Attrition compounds across phases, not just once at the end
An aggregate approval rate hides where a program is most likely to die, and each phase fails for a different reason.
What BIO's phase transition rates imply for 100 Phase I entrants
100% of programs entering Phase 1 trials, the reference cohort for the funnel.
Four stages, four different kinds of evidence
Each stage buys a different type of uncertainty reduction, and a shortcut in one does not substitute for evidence in another.
Disease model and target
Human genetics, biology, and causal evidence define an intervention hypothesis.
- Measure
- Target validation
- Failure boundary
- A model that fails to translate to human biology invalidates everything built on it.
Where the frontier moves
Human genetics and patient-derived data replacing animal-model-only validation.
Molecule and preclinical
Design, synthesis, assays, exposure, toxicology, and formulation create a candidate.
- Measure
- Nomination rate · quality
- Failure boundary
- False confidence from a model that generates plausible molecules without predicting clinical success.
Where the frontier moves
Predictive assays benchmarked prospectively, not just retrospectively.
Clinical evidence
Trials estimate safety, dose, efficacy, and benefit-risk in people.
- Measure
- Phase success · time
- Failure boundary
- Recruitment and endpoint selection can sink an otherwise sound program.
Where the frontier moves
Adaptive trials and biomarkers that reduce sample size without weakening statistical validity.
CMC and delivery
A reproducible product, approval package, supply chain, and care pathway reach patients.
- Measure
- Release yield · access
- Failure boundary
- Manufacturing and adoption failures can strand an approved therapy from patients.
Where the frontier moves
Platform manufacturing that carries validated processes across products.
Only an experiment can tell you what happens in a person
No amount of computation removes the need to observe a therapy in relevant biology and, ultimately, people. Better priors can reduce experiments; they cannot make unmeasured risk disappear.
Cheap ideas make trials the scarce resource
As virtual candidates multiply, high-quality assays, representative models, clinical sites, patients, endpoints, and manufacturing comparability become the constrained resources.
Human-grounded targets
Use genetics and patient data to improve causal confidence before committing to a molecule.
Predictive assays
Benchmark models against prospective experimental outcomes, not retrospective curve-fitting.
Adaptive trials
Learn and reallocate resources mid-trial while protecting statistical validity.
Platform reuse
Carry validated manufacturing processes, analytics, and delivery systems across products.
Who is building what
High-content phenotypical screening, microfluidic organ-on-chip models, free-energy molecular dynamics, and synthetic control arms combat late-stage clinical attrition. Search the record, or filter by discovery platform.
Recursion PharmaceuticalsPhenomics & BioHive-2High-throughput automated fluorescence microscopy imaging millions of gene-knockout and chemical cellular perturbations, analyzed with computer vision
- Reported evidence
- Built one of the world's largest biological phenomics datasets; advanced multiple oncology and rare disease candidates into Phase II trials.
- Announced next step
- Industrializing drug discovery by systematically mapping disease phenotypes rather than isolated single-target hypotheses.
- Unresolved risk
- Translating statistical image morphology similarities into clinically validated disease mechanisms that translate into human efficacy.
Emulate BioHuman Emulation System (Organ-Chip)Microfluidic cell culture chips with living human cells subject to dynamic mechanical strain and fluid flow, recreating organ microenvironments
- Reported evidence
- Landmark multi-center study published in Science Translational Medicine demonstrated 87% sensitivity and 100% specificity predicting human drug-induced liver injury.
- Announced next step
- Universal adoption of Liver-Chip and Intestine-Chip systems across pharmaceutical discovery pipelines to replace legacy animal toxicology.
- Unresolved risk
- Throughput scalability compared to 384-well microplates, chip consumable cost, and vascular endothelial cell sourcing consistency.
Insilico MedicinePharma.AI & RentosertibEnd-to-end generative AI platform discovering novel biological targets (PandaOmics) and designing novel small-molecule chemical structures (Chemistry42)
- Reported evidence
- Discovered novel antifibrotic target and advanced Rentosertib (ISM001-055) from initial target hypothesis through to Phase II clinical trials in idiopathic pulmonary fibrosis.
- Announced next step
- Validating the first fully AI-discovered target and AI-designed molecule through Phase II/III randomized clinical outcomes.
- Unresolved risk
- Clinical efficacy readouts in complex multifactorial human diseases where earlier computational target predictions cannot guarantee disease reversal.
