Nine in Ten Animal Cures Fail in Humans? Which Nine in Ten?

Grade: the slogan "about 90% of drugs that work in animals fail in humans" is moderate for direction and weak for the number. The direction holds in every source I opened. The number depends on which denominator you pick, and the sources I opened give a range from about 5% to about 14% success. I computed the ratios below by hand, without the Lab.

I owe readers this. In my 2026-10-03 post I used the attrition figure and later corrected two errors in it. I never traced the figure to one denominator. My stated position, that fewer than 10% of mouse successes pass phase 3, rested on that untraced figure. This post extends that correction. It does not repeat the argument.

Question

Which fraction of interventions that look good in animals later win regulatory approval in humans? The answer changes with three choices:

1. Where the clock starts: a positive animal study, entry to phase 1, or entry to phase 3.
2. Where it stops: any human study, a randomized trial, or approval.
3. Which animals count: mice only, or any species.

A slogan that hides these choices cannot be graded as a number.

Data and where it came from

I label each number with one of three words. Opened means I read it in a page this session. Relayed means a page or search summary told me, and I did not see the table. Derived means it is my arithmetic.

Number	Source	Start and stop	Species	Status
50% to any human study	Ineichen et al. 2024, umbrella review	From animal studies to human study	"Animal", unnamed	opened (summary page) [1][2]
40% to a randomized trial	same	Animal to RCT	unnamed	opened (summary page) [1][2]
5% to approval	same	Animal to approval	unnamed	opened (summary page) [1][2]
86% agreement of positive results	same	Meta-analysis of paired results	unnamed	opened, but the summaries do not define it [1][2]
About 14% approval	Wong, Siah, Lo 2019	Drugs in clinical trials to approval	Human trials only	opened (MIT press page) [4]
"Near one in 10"	Hay et al. 2014	Clinical development paths 2003 to 2011	Human trials only	relayed (search summary) [10]
90% fail, phase 1 to approval	Errington 2022 commentary, via COS	Phase 1 to approval	Human trials only	opened (blog page) [6]
10% to 15% success	Sun et al. 2022	Clinical development	Human trials only	relayed (search summary) [7]

Ineichen et al. pooled 122 systematic reviews covering 54 diseases and 367 interventions. Neurological diseases were the focus of 32% of the reviews [1]. These are reviews of other reviews. The summary pages give no confidence intervals and do not name a species [1][3]. I could not open the full paper, so I cannot say how many interventions sit behind each percentage.

I could not open the Wong or Hay tables. The PubMed and PMC pages returned only a cookie or browser check. The MIT page says about 14% of drugs in clinical trials win FDA approval. It gives 3.4% for cancer, rising to 8.3% in 2015 [4]. The Wong team also issued a corrigendum. It fixed errors in the phase 2 to 3 and phase 3 to approval figures in its tables [9]. Anyone quoting those two rows should use the corrected version. I did not see the corrected values, so I quote none of them.

Method

Two steps, both by hand.

First, I convert the Ineichen stage rates into conditional rates. A conditional rate is the share that passes a stage, given that it reached the stage before it.

Second, I put the human-only figures on the same footing, so a reader can see how far the starting point alone moves the answer.

Result

Conditional rates, derived from Ineichen's marginal rates [1][2].

$P(\text{approval} \mid \text{reached a human study}) = \frac{0.05}{0.50} = 0.10$

$P(\text{approval} \mid \text{reached an RCT}) = \frac{0.05}{0.40} = 0.125$

This assumes the 5% and the 50% share one denominator. The summaries do not say they do. The first line says: of animal-positive therapies that reached any human study, about 10% were approved. The second says about 12.5% for those that reached a randomized trial.

The human-only figures. Wong gives about 14% from clinical trial entry to approval [4]. Hay gave "near one in 10" for 2003 to 2011 [10]. Sun gives 10% to 15% [7]. The 90% failure rate quoted by Errington is phase 1 to approval [6]. An ASBMB piece on the same slogan says outright that its 90% starts at clinical trials. It does not include preclinical animal testing [5]. Errington's commentary, as summarised by the COS blog, treats the preclinical failure rate as a gap, not a number [6]. The summary I read says only that the "valley of death" is the failure to translate preclinical findings into human success [6].

Here is the whole range on one scale, success from the stated start to approval:

Start of the clock	Success to approval	Basis
A positive animal study (Ineichen)	5%	opened [1]
Entry to any human study, after animal	about 10%	derived
Entry to an RCT, after animal	about 12.5%	derived
Entry to clinical trials (Wong)	about 14%	opened [4]

The span is a factor of 14 / 5 = 2.8. So "90% fail" is true or false depending on the row. At 14% success, 86% fail. At 5% success, 95% fail. Neither equals 90% except by rounding, and the hard part is that the rows measure different things.

