What are the rodent C-cell findings behind the thyroid warning, and do they transfer to humans?

The US labelling for several drugs in this class carries a boxed warning about thyroid C-cell tumours, and it is stated fairly starkly — tumours in both rats and mice, at clinically relevant exposures, with human relevance described as undetermined. That is an alarming sentence to read and then be told not to worry about.

What I would like is the actual content of the finding rather than the label summary. Which studies, which species, what doses relative to human exposure, and what kind of lesions. Then the counter-argument, because everyone who discusses this says the mechanism is species-specific and nobody I have read explains why it would be. "Rodents are different" is not a mechanism.

And finally, what the human data now say. The class has been in wide use for over fifteen years, so there ought to be pharmacoepidemiology by now, and I have seen it cited in both directions — one French study reporting an increased risk of thyroid cancer, and a Scandinavian one reporting none. I cannot tell whether those are contradictory or compatible.

The finding is real, reproducible and dose-dependent in rodents. The species-specificity argument is not hand-waving — it rests on a measured difference in receptor expression on the target cell, plus a negative primate study, plus fifteen years of human calcitonin data. Neither of those facts cancels the other, and the label's "human relevance undetermined" is an accurate description of where the evidence sits.

What the rodent studies actually showed

Two-year carcinogenicity studies in rats and mice produced dose- and duration-dependent increases in thyroid C-cell lesions, progressing across the usual sequence: hyperplasia, adenoma, then in some cases medullary carcinoma. The findings were reproducible across the class — liraglutide, exenatide extended-release, dulaglutide, semaglutide and tirzepatide all carry the resulting warning — which argues for a receptor-mediated class effect rather than molecule-specific toxicity.

The mechanistic work that accompanied those studies is the reason the story is more interesting than the label suggests. GLP-1 receptor agonist administration in rodents produced a prompt, dose-dependent rise in plasma calcitonin, C-cell activation and proliferation, and the effect required the GLP-1 receptor: it was absent in receptor-knockout animals. Rodent C cells express the GLP-1 receptor at readily detectable density, and chronic agonism produces sustained secretory stimulation followed by hyperplasia [1].

So the rodent finding has a complete mechanism: receptor on cell, agonist activates receptor, cell secretes and proliferates, chronic proliferation produces neoplasia. That completeness makes it credible and also makes the species question answerable, because each step can be checked elsewhere.

Why the mechanism appears species-specific

Four separate lines, and they point the same way.

• Receptor expression on the target cell. The central point. In the same body of work, GLP-1 receptor expression on primate and human thyroid C cells was low to undetectable, in contrast to clear expression in rodent C cells. No receptor, no receptor-mediated proliferation: the step driving the whole cascade is missing or greatly attenuated.
• The primate study. Cynomolgus monkeys dosed for 20 months at exposures well above human therapeutic levels showed no C-cell proliferative changes and no sustained calcitonin elevation [1]. A negative long-term primate study at supratherapeutic exposure is far more informative about humans than a positive rodent study.
• C-cell biology differs quantitatively. Rodent thyroid contains a considerably higher proportion of C cells than human thyroid, and laboratory rat strains have a high spontaneous background rate of C-cell hyperplasia and adenoma that rises steeply with age. A two-year rat study therefore runs against a high and rising baseline, and modest increases in a common spontaneous lesion are the sort of finding that most often fails to replicate across species.
• Human calcitonin data. Across the large clinical programmes, serum calcitonin was measured serially and did not show the sustained treatment-related rise the rodent mechanism predicts. If the mechanism operated in humans, calcitonin is the biomarker that should move first, and it did not.

That is the argument. It is a good argument. It is not proof, because the honest limitation is exposure duration: rodent carcinogenicity studies cover most of a rodent lifespan, and human data cover fifteen to twenty years of a much longer one, with most of that exposure in older people with diabetes rather than in people starting at 30 and continuing for four decades. That gap is why the label says "undetermined".

The two pharmacoepidemiology studies

Less contradictory than they appear, and the discrepancy is instructive.

A French nested case-control analysis within the national claims database reported an increased risk of thyroid cancer with GLP-1 receptor agonist use, with adjusted hazard ratios of roughly 1.6 for all thyroid cancer and higher for medullary thyroid carcinoma specifically, at one to three years of exposure [2]. A large Scandinavian cohort study using the national registries of Denmark, Norway and Sweden, with an active comparator design, found no increased risk, with a hazard ratio close to 1.0 and an upper confidence bound that excluded any large effect [3].

Three reasons those can both be honest results:

• Detection bias. People starting a new injectable therapy get more clinical contact, more neck examinations and more incidental thyroid imaging. Thyroid cancer is uniquely susceptible to this because the reservoir of clinically silent, indolent papillary microcarcinoma in the general population is enormous, and any increase in looking produces an increase in finding — comfortably large enough to generate a hazard ratio of 1.6.
• Comparator choice. Comparing against non-users mixes the drug effect with everything that distinguishes people who get prescribed a new therapy. The Scandinavian study used an active comparator, the stronger design for exactly this reason.
• Latency. One to three years is short for a carcinogenic effect on a solid tumour. An association appearing that quickly is more consistent with detection than with induction.

