For boards and investors

Six Questions a Board Should Be Able to Answer Before Approving the Next Cohort

Published 2026-09-04 · Regulatory references are FDA guidance documents, identified below as final or draft. Guidance is non-binding and does not establish legally enforceable responsibilities. Not legal, regulatory or investment advice.

Before the board reads anything, a patient has already answered a harder question.

Every person who enrols in a first-in-human oncology trial is betting on the intentions of the people who designed it — that someone decided what the experiment needed to resolve, and that this was a defensible way to resolve it. They cannot audit that. They are not in the room where the next cohort is approved. Their interests are represented there only if the people at the table choose to represent them.

The six questions below are the board's version of the question the patient has already answered. They are also, taken together, the oncology-specific form of the question every director is actually asking when the pre-read arrives: are we all right?

Each one closes on the same test: what does a failure here do to the commitment you made at the last raise — the value inflection you promised, on the timeline you promised, with the cash you have?

These are the gates I run in the FIH Diligence Read, published here in full. The list is not the instrument. Asking these six is straightforward; grading the answers is the work, and a board is structurally poorly placed to grade its own.

A note on base rates before they do any work. Wong, Siah and Lo, using 406,038 clinical trial entries across 21,143 compounds, put oncology's Phase 1-to-approval probability at 3.4%. Earlier and smaller industry-curated samples put it at 5.1% and 6.7%. That three careful estimates of the same quantity differ by a factor of two is itself the argument: the prior is unfavourable and imprecisely known, so the burden sits with the evidence in front of you. None of these numbers is the probability that your asset fails. Diligence exists because assets are not interchangeable draws.

1. A mechanism that cannot be wrong is not yet a mechanism

Ask the team to state the causal chain from target engagement to clinical benefit, and then ask what result would have falsified it.

A weak answer restates the hypothesis with the responses appended, as though the responses confirmed it. A strong answer names the step the trial actually measured and the observation that would have counted against it.

There is a reason to put this first beyond intuition. When AstraZeneca reviewed its small-molecule portfolio from 2005 to 2010 and asked what separated the projects that worked from the ones that did not, the first of the five determinants they identified was the right target. The follow-up paper describes what they changed in response, and target validation is the first thing on the list.

Two mechanism classes show what that failure costs when nobody asks the falsifying question early. Jentzsch and colleagues traced every clinical trial of an inhibitor of the insulin-like growth factor-1 receptor: 16 inhibitors, 183 trials, 12,396 patients, an estimated $1.6bn to $2.3bn, and no trial in any tumour type showing efficacy that permitted approval. Their analysis of the preclinical work that launched those trials found half the published in vivo data showed less than 50% inhibition of tumour growth. The evidence was mixed on its own terms before the first patient enrolled. What was missing was not data but the question of what would have counted against the hypothesis.

The same group followed TIGIT a decade later. Twenty-one companies, thirty inhibitors, 220 trials, 48,731 patients enrolled between May 2016 and September 2025, an estimated $3.1bn to $3.6bn — and no marketing authorisation anywhere as of November 2025. Eight of the thirty programmes have been terminated, having enrolled 14,901 patients between them. Those are people who consented to an experiment on the understanding that someone had asked what it would take to be wrong.

Neither paper shows the biology was wrong. TIGIT candidates remain in late-stage trials, and absence of approval to date is not proof of absence of effect. What both show is the cost of scale without falsifiability — and they raise a second question a board should ask when an asset sits in a crowded class. Is the crowd evidence, or is it correlated error? Twenty-one companies pursuing one target is not twenty-one independent validations of it.

What this gate tests is narrower than it first appears, and it is not the same test as the fourth. Here the question is whether the mechanism is specified tightly enough to be wrong. In the fourth the question is whether the observed signal discriminates between that mechanism and simple selection. A programme can have a well-specified mechanism and an uninterpretable signal, or a vague mechanism and a clean one.

Against the commitment: a mechanism that cannot be falsified cannot be de-risked either. Whatever inflection you promised is being promised on a story no experiment can retire.

2. Dose is what you gave. Exposure is what the patient got.

These are treated as interchangeable in decks where they are not interchangeable in the patient.