Relay TherapeuticsDynamo PlatformAtomic-scale molecular dynamics simulations tracking protein conformational motion, revealing transient allosteric binding pockets invisible in static crystallography
- Reported evidence
- Discovered and advanced RLY-2608 (PI3Kα allosteric mutant-selective inhibitor) into clinical trials with high mutant selectivity over wild-type.
- Announced next step
- Targeting traditionally 'undruggable' oncogenic proteins by stabilizing specific inactive conformational states.
- Unresolved risk
- Computational supercomputing costs for microsecond-scale molecular dynamics simulations and accurately modeling membrane-bound protein dynamics.
SchrödingerFree Energy Perturbation (FEP+)Physics-based computational chemistry software calculating accurate relative and absolute binding free energies (ΔΔG) across candidate molecular libraries
- Reported evidence
- Widely utilized across global pharmaceutical discovery teams; demonstrated predictive accuracy within ~1 kcal/mol of experimental binding affinities.
- Announced next step
- Exploring trillions of virtual chemical compounds with high physical precision before synthesizing any physical molecules in the lab.
- Unresolved risk
- Parameterization challenges for uncommon metalloproteins, macrocycles, and water-mediated hydrogen bonding networks.
Unlearn.AIDigital Twin GeneratorGenerative machine learning models trained on historical clinical trial registries generating prognostic digital twins of clinical trial participants
- Reported evidence
- Qualified by the European Medicines Agency (EMA) for use in Phase II/III trials; enables smaller control arms without compromising statistical power.
- Announced next step
- Slashing required human patient enrollments in placebo control arms by up to 30–50% in Alzheimer's and ALS trials.
- Unresolved risk
- Model bias if historical clinical trial cohorts lack diverse ancestral, socioeconomic, and comorbidity representation.
CN BioPhysioMimix Multi-Organ SystemMulti-organ microfluidic plates interconnecting liver, gut, and kidney compartments to model systemic drug absorption, metabolism, and secondary metabolite toxicity
- Reported evidence
- Adopted by FDA regulatory research labs; demonstrated accurate modeling of oral prodrug gut absorption and hepatic first-pass clearance.
- Announced next step
- Routine benchtop testing of human multi-organ pharmacokinetics and safety pharmacology.
- Unresolved risk
- Maintaining balanced common culture media supporting multiple distinct cell types without de-differentiating hepatocytes.
FDA Modernization Act 2.0 / NCATSNon-Animal Testing AlternativesFederal legislation removing the 1938 mandate requiring animal testing for new drug applications, authorizing qualified organ-chips and computational models
- Reported evidence
- Enacted into US federal law; FDA established qualification programs for New Approach Methodologies (NAMs) in investigational new drug (IND) submissions.
- Announced next step
- Establishing formalized regulatory clearance pathways for human organ-chips and in silico toxicology models to replace animal studies.
- Unresolved risk
- Institutional regulatory conservatism within clinical review divisions requiring decades of parallel data before discarding legacy animal models.
Preclinical target validation and in vitro IC50 efficacy do not predict human Phase II/III clinical trial survival; human organ complexity and idiosyncratic drug toxicity remain the primary failure boundaries.
The optimistic view, with conditions
Discovery becomes an evidence-optimization system
The winning platforms will not boast the most generated molecules; they will produce better-calibrated stop/go decisions and more approvals per dollar and patient-year.
Filter earlier, not just faster
A model that predicts Phase 2 failure before a candidate enters Phase 1 saves the most expensive years.
Target the steepest drop
In BIO's 2011–2020 analysis, Phase II advanced 28.9% of programs that reached it, the lowest transition rate in the funnel.
Measure approvals per dollar
Not candidates generated, not molecules synthesized: approved therapies per dollar spent across the whole portfolio.
What actually lowers the cost of an approved medicine
- Better stop/go decisionsCalibrated models that kill weak programs before Phase 2, not after.
- Human-grounded targetsGenetics and patient data that improve causal confidence ahead of the clinic.
- Validated predictive assaysModels benchmarked against real prospective trial outcomes, not retrospective fit.
- Adaptive trial designMid-trial learning and reallocation that preserves statistical validity.
- Platform manufacturingReusable processes and delivery systems that don't restart from zero each product.
“Billions” depends on what is counted
A 2016 Tufts estimate put capitalized cost at about $2.6 billion per approved medicine. Wouters and colleagues’ 2020 analysis estimated a $985 million median and $1.34 billion mean after accounting for failed trials and capital cost. A 2025 JAMA Network Open study found a $708 million median and $1.31 billion mean on its own sample and assumptions. These are alternative samples and accounting methods, not a clean chronological fall in the price of creating the same drug.
