When a Biomarker Isn't a Target
Six Drugs That Moved the Number and Changed Nothing

In 2017, a drug called evacetrapib more than doubled the HDL cholesterol of 12,092 people with vascular disease and cut their LDL cholesterol by nearly a third. The trial was stopped early — not for success, but because it was not going to show one. Over a median of 26 months, 12.9% of the people taking it had a cardiovascular event, against 12.8% of those taking placebo.
The short answer
A biomarker that reliably predicts a disease is not the same thing as a lever that changes it. Cholesterol transport, heart rhythm, homocysteine, triglycerides and lipoprotein(a) have all been moved dramatically by drugs that then failed to prevent a single extra heart attack — and in two cases caused harm. The test for whether a number is worth treating is not whether it predicts risk. It is whether moving it, in a randomised trial, moved the outcome.
The newest example arrived this month
On 7 November 2026 the full results of Lp(a)HORIZON were published in the New England Journal of Medicine.[1] Pelacarsen, an antisense drug built to lower lipoprotein(a), was given to 8,350 people with established cardiovascular disease and an Lp(a) of at least 70 mg/dL. It did exactly what it was designed to do: at twelve months the median Lp(a) in the treated group was 64.4 nmol/L, against 233.4 nmol/L on placebo, down from a median of 236 nmol/L at baseline.
Over a median of 4.7 years, a major cardiovascular event occurred in 511 patients (12.3%) on pelacarsen and 503 (12.1%) on placebo — a hazard ratio of 1.02, with a confidence interval from 0.90 to 1.15 and a P value of 0.76. Among the 6,540 patients with the highest baseline levels, the event rate was 12.3% in both arms.
Lipoprotein(a) is not a weak marker. It is one of the best-validated risk predictors in cardiology, with genetic evidence that it is causally involved in atherosclerosis. None of that was wrong. What the trial showed is that a number can be causally involved in a disease and still not be a useful thing to aim a drug at — at least not this drug, at this stage of the disease, over this time frame. We covered the full result and its safety profile in a separate piece on what Lp(a)HORIZON actually showed, and what still lowers risk when your own Lp(a) is high.
Lp(a)HORIZON is only the most recent entry on a long list. It is worth looking at the rest of it, because the pattern is old, expensive and — once you have seen it a few times — fairly easy to spot.
The first time, the drug worked and people died
The canonical case is the Cardiac Arrhythmia Suppression Trial, published in 1991.[2] The reasoning behind it was impeccable. Extra ventricular beats after a heart attack predict sudden cardiac death. Encainide and flecainide suppress extra ventricular beats. Therefore, suppressing them should prevent sudden death.
CAST enrolled 1,498 patients in whom the ectopy had actually been suppressed by the drug, and randomised them to continue the drug or switch to placebo. After a mean follow-up of ten months, 89 patients had died. Fifty-nine died of arrhythmia — 43 on the drug against 16 on placebo (P = 0.0004). Another 22 died of non-arrhythmic cardiac causes — 17 on drug against 5 on placebo (P = 0.01). Encainide and flecainide were withdrawn from the trial for excess mortality.
The drugs did precisely what they were prescribed to do to the surrogate, and killed people doing it. The authors' closing sentence has aged into one of the most quoted lines in trial methodology: the mechanisms underlying the excess mortality "remain unknown."
Five years later, Fleming and DeMets set out the general argument in the Annals of Internal Medicine, under a title that is still the shortest summary of the problem: "Surrogate end points in clinical trials: are we being misled?"[3]
What happened when each biomarker was moved
Primary-endpoint hazard or rate ratios with 95% confidence intervals. Eight of the ten attempts to treat a biomarker produced an interval that crosses 1.0 — no detectable effect. The bottom bar is not a single trial: it is the Cholesterol Treatment Trialists' meta-analysis of 26 statin trials, expressed per 1.0 mmol/L of LDL cholesterol lowering, included here as a positive control to show what a surrogate that does hold up looks like. The populations, endpoints and follow-up differ between every bar, so read each one against 1.0 rather than against its neighbours.
