Clinical and regulatory success probabilities turn a biotech pipeline from a list of possible products into a set of investable assumptions. They estimate the chance that a drug programme clears its remaining development hurdles, reaches approval and supports the cash flows in a valuation. A small change can move the result sharply. Yet the percentage entered into a spreadsheet often receives less scrutiny than peak sales. A sound estimate starts with a relevant historical benchmark, adjusts for the evidence available today and separates clinical success from the risks that remain after a positive trial. Within a risk-adjusted net present value model, these probabilities should govern both potential revenue and the development spending required to reach it. They are assumptions to test, rather than facts to inherit from management.
Define the Event Before Estimating Its Probability
“Probability of success” can refer to several different events. A trial might succeed by meeting its primary endpoint. A development stage might succeed when the sponsor advances the programme. Regulatory success might mean approval during the first review, eventual approval after further work, or approval for a commercially useful patient population. Those events are related, but their probabilities are different. Write the intended event beside every estimate. “Approval in the United States for adults with the studied disease, following the planned pivotal programme” is a usable definition. “Drug succeeds” is not. The distinction matters when comparing Phase 1, Phase 2 and Phase 3 programmes, because the remaining evidence requirements change as development progresses.
A stage transition probability is conditional: it measures the chance of clearing the next hurdle from the current starting point. Cumulative probability measures the chance of clearing every remaining hurdle. Suppose a programme has a 60% chance of producing evidence sufficient for submission and a 90% chance of approval conditional on that outcome. Its approval probability is 54%, calculated as 60% × 90%. This multiplication does not require the events to be independent; the second probability is already conditional on the first event occurring. Trouble starts when analysts multiply unrelated averages without checking their definitions, or apply an approval probability to a valuation that has already been adjusted for the same development risk.
Use Historical Success Rates as a Starting Point
A historical benchmark supplies a starting assumption, often called a prior. The BIO, Informa Pharma Intelligence and QLS Advisors study covering 2011–2020 reported the following transition rates for company-sponsored programmes intended to support FDA registration. Its cumulative likelihood of approval from Phase 1 was 7.9%. Importantly, the filing-to-approval figure included resubmissions; it was not a first-review approval rate. These are dated cohort estimates, not universal probabilities for drugs being developed now. The definitions and figures appear in the clinical development success rates report for 2011–2020.
| Development transition | Historical success rate | Event measured |
|---|---|---|
| Phase 1 to Phase 2 | 52.0% | Advancement to the next phase |
| Phase 2 to Phase 3 | 28.9% | Advancement to the next phase |
| Phase 3 to regulatory filing | 57.8% | Progression to an NDA or BLA submission |
| Regulatory filing to approval | 90.6% | Approval, including resubmissions |
Multiplying these four transition rates gives approximately 7.9%. That calculation shows how attrition compounds across development. It does not establish that every Phase 1 asset deserves a 7.9% estimate. A programme with persuasive human evidence and a well-characterised target presents a different question from one relying on animal results and an untested biological hypothesis. Nor should a benchmark that includes eventual approval be used to predict approval by a particular decision date. A model can allow a reasonable chance of eventual approval while assigning a much lower chance to approval on schedule.
Different datasets can produce different answers without either calculation being wrong. Observation periods, programme definitions, missing records and the treatment of unfinished development paths all affect the denominator. A 2025 Nature Communications study of clinical trial success rates examined development programmes through 2023 and showed how estimates vary with methodology, time window and the groups being analysed. Use those differences to question comparability. Do not average published percentages simply because they share a label. Recent publication also does not mean recent outcomes: a paper published in 2025 may still depend on programmes that began many years earlier.
Choose a Reference Class That Resembles the Programme
Start with the disease and development stage, then consider the treatment technology, patient population and intended regulatory route. A broad oncology average may be a poor match for a treatment directed at a measurable alteration in a narrowly defined cancer population. An indication expansion for an approved medicine also differs from the first human test of a new compound, although prior approval does not establish efficacy in another disease. Keep the comparison group large enough to be informative. A tiny group of seemingly perfect matches can produce a fragile estimate, especially if several programmes share the same target or sponsor. Record the sample size, observation period and outcome definition before making any adjustment. If the dataset does not disclose them, use a wider uncertainty range.
Adjust the Benchmark for the Evidence That Matters
Biology and Human Target Validation
Ask what supports the proposed connection between the drug’s action and a useful patient outcome. Human genetic evidence, results from other drugs acting on the same target and direct measurements of target engagement can inform that assessment. They answer different questions. Reaching a target does not prove that changing it helps patients, and a validated target does not prove that a particular molecule can act on it safely at an effective dose. Separate target risk from molecule risk. The scientific rationale behind drug targets and mechanisms of action provides a framework for assessing that distinction. An adjustment should identify which uncertainty the evidence reduces, rather than awarding a general premium for an attractive scientific story.
