Chat GPT summarizes the data on the recently released SERENA-6 trial and ODAC-FDA advisory board. OncLive here. Fierce Biotech here. Precision Medicine Online (Turna Ray) here.
###
Chat GPT:
SERENA-6 was a Phase III trial in HR-positive, HER2-negative advanced breast cancer testing whether therapy should be changed when an ESR1 mutation first appears in circulating tumor DNA, before conventional radiologic progression. Patients on first-line aromatase inhibitor plus CDK4/6 inhibitor therapy were serially monitored with Guardant360 ctDNA testing. When an ESR1 mutation was detected but imaging had not yet shown progression, patients were randomized.
A key point is that the ESR1 test was used in both arms. Thus, both groups had the same molecular starting point: known ESR1 mutation detection without radiographic progression. In the experimental arm, patients switched from the aromatase inhibitor to camizestrant while continuing the CDK4/6 inhibitor. In the control arm, the therapy change was withheld, even though the ESR1 mutation was known, and patients continued the aromatase inhibitor plus CDK4/6 inhibitor until conventional progression.
The reported PFS difference was substantial: median PFS was about 16 months in the camizestrant-switch arm versus 9.2 months in the control arm, a delta of about 6.8 months. Conceptually, however, this was an unusual PFS interval. The start point was molecular progression or molecular resistance, defined by ESR1 mutation detection in blood, while the end point was conventional progression, mainly radiologic progression or death. In other words, the trial measured whether acting at a molecular warning point delayed later radiographic progression.
The FDA and ODAC concern was not that one arm had earlier detection than the other; both arms began from the same ESR1-positive molecular trigger. Rather, the concern was whether prolonging the interval from molecular progression to radiologic progression is, by itself, a sufficiently established clinical benefit to support approval of a new treatment-switching paradigm. ODAC voted 6–3 that the evidence did not convincingly show clinically meaningful benefit. FDA reviewers emphasized uncertainty about the value of PFS in this setting and pointed to the lack of mature overall survival evidence as a major unresolved issue.
###
If they want OS, that's really hard. These are early patients. It's their first progression. They will go years, on various and multiple lines of therapy, clinical trials, hospice, etc, between now and OS. So OS will be very very noisy and years out.
###
Yes — that is exactly the practical counterargument to FDA’s desire for OS.
For this setting, OS is a very blunt and delayed endpoint. These are not terminal last-line patients with a short survival horizon. They are metastatic HR+/HER2− breast cancer patients around the first-line-to-second-line transition, often with years of subsequent management ahead of them. After the SERENA-6 randomization point, they may receive multiple later endocrine regimens, targeted agents, chemotherapy, clinical trials, supportive care, and eventually hospice. By the time OS matures, the signal attributable to the early camizestrant switch may be badly diluted.
So there is a real tension:
FDA’s instinct:
“Show us that acting on molecular progression improves something hard and durable — preferably OS.”
Sponsor/clinician counterargument:
“OS is too remote, confounded, and slow in this disease setting. A randomized improvement in time from molecular progression to radiographic progression, plus quality-of-life or chemotherapy-delay signals, may be the most practical evidence of benefit.”
That is why PFS, PFS2, time to chemotherapy, time to symptom deterioration, quality of life, and treatment-sequence endpoints matter so much here. They are not perfect, but they may be more sensitive to the actual clinical question than OS.
The most persuasive pro-approval framing would be:
Once ESR1 molecular resistance is detected, continuing the aromatase inhibitor is biologically inferior. SERENA-6 randomizes patients at that molecular inflection point and shows that switching to camizestrant substantially delays radiographic progression, delays symptom deterioration, and may delay chemotherapy, without an apparent OS detriment. Requiring mature OS in this setting may be asking for an endpoint that is both impractically delayed and poorly matched to the intervention.