Target choice is one measurable way to change attrition. A 2024 study of clinical mechanisms estimated a 2.6-fold higher probability of success for mechanisms with human genetic support than those without, under its target–indication definitions. That is an association across a selected pipeline, not a promise that every genetics-backed drug succeeds. The relevant cost metric is spending per clinically useful, approved outcome after failures, trial duration and manufacturing are included, which is why an early hit or a cheaper assay does not by itself collapse total development cost.
The funnel changes by disease, and AI has a clinical test
Scannell and colleagues’ Eroom’s law analysis found that approvals per inflation-adjusted billion dollars of R&D roughly halved every nine years between 1950 and 2010. It is a historical productivity estimate, not a physical law or a forecast. The BIO 2011–2020 cohort estimates 7.9% Phase I-to-approval success overall, versus 5.3% for oncology and 9.3% for non-oncology indications. Figure 2 uses this same 2011–2020 analysis. Its rates describe programs across indications and should not be read as the odds for a specific drug or as a current forecast.
Rentosertib’s published randomized Phase IIa trial is a useful test of AI-enabled discovery: its 12-week, small IPF study reported a dose-related lung-function signal, but does not establish registrational efficacy or a lower portfolio cost. Clinical development still needs enough patient-years to resolve an effect and adverse events. For a simple two-arm trial, required sample size scales approximately with the outcome variance divided by the square of the effect size; halving the effect size needs roughly four times as many participants at the same power and significance level. The resulting research cost is different from a medicine’s sale price, which also reflects manufacturing, exclusivity, competition and payment rules. The FDA’s 2025 roadmap advances validated alternatives to some animal tests; it does not remove clinical evidence requirements.
Sources, method, and boundaries
FDA figures are broad educational ranges rather than a single audited industry cohort. Figure 2 uses BIO, Informa Pharma Intelligence, and QLS Advisors' 2011–2020 transition rates. Its 100-program funnel is a derived illustration, not one observed cohort. The earlier BIO 2006–2015 analysis reported 9.6% overall versus 7.9% in 2011–2020; the studies differ in observation window and accounting of dormant programs. Rates vary by therapeutic area and modality. The portfolio framing includes failures and distinguishes discovery throughput from approval productivity.
- Phase transition rate
- The share of programs entering a given trial phase that advance to the next phase.
- CMC
- Chemistry, manufacturing, and controls, the evidence that a product can be reproducibly made and tested.
- Biomarker enrichment
- Selecting trial participants most likely to respond, reducing required sample size when analytically valid.
Read More
Essays, books, talks, and research that shaped this field’s arguments. Influence is not endorsement; company communications and advocacy are labeled. Some publisher links require a subscription.
- Book chapter
Jack Scannell — Eroom’s Law and the decline in the productivity of biopharmaceutical R&D
The later essay framing of pharmaceutical productivity and what AI would have to change.
- Essay collection
Derek Lowe — In the Pipeline
The working chemist’s commentary on attrition, clinical evidence, and AI discovery claims.
- Investor writing
a16z — Bio + Health
Influential founder and investor conversations; compare their platform cases with clinical attrition data.
- Paper / report
Scannell et al., “Diagnosing the decline in pharmaceutical R&D efficiency” (Nature Reviews Drug Discovery, 2012)
The “Eroom’s Law” essay; one of the most influential framings of declining drug R&D productivity.
- Paper / report
DiMasi, Grabowski & Hansen, “Innovation in the pharmaceutical industry: New estimates of R&D costs” (Journal of Health Economics, 2016)
A central and contested empirical reference in the debate over the cost of developing an approved drug.
Across the fields: learning curves, deployment, and rebound
- Theodore Wright — Factors Affecting the Cost of Airplanes (1936)
The original experience-curve formulation: costs change with accumulated production.
- Kenneth Arrow — The Economic Implications of Learning by Doing (1962)
The economics of productivity improvement through production experience.
- William Stanley Jevons — The Coal Question (1865)
The classic rebound argument: lower effective costs can expand total demand.
- Arnulf Grübler — The costs of the French nuclear scale-up: A case of negative learning by doing (2010)
A counterexample to assuming that greater deployment always lowers costs.



















