side effects in humans?
Mixed. In an industry database of 182 molecules, animal studies correctly cleared drugs that proved safe in an organ 86% of the time, but when animals showed a toxicity, people developed it only 43% of the time. An older study found animals had flagged 71% of human toxicities when a rodent and a non-rodent were both tested.
Is the FDA ending animal testing?
Not entirely. Since 2022 the law no longer requires animal tests by default, and the FDA has issued plans and draft guidance to replace them where human-based methods are as good or better. A final rule effective February 4, 2027 replaces animal testing language in FDA regulations with broader terms such as nonclinical tests. The agency says animal studies may still be needed for some questions.
Do AI-designed drugs succeed more often in trials?
Early data says they pass Phase 1 more often, at 80 to 90% in a Boston Consulting Group analysis published in 2024, but pass Phase 2 at about 40%, close to the industry average. AI appears to help design tolerable molecules; it has not yet shown it picks targets that work better against disease.
How much does it cost to develop a new drug?
A 2020 JAMA study estimated a median of $985 million and a mean of $1.34 billion per approved drug, including the cost of failed programs. Cancer and immune drugs had a median estimate of $2.77 billion.
Sources
- Sun D, Gao W, Hu H, Zhou S. Why 90% of clinical drug development fails and how to improve it? Acta Pharmaceutica Sinica B, 2022. PMID 35865092. DOI
- Olson H, et al. Concordance of the toxicity of pharmaceuticals in humans and in animals. Regulatory Toxicology and Pharmacology, 2000. PMID 11029269. DOI
- Monticello TM, et al. Current nonclinical testing paradigm enables safe entry to First-In-Human clinical trials: The IQ consortium nonclinical to clinical translational database. Toxicology and Applied Pharmacology, 2017. PMID 28893587. DOI
- Atkins JT, et al. Pre-clinical animal models are poor predictors of human toxicities in phase 1 oncology clinical trials. British Journal of Cancer, 2020. PMID 32868897. DOI
- Wouters OJ, McKee M, Luyten J. Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009-2018. JAMA, 2020. PMID 32125404. DOI
- Suntharalingam G, et al. Cytokine storm in a phase 1 trial of the anti-CD28 monoclonal antibody TGN1412. New England Journal of Medicine, 2006. PMID 16908486. DOI
- Jayatunga MKP, et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024. PMID 38692505. DOI
- BIO, Informa Pharma Intelligence and QLS Advisors. Clinical Development Success Rates and Contributing Factors 2011 to 2020, 2021. BIO
- RAPS. FDA to remove animal testing terminology from its regulations, September 21, 2026. RAPS
- RAPS. FDA drafts guidance on animal testing alternatives, March 18, 2026. RAPS
- Healio. FDA efforts to reduce animal testing signal shift toward AI, organ-on-chip study methods, March 2026. Healio
- Clinical Trials Arena. FDA outlines plan to phase out animal testing in new draft guidance, March 19, 2026. Clinical Trials Arena
- Silicon Valley Girl. Interview with Max Jaderberg, President of Isomorphic Labs, November 10, 2026. YouTube
Compiled by Healthcare Discovery from peer-reviewed studies retrieved from PubMed, an industry success-rate report, regulatory news coverage and a public interview. Every PMID and DOI on this page was checked against PubMed on November 16, 2026. This article is for general education and is not medical advice; talk with your care team before joining or leaving any clinical trial. No company named on this page reviewed or paid for it.