The third of AstraZeneca's five determinants is the right tissue — whether the drug achieves adequate exposure at the site of action. It sits alongside the right target as a separate question for a reason: a correct target reached at insufficient concentration produces the same clinical result as a wrong target, and the two failures are indistinguishable from the response data alone.

Ask what exposures were achieved, in whom, with what between-patient variability, and whether any relationship exists between exposure and the effect being claimed. "We dosed at the recommended Phase 2 dose" answers a different question.

Against the commitment: if exposure was never characterised, the dose carried into the registrational programme is an assumption, and the timeline built on it inherits that assumption.

3. A tolerated dose is not an optimised dose, and the agency has said so

Oncology has historically selected doses by pushing to the maximum tolerated dose, a convention inherited from cytotoxic chemotherapy, where more drug generally meant more cell kill. FDA's final guidance of August 2024 describes how that was done: "The MTD was identified by evaluating stepwise, increasing doses in a small number of patients at each dose level for short periods of time until a prespecified rate of severe or life-threatening dose-limiting toxicities (DLTs)."

For targeted agents and immunotherapies that logic does not hold, and the guidance sets out what it recommends instead: "A recommended trial design to compare multiple dosages is a randomized, parallel dose-response trial," because randomisation "promotes comparability of patients receiving each dosage, minimizing bias in estimation of dose- and exposure-response relationships." The guidance is final and non-binding; its operative verb is should, and a board should hear the difference between a recommendation and a requirement.

What the initiative is actually about is worth stating plainly, because it is easy to file as a design-standards question. It came from watching patients take more drug than they needed, for longer than they needed to, and carry toxicity that bought them nothing.

FDA's own comparison makes the point. In DESTINY-Lung02, two dosages of trastuzumab deruxtecan were compared under randomisation. Response rates were 58% and 43% in favour of the lower dosage; interstitial lung disease and pneumonitis ran 6% against 14%. The higher dose bought toxicity and not response. A programme reasoning from dose escalation alone would have carried the wrong one forward.

So the question is not what the dose is. It is what it was compared against. And as the guidance notes, "conducting clinical trials to compare multiple dosages may be challenging to complete once a drug is approved for a given indication" — which is the argument for settling it now.

Against the commitment: a dose that was tolerated rather than compared is an unpriced regulatory obligation sitting inside the registrational plan, and it is not cheaply discharged later.

4. The surrogate is not the endpoint

FDA's public listing carries 35 withdrawn cancer accelerated approvals, the earliest withdrawn in November 2011 and the most recent in September 2026. Alongside them sit 125 accelerated approvals that converted to traditional approval on verified clinical benefit, and 57 still under confirmatory obligation.

Give a board the numerator alone and you have made an argument by selection. The honest reading is that early evidence a regulator found sufficient did not survive confirmation in a substantial minority of resolved cases — enough to take seriously, not enough to treat accelerated approval as a broken mechanism. The category is mixed in both directions. Some rested on a response rate; bevacizumab in first-line metastatic breast cancer rested on progression-free survival. I have not counted the split, and will not assert one. And not all were withdrawn because confirmation failed — some confirmatory trials were never completed, and some withdrawals were commercial decisions. Withdrawal is a mixed category, and a board told otherwise is being sold a cleaner story than the record supports.

There is now a direct measurement of how far early signals travel. Li and colleagues matched 51 pairs of early-phase and Phase III trials drawn from more than 9,600 PD-1/PD-L1 studies, holding regimen, cancer type, treatment line and biomarker strategy constant across each pair. Early-phase response rates systematically overestimated the Phase III result — an odds ratio of 1.66, with a 95% confidence interval of 1.43 to 1.92. The eligibility factor significantly associated with the gap was the exclusion of patients with autoimmune disease. The early trials were not wrong about their own populations. They were run in populations that the Phase III could not reproduce.

So the gate: what is the denominator, how did patients enter it, how many responses were confirmed rather than counted once, how long did they last, and what would the comparator plausibly have produced in the same population?

