Why so many trials fail? And no, it's not what you think…

1. We know how many fail. We rarely ask why.

The biopharma industry has spent two decades measuring one thing: how many drugs that enter clinical development reach the market. The answer, consistent across every major study, is a single digit. The earliest influential academic framing came from Kola and Landis in 2004, who reviewed pharmaceutical R&D productivity and established that attrition was a systemic, measurable problem [1]. Since then, the estimates have refined but the conclusion has not changed. DiMasi et al. 2010 reported a 19.0% Phase-1-to-approval rate for lead indications [2]. Hay et al. 2014 found 10.4% across all indications [3]. Wong, Siah, and Lo estimated 13.8% across all indications and 21.6% for lead indications in their landmark 2019 analysis of 406,038 trial entries [4]. The BIO/QLS 2021 report found 7.9% across all indications from 2011–2020 [5]. Schuhmacher et al. 2025 reported 14.3% across 18 leading pharmaceutical companies [6]. Zhou et al. 2025 found approximately 5% for all clinical development programs in recent years [7]. Renny's 2025 global clinical attrition analysis showed 6.7% [8].

The spread from 5% to 21% reflects different methodologies, time periods, and denominator definitions — molecular entities versus development programs, lead versus all indications, industry-sponsored versus all sponsors. But the direction is consistent: 86–92% of programs that enter Phase 1 never reach approval.

The industry has invested heavily in understanding and improving that number. Billions have gone into better trial design, adaptive protocols, real-world evidence, biomarker-driven patient selection, AI for site optimization, and decentralized trials. These are execution interventions — they make trials faster, cheaper, and better-run. Yet the success rate has barely moved. The BIO/QLS report found 7.9% for 2011–2020, down from earlier estimates. More recent global clinical attrition data shows 6.7% [8]. Deloitte's 2025 benchmark found the average cost to develop a drug reached $2,671 million, with R&D return at just 2.9% when GLP-1 assets are excluded [9]. Bain reported that trial timelines have grown more than a third over the past decade [10].

The assumption behind these investments is straightforward: trials fail because the drug doesn't work or isn't safe. Biology is hard, attrition is inevitable, and the solution is better science and better trial execution. That assumption is wrong — or more precisely, it is incomplete in a way that matters. We analyzed 421,513 clinical trials to find out what actually kills programs. The data tells a different story: efficacy and safety account for 16.3% of failures. The rest is decisions — decisions in designing the program, directing the portfolio, and executing the trial. The science is working. The decisions aren't.

2. What the data actually shows — 421,513 trials analyzed

We analyzed 421,513 interventional trials registered in ClinicalTrials.gov via the AACT database [11]. This is the full population of interventional studies, not a sample. The status breakdown: 236,418 completed (56.1%), 86,505 ongoing, 57,939 of unknown status, and 40,651 terminated or withdrawn (9.6%).

Of those 40,651 terminated or withdrawn trials, 36,000 had a disclosed reason for stopping. We classified those reasons into eight named categories using keyword matching against the free-text “why_stopped” field, excluding COVID-19 pandemic terminations, never-started trials, misclassified completions, and non-informative boilerplate. This left 26,172 trials with a classifiable named failure cause — no “other” bucket in the named denominator, and all exclusions reported separately.

The distribution upends the conventional narrative:

The data separates into three decision pillars and one biology pillar. Biology — efficacy and safety combined — accounts for 16.3% of failures. Designing — poor enrollment and study design — accounts for 45.8%. Directing — business/strategic and regulatory — accounts for 33.5%. Executing — staff departure and drug supply — accounts for 4.5%.

Nearly half of all failures stem from designing decisions: choosing the wrong patient population, writing eligibility criteria too narrow for the disease prevalence, designing a protocol that cannot enroll, selecting endpoints that don't match the disease biology. Another third stem from directing decisions: portfolio prioritization, competitive displacement, regulatory pathway missteps, strategic reviews that kill programs for business reasons. Together, designing and directing account for 79.3% of failures. Biology accounts for 16.3%. Execution accounts for 4.5%.

Even if you restrict to industry-sponsored trials (which have a higher termination rate of 11.96% vs. 8.86% for academic trials), the pattern holds. The implication is uncomfortable: the industry has been investing as if biology is the bottleneck, when the data says the bottleneck is the decisions made around the biology.