Sources: references 1, 2, 4–11.
The HDL hypothesis: four drugs, and one honest exception
For twenty years, low HDL cholesterol was among the most reliable predictors in cardiovascular epidemiology. Observationally, each standard deviation of higher HDL came with roughly 38% lower odds of myocardial infarction.[4] Inhibiting cholesteryl ester transfer protein raises HDL sharply, so four such drugs went into large outcome trials.
Torcetrapib was tested in ILLUMINATE, in 15,067 patients at high cardiovascular risk.[5] At twelve months HDL cholesterol was 72.1% higher and LDL cholesterol 24.9% lower. Cardiovascular events rose: hazard ratio 1.25 (1.09–1.44, P = 0.001). Death from any cause rose: hazard ratio 1.58 (1.14–2.19, P = 0.006). The trial was stopped. Systolic blood pressure had risen 5.4 mm Hg, potassium had fallen, and sodium, bicarbonate and aldosterone had risen. The usual retelling is that torcetrapib failed because of an off-target blood-pressure effect — but the authors were careful, and so should we be: they wrote that although there was evidence of an off-target effect, they could not rule out adverse effects related to CETP inhibition itself.
Dalcetrapib was tested in dal-OUTCOMES, in 15,871 patients after a recent acute coronary syndrome.[6] HDL cholesterol rose by 31% to 40% over the trial, against 4% to 11% on placebo, with minimal effect on LDL. At a prespecified interim analysis with 1,135 primary events — 71% of the number projected — the data monitoring board stopped it for futility. Event rates were 8.0% and 8.3%; hazard ratio 1.04 (0.93–1.16), P = 0.52.
Evacetrapib was tested in ACCELERATE, in 12,092 patients with high-risk vascular disease.[7] At three months, HDL cholesterol was 133.2% higher than baseline against a 1.6% rise on placebo, and LDL cholesterol was 31.1% lower against a 6.0% rise. The trial was stopped early for lack of efficacy after 1,363 of the planned 1,670 events. A primary event occurred in 12.9% of the evacetrapib group and 12.8% of the placebo group: hazard ratio 1.01 (0.91–1.11), P = 0.91. A drug that more than doubled HDL cholesterol and cut LDL cholesterol by a third changed the event rate by one tenth of one percentage point.
Anacetrapib is the exception, and the reason it is an exception is the whole argument of this article. REVEAL randomised 30,449 adults with atherosclerotic vascular disease, already on intensive atorvastatin, and followed them a median of 4.1 years.[8] The primary outcome occurred in 10.8% on anacetrapib against 11.8% on placebo: rate ratio 0.91 (0.85–0.97), P = 0.004. A real, modest benefit. At the trial midpoint HDL cholesterol was 43 mg/dL higher — a relative difference of 104% — and non-HDL cholesterol was 17 mg/dL lower, a relative difference of 18%.
A benefit of roughly 9% is about the size you would predict from an 18% reduction in non-HDL cholesterol alone. The 104% rise in HDL appears to have bought nothing. This is an inference from the size of the effect, not a mediation analysis the trial published — but it is the straightforward reading, and it means REVEAL is better understood as a successful LDL-type trial than as a successful HDL trial.
The HDL hypothesis, biomarker beside outcome
What each drug did to HDL cholesterol
What each drug did to events
Read down each panel, not across — the two panels are in different units. On the left, how far each drug moved the biomarker it was built to move. On the right, what happened to the clinical outcome in the same patients. The dalcetrapib bar on the left shows the published range of 31% to 40% over the course of dal-OUTCOMES, not a confidence interval; every interval on the right is a 95% confidence interval. Anacetrapib is shown in plum because it is the one trial with a benefit — and its non-HDL cholesterol fell 18%, which is the likelier explanation.
Sources: references 5–8.
