Why Drug Discovery Still Fails Late

Cheaper prediction and screening shrink the front of the funnel. Biology, clinical evidence, manufacturing, and the cost of failures still set the approved-drug curve.

Last updated September 2026

The argument

Drug discovery cost is portfolio arithmetic: total spending across successful and failed programs divided by approved therapies. AI can make hypotheses and molecules cheaper without shortening the slow, safety-critical evidence chain unless it improves the probability of clinical success.

  • FDA materials describe development as roughly a decade or more and estimate that fewer than 10–12% of drugs entering trials are approved.
  • Typical Phase 3 trials involve about 300–3,000 participants and can last one to four years.
  • Candidate generation is only one stage among synthesis, preclinical work, trials, review, manufacturing, and delivery.
  • A better early filter creates value by stopping weak programs before expensive trials, not by maximizing candidates.

Measured results, derived quantities, projections, targets, and editorial inference are identified by context. Announced capacity is never treated as operating performance.

Part I: What changed

The expensive experiment is a human outcome

Models can rank molecules and assays can measure mechanisms, but efficacy and safety emerge in diverse patients over time. The development system buys evidence under uncertainty, and late failures carry the cost of every earlier stage.

Three numbers that locate the frontier

<10%FDA primer estimate of trial entrants eventually approved.
~10 yearsFDA primer's order-of-magnitude first-human-to-approval timeline.
300–3,000Typical Phase 3 participant range in FDA's process guide.

Development cost estimates vary with indication, capital cost, failure accounting, trial design, and whether discovery and post-approval work are included.

Part II: The measurable curve

Measure cost per approved therapeutic hypothesis

The curve bends when tools improve calibrated decisions: which target to pursue, which molecule to nominate, which patient to enroll, and which program to stop. Faster generation without better selection can increase downstream burden.

Biomarkers can enrich for responders and reduce sample size only when analytically and clinically valid.

Platform reuse can amortize manufacturing and regulatory knowledge across products.

Part III: The physical stack

The headline metric sits on a system

Each layer can become the bottleneck even when the layer before it improves.

01

Disease model and target

Human genetics, biology, and causal evidence define an intervention hypothesis.

Measure
Target validation
Failure mode
Model translation
02

Molecule and preclinical

Design, synthesis, assays, exposure, toxicology, and formulation create a candidate.

Measure
Nomination rate · quality
Failure mode
False confidence
03

Clinical evidence

Trials estimate safety, dose, efficacy, and benefit-risk in people.

Measure
Phase success · time
Failure mode
Recruitment and endpoints
04

CMC and delivery

A reproducible product, approval package, supply chain, and care pathway reach patients.

Measure
Release yield · access
Failure mode
Manufacturing and adoption
Part IV: The floor

Uncertainty must be reduced with observations

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.

portfolio spend÷approved useful medicines=cost per success
Part V: The bottleneck shift

Cheaper ideation moves scarcity into validation

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.

Predictive assays

Benchmark models against prospective experimental outcomes.

Adaptive trials

Learn and reallocate while protecting statistical validity.

Platform reuse

Carry validated processes, analytics, and delivery across products.

An 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.

Sources, method, and boundaries

FDA figures are broad educational ranges rather than a single audited industry cohort. The portfolio framing includes failures and distinguishes discovery throughput from approval productivity.

FDA: Primer on Drug DevelopmentFDA: Step 3—Clinical ResearchFDA: Drug development and review definitions