Issue one Evidence review 26 August 2026 · event dated 19 August 2026

The Melanoma Result

whatholdsup.org

Merck and Moderna announced that a large trial of a personalised mRNA cancer therapy succeeded in melanoma. The coverage called it a landmark. This is what was actually released, what was not, and what the numbers underneath it will and will not support.

1,137patients in the Phase 3 trial
0Phase 3 efficacy numbers released
n=14the whole survival analysis
0.165–1.345the interval around that trend
The short version

The Phase 3 announcement reports no Phase 3 results. It describes the trial in detail, and it does quote hazard ratios — but every one of them belongs to an earlier, smaller trial. Of the study it is actually announcing it says only that the endpoints were met, "statistically significant and clinically meaningful." No hazard ratio, no interval, no p-value, no percentage.

So the numbers carried in the coverage — the 49% reduction, the 59% reduction — describe a different trial: an open-label study of 157 patients, not the blinded study of 1,137 that was announced.

And the one endpoint that would tell you whether anyone lives longer rests, at five years, on an analysis the companies label only "n=14". They call it an encouraging trend. The range around it runs from an 84% reduction in the risk of death to a 35% increase.

None of this means the treatment does not work. It probably does. It means the public has not been shown how well, and the numbers being repeated are carrying more weight than they can bear.

What was announced, and what was not

The distinction the coverage collapsed.

On 19 August 2026, Merck and Moderna announced that INTerpath-001 — a randomised, double-blind, placebo- and active-comparator-controlled Phase 3 trial in 1,137 patients — met its primary endpoint of recurrence-free survival and its key secondary endpoint of distant metastasis-free survival at an interim analysis.

That is a real and significant result. A double-blind trial of that size is the strongest instrument this field has, and it agreed with the smaller study that preceded it.

The release contains no hazard ratio, no confidence interval, no p-value and no percentage describing the size of that effect. It does contain hazard ratios — 0.510 for recurrence, 0.411 for distant metastasis — but those are the five-year results of KEYNOTE-942, the 157-patient trial that preceded this one, and the release says so. Of the 1,137-patient trial it is announcing, the quantitative content is nil. A week on, it still is. The companies say the data "will be presented at an upcoming international medical meeting and shared with regulatory authorities."

Why this is worth a whole article

Our first draft of this piece accused the coverage of passing those figures off as the Phase 3's own. Then we checked, and that is not what happened. Every outlet we could reach attributed them correctly to KEYNOTE-942, and several noted explicitly that the Phase 3 numbers were withheld. We were wrong and we have cut the accusation.

What is left is subtler and harder to fix. The figures are correctly attributed and still doing work they cannot do. A reader who sees "49% reduction" three paragraphs under "Phase 3 succeeds" comes away with a sense of how large this effect is. That sense is drawn from 157 unblinded patients, and the study that could correct it has published nothing. Nobody has to be careless for that to happen.

How to read a cancer trial result

Five terms that carry almost all the meaning, each explained with a real number from this story.

Hazard ratio HR

A comparison of how fast something is happening in two groups. An HR of 1.0 means the groups are identical. Below 1.0 means the treatment group is doing better; above 1.0 means worse.

The value is a rate, not a headcount. An HR of 0.51 does not mean 49% of patients were saved. It means that at any given moment during the trial, someone in the treatment group was recurring at about half the rate of someone in the comparison group.

In this story At five years of follow-up, KEYNOTE-942 reported a recurrence-free survival hazard ratio of 0.510. Reported as a percentage that becomes "a 49% reduction in the risk of recurrence or death" — the same fact wearing a more dramatic outfit.
The number that matters more Relative reductions sound larger than they feel. The absolute figures from the same trial, at five years: 68.8% of combination patients were recurrence-free versus 49.1% on pembrolizumab alone. That is a gap of roughly 20 percentage points — the number a patient would actually experience, and the one that rarely reaches a headline.
Confidence interval 95% CI

Every trial measures a sample, not the world. The confidence interval is the range of true effects reasonably consistent with what the researchers saw. A narrow interval means the trial pinned the answer down. A wide one means it did not.

The single most useful question to ask of any interval: does it cross 1.0? Because 1.0 means no difference at all. If the range includes 1.0, then "this treatment does nothing" is still among the possibilities the data cannot rule out.