The Strength and Relevance of Clinical Evidence
Clinical results deserve more weight when the study can distinguish a treatment effect from plausible alternative explanations. Examine the comparator, randomisation, blinding and handling of missing observations. Then check whether the next trial will test the same proposition. A positive result in a selected patient group may provide weak support for success in a broader population. Changes in dose, background treatment or follow-up can also weaken the connection. The practical question is whether the evidence survives the conditions planned for the next study. Use the clinical trial design to identify those differences before raising the success estimate. A company announcing that it is moving into Phase 3 confirms a development decision, not that the underlying evidence has become stronger.
Do not substitute statistical power for programme success probability. A trial designed with 90% power has a 90% chance of rejecting the null hypothesis under the assumed treatment effect and other design assumptions. It does not have a 90% chance of approval. The true effect may be smaller than assumed, safety may be unacceptable, or the evidence may fail another requirement. A p-value is not the probability that the drug works, either. Review effect sizes, confidence intervals and trial statistics to judge how much uncertainty remains around the observed benefit, especially when an early study supplies the assumptions for a larger one.
Safety, Exposure and Treatment Duration
A clean early safety report should change only the risks that the study could reasonably examine. Consider how many people received the proposed dose, how long they remained on treatment and whether the studied population resembles intended users. A short study provides little reassurance about problems that require longer exposure to emerge. The relevant tolerance for harm also depends on the benefit being pursued and the available alternatives. Avoid reducing safety to a count of adverse events. Assess severity, treatment discontinuations and whether patients can stay on an effective dose. A structured review of clinical safety data and benefit–risk balance can help distinguish a manageable treatment burden from a risk that threatens the entire programme.
Update Probabilities Without Manufacturing Precision
Bayesian reasoning offers a disciplined principle: start with an initial estimate, then revise it in proportion to the strength and relevance of new evidence. Formal models can combine earlier trial results with uncertainty about the true treatment effect and the design of a future study. Published research on Bayesian assessment of late-stage development success develops this approach using Phase 2 evidence, industry benchmarks and other information. An investor does not need to reproduce a sponsor’s statistical model to apply the discipline. The useful habit is to state the starting assumption, identify the new evidence and explain why it changes the forecast. Replacing 30% with 65% because the presentation looked persuasive is not an update that can be audited.
A hypothetical odds calculation shows the mechanics. Start with a 30% probability, equivalent to odds of 0.30 ÷ 0.70. If evidence were three times as likely under eventual success as under failure, multiplying those odds by three would produce a revised probability of 56.25%. The difficult part is the evidence multiplier. It cannot be obtained by reading a p-value or borrowing the company’s enthusiasm. Without a defensible model for that multiplier, report a reasoned range instead. A base estimate of 55%, tested against 40% and 70%, communicates the uncertainty more honestly than a figure carried to two decimal places. Decimal places are cheap; supporting evidence is harder to obtain.
Update the risks that the result actually addresses. A convincing efficacy readout may raise clinical success probability while leaving manufacturing readiness unchanged. A competitor’s result may strengthen confidence in the target but reveal a tougher commercial standard. Before changing the model, compare the announcement with the trial protocol, reported analyses and available results. The discipline used when reading clinical trial results matters more than the speed of the spreadsheet revision. Keep a dated record of the old estimate, the new estimate and the reason for the change. That record makes it harder to rewrite the original investment thesis after an inconvenient outcome.
Estimate Regulatory Success Separately From Trial Success
A positive primary endpoint does not settle the approval decision. FDA review examines whether the submitted evidence supports the proposed use, including the benefit–risk balance, study analyses and product manufacturing. Manufacturing deficiencies can delay or prevent approval even when the clinical case appears persuasive. These are separate review questions within the FDA drug review process. For valuation purposes, define regulatory probability conditional on the evidence the programme is expected to produce. An estimate made before a pivotal result should not assume the same regulatory position as one made after a complete, consistent package of results has become available.
Distinguish permanent failure from a repairable setback. A request for further analysis, a manufacturing correction and a requirement for another efficacy trial have different implications for cost and timing. They should not all be represented by subtracting ten percentage points from approval probability. Build separate branches when the distinction changes value: approval on schedule, approval after further work, and abandonment. The treatment of regulatory setbacks and additional studies should reflect what must be repaired and whether the sponsor has a credible route to doing so. A delay branch can retain substantial approval probability while losing value through extra expenditure and a later launch.
Accelerated Approval Leaves Another Evidence Hurdle
For qualifying serious conditions, FDA accelerated approval can rely on a surrogate or intermediate clinical endpoint considered reasonably likely to predict clinical benefit. Confirmatory studies remain necessary, and failure to verify benefit can lead to withdrawal or a changed indication under the FDA accelerated approval framework. Model initial approval and the durability of the approved indication as separate events when relevant. Revenue may begin before the confirmatory evidence is complete, so applying one probability to every future year can conceal an important difference between early sales and sustained access. Confirmatory study costs also belong in the model. Earlier market entry changes the timing of risk; it does not remove the unresolved clinical question.