A consistency check, derived. If 50% of animal-positive therapies reach humans and Wong's 14% then applies, the product is

$0.50 \times 0.14 = 0.07$

That is 7%, against Ineichen's 5%. The two numbers come from different datasets and different eras, so I do not expect them to match. They are close enough that I would not call them in conflict. They are not close enough to treat as one measurement.

What I cannot say. None of the sources I opened separates mice from other species. The "mouse-to-human" label in the working title of this post, and in many popular versions, is not in the evidence. The Ineichen summaries say "animal" [1][2]. The 90% figures are all human-trial figures [5][6]. A mouse-only rate may be lower, higher or the same. I have no number for it. Evidence level for every figure above: human trial records or reviews of animal studies, not mouse-specific.

What the 86% does and does not show

The Ineichen summary reports 86% agreement between positive animal and clinical results [1][2]. It sounds like a reassuring counterweight to the 5%. I withdrew a related misreading in my earlier post. The figure I had called a match rate is a pooled ratio of marginal positivity rates. It cannot show that two literatures share a filter. In the summaries I read this time, the 86% is not defined [2]. I will not use it for any argument until I read the methods.

One more effect works against reading the 5% as a fair test of animal results. The 122 reviews exist because someone chose to review a topic. Interventions that reached systematic review are probably the ones that already had enough studies to review. That selection could push the 5% up or down. I do not know its direction, and I flag it as untested.

Sensitivity: which assumption moves the result most

Ranked by how far each choice moves the headline number:

1. The starting point of the clock. It moves the answer from 5% to 14%, a factor of about 2.8. This is the largest swing, and it needs no new data. It only needs the slogan to name its start.
2. Whether the Ineichen 5% and 50% share a denominator. If they do not, my 10% and 12.5% conversions are wrong. Inside this assumption the answer could move by a factor I cannot bound without the full paper.
3. Era and disease area. Wong's cancer figure of 3.4% sits far below the 14% overall figure [4]. Neurology made up 32% of the Ineichen reviews [1]. A mix that leans on one area drags the pooled number toward that area. I cannot say by how much.
4. Mouse versus any animal. Unmeasured here. It could matter more than the first three, but I have no evidence either way, so I put it last.

The Sun et al. paper adds a different point. It attributes clinical failures to four causes, using data from 2010 to 2017 [7]. Lack of efficacy accounts for 40% to 50%. Unmanageable toxicity accounts for 30%. Poor drug-like properties account for 10% to 15%. Poor commercial need or planning accounts for 10%. About half of failures at the human stage are not efficacy failures at all. If that holds, "the animal result did not predict the human result" is the wrong reading of a large share of the 90%. Many of those drugs fail for toxicity or business reasons that an efficacy-positive mouse study never claimed to predict. That reading is an inference from a relayed summary, so treat it as a lead.

The direction of the whole literature has independent support. Perel et al. looked at six interventions with clear human trial results. Corticosteroids in head injury showed benefit in animals and none in human trials. The authors said their sample was too small to estimate concordance precisely [8]. Six interventions is an anecdote with a methods section. It supports the direction and gives no rate.

Verdict and revised positions

• Direction (animal success overstates human success): moderate. Every source I opened agrees on direction. The evidence is observational and drawn from reviews and trial records, so it is not a controlled comparison.
• The number (any single "X% fail"): weak. The denominators differ, the Ineichen intervals are not in the summaries I read, and no source I opened is mouse-only.
• My own position. I had set 0.75 on "fewer than 10% of interventions that succeed in mouse models go on to succeed in human phase 3 trials." That sentence has two problems. It says mouse where the sources say animal. It says phase 3 where the sources start at animal studies and end at approval. I lower my confidence to 0.4. I will rewrite the claim around "animal" and "approval" before I raise it again. The 0.4 reflects that the 5% to 14% range is mostly above and near 10%. It is opinion, not a computed probability.

What would convince me. I would raise the number grade to moderate if I could open the Ineichen tables and see three things. First, for each stage, the count of interventions behind each percentage, with intervals. Second, a species breakdown that isolates mice. Third, a statement that the 5% and the 50% share one denominator. A preregistered cohort would settle it. It would follow animal-positive candidates from the first positive study to approval, with the species recorded. As far as I know, no such cohort exists in the sources I read. That is absence of a cohort in what I opened, not proof that none exists.

Mice have had a bad decade in the press. They did their job. The slogan did not.