Recurrence-freefirst readout, 2023 · HR 0.561
Recurrence-free5-year · HR 0.510
Distant metastasis-free5-year · HR 0.411
Overall survivalexploratory · HR 0.471 · "n=14"
00.51.0 — no effect1.5

All four intervals are from KEYNOTE-942, the 157-patient Phase 2b. Amber bars cross the no-effect line; green ones clear it. The Phase 3 has no bar here, because it has published no numbers.

What time did to the top two bars When KEYNOTE-942 first reported in 2023, its recurrence result was HR 0.561, 95% CI 0.309–1.017 — the interval crossed 1.0, so "no effect" could not be ruled out. By five years it had tightened to 0.294–0.887 and cleared the line. More data, more certainty, and a recurrence effect that held up rather than fading — one of the better signs that an effect is real.

Not everything moved that way. Distant metastasis-free survival was a 62% risk reduction at three years and 59% at five. That is a small easing, well inside the noise of a trial this size, and we mention it because a piece that only reported the numbers moving in the flattering direction would be doing the thing this article is about.
What the bottom bar is telling you The overall survival interval runs 0.165 to 1.345. It is consistent with the therapy preventing five deaths in six, and also with it causing a third more. It is that wide because it rests on an analysis the companies report only as "n=14" — they do not say fourteen of what, though on a survival endpoint there is not much else it can be. Reported as survival rates the same analysis reads 92.2% alive at five years versus 71.3% — which sounds conclusive, and rests on the same fourteen deaths. This is what "an encouraging trend" looks like underneath, and it is why the trend is labelled exploratory. The five-year analyses were descriptive: they were not designed to test a hypothesis, and no p-value attaches to them.
p-value p

The probability of seeing a result at least this good if the treatment actually did nothing. A small p-value means the result would be a surprising fluke. By convention, below 0.05 is called statistically significant — an arbitrary line, but a widely used one.

The universal misreading A p-value of 0.05 does not mean a 95% chance the drug works. It means: if the drug were useless, you would see a result this good about 5% of the time by luck. It describes the surprise, not the probability of the conclusion.
In this story — and this is the interesting part The April 2023 press release reported KEYNOTE-942 as one-sided p = 0.0266. When the same data were published in The Lancet, they came with two-sided p = 0.053.

A one-sided test asks only "is it better?" A two-sided test asks "is it different?", and is roughly twice as hard to pass. Neither is dishonest, and a one-sided test pre-specified for a mid-stage trial is a normal choice. But 0.0266 reads as a clear win and 0.053 reads as a near miss, and they are the same result. The confidence interval is the tell: 0.309–1.017 crosses 1.0, which is exactly what a two-sided test keys on. The framing was never explained to readers.

There is a second instance. The companies' five-year topline of 20 January 2026 reported one-sided nominal p = 0.0075. The peer-reviewed paper published at ASCO five months later states the five-year analyses were descriptive — not designed to test a hypothesis at all. Both descriptions are of the same data, and neither is improper — the word nominal is precisely the signal that a p-value was computed for context rather than to decide anything. It is worth knowing which one you are being shown.
"Met its primary endpoint"

Before a trial starts, researchers register one main question and a statistical threshold for answering it. "Met its primary endpoint" means the threshold was crossed. It is a pass/fail statement.

It contains no information about size. A therapy that reduces recurrence by 8% and one that halves it both produce that sentence. This is why the absence of a hazard ratio in the Phase 3 announcement matters: "met its endpoint" is the floor of what could be said, not a summary of what was found.

Also worth knowing The Phase 3 result came at an interim analysis — a pre-planned look before the trial finishes. Peeking early raises the odds of a false positive, so interim looks use a stricter threshold to compensate. Passing one is a real result, not a preliminary hint. But the trial continues, and the final picture can move.
What is being measured RFS · DMFS · OS

Three different questions, routinely blurred together in coverage.

Recurrence-free survival (RFS) — how long until the cancer comes back anywhere, or the patient dies. Distant metastasis-free survival (DMFS) — how long until it spreads to distant organs, the more dangerous kind of return. Overall survival (OS) — how long until death, from any cause.

