Jump to thr srticle with DOI: 10.13140/RG.2.2.32578.75207 here
Why drug programmes fail, and how asking questions in the right sequence could prevent it
J.Konstapel,Leiden,30-10-2026.
Aanleiding FD : in Kopenhagen spuiten de biotechbedrijven uit de grond, nu nog zorgen dat ze ook blijven
Marie Cadot is oprichter van Alvara Therapeutics.
Met A alvara Therapeutics gaat ze verder waar haar academische onderzoek eindigde. Ze wil effectiviteit en mogelijke bijwerkingen gaan testen. ‘Helemaal tot fase 1, waarin je een medicijn test op gezonde proefpersonen’, vertelt ze. Maar nu ervaart ze plots de keerzijde van haar eerdere succes.
Every medical research programme that aims to end a disease must eventually answer three questions:
- Which agent should we develop?
- Who should receive it?
- What should we measure to know whether it worked?
The modern pharmaceutical industry answers these questions in exactly that order — molecule first, patient population and measurement last. This essay argues that the order is inverted, and that this inversion is the single largest avoidable cause of failure in medicine today.
1. The problem: an inverted sequence
The second and third questions determine whether the first can be answered at all. Consider two ways a trial becomes unreadable:
- Averaged opposites. If a drug helps one group of patients and harms another, a trial run across both produces an average of two opposite effects. The result looks like “the drug did not work” — indistinguishable, on paper, from a genuinely inert compound.
- Measuring the mechanism. If the chosen endpoint is a quantity the drug moves directly (for example, a kidney drug whose endpoint is a kidney-function measure the drug itself raises), the trial measures the mechanism rather than the disease. Such a trial cannot fail — and therefore cannot be believed.
In both cases the molecule is blameless, but the trial file says the compound failed. The programme is closed, and nobody learns why.
This is not an argument that molecules don’t matter. It is an argument about sequence. The first question is expensive and glamorous — the entire discovery apparatus is built to answer it. The second and third are cheap and dull, and can often be answered from data that has already been collected. Yet they are asked last, when the protocol is drafted and the population is defined as simply “whoever carries the diagnosis.” By then the molecule has been chosen, and the design is a formality around it.
2. The evidence: attrition and its causes
How bad is the failure rate? Across 12,728 phase transitions recorded between 2011 and 2020, a programme entering phase 1 has only a 7.9% chance of reaching approval:
| Transition | Success rate |
|---|---|
| Phase 1 → phase 2 | 52.0% |
| Phase 2 → phase 3 | 28.9% |
| Phase 3 → filing | 57.8% |
| Filing → approval | 90.6% |
The trough is unmistakable: phase 2 to phase 3, where seven programmes in ten stop. This is precisely the step where a mechanism that looked active in a small study must show an effect in a diagnosis-defined population, against an endpoint a regulator accepts — exactly where the second and third questions arrive unanswered.
The recorded reasons for failure point the same way. Of 640 novel drugs entering pivotal trials between 1998 and 2008, 344 (53.8%) failed:
| Reason | Share of failures |
|---|---|
| Inadequate efficacy | 56.7% |
| Safety concerns | 17.2% |
| Commercial or other | 21.5% |
| Not identified | 4.7% |
Over half of late-stage failure is recorded as “the drug didn’t work” — the very category where an unreadable trial lands. The record cannot distinguish an inert compound from an averaged pair of opposite effects.
Two further figures deserve attention. Results were published for only 138 of the 344 failures (40.1%) — three fifths of what was learned was never written down. And of the 74 agents dropped for commercial reasons, only 6 (8.1%) were published. A mechanism can therefore be abandoned twice, by two different companies, for the same unmeasured reason — and the second cannot read about the first.
The money sits on the far side of the mistake. Average out-of-pocket spending per phase has been estimated at $25.3M (phase 1), $58.6M (phase 2) and $255.4M (phase 3) — phase 3 is 75.3% of the clinical outlay, and it is committed after the population and endpoint have been fixed. Three quarters of the money executes two decisions that were made for nothing.