Genetics had already said so
The strongest evidence that HDL cholesterol was a marker rather than a mechanism arrived before the last two trials reported. In a Mendelian randomisation study published in the Lancet in 2012, Voight and colleagues used naturally occurring genetic variants as a lifelong randomisation.[4]
Carriers of one variant in the endothelial lipase gene, LIPG Asn396Ser, present in 2.6% of people, have HDL cholesterol 0.14 mmol/L higher than non-carriers and otherwise similar risk factors. From the observational data, that difference should translate into 13% lower risk of myocardial infarction (odds ratio 0.87, 0.84–0.91). The observed odds ratio in carriers was 0.99 (0.88–1.11), P = 0.85. A genetic score of 14 HDL-raising variants gave the same answer: one standard deviation of genetically higher HDL carried an odds ratio of 0.93 (0.68–1.26), P = 0.63, against 0.62 (0.58–0.66) in the observational data.
The reason this study is persuasive rather than merely nihilistic is its positive control. The same method, applied to a genetic score for LDL cholesterol, reproduced the observational association almost exactly: one standard deviation of higher LDL cholesterol carried an observational odds ratio of 1.54 (1.45–1.63) and a genetic odds ratio of 2.13 (1.69–2.69). The technique can detect a causal lipid. It simply did not detect one here.
Niacin moved three numbers at once
Niacin is the rare drug that improves nearly every line of a lipid panel simultaneously, which made it the obvious way to test whether improving the panel improves the patient.
AIM-HIGH randomised 3,414 patients with established cardiovascular disease and low HDL to extended-release niacin or placebo, on top of simvastatin with ezetimibe added as needed to hold LDL cholesterol between 40 and 80 mg/dL.[9] At two years, niacin had raised median HDL cholesterol from 35 to 42 mg/dL, lowered triglycerides from 164 to 122 mg/dL, and lowered LDL cholesterol from 74 to 62 mg/dL. The primary endpoint occurred in 16.4% on niacin and 16.2% on placebo: hazard ratio 1.02 (0.87–1.21), P = 0.79. The trial was stopped after a mean of three years for lack of efficacy.
HPS2-THRIVE was larger and more definitive: 25,673 adults with vascular disease, randomised to extended-release niacin with laropiprant or placebo, median follow-up 3.9 years.[10] LDL cholesterol was 10 mg/dL lower and HDL cholesterol 6 mg/dL higher than placebo. Major vascular events occurred in 13.2% against 13.7%: rate ratio 0.96 (0.90–1.03), P = 0.29.
That is the null result. The part that matters more is the harm column, and it is the reason this trial is worth a figure of its own. Niacin-laropiprant produced significant absolute excesses of serious disturbances in diabetes control (3.7 percentage points), new diagnoses of diabetes (1.3), gastrointestinal serious adverse events (1.0), musculoskeletal (0.7), skin (0.3), and — the authors' own word is "unexpectedly" — infection (1.4) and bleeding (0.7).
"No benefit" and "no harm" are different findings. A drug with a null efficacy result and a populated harm column is not neutral. It is a net negative.
HPS2-THRIVE: no benefit, and not no harm
Absolute excess of serious adverse events on extended-release niacin with laropiprant compared with placebo, in percentage points, among 25,673 adults followed a median of 3.9 years. Every excess shown reached statistical significance in the trial. Over the same period the primary endpoint — major vascular events — showed no significant difference at all: 13.2% against 13.7%, rate ratio 0.96 (0.90–1.03). The infection and bleeding excesses were described by the investigators as unexpected.
Source: reference 10.
Homocysteine, and what chance looks like
Higher homocysteine predicts coronary disease and stroke, and folic acid with vitamins B6 and B12 reliably lowers it. HOPE-2, led from McMaster University in Hamilton, randomised 5,522 patients aged 55 or older with vascular disease or diabetes to five years of that combination or placebo.[11]
The intervention worked on the biomarker: mean homocysteine fell 2.4 µmol/L on treatment and rose 0.8 µmol/L on placebo. The primary composite of cardiovascular death, myocardial infarction and stroke occurred in 18.8% against 19.8%: relative risk 0.95 (0.84–1.07), P = 0.41.
Two secondary outcomes are worth printing, precisely because they point in opposite directions. Fewer patients on active treatment had a stroke (relative risk 0.75, 0.59–0.97). More were hospitalised for unstable angina (relative risk 1.24, 1.04–1.49). Neither was the primary endpoint; both are nominally significant. Two nominally significant secondary results running in opposite directions, in a trial whose primary endpoint was null, is close to a textbook illustration of what chance looks like when you measure many things. The honest reading is to report both and conclude neither.
Triglycerides, and the number that didn't move
The newest and cleanest case is PROMINENT, published in 2022.[12] It enrolled 10,497 patients with type 2 diabetes, triglycerides between 200 and 499 mg/dL, HDL cholesterol at or below 40 mg/dL and LDL cholesterol at or below 100 mg/dL, and randomised them to pemafibrate or placebo for a median of 3.4 years.
At four months, pemafibrate had lowered triglycerides by 26.2%, VLDL cholesterol by 25.8%, remnant cholesterol by 25.6% and apolipoprotein C-III by 27.6%. The primary endpoint occurred in 572 patients on pemafibrate and 560 on placebo: hazard ratio 1.03 (0.91–1.15). Adverse renal events and venous thromboembolism were more common on the drug; non-alcoholic fatty liver disease was less common.
And then there is the one lipid number in that list that moved the wrong way. Apolipoprotein B rose 4.8%.
ApoB counts the atherogenic particles rather than the cholesterol inside them, and it is the lipid measure that Mendelian randomisation says actually carries the risk. In an analysis of 63 studies, 654,783 participants and 91,129 cases of coronary heart disease, Ference and colleagues compared a triglyceride-lowering genetic score in the LPL gene against an LDL-lowering score in the LDLR gene.[13] Per 10 mg/dL lower apoB, the LPL score corresponded to 69.9 mg/dL lower triglycerides and 0.7 mg/dL higher LDL cholesterol; the LDLR score to 14.2 mg/dL lower LDL cholesterol and 1.9 mg/dL lower triglycerides. Completely different lipid panels — and almost identical risk reductions: odds ratio 0.771 (0.741–0.802) and 0.773 (0.747–0.801). When all three were modelled together, the associations of triglycerides (1.014, 0.965–1.065) and LDL cholesterol (1.010, 0.967–1.055) with coronary disease became null, while apoB remained at 0.761 (0.723–0.798).
So in PROMINENT, the surrogate that failed and the surrogate that works were measured in the same patients at the same visit. Four lipid measures improved by about a quarter each; apoB, the one with the strongest causal credentials, went slightly the wrong way; and the outcome did not move. That is about as direct a demonstration as this literature offers that it is not the number on the report that matters, but which number.
This is not only a cholesterol problem
Cardiology has the best-documented examples because its outcomes are countable and its trials are enormous. But the structure is general, and oncology provides the broadest audit of it.
The US Food and Drug Administration's accelerated approval pathway lets a cancer drug reach patients on the strength of a surrogate measure — tumour shrinkage, or progression-free survival — with clinical benefit to be confirmed afterwards. A 2025 analysis in the Journal of the National Cancer Institute examined every cancer indication that had been converted from accelerated to regular approval between 1992 and 2021, following them through December 2024.[14]
Of 77 confirmatory trials, 25 (32%) showed an overall survival benefit at the point of conversion to full approval. The other 52 (68%) relied on surrogate measures. After a median of five further years of follow-up, a survival benefit emerged for 7 more indications (9%). Quality-of-life benefit was demonstrated in 9 confirmatory trials (12%) at conversion and 4 (5%) afterwards.
This is not an argument that those drugs do not work. Many clearly do, and 47% of the scoreable trials met the European Society for Medical Oncology's threshold for substantial clinical benefit. It is an argument that the gap between "the number improved" and "the patient lived longer or better" is a real gap, that it stays open for years, and that it does not close on its own.
What a surrogate that holds up looks like
It would be easy to read all of this as a case against treating numbers at all. That would be the wrong lesson, and the evidence refutes it directly.
LDL cholesterol is a surrogate endpoint in exactly the same technical sense as HDL cholesterol or homocysteine. The difference is that when it was tested, it held. The Cholesterol Treatment Trialists' Collaboration pooled individual participant data from 26 randomised trials in roughly 170,000 people and found that each 1.0 mmol/L reduction in LDL cholesterol was associated with a 22% lower rate of major vascular events (rate ratio 0.78, 0.76–0.80) and a 10% lower rate of death from any cause (0.90, 0.87–0.93).[15] The proportional benefit was the same in every type of patient studied, including those who started with an LDL cholesterol below 2 mmol/L, and the relationship held across the whole range tested, with no threshold.
That is what a validated target looks like: a dose–response relationship, reproduced across dozens of trials in different populations, with genetic evidence agreeing, and the size of the benefit matching the size of the change.
None of the failures above were failures of science. They were science. The only way anyone could know that raising HDL cholesterol by 133% does nothing was to raise it by 133% in 12,092 people and count the events. The alternative — treating everyone on the strength of a plausible mechanism — is exactly what CAST was doing in 1991, and it is what killed those patients. The trials are the correction mechanism, and they worked.
Five questions to ask about any number
Next time a supplement, a drug or a clinic promises to move one of your biomarkers, these five questions do most of the work. They are the same ones a trialist would ask.
- Has the number been moved by something, in a randomised trial, with hard outcomes counted? Or has it only ever been measured and correlated? An association, however strong and however often replicated, has never once been sufficient.
- Did the thing that actually changed track the thing that matters? REVEAL's benefit tracked its 18% fall in non-HDL cholesterol, not its 104% rise in HDL. Look past the number being advertised to the rest of the panel.
- Does genetics agree the marker is causal? Mendelian randomisation is not perfect, but it is a lifelong randomisation that confounding cannot easily explain. For LDL cholesterol and apoB it says yes. For HDL cholesterol it says no.
- Is the proposed mechanism the only thing the intervention does? Torcetrapib also raised blood pressure 5.4 mm Hg and disturbed aldosterone. Few interventions are as clean as their mechanism diagram.
- Is anyone reporting the harm column? HPS2-THRIVE's seven significant excesses only look important once you accept that the benefit column is empty.
And one more worth holding on to: a null outcome trial does not retire a risk marker. Lp(a) still predicts cardiovascular risk, and so does a low HDL cholesterol. What those trials established is narrower and more useful — that the job shifts from treating the number to treating everything the number is telling you about. If you want the practical version of that argument, it is in our piece on longevity and cardiovascular risk and in the cardiovascular biomarker database.
The practical bottom line
- A biomarker earns the status of "target" only by being moved in a randomised trial that then counts hard outcomes. Until that has happened, it is a predictor, and predictors are for estimating risk, not for aiming drugs at.
- Treat a dramatic percentage change in a biomarker as a claim about pharmacology, not about health. Evacetrapib's 133% is a real number about a real molecule. It told us nothing about heart attacks.
- When you are shown a benefit, ask which measured thing it tracked. The advertised number and the responsible number are often not the same one.
- A null efficacy result with a populated harm column is a negative result, not a neutral one.
- A failed drug does not invalidate the marker. Lp(a), homocysteine and low HDL all still predict risk. They just are not, on current evidence, the lever.
- None of this is a reason to stop a medication you are taking, or to start one. It is a way of reading the claims made about new ones.
Frequently asked questions
What exactly is a surrogate endpoint?
It is a measurement used in place of the outcome you actually care about — a lab value, an imaging finding or a short-term event standing in for heart attack, disability or death. Surrogates make trials faster and cheaper because they move sooner and more often. They are legitimate and often necessary. The problem is that a surrogate is only as good as the assumption that moving it moves the real outcome, and that assumption has to be tested rather than reasoned.
Does this mean my cholesterol results don't matter?
No — and LDL cholesterol in particular is the clearest example of a surrogate that was tested and held up, across 26 randomised trials in roughly 170,000 people. Risk prediction is a genