Drug development

How Are Drug Approval Probabilities Calculated?

Drug approval probabilities are usually built from historical phase transition rates, multiplied together. How programs are counted changes the answer, and estimates for individual drugs require models that account for the evidence.

Key points

  • The most common method multiplies historical success rates for each remaining phase. For 2011–2020 industry data, that gives a 7.9% chance of approval for a drug entering Phase 1.
  • Published estimates of the same quantity range from about 8% to about 14%, depending on the data source and how programs are counted.
  • An average success rate describes a group of programs. It is not a prediction for any single drug.
  • Drug-specific estimates require a model that adjusts the historical baseline for evidence such as disease area, trial results, and patient selection, using only information available at the time.
  • A probability estimate is only trustworthy if it has been tested for calibration: events given a 20% chance should happen about 20% of the time.

When you read that a drug has "a 15% chance of approval," that number came from somewhere. Most published figures come from the same basic approach: count how often similar drugs succeeded in the past. The details of that counting matter more than most people realize, and turning a group average into an estimate for one drug requires another step entirely.

Step 1: Phase transition rates

The foundation is the phase transition success rate: of the programs that completed a given phase, the share that advanced to the next one.

Researchers assemble a database of development programs, usually one record per drug–indication pair, and observe each program's fate:

  • Advanced: the program moved to the next phase.
  • Suspended: the program was stopped, terminated, or abandoned.
  • Still in progress: the outcome is not yet known, so the program is usually excluded from that phase's rate.

The success rate for a phase is the number that advanced divided by the number that either advanced or were suspended.

Step 2: Multiply to get the likelihood of approval

To estimate the chance that a drug entering a given phase is eventually approved, the rates for each remaining phase are multiplied together. Using the 2011–2020 BIO, Informa Pharma Intelligence, and QLS Advisors data:

Phase transition Rate
Phase 1 → Phase 2 52.0%
Phase 2 → Phase 3 28.9%
Phase 3 → application 57.8%
Application → approval 90.6%
Likelihood of approval from Phase 1 52.0% × 28.9% × 57.8% × 90.6% ≈ 7.9%

The same logic gives about 15% from Phase 2 and about 52% from Phase 3.

Why published estimates disagree

Different research groups, using careful methods, report noticeably different overall success rates:

Study Data period Likelihood of approval from Phase 1
BIO, Informa, QLS (2021) 2011–2020 7.9%
BIO, Biomedtracker, Amplion (2016) 2006–2015 9.6%
Hay et al., Nature Biotechnology (2014) 2003–2011 10.4%
Wong, Siah, and Lo, Biostatistics (2019) 2000–2015 13.8%

The 2016 and 2014 figures are as reported in the 2021 BIO analysis and in Wong et al. (2019).

The main reasons for the differences:

  • The phase-by-phase method versus the path-by-path method. The BIO reports compute each phase's rate separately and multiply them. Wong, Siah, and Lo follow each program's complete path from start to finish, which treats programs that skip phases or run phases in parallel differently.
  • Data source. Industry databases built from company announcements capture different programs than databases built from trial registries such as ClinicalTrials.gov.
  • Inferring silent failures. Many programs are never formally announced as discontinued. Studies differ in how long a program must be inactive before it is counted as failed.
  • Time period. Success rates change over time as the mix of diseases and drug types in development shifts.

None of these estimates is simply "right." Each answers a slightly different question.

Step 3: From a group average to a single drug

A historical rate describes a population of programs. Using it as the probability for one specific drug assumes that drug is a typical member of the group, which is rarely true. Two programs in the same phase can differ in:

  • disease area and indication,
  • drug modality (small molecule, antibody, gene therapy, and so on),
  • whether patients are selected with a biomarker,
  • whether the target has human genetic support,
  • how strong and well-controlled the previous trial results were, and
  • whether the drug has already been approved for another use.

The first step is to choose a more relevant comparison group, for example Phase 2 oncology programs rather than all Phase 2 programs. Splitting the data too finely creates small groups with unreliable rates. Statistical models solve this by estimating the effect of each factor while sharing information across related groups.

Models used for drug-specific estimates

Common approaches include:

  • Logistic regression, which estimates how each characteristic raises or lowers the odds of success.
  • Hierarchical (multilevel) models, which let estimates for small groups, such as a rare indication, borrow strength from larger related groups without being treated as average.
  • Bayesian models, which start from the historical rate as a prior and update it with drug-specific evidence. Their output is a full probability distribution, so every estimate comes with a measure of uncertainty.
  • Machine learning models, which can detect complex patterns in large datasets but are harder to interpret.

The rules that make an estimate trustworthy

No hindsight

A model must be tested using only information available on the date each prediction would have been made. If it accidentally learns from later events, such as a trial result announced after the prediction date, it will look far more accurate than it really is.

Calibration

A well-calibrated model's probabilities match reality. Among all programs it gives a 30% chance, about 30% should succeed. Calibration is checked by grouping past predictions and comparing predicted with observed outcomes.

Discrimination

A useful model should also rank programs well, placing those that eventually succeed above those that fail more often than a simple historical average does.

Uncertainty

A single bare percentage hides how much is unknown. A well-reported estimate gives a range, for example "18% (80% interval: 11% to 27%)."

What a probability does not mean

  • A 15% probability is not a forecast that a drug will fail. Roughly 15 in 100 comparable programs would be expected to succeed.
  • An approval probability is not a prediction of commercial success or of a company's share price.
  • A probability is only as good as the data and model behind it. It should always come with an explanation of where it came from.

Spectral Genomics' own approach, including how we plan to validate it before publishing drug-specific estimates, is described in our approval probability methodology.

Sources

  1. Biotechnology Innovation Organization, Informa Pharma Intelligence, QLS Advisors. Clinical Development Success Rates and Contributing Factors 2011–2020 (2021).Industry report
  2. Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20(2):273–286.Peer-reviewed
  3. Hay M, Thomas DW, Craighead JL, Economides C, Rosenthal J. Clinical development success rates for investigational drugs. Nature Biotechnology. 2014;32:40–51.Peer-reviewed
  4. Minikel EV, Painter JL, Dong CC, Nelson MR. Refining the impact of genetic evidence on clinical success. Nature. 2024;629:624–629.Peer-reviewed

This article is for education. It is not medical advice and not investment advice. Figures are drawn from the sources above and describe historical averages, not the prospects of any particular drug. See our editorial standards and disclaimer.