3. The one design feature that moves the number
In the same registry of 12,728 transitions, programmes whose trials used a patient preselection biomarker reached approval at 15.9%, against 7.6% for those that did not — roughly a twofold difference. Put in programmes: without preselection, 13.2 programmes must enter phase 1 for each approval; with it, 6.3.
Yet only 6.03% of transitions (767 of 12,728) used one. The single design feature with the largest published association with success appears in one programme in seventeen.
An independent analysis of twenty years of oncology trials found even larger effects: drugs tested in biomarker-based trials were almost five times more likely to be approved — twelvefold in breast cancer, eightfold in melanoma, sevenfold in lung cancer (colorectal showed no benefit).
Two honest cautions. First, this is association, not proven causation: programmes with usable markers tend to sit in better-understood biology, which raises success on its own. Second, oncology markers usually identify the drug’s target itself — an easier problem than finding the boundary at which an agent that removes nothing reverses sign. The oncology figures establish direction, not a transferable effect size.
What survives both cautions is this: a programme that cannot name its marker cannot tell a null result from an averaged pair of opposite results. That distinction is worth everything, and the next section shows what it cost when the answer became known too late.
4. Five trials where the order decided the outcome
| Trial | Fixed first | Outcome |
|---|---|---|
| BEACON (bardoxolone, stage 4 kidney disease, n = 2,185) | Molecule | Stopped for harm at 9 months |
| AYAME (bardoxolone, diabetic kidney disease) | Molecule; endpoint = eGFR decline | Endpoint met, programme abandoned |
| CANTOS (canakinumab after heart attack, n = 10,061) | Entry by marker — but alpha split across doses | Effect real; the comparison that mattered unpowered |
| HARP-2 (simvastatin in ARDS, n = 540) | Molecule | Null overall; benefit recovered in one class afterwards |
| SAVE-MORE (anakinra in COVID-19) | Entry by marker, before randomisation | Effect found where others expected none |
BEACON. The primary endpoint was nothing (hazard ratio 0.98). But in the same patients, heart-failure hospitalisation or death ran at 96 against 55 events (HR 1.83) — one extra event for every 26 patients treated in nine months. The monitoring committee looked for characteristics predicting this harm and found none. A later analysis did find them: elevated BNP at baseline and prior heart-failure hospitalisation. The marker existed in the trial’s own baseline data. It was not a missing measurement; it was an unasked question.
AYAME. The same mechanism met its primary endpoint (a 30% decline in eGFR or end-stage renal disease) — and the sponsor abandoned the programme anyway. A trial that wins on its endpoint and is then discarded has measured the wrong thing. eGFR is the wrong endpoint for a drug whose mechanism moves eGFR directly.
CANTOS. Entry correctly required elevated hsCRP. But the statistical allowance was split across three doses of one antibody rather than across patient strata. The 150 mg dose passed its threshold by a margin of 0.0004; the 300 mg dose (HR 0.86) was declared non-significant. Meanwhile, splitting patients by on-treatment hsCRP revealed mortality hazard ratios of 0.69 below 2 mg/L versus 1.05 above — a 3.6-fold difference in numbers needed to treat. The alpha was spent on the dimension that did not matter.
HARP-2. A 540-patient trial of simvastatin in ARDS was filed as negative. A later analysis — using baseline data only, no outcomes, no new patients — split the population into two classes (hyponinflammatory, 28-day mortality 17%; hyperinflammatory, 39%). Within the hyperinflammatory class, simvastatin improved survival (P = 0.008). The benefit had been in the trial all along, averaged against its own absence.
SAVE-MORE. The one trial done in the right order. Anakinra was given to hospitalised COVID-19 patients selected by plasma suPAR ≥ 6 ng/mL, before randomisation, with the severely hypoxic excluded. The primary outcome improved dramatically (odds ratio 0.36) and 28-day mortality fell 54%. The effect was found in patients conventionally considered “too mild” for anti-inflammatory treatment — because entry was decided by a measurement, not a clinical stage.