We should be honest about what this methodology can and cannot show. The classification relies on keyword matching of free-text fields, which means some trials may be misclassified. “Completed” does not equal “succeeded” — a trial can complete its protocol without the drug advancing to the next phase or to approval. AACT is a trial registry, not a cohort tracking system, so we cannot follow individual compounds from Phase 1 through approval within this dataset. What the data does show, clearly and consistently, is that among trials that stop before completion, the reasons are overwhelmingly not about the drug failing to work.

(Failure pattern analysis of 421,513 clinical trials powered by ARiDA, INBISTRA's multi-agent AI platform for biopharma strategic intelligence.)

3. Where programs die and why it matters

The phase-by-phase breakdown reveals where designing, directing, and biology failures concentrate.

In our data, efficacy failures rise with phase: 7.1% of Phase 1 terminations, 11.9% of Phase 2, 15.7% of Phase 3, and 6.9% of Phase 4 [11]. This makes sense — later phases test efficacy more rigorously, so biology failures concentrate there. But even at Phase 3, efficacy accounts for only 15.7% of terminations. Business/strategic decisions are the dominant cause at Phase 1 (36.7%) and remain substantial at Phase 2 (24.2%) and Phase 3 (24.0%). Poor enrollment peaks at Phase 2 (40.7%), consistent with Phase 2 being where patient populations narrow and eligibility criteria tighten.

Published phase transition rates — which we cite from the literature rather than compute from AACT — tell a complementary story. The BIO/QLS 2011–2020 report found Phase II had the lowest transition rate at 28.9%, compared to 52.0% for Phase I and 57.8% for Phase III [5]. Hay et al. reported an even steeper Phase II cliff at 32.4% [3]. DiMasi et al. reported Phase II-to-III transition at 45.0% for lead indications [2]. Wong et al. reported a higher transition rate for the second-to-third-phase transition across all indications [4]. The variation across studies reflects different time periods and methodologies, but the direction is consistent: Phase II is where most programs die.

Phase 2 is where designing failures and biology failures collide. It is the first phase where efficacy is tested in a meaningful way, and it is the phase where eligibility criteria narrow most aggressively. A program that entered Phase 1 with a broad population and a plausible mechanism encounters Phase 2 with a specific patient subset, a defined endpoint, and a competitive landscape that may have shifted since the program was designed. If the trial was designed for a population that cannot sustain enrollment, or if the competitive landscape has moved, the program dies — not because the molecule failed, but because the designing and directing decisions made around it were wrong.

Poor enrollment as the single largest failure cause — 39.5% of named terminations [11], and 60–70% of trial sites fail to enroll their initial targets according to industry data [12] — deserves scrutiny. Enrollment failure is often framed as an execution problem: sites underperform, recruitment is slow, protocols are burdensome. But enrollment failure is fundamentally a designing problem. If a trial targets a patient population so narrow that the eligible cohort cannot sustain accrual, the trial was designed into a structurally unwinnable position before it began. The drug may work. The trial will still fail.

4. The trend that changes the conversation

The temporal trend in our data is the most striking finding and the one most at odds with the conventional narrative.

We excluded COVID-19-related terminations (1,983 trials) from the temporal analysis and examined the distribution of failure causes across three periods:

Directing failures — business and strategic terminations — have nearly doubled over the past decade, from 21.3% in 2010–2015 to 41.8% from 2022 onward [11]. In the most recent period, they are the single largest named failure cause — surpassing poor enrollment for the first time. Meanwhile, biology failures — efficacy terminations — have dropped from 9.2% to 5.5%. The science is working better. The directing is getting worse.

This trend is not unique to our dataset. Hwang et al. found that 22% of late-stage development failures between 1998 and 2008 were attributable to commercial reasons [13]. Iannantuono et al. reported in 2026 that 51.1% of Phase 1 solid-tumor early terminations were sponsor or strategic decisions — not efficacy or safety [14]. An analysis of pharmaceutical R&D portfolio decision-making found that leading biopharma companies routinely discontinue 21–22% of their pipeline programs annually, with approximately 50% of discontinued assets now terminated in Phase I [15]. A 2025 Nature Communications analysis found that success rates have declined substantially over two decades, partly because companies have adopted “quick-kill” strategies that terminate inferior candidates earlier [7].

The pattern is clear: companies are making more directing decisions, earlier. Some of this is healthy — killing bad programs before they consume late-stage capital is rational. But the volume and trajectory of strategic terminations suggest something else: companies are directing their portfolios without the intelligence base to do it well. They are killing programs because of competitive pressure, portfolio repositioning, and resource constraints — often after the decision should have been made, and sometimes before the data warranted it.

5. Decision quality is measurable and actionable

If the problem is decisions in designing, directing, and executing — not biology — the question becomes: can we measure decision quality, and does improving it change outcomes? The literature says yes, and the effect sizes are large.

Designing: biomarker selection. Wong et al. found that oncology trials using biomarkers for patient selection had substantially higher approval rates than those without biomarker selection [4]. The BIO/QLS report found that biomarker-selected programs across all indications had a 15.9% LOA, more than double the 7.9% for programs without biomarkers [5]. This is not a marginal improvement. It is the difference between a program that is likely to fail and one that has a meaningful chance of success. And it is a designing decision — choosing the right patients — not a biology breakthrough.

Designing: indication order. DiMasi et al. showed that for cancer drugs, the first indication pursued had a 9.0% approval success rate, the second 8.2%, and the third 6.9% [16]. More striking was the conditional dependency: if the first indication succeeded, the second had a 54.9% success rate. If the first indication failed, the second had a 2.5% success rate [16]. A 22-fold swing based on indication sequencing. The first indication you design a program around materially shapes the probability that the drug succeeds at all.

Designing: target and modality. Yamaguchi et al. evaluated 3,999 compounds and found an overall success rate of 12.8%, but with large variation by target, action, and modality [17]. Stimulant action or enzyme target combined with biologic modality reached 34.1% and 31.3% success rates — nearly three times the aggregate [17]. The choice of what to develop — which target, which mechanism, which modality — is itself a designing decision with measurable downstream consequences.

Directing: AstraZeneca's 5R framework. The most compelling case study is AstraZeneca. The company restructured its R&D around five principles — Right target, Right tissue, Right safety, Right patient, Right commercial potential — and established what it called a “truth-seeking culture” where scientists were encouraged to ask “killer questions” early [18]. In five years, the proportion of pipeline molecules advancing from preclinical investigation to completion of Phase III improved from 4% to 19% — a fivefold improvement [18]. This was not a biology breakthrough. It was a decision discipline across designing (right target, right patient) and directing (right commercial potential, killer questions). AstraZeneca improved its success rate by being more rigorous about the decisions made around the science, not by being better at the science itself.

These are decision interventions, not biology interventions. They operate on how programs are designed, how portfolios are directed, and how trials are executed. They have 3–5x more leverage on success rates than pure execution improvements like adaptive trial designs or AI-assisted site selection. And yet the industry's investment thesis has been overwhelmingly weighted toward execution tooling — faster trials, better sites, decentralized protocols — while the designing and directing decisions that determine whether a program should exist at all receive comparatively little systematic support.

6. What it looks like in practice

The abstract statistics become tangible when you look at specific trials. The ClinicalTrials.gov registry, validated against our AACT analysis with a live API pull of 1,000 recent terminations, provides concrete examples of each failure mode.

Directing failure — the drug worked, the portfolio won. ViiV Healthcare terminated a Phase 2b study of GSK3640254, an HIV maturation inhibitor, after enrolling 161 treatment-naive patients. The reason: “Company decision to stop compound development. The decision is not based on any safety or efficacy concerns. It reflects the company strategy for portfolio progression” [19]. AKARI Therapeutics went further — terminating a Phase 3 study of nomacopan in pediatric transplant-associated thrombotic microangiopathy after enrolling only 10 patients. The stated reason: “The early termination of this study is a business decision following a portfolio reprioritization plan. The decision is not related to any Efficacy, Safety or Clinical concerns” [20]. A Phase 3 trial killed for directing reasons after 10 patients. Incyte stopped its oral PD-L1 inhibitor INCB086550 after 16 patients, explicitly to “prioritize another oral PD-L1 inhibitor with a more favorable profile” [21] — an internal portfolio horse-race, not a scientific failure.

Directing failure — the landscape moved. A Columbia University biomarker study of pembrolizumab in frontline NSCLC was terminated because “Frontline pembrolizumab approved in NSCLC as monotherapy and in combination with chemotherapy representing a new standard of care” [22]. The trial enrolled 19 patients before a competitor's regulatory approval made its design obsolete. This is not a failure of the drug or the investigator. It is a directing failure — the competitive landscape was not anticipated when the program was designed.

Designing failure — the population didn't exist. A Weill Cornell Phase 1 trial of intra-arterial chemotherapy for atypical choroid plexus papilloma enrolled a single patient before terminating for low accrual [23]. The condition is so rare that the eligible population cannot sustain a trial regardless of the therapy's merit. A Northwestern University trial of ivosidenib plus FLAG chemotherapy for IDH1-mutant relapsed/refractory AML enrolled 2 patients before stopping [23]. Molecular selection narrowed an already small relapsed/refractory population below feasibility. These are not execution failures. They are designing failures — the wrong trial was designed for the wrong population size.

Biology failure — but in a questionable indication. GSK's Phase 2 TRANSFORM study of GSK3915393 in idiopathic pulmonary fibrosis enrolled 158 patients but stopped at interim analysis for futility, with “no clinical efficacy of the investigational drug” [24]. IPF has a long history of late-stage failures across multiple mechanisms. The question is not whether this particular drug failed — it did — but whether the indication choice was well-supported by the preclinical and mechanistic evidence. When biology failures cluster in specific indications with histories of repeated failures, the problem is not that biology is hard in general. It is that the designing decision — which biology to test in which disease — was not disciplined enough.

7. The implication — and what to do about it

The industry invests billions in execution improvement: adaptive trial designs, decentralized trials, real-world evidence, AI for site selection and patient recruitment, biomarker-driven adaptive protocols. These investments are real and they matter. Deloitte's 2025 benchmark found the average cost to develop a drug from discovery to launch reached $2,671 million, with R&D rate of return at just 2.9% when GLP-1 assets are excluded [9]. Bain reported that trial timelines have grown more than a third over the past decade and that only 28% of AI pilots reach production [10]. The productivity problem is genuine.

But the data suggests the industry is investing disproportionately in the wrong part of the problem. Execution improvements — 4.5% of failures — make trials faster, cheaper, and better-run. They do not change which programs enter the pipeline, which indications are pursued first, which patient populations are targeted, or when to kill. Designing and directing decisions — 79.3% of failures — determine whether a program should exist at all, whether the trial is designed for a population that can sustain it, and whether the competitive and regulatory landscape supports continued investment. These decisions have 3–5x more leverage on the overall success rate, based on the effect sizes documented in the literature. AstraZeneca proved it: a fivefold improvement in preclinical-to-Phase III completion through decision discipline across designing and directing, without any fundamental breakthrough in biology [18].

The rise of directing failures tells us the industry already knows this, at some level. Companies are already making portfolio decisions — they are killing programs at twice the rate of a decade ago. But they are making those decisions reactively, after spending years and millions on programs that were structurally challenged from the start. The question is not whether to make designing and directing decisions. It is whether to make them with intelligence before capital is committed, or after it is burned.

Better designing and directing does not just improve the success rate. It redirects capital away from programs that were never going to work. Every dollar spent on a program terminated for strategic reasons after Phase 1 is a dollar that could have funded a program with a higher probability of success. Every trial that enrolls one patient in a population of three is a trial that should never have been designed. Every Phase 3 killed for “business reasons” after 10 patients is a directing failure that happened years too late.

The 86–92% of clinical programs that never reach approval — the number the industry has spent two decades trying to improve — is not a property of drug development. It is a property of the decisions made in designing programs, directing portfolios, and executing trials. The data shows that biology is the smallest piece of that problem. The science is getting better — efficacy terminations are dropping. What is rising is the share of programs killed for reasons that have nothing to do with whether the drug works. That is not a biology problem. It is a decision problem. And decision problems are solvable.