The reasonable synthesis: the human absolute risk, if any, is small enough that observational studies with different designs disagree about its sign, and no signal has emerged from the randomised programmes. That is a very different situation from a confirmed hazard, and it is also not a clean all-clear.

Where the warning genuinely bites

A personal or family history of medullary thyroid carcinoma, or of multiple endocrine neoplasia syndrome type 2, is a labelled contraindication rather than a caution, and that is the part that should change behaviour. It is a precautionary exclusion in a group whose C cells are already the problem, and the relevant test in that setting is germline RET testing, not a calcitonin level.

edited 17 Oct 2024 by plunger_stop — corrected a unit error in the worked example

One point on how to read a rodent carcinogenicity finding in general, since it generalises well beyond this class and it is the part that makes the label sentence less alarming without dismissing it.

Two-year rodent bioassays are regulatory screening instruments, not predictions. They are deliberately conducted at doses producing exposures many multiples of the human therapeutic exposure, in inbred strains with high and characteristic spontaneous tumour rates, over essentially a full lifespan. Their purpose is to be sensitive — to fail safe — and the price of that sensitivity is a substantial false-positive rate for human relevance.

The questions that determine whether a positive finding transfers:

• Is there a plausible mechanism, and is it present in humans? Here the mechanism is unusually well characterised, and its first step is measurably attenuated in humans. This is the strongest possible version of a species-difference argument, because it is specific rather than a general appeal to metabolic differences.
• Does the exposure multiple matter? Findings only at 50-fold human exposure are weaker evidence than findings at the therapeutic multiple. For this class, the rodent C-cell findings occurred across a range that includes clinically relevant multiples, which is why the warning is boxed rather than merely mentioned.
• Is the tumour type one with a high spontaneous rate in that strain? Rat C-cell lesions are. A treatment-related increase in a common background lesion is weaker evidence than the appearance of a rare one.
• Is there a biomarker that should move in humans, and does it? Calcitonin. It does not.
• Is there a negative study in a species closer to humans? Yes, the 20-month primate study.

Scored against that framework, this finding sits in an unusual place: strong and mechanistically complete in rodents, and with a specific, measured reason to think the mechanism does not operate in humans. Contrast it with the more common situation where a rodent tumour appears with no mechanism at all and the argument for irrelevance is purely statistical.

A note on the related biology, since it comes up. There is a separate literature on GLP-1 receptor expression in human thyroid tissue, and some studies have reported low-level expression, including in some papillary carcinoma specimens. The honest position is that "undetectable" is assay-dependent, and low-level expression is not the same as functional proliferative signalling. This is a real area of continuing investigation rather than a settled question, which is again what the label sentence is describing.

None of this is a reason to do anything differently in the absence of the specific family history. It is a reason to understand that the boxed warning is describing genuine uncertainty about a small potential risk rather than a demonstrated hazard, and those are the two things people most often confuse.

Worth adding what the randomised programmes reported, because the boxed warning is derived from animal data and the human randomised evidence tends to get skipped over in favour of the observational studies.

Across the large cardiovascular and obesity outcome trials — thousands of participants, follow-up in the three-to-five-year range, with prespecified adjudication of neoplasms — thyroid neoplasms have been uncommon and not clearly imbalanced. SELECT randomised over 17,000 participants with a mean follow-up near 40 months and reported no thyroid malignancy signal [1]. The same holds across the SUSTAIN, STEP and SURMOUNT programmes.

Two caveats on how much weight that should carry, both of which cut against over-reading the reassurance:

• Power. Medullary thyroid carcinoma has an incidence on the order of 1 to 2 per 100,000 per year. In 17,000 participants over 3.3 years, that is roughly 17,000 × 3.3 × 0.000015 = 0.8 expected cases in each arm. A trial expecting less than one event per arm cannot exclude a doubling, a tripling, or a tenfold increase in that specific tumour. The absence of a signal is nearly uninformative about medullary carcinoma specifically.
• Duration. Three to five years is short relative to the latency of solid tumours and very short relative to the exposure duration of someone starting this class in their thirties for weight management.

The arithmetic in that first bullet is the honest limitation of the entire human dataset. It is not that the trials looked and found nothing; it is that for the specific rare tumour implicated by the rodent mechanism, no realistically sized randomised programme could have found anything. This is why the label's position has not changed despite the accumulation of trial data, and why long-term registries rather than trials are where any answer will eventually come from.

Papillary thyroid carcinoma is a different matter: it is far more common, so the trials do have some power, and it is also the tumour most subject to detection bias in observational data. The mechanism from the rodent studies is C-cell specific and says nothing about papillary carcinoma, which is a follicular-cell tumour. Anyone citing the boxed warning as a reason to expect papillary carcinoma has conflated two unrelated cell types.

None of the above is medical advice. If you have thyroid nodules, a neck lump, hoarseness or dysphagia, those need assessment on their own merits and the drug history is a secondary consideration.