5. The procedure: six gates, molecule last
The remedy is not a new discovery method. It is an order of questions, each with a stop rule, all answerable before the money is committed:
Gate 1 — Name the operation. Does the agent remove something, supply something missing, protect a self-completing process, or change a proportion between two things that both remain? These are the only four ways a disease state ends. The first three can be enrolled by diagnosis; the fourth cannot — and the fourth is where sign-reversal lives.
Gate 2 — Name an endpoint the agent does not move directly. A trial that measures its own mechanism cannot fail and cannot be believed. If the only acceptable endpoint is one the mechanism moves, do not run the trial yet.
Gate 3 — Name the marker and the direction of reversal. State in advance which measurable quantity divides the patients the agent helps from those it harms. This is a falsifiable statement and costs nothing. A programme that cannot make it has merely moved the question to after the trial.
Gate 4 — Look for that marker in data that already exists. Predecessor trials have baseline data; it can be reanalysed without enrolling anyone. HARP-2 recovered a benefit from a file marked negative; the bardoxolone heart-failure predictors were found in BEACON’s own baseline data.
Gate 5 — Fix the strata before allocation, and spend the statistical allowance on them. Not on doses. A dose can be found afterwards in anyone still alive to take it; a stratum cannot be recovered once the result is filed.
Gate 6 — Now choose the molecule. The marker names the property to select on, and it is usually not the property the chemistry was optimised for. For a mechanism whose predecessor died of volume retention, the criterion is the absence of a volume signal — not potency. Choosing the molecule first means choosing on the only criterion available before the marker is named, and potency was never what killed the predecessor.
What it costs
Gates 1–3 cost the time of people already employed: three sentences, either available or not. Gate 4 costs a statistician and access to data already paid for. Gates 5–6 are two paragraphs in a protocol. Set against $255.4 million for an average phase 3, the entire procedure is cheaper than the statistical reporting of the trial it governs.
Be exact about what it does: it does not raise the probability that a molecule is active. It raises the probability that the trial can read what the molecule did. For a portfolio, the arithmetic is stark: 13.2 programmes per approval becomes 6.3. And the procedure fails informatively — a stratified null proves the marker wrong, a result about the disease that can be written down and used.
6. Who this is worth most to
A company with one budget. A start-up with €500,000 of seed capital and 200 candidate molecules is framing its choice as €2,500 per molecule — along the one axis the budget cannot reach. Its mechanism’s predecessor died twice in phase 3: once for harm in a stage-defined population, once from an endpoint that moved while the disease did not. Neither death was a chemistry problem, and no selection among 200 molecules would have prevented either. The gates reorder that spending without increasing it.
A continent that will not win on capital. Europe’s biotech debate centres on money — a €10 billion initiative against roughly €200 billion invested in the United States. That race is not winnable. But what is missing from fourth-operation medicine is not capital; it is a design: no standard trial architecture for sign-reversing agents, no accepted entry-marker class, no regulatory precedent for a label naming a stratum rather than a diagnosis. Establishing those is a methodological act, not an investment. Whoever writes that standard defines the category — and being poorer does not lose that race. It is currently unused, while the argument is about money.
7. Limitations
Every figure in the original was recomputed from reported counts before being quoted, and the phase-transition chain reproduces the stated approval likelihood to within 0.03 percentage points — the strongest internal check available. Honest caveats remain: the preselection comparison is a registry association, not demonstrated causation; the oncology multiples come from a different kind of marker and did not hold in colorectal cancer; the failure-reason study covers 1998–2008 and proportions may have shifted; per-phase costs are portfolio averages used only for their ratio; CANTOS’s stratification used a three-month on-treatment measurement (weaker than baseline-only classification) and carried excess fatal infections; BEACON’s predictors come from a post-hoc analysis of a terminated trial. And the procedure itself is an ordering of existing practice, not a discovery — every gate is something somebody already does. The claim is only that doing them in this sequence is cheaper than doing them in the usual one.
Annotated References
- BIO, Informa Pharma Intelligence & QLS — Clinical Development Success Rates and Contributing Factors 2011–2020.
The core industry registry. Source for the 12,728 phase transitions, the 7.9% overall likelihood of approval, the phase-by-phase success table, and the preselection-biomarker comparison (15.9% vs 7.6%; 767 preselected transitions, 6.03%). Used for the essay’s central quantitative claims. - University of Toronto, on the Cancer Medicine biomarker analysis — Predictive biomarkers and the success rate of cancer drugs.
Independent check on the preselection effect, with twenty years of oncology data across four cancers. Provides the 5×, 12×, 8× and 7× multiples and the negative colorectal finding — establishing direction and order of magnitude, with the transferability caution stated in the essay. - Failure of investigational drugs in late-stage clinical development and publication of trial results, JAMA Internal Medicine.
Source for the 640 drugs entering pivotal trials 1998–2008, the 344 (53.8%) failures, the reason breakdown (56.7% inadequate efficacy), and the publication-gap figures (40.1% of failures published; 8.1% of commercial failures). Grounds the claim that most “efficacy” failures are recorded without distinguishing unreadable from inert. - DiMasi et al., Journal of Health Economics 47 (2016) 20–33, as tabulated by Knowledge Ecology International.
Per-phase out-of-pocket cost estimates ($25.3M / $58.6M / $255.4M). Used only for the ratio between phases — phase 3 as 75.3% of clinical outlay — because the dollar year is not stated explicitly in the source. - BEACON primary publication (NEJM 2013) and the mechanisms analysis (American Journal of Nephrology).
The NEJM paper supplies the trial’s primary result (HR 0.98), the eGFR and albuminuria findings, and the heart-failure harm (96 vs 55; HR 1.83). The American Journal of Nephrology analysis supplies the baseline BNP and prior-hospitalisation predictors and the urine-volume substudy — the evidence that the marker existed in data collected before the trial was stopped. - AYAME phase 3 results and discontinuation decision (Kyowa Kirin).
Shows the endpoint-met-but-abandoned pattern: primary and key secondary endpoints won, development discontinued. The essay’s example of a trial that measured its own mechanism. - CANTOS trial documentation (WikiJournalClub; American College of Cardiology) and the on-treatment hsCRP analyses (Residual Risk Reduction Initiative; TCTMD).
Supplies entry criteria (hsCRP ≥ 2 mg/L), the per-dose hazard ratios against prespecified thresholds (150 mg passing by 0.0004; 300 mg rejected), the on-treatment hsCRP mortality split (HR 0.69 vs 1.05; NNT 16 vs 57), and the fatal-infection excess. The essay’s example of statistical allowance spent on the wrong dimension. - Calfee et al., ARDS subphenotypes and differential response to simvastatin — HARP-2 secondary analysis.
The latent-class reanalysis of baseline-only data that recovered a benefit (P = 0.008 in the hyperinflammatory class) from a trial filed as negative. The essay’s benchmark for what Gate 4 costs and yields. - SAVE-MORE phase 3 (early anakinra guided by suPAR).
The trial run in the right order: marker-based entry before randomisation, suPAR ≥ 6 ng/mL, severely hypoxic excluded, ordinal outcome OR 0.36, 28-day mortality 3.2% vs 6.9%. Proof that the sequence works in practice. - Het Financieele Dagblad, Made in Europe series, 30 October 2026 (Alvara Therapeutics); European Commission and EIB Group biotech initiative.
The contemporary worked examples: the €500,000 seed / 200-molecule start-up, and the €10 billion European mobilisation set against ~€200 billion US investment. Illustrate who benefits most from the reordering. - J. Konstapel, Constable Research, Leiden — Research Directions to Cure All Diseases (29 Oct 2026); Depth of Discharge (29 Oct 2026); The Vacuum.Net Theory, 9th ed. (27 Oct 2026).
The framework works this essay operationalises: the four operations of Gate 1, the stratification argument, and the broader research programme within which the six-gate procedure sits.
Article
Based on “The Order of Questions,” J. Konstapel, Constable Research, Leiden, 30 October 2026.