Sources

1. [1] Kola I, Landis J. “Can the pharmaceutical industry reduce attrition rates?” Nature Reviews Drug Discovery 2004;3(8):711-716. https://pubmed.ncbi.nlm.nih.gov/15286737/
2. [2] DiMasi JA, Feldman L, Seckler A, Wilson A. “Trends in risks associated with new drug development: success rates for investigational drugs.” Clinical Pharmacology & Therapeutics 2010;87(3):272-277. https://pubmed.ncbi.nlm.nih.gov/20130567/
3. [3] Hay M, Thomas DW, Craighead JL, Economides C, Rosenthal J. “Clinical development success rates for investigational drugs.” Nature Biotechnology 2014;32(1):40-51. https://pubmed.ncbi.nlm.nih.gov/24406927/
4. [4] Wong CH, Siah KW, Lo AW. “Estimation of clinical trial success rates and related parameters.” Biostatistics 2019;20(2):273-286. https://pubmed.ncbi.nlm.nih.gov/29394327/
5. [5] DrugPatentWatch. “A Strategic Investor's Guide to Pharmaceutical Portfolio Risk Assessment.” December 11, 2025. https://www.drugpatentwatch.com/blog/a-strategic-investors-guide-to-pharmaceutical-portfolio-risk-assessment
6. [6] Schuhmacher A, Hinder M, Brief E, Gassmann O, Hartl D. “Benchmarking R&D success rates of leading pharmaceutical companies: an empirical analysis of FDA approvals (2006-2022).” Drug Discovery Today 2025;30(2):104-112. https://pubmed.ncbi.nlm.nih.gov/39805539/
7. [7] Zhou Y et al. “Dynamic clinical trial success rates for drugs in the 21st century.” Nature Communications 2025;16:64552. https://pubmed.ncbi.nlm.nih.gov/41162353/
8. [8] Renny J. “Global Clinical Attrition by Modality (2005–2025).” LinkedIn, September 11, 2025. https://www.linkedin.com/pulse/global-clinical-attrition-modality-20052025-joe-renny-xqise
9. [9] Deloitte. “Navigating the GLP boom: Measuring the return from pharmaceutical innovation 2025.” https://www.deloitte.com/us/en/industries/life-sciences-health-care/perspectives/navigating-the-glp-boom.html
10. [10] Bain & Company. “How to break through pharma's innovation bottleneck with AI.” 2025. https://www.bain.com/insights/how-to-break-through-pharmas-innovation-bottleneck-with-ai
11. [11] AACT Internal Database: Failure pattern analysis of 421,513 interventional clinical trials from ClinicalTrials.gov (AACT), 2026. https://aact.ctti-clinicaltrials.org/
12. [12] Clinical Trials Arena / HHS ASPE. “The unsustainability of clinical trials and need for transformative change.” 2025. https://www.clinicaltrialsarena.com/sponsored/the-unsustainability-of-clinical-trials-and-need-for-transformative-change
13. [13] Hwang TJ, Carpenter D, Lauffenburger JC, Wang B, Franklin JM, Kesselheim AS. “Failure of investigational drugs in late-stage clinical development and publication of trial results.” JAMA Internal Medicine 2016;176(12):1826-1833. https://pubmed.ncbi.nlm.nih.gov/27723879/
14. [14] Iannantuono GM et al. “Temporal patterns and factors associated with early termination in Phase 1 clinical trials for solid tumors.” Journal of the National Cancer Institute 2026. https://pubmed.ncbi.nlm.nih.gov/41902835/
15. [15] DrugPatentWatch. “Mastering Strategic Decision-Making in the Pharmaceutical R&D Portfolio.” March 3, 2026. https://www.drugpatentwatch.com/blog/decision-making-product-portfolios-pharmaceutical-research-development-managing-streams-innovation-highly-regulated-markets
16. [16] DiMasi JA, Reichert JM, Feldman L, Malins A. “Clinical approval success rates for investigational cancer drugs.” Clinical Pharmacology & Therapeutics 2013;94(3):329-335. https://pubmed.ncbi.nlm.nih.gov/23739536/
17. [17] Yamaguchi S, Kaneko M, Narukawa M. “Approval success rates of drug candidates based on target, action, modality, application, and their combinations.” Clinical and Translational Science 2021;14(3):1117-1125. https://pubmed.ncbi.nlm.nih.gov/33831276/
18. [18] AstraZeneca. “Transforming AstraZeneca's R&D productivity.” 2016. https://www.astrazeneca.com/what-science-can-do/topics/disease-understanding/transforming-astrazenecas-rd-productivity.html
19. [19] ClinicalTrials.gov. NCT04493216: ViiV Healthcare, GSK3640254 Phase 2b in HIV-1. https://clinicaltrials.gov/study/NCT04493216
20. [20] ClinicalTrials.gov. NCT04784455: AKARI Therapeutics, nomacopan Phase 3 in transplant-associated thrombotic microangiopathy. https://clinicaltrials.gov/study/NCT04784455
21. [21] ClinicalTrials.gov. NCT04629339: Incyte Corporation, INCB086550 Phase 2 in solid tumors. https://clinicaltrials.gov/study/NCT04629339
22. [22] ClinicalTrials.gov. NCT02710396: Columbia University, pembrolizumab biomarker study in NSCLC. https://clinicaltrials.gov/study/NCT02710396
23. [23] ClinicalTrials.gov. NCT04994977: Weill Cornell, intra-arterial chemotherapy for choroid plexus tumors (1 patient enrolled). NCT04250051: Northwestern University, ivosidenib in IDH1-mutant AML (2 patients enrolled). https://clinicaltrials.gov/study/NCT04994977
24. [24] ClinicalTrials.gov. NCT06317285: GlaxoSmithKline, GSK3915393 Phase 2 in idiopathic pulmonary fibrosis. https://clinicaltrials.gov/study/NCT06317285