But FDA’s reply would be:
If we approve this, we are not just approving a drug. We are approving a new rule of oncology practice: serial ctDNA monitoring plus treatment switching before conventional progression. That paradigm could expose many patients to earlier therapy changes and repeated testing. For that, a higher evidentiary bar may be appropriate.
So I think your instinct is right: OS is scientifically attractive but operationally unrealistic and probably noisy here. The more reasonable debate is whether the package of randomized PFS improvement, patient-reported outcomes, delayed deterioration, PFS2, and delayed chemotherapy is enough to validate the early-switching paradigm.
###
We often hear, “We need OS,” but that standard can break down in several settings.
First, it can break down in early-stage trials, where death may be far downstream from the intervention, after many years of subsequent treatments, competing risks, and loss to follow-up. PSA screening trials are a good example: the relevant endpoint may unfold over decades, not months.
Second, it can break down in screening studies, such as colonoscopy. Overall survival is affected by every possible cause of death, while the screening intervention is aimed at preventing only one category of death. If colon cancer accounts for, say, 5% of deaths, and screening reduces those deaths by 20%, the result is only about a 1% shift in all-cause mortality. That small signal is then embedded within the remaining 99% of deaths from other causes, often over a long biological timeline from polyp to cancer to death.
By contrast, in a metastatic breast cancer trial, OS is still noisy, but at least most deaths are related to the disease being treated. In screening, the intervention may be clinically important while still being very difficult to prove through overall survival for straightforward, intuitive reasons.
###
Exactly. “We need OS” is rhetorically powerful but often methodologically naïve unless the clinical setting is one where the intervention’s effect on death should be large, near-term, and not heavily diluted by later events.
Your examples are two classic ways OS can become the wrong hammer.
1. Early-line oncology: OS is far downstream and contaminated
In metastatic HR+/HER2− breast cancer, especially around first progression, OS is often years away and mediated by many subsequent choices:
next endocrine therapy, CDK4/6 sequencing, PI3K/AKT/mTOR drugs, antibody-drug conjugates, chemotherapy, trials, local therapies, access differences, crossover, physician preference, and attrition.
So even if early camizestrant switching truly delays clinically meaningful progression, the OS signal may be washed out by later treatment heterogeneity.
This is why FDA’s desire for OS is understandable but not always realistic. The more relevant question may be:
Did the intervention improve the patient’s disease course at the point where it was applied?
For SERENA-6, that means: after ESR1 molecular progression, did switching therapy delay radiographic progression, symptom deterioration, chemotherapy, or next major treatment transition? That is a much more proximate question than “did it change death years later after several additional lines of therapy?” The ODAC article notes exactly this tension: OS was immature, and final OS was not expected for years.
2. Screening: all-cause mortality is diluted by unrelated deaths
Your colonoscopy example is very good.
Suppose, roughly:
Colorectal cancer = 5% of all deaths
Screening reduces colorectal cancer death by 20%
Then the expected effect on all-cause mortality is:
5% × 20% = 1% relative reduction in total deaths
That is before accounting for adherence, crossover, competing risks, lag time from adenoma to cancer to death, imperfect screening quality, and unrelated mortality. So even a highly valuable screening intervention can have a tiny effect on all-cause mortality.
This is why requiring all-cause OS for screening can become absurdly insensitive. The intervention is not supposed to prevent heart attacks, dementia, trauma, renal failure, suicide, sepsis, or other cancers. It is supposed to reduce one cause-specific pathway.
So the right endpoint may be colorectal cancer incidence, advanced adenoma detection, interval cancer rate, colorectal cancer-specific mortality, stage shift, or cancer-specific mortality, depending on the question.
The broader principle
A useful rule is:
The farther the endpoint is from the intervention’s mechanism, the more noise you accumulate.
OS is maximally hard because it sits at the end of every possible pathway. It is clean in one sense — death is hard to misclassify — but dirty in another sense: it is causally overdetermined.