One refinement worth carrying, and in oncology it is not a small one. Wong, Siah and Lo separated trials that use a biomarker to select patients from those that do not. In oncology the gap is 1.6% probability of success without, against 10.7% with — nearly sevenfold, on 9,349 and 1,136 Phase 1 transitions respectively, and oncology is the one therapeutic area where the biomarker sample is large enough to read. Two things to hold alongside it. Those figures come from a narrower window and a different computation than the 3.4% above — 2005 to 2015, phase-by-phase, rather than 2000 to 2015 path-by-path; on the narrower basis oncology overall is 2.1% rather than 3.4%. And this is an association in observational data, not a demonstrated cause. The authors also reran it with a wider definition, counting trials whose objective was to evaluate or identify a novel biomarker rather than to select patients using an established one. On that definition the advantage disappears entirely. So selection is not the enemy of a credible signal; unreproducible selection is — and a marker still being validated is not the same asset as a marker already deciding who receives the drug. The question is not whether the population was enriched. It is whether the same enrichment can be applied to the population you intend to treat.

The failure runs in both directions. A coherent signal in a well-characterised population can be discounted because the package is disorganised — the evidence exists but was never assembled into a form a committee can trust. Under-reading costs as much as over-reading; it is simply less embarrassing.

The two-chair version of this argument — how the same trial reads to the developer and the investor — is set out in What a First-in-Human Trial Is Actually For.

Against the commitment: if the signal will not survive a control arm, the inflection you promised will not be recognised by the market that has to price it.

5. The DLT tally is not the safety database

A dose-limiting toxicity window is short by construction. An agent given continuously for months produces a profile that a first-cycle window cannot see.

Ask for exposure-adjusted follow-up rather than patient counts. Ask about dose reductions, dose interruptions, and discontinuations for reasons other than progression. Those three numbers are the record of what patients could actually live with, which is what determines whether anyone stays on the drug long enough to benefit from it. A deck built around a DLT tally has nowhere to put them.

Against the commitment: tolerability failures surface in Phase 2 and 3 as enrolment and retention problems, which show up on the timeline before they show up in the data.

6. The next experiment should resolve the uncertainty, not advance the programme

The first five gates identify what is unknown. This one asks whether the unknown is tractable, and whether the next spend is aimed at it.

Some uncertainties resolve with a well-designed cohort. Some need a randomised comparison the company cannot fund alone. Some cannot be resolved at this stage at any price — which is a finding, not a failure. It tells you the next dollar should buy something else.

When the planned experiment is the one that most advances the programme rather than the one that most reduces uncertainty, two things are true at once and they are the same thing. The capital is buying progress rather than information. And patients have been asked to enrol in an experiment that had no honest question it was trying to answer. The argument for rigour that a board makes to its investors and the argument it owes the people taking the risk are not in competition. They converge, and where they diverge the programme has told you something.

Against the commitment: this is the question the commitment turns on. Can the next experiment deliver the inflection you promised, with the cash on hand, and would achieving it actually change what the asset is worth?

What a gate failure is, and what a defensive answer is

A failed gate is not a recommendation to pass. It is a named condition with a route to reopening it. "Dose credibility fails; a randomised two-dose expansion in the intended population would reopen it" is actionable. "Concerns about dose" is not.

I hold a firm view on what follows from that, formed over thirty years: programmes that adapt on emerging evidence outperform programmes that defend the plan they started with. The closest thing to a controlled demonstration I know of is AstraZeneca's published before-and-after. After making evidence-based decision discipline explicit and applying it to target validation, exposure and dose, their success rate from candidate nomination to Phase III completion rose from 4% in 2005–2010 to 19% in 2012–2016. It is one company's own account of its own portfolio, not a randomised comparison, and it should be read as such. But a failed gate is the cheapest form of that evidence, and it arrives while it can still change something.

The second thing to grade is not in the answer at all. When one of these six draws a defensive answer, the board has learned something independent of the science — because the same reflex that makes a management team defend a dose in a board meeting makes them defend it in a pre-IND meeting, and the agency is a less forgiving audience. Directors who find candor punished stop offering it, and the reservations they were carrying surface at the next raise instead, when nothing can be done about them. A board that wants these six questions answered honestly has to make honest answers safe first. Otherwise it gets six confident answers and no information.

What I would not conclude from any of this

I would not conclude that a programme passing all six is a good investment. These gates test whether the evidence supports the claims being made about it. They do not price the asset, judge the market, or assess the team, and a programme can clear every gate and fail for reasons no first-in-human dataset could have exposed.

I would not conclude that the base rate applies to any particular asset. It describes a population, imprecisely.