Apply Probabilities to Cash Flows at the Correct Stage
A useful risk-adjusted valuation weights each cash flow by the probability that it will occur. Revenue depends on reaching the market. Development costs depend on reaching the stage when the spending becomes necessary. Costs already incurred are sunk, while spending committed before the next result should not be reduced by the chance of eventual approval. Consider a fictional programme entering its final clinical stage. Assume a 60% chance of producing a submission-ready package and a 90% chance of approval conditional on that package. Its cumulative approval probability is 54%. Keep commercial assumptions fixed initially so the effect of changing the clinical probability remains visible.
Assume the commercial cash flows, after launch and operating costs, have a present value of $500 million conditional on approval. The programme must incur $40 million of remaining trial expenditure before the result, regardless of outcome. Another $5 million of submission and review expenditure is incurred only if the trial programme succeeds. All amounts below are discounted to today, and the example assigns no salvage value to failure. The expected submission cost is therefore 60% × $5 million, or $3 million. The calculation is:
Risk-adjusted programme value = (54% × $500 million) − $40 million − (60% × $5 million) = $227 million.
| Illustrative assumption | Lower clinical case | Base case | Higher clinical case |
|---|---|---|---|
| Clinical programme success probability | 40% | 60% | 80% |
| Approval probability after clinical success | 90% | 90% | 90% |
| Cumulative approval probability | 36% | 54% | 72% |
| Probability-weighted commercial value | $180 million | $270 million | $360 million |
| Committed trial cost | $40 million | $40 million | $40 million |
| Expected submission and review cost | $2 million | $3 million | $4 million |
| Risk-adjusted programme value | $138 million | $227 million | $316 million |
The range is wide despite identical sales assumptions. That is useful information: it identifies clinical probability as a major driver of value. Next, allow the assumptions to interact. Weaker efficacy could support a smaller eligible population, lower adoption or further trials rather than simple failure. A delayed approval could shorten the useful selling period. These cases belong in valuation sensitivity and scenario analysis, rather than being concealed inside one percentage. Avoid adding an arbitrary discount-rate premium for the same failure risk already represented by explicit probabilities. Discounting and probability weighting serve different purposes, and applying overlapping adjustments can obscure what the valuation assumes.
Handle Shared Risks Across Trials and Indications
Two pivotal trials are not automatically independent bets. They may share the same uncertain treatment effect, dose or vulnerability in the endpoint. If that common assumption is wrong, both can disappoint. Multiplying two standalone success estimates can misstate the chance of completing the programme. Where several studies are required, estimate success conditional on the shared evidence or use a programme-level model that allows for dependence. Also check the definition of success: requiring both studies to meet their objectives differs from requiring at least one persuasive result. Multiplying conditional probabilities remains valid. Treating marginal probabilities as though they were conditional is the mistake.
The same issue applies across indications and pipeline assets. Approval in one disease can inform safety, dosing and manufacturing assumptions elsewhere, but it does not establish efficacy in another condition. Programmes using a shared delivery system may also share a failure mode, even if their target diseases differ. A company with several assets can therefore have less risk spreading than the asset count suggests. This is why diversification across technologies, diseases and development stages needs to consider common dependencies. Correlation affects the distribution of company outcomes and the chance of several failures occurring together; it does not change the arithmetic rule that expected asset values can be added when their individual estimates are sound.
Make the Probability Estimate Auditable
Keep technical success separate from the company’s capacity to finance it. A drug can retain scientific promise while its sponsor faces expensive financing, dilution or a forced partnership. Conversely, ample cash does not improve an unconvincing efficacy result. If a historical dataset counts programmes discontinued for commercial or funding reasons as failures, recognise that its benchmark may already contain some nontechnical attrition. Do not add the same risk again without adjustment. Assess cash burn and runway through the next milestone alongside the development forecast, then show how financing changes shareholder value. This keeps a forecast about the medicine from becoming an accidental forecast about the stock.
A working probability record should contain enough detail for another reader to reconstruct the judgement. Use a short decision log rather than a collection of unexplained percentage changes:
- Event and timing: define the outcome, jurisdiction and deadline being estimated.
- Historical starting point: record the dataset, comparison group, sample size and treatment of unfinished programmes.
- Evidence adjustments: identify the uncertainty each result reduces or increases.
- Remaining dependencies: note shared assumptions across trials, indications and manufacturing.
- Revision triggers: specify which forthcoming evidence would change the estimate and which valuation branches it would affect.
Support the record with retrievable documents, distinguishing registry entries and regulatory materials from sponsor interpretation. A repeatable process for researching trial registries, scientific papers and financial filings makes revisions easier to defend. Then test the investment at the lower end of the plausible probability range. If the valuation only works when every judgement favours success, the model is revealing dependence on optimism rather than a margin for error. Leave the uncertainty visible. Before the next readout, write down what a persuasive result must show, what an ambiguous result would mean and which outcome would invalidate the current estimate.