For OS to be a good primary endpoint, you generally want:
- A high fraction of deaths caused by the target disease.
- A short time horizon.
- Few effective post-progression treatments.
- Limited crossover.
- A large expected treatment effect.
Late-line aggressive oncology often fits that. Early-line indolent cancers often do not. Screening trials often do not.
The better formulation
Instead of reflexively saying, “We need OS,” the more sophisticated question is:
What endpoint is close enough to the intervention’s mechanism to be sensitive, but clinically meaningful enough not to be a surrogate fantasy?
For SERENA-6, that probably means the debate should center on radiographic PFS after molecular progression, time to symptom deterioration, quality of life, time to chemotherapy, PFS2, and whether later outcomes show no harm — not necessarily mature OS as the only acceptable proof.
For colonoscopy, it means colorectal cancer-specific outcomes may be far more rational than all-cause mortality.
So yes: OS is the gold standard only when the biology, timeline, and causal pathway make it a usable standard. In many settings, it becomes a gold-plated fog machine.
###
I once used a deliberately stark example.
Imagine a stadium with 10,000 people. A gunman enters and kills 40 people before being stopped. Compare that with another stadium of 10,000 people where no gunman appears. In a formal statistical analysis of overall survival, the difference between the two stadiums might not be statistically significant, especially if follow-up includes all the ordinary causes of death that occur afterward.
But that does not mean the intervention was unimportant. If an action prevented those 40 deaths, most people would regard it as profoundly worthwhile, even if it did not produce a clean or statistically significant OS signal.
The point is not to trivialize OS, but to show its limits. A real, meaningful benefit can be diluted when measured against all-cause mortality, especially when the event being prevented is only one component of overall death risk.
###
Yes — it’s a brutal example, but analytically excellent.
It shows the absurdity of using all-cause mortality as the required endpoint when the intervention targets a small, specific, high-value cause of death.
In your stadium example:
Stadium A: 10,000 people, gunman kills 40.
Stadium B: 10,000 people, no gunman.
The gunman causes a 0.4% absolute mortality difference.
In ordinary statistical terms, especially if background mortality over the observation window is low or variable, you might not get a “statistically significant OS difference” unless the trial is enormous or repeated many times. But morally, clinically, and operationally, preventing 40 deaths is obviously meaningful.
So the right question is not:
Did all-cause mortality change enough to satisfy a conventional OS endpoint?
The right question is:
Did the intervention prevent the deaths it was designed to prevent, without causing offsetting harms?
That maps directly to screening and some early oncology settings.
For colonoscopy, the intervention is not designed to reduce all deaths. It is designed to reduce colorectal cancer incidence and colorectal cancer mortality. If colorectal cancer is only a small fraction of all deaths, then all-cause mortality is an insensitive endpoint.
For SERENA-6, the intervention is not designed to transform the entire survival arc of metastatic breast cancer from first progression to death years later through multiple treatments. It is designed to act at a specific biological inflection point: ESR1-mediated endocrine resistance before radiographic progression. The proximate question is whether acting there delays radiographic progression, symptoms, chemotherapy, or later treatment failure without harm.
Your stadium example captures the central fallacy:
An endpoint can be “hard” and still be poorly matched to the causal claim.
All-cause mortality is hard, objective, and important. But it can be too blunt to detect a real, valuable effect on a specific pathway.
A polished version for policy writing might be:
Requiring all-cause mortality can be misleading when the intervention targets a narrow causal pathway. If 40 deaths occur in a stadium of 10,000 because of a discrete preventable event, the absence of a statistically significant difference in all-cause mortality would not mean the intervention was unimportant. It would mean the endpoint was too diluted to capture the effect. The same logic applies to screening and to early-line oncology trials, where the clinically relevant benefit may occur along a specific pathway long before all-cause survival can reliably measure it.
Or even shorter:
OS is often called the hardest endpoint, but “hard” does not always mean “fit for purpose.”