I would not conclude that a board running these questions itself does not need an independent read. It needs one less often. What it cannot do from the inside is grade its own answers — the questions are cheap to ask and expensive to mark, and the people best placed to mark them are the people who wrote the answers.

And I would not conclude that the six are complete. They are the ones that have earned their place across thirty years of development decisions and seven anticancer approvals worldwide, including by the FDA, EMA, PMDA, and NMPA. If a programme fails in a way these six would not have caught, that is information about the instrument, and it will be recorded as such.

---

A board can run these six questions. Grading them is a different job, and an independent read is how a board gets that adjudication without spending director-hours it does not have. That is what the FIH Diligence Read is: ten business days, fixed scope, fixed fee — the six gates above, nine scored domains with a confidence overlay, a board-ready synthesis, and a live session where your thesis is argued against rather than summarised back to you. If the decision in front of you is an FDA interaction rather than a capital one, that is the Calibration Sprint's terrain — same discipline, different door.

Sources

• FDA — Optimizing the Dosage of Human Prescription Drugs and Biological Products for the Treatment of Oncologic Diseases — final guidance, August 2024, docket FDA-2022-D-2827. Quotations read from the guidance text: MTD description, II. Background; post-approval dosage comparison, II. Background; recommended trial design and randomisation rationale, III.B.
• FDA — Withdrawn Cancer Accelerated Approvals (35 entries), Ongoing (57), and Verified Clinical Benefit (125). Counts read 2026-09-04; the three tables carry different currency dates, so the figures are given separately rather than summed.
• D. Cook, D. Brown, R. Alexander, R. March, P. Morgan, G. Satterthwaite, M. N. Pangalos — Lessons learned from the fate of AstraZeneca's drug pipeline: a five-dimensional framework, Nature Reviews Drug Discovery 2014;13(6):419–431. The five R's: right target, right patient, right tissue, right safety, right commercial potential.
• P. Morgan, D. G. Brown, S. Lennard, et al. — Impact of a five-dimensional framework on R&D productivity at AstraZeneca, Nature Reviews Drug Discovery 2018;17(3):167–181. Candidate nomination to Phase III completion: 4% (2005–2010) to 19% (2012–2016). Single-company, self-reported, before-and-after — not a randomised comparison.
• X. Li, Y. Zhou, B. Xu, et al. — Comparison of efficacy discrepancy between early-phase clinical trials and phase III trials of PD-1/PD-L1 inhibitors, Journal for ImmunoTherapy of Cancer 2024;12(1). 51 matched pairs; ORR overestimation OR 1.66 (95% CI 1.43–1.92).
• V. Jentzsch, L. Osipenko, J. W. Scannell, J. A. Hickman — Costs and Causes of Oncology Drug Attrition With the Example of Insulin-Like Growth Factor-1 Receptor Inhibitors, JAMA Network Open 2023;6(7):e2324977. 16 inhibitors, 183 trials, 12,396 patients, no approval; half of the published in vivo preclinical data showed less than 50% tumour growth inhibition.
• E. H. Zhang, L. Osipenko, J. A. Hickman — Human and financial consequences of herding in oncology drug development: clinical trials of TIGIT inhibitors, BMJ Oncology 2026;5:e001037. 21 companies, 30 inhibitors, 220 trials, 48,731 patients, no marketing authorisation as of November 2025.
• C. H. Wong, K. W. Siah, A. W. Lo — Estimation of clinical trial success rates and related parameters, Biostatistics 2019;20(2):273–286. Table 2 (2000–2015, path-by-path): oncology Phase 1 to approval 3.4%. Table 3 (2005–2015, phase-by-phase): oncology 1.6% without a patient-selection biomarker vs 10.7% with. Widening the definition to trials evaluating a novel biomarker removes the advantage. Data end October 2015; the 5.1% and 6.7% comparisons are as reported there.
• FDA — Project Optimus presentation by Mirat Shah, MD, Clinical Lead, on the dosage-optimization guidance. Source of the DESTINY-Lung02 comparison (trastuzumab deruxtecan, 5.4 vs 6.4 mg/kg, randomised 2:1).

← Back to Writing  ·  What an FIH trial is actually for →  ·  The FIH Diligence Read →