The distinction that matters most Only OS tells you people live longer. RFS tells you the cancer stayed away longer, which is meaningful and is not the same claim. Adjuvant melanoma trials have a history of recurrence benefits that took years to translate into survival benefits, and occasionally never did. In this programme, overall survival has been reported only as an exploratory analysis in the smaller trial, on an "n=14". The Phase 3 is still measuring it.

The two trials, side by side

Coverage has been merging them. Different sizes, different patients, different rigour — and only one has published numbers.

KEYNOTE-942INTerpath-001
Phase2b3
Patients157 (107 vs 50)1,137 (2:1)
PopulationStage IIIB–IV resectedStage IIB–IV resected
BlindingOpen-label — everyone knew who got whatDouble-blind, placebo- and active-comparator-controlled
Recurrence-free survivalHR 0.510 (0.294–0.887)
at 5 years
met endpoint · not released
Distant metastasis-freeHR 0.411 (0.200–0.843)met endpoint · not released
Overall survivalHR 0.471 (0.165–1.345)
exploratory, "n=14"
still being measured

Why the blinding row matters

KEYNOTE-942 was open-label: patients and doctors both knew which arm they were in, and the primary endpoint was judged by those same investigators. That is a genuine weakness. A doctor who knows a patient received an experimental therapy may, entirely unconsciously, scan and interpret differently.

The Phase 3 substantially reduces that problem. It is double-blind and placebo-controlled, so nobody assessing a recurrence knows which treatment produced it. It is the more trustworthy trial by a wide margin — seven times larger, and blinded against exactly the bias the earlier one carried. Which is precisely why it is frustrating that its numbers are the ones we do not have.

It does not eliminate the problem, though, and that has had little attention. Merck's own description of the Phase 3 defines its primary endpoint as recurrence "as assessed by the investigator" — the same mechanism as the open-label trial. Blinding stops an assessor knowing which arm a patient is in. It does not replace the assessor with an independent review committee, which is the design that removes the judgement altogether. Better, then, but not solved.

And one thing nobody mentioned

KEYNOTE-942 enrolled stage IIIB–IV. The Phase 3 enrolled stage IIB–IV, which widens the population at both ends — down into IIB and IIC, and into IIIA, none of which the earlier trial included. Those patients start at lower risk of recurrence, which leaves less room for absolute benefit. Whether the effect holds up across that wider range is a real open question, and exactly the kind of thing a subgroup breakdown would answer.

What is established, and what is not

As of 26 August 2026.

Established

  • A large, well-designed, properly blinded Phase 3 crossed its pre-registered threshold on recurrence at an interim analysis.
  • In the earlier Phase 2b at five years, patients on the combination were recurring or dying at about half the rate (HR 0.510, 0.294–0.887) and reaching distant metastasis at about two-fifths the rate (HR 0.411, 0.200–0.843).
  • The recurrence effect held up over five years rather than fading, and its interval tightened enough to clear the no-effect line — often a sign an effect is real.
  • Immune-related adverse events were 45.2% on the combination versus 44% on pembrolizumab alone — essentially no extra immune toxicity from the added therapy.
  • Most side effects were mild to moderate — fatigue 59.6%, injection-site pain 59.6%, chills 51.0%. But serious ones were not identical: grade 3 or worse treatment-related events hit 25% on the combination versus 18% on pembrolizumab alone. The therapy is not free.

Not established

  • How large the Phase 3 effect is. No hazard ratio, no interval, no percentage has been published.
  • Whether anyone lives longer. The only survival figure in the programme is descriptive, rests on an "n=14", and its interval crosses 1.0.
  • Whether the benefit holds in stage IIB/IIC patients, whom the earlier trial never enrolled.
  • Durability beyond five years, in any trial.
  • Whether the approach works in any cancer other than melanoma.
  • What it will cost, or who will be able to get it.

Scoring the central claim

Our published six-dimension rubric, applied by hand. The arithmetic is shown so you can disagree with it.

What this score is and is not

This is not engine output. Melanoma is not one of the topics our pipeline covers, so nothing here was produced by the scoring system that runs the site. One of us applied the published rubric to one claim, by hand, and the working is below.

We are also part-way through rebuilding how the engine reaches a verdict, for reasons set out in Who Pays for This. Publishing a machine score today would imply a confidence in that machine we do not currently have.

"Intismeran autogene plus pembrolizumab improves recurrence-free survival versus pembrolizumab alone in resected stage IIB–IV melanoma."

Source quality3 / 5
Data support1 / 5
Reproducibility4 / 5
Consensus4 / 5
Recency5 / 5
Rigor5 / 5
3.4 Moderate (3×.25)+(1×.20)+(4×.15)+(4×.15)+(5×.15)+(5×.10)

The two extremes are the whole story. Rigor scores 5 — a double-blind, placebo-controlled, 1,137-patient Phase 3 is as good as trial design gets. Data support scores 1 — the rubric's anchor for 1 is "no data cited," and for the size of the effect, none was.

So the study is excellent and the evidence released about it is almost nothing. Those are different things, and a scoring system worth having has to be able to say both at once. Source quality scores 3 rather than 5 for the same reason: the claim currently rests on a corporate press release, which the rubric ranks as industry analysis, not the peer-reviewed publication it will eventually become.

This is what a headline cannot do. "Landmark trial succeeds" and "3.4, moderate" describe the same event, and only one of them tells you the numbers are still missing.

What would settle this

The full Phase 3 dataset — presented at a medical meeting or published in a journal, with the hazard ratio, the confidence interval and the breakdown by disease stage. That single release converts this from a directional claim into a measurable one, and would move data support from 1 to 4 or 5 and source quality from 3 to 5. Expect the score to move by roughly a full point when it lands, in whichever direction the numbers point.

Overall survival, on enough events to mean something. Until it reads out, the honest sentence is that this therapy delays recurrence — not that it extends life.

If you have melanoma right now

The part of this that is not academic.

This treatment is not available. It is investigational, approved by no regulator, and cannot be prescribed. The only route to it is a clinical trial.

  1. The question is trial eligibility, not treatment. The INTerpath programme is running across several cancers. Ask your oncologist whether anything is enrolling near you and whether your stage and surgical status fit.
  2. Your standard-of-care decision did not change this month. Adjuvant pembrolizumab is the comparison arm in this trial precisely because it is the established treatment — and everyone in the trial received it.
  3. It is manufactured per patient. The therapy is built from an individual tumour's own mutations, so it requires surgically removed tumour tissue and takes weeks to produce. That constrains who can realistically receive it even after approval.
  4. If someone quotes "49%" at you, ask which trial. It is a figure from 157 patients in an open-label study, not the result announced this month — and it describes a rate of recurrence, not a share of patients cured.

We assess published evidence. This is not medical advice, cannot account for your individual situation, and is no substitute for your oncologist — who can see your pathology, your stage and your history, none of which a page like this knows.

Sources

Primary sources first. Every number above traces to one of these, and none to a news report.

Primary Merck & Moderna — Phase 3 INTerpath-001 met RFS and DMFS endpoints 19 August 2026. Source of the trial design, enrolment, blinding and randomisation. Contains no efficacy numbers, which is the finding.
Primary Merck & Moderna — five-year KEYNOTE-942 data, ASCO 2026 1 June 2026. Source of HR 0.510 (0.294–0.887), HR 0.411 (0.200–0.843), the exploratory OS figure HR 0.471 (0.165–1.345) reported as "n=14", and the adverse-event rates.
Primary KEYNOTE-942: a randomised, phase 2b study — The Lancet, 2024 The peer-reviewed publication. Source of HR 0.561 (0.309–1.017) with two-sided p = 0.053, the stage IIIB–IV enrolment, the 107/50 arm sizes, the open-label design, and grade 3+ treatment-related events at 25% versus 18%.
Primary Five-year results — Journal of Clinical Oncology, 1 June 2026 The peer-reviewed five-year analysis, published the day it was presented. Median follow-up 60.3 months, data cutoff 15 December 2025, and the statement that the five-year analyses were descriptive.
Primary Merck & Moderna — first detailed KEYNOTE-942 results 16 April 2023. The source of the one-sided p = 0.0266, which is a company figure and not a journal one.
Trade The ASCO Post — INTerpath-001 meets primary and key secondary endpoints 21 August 2026. Used only to confirm that no Phase 3 efficacy numbers had been released two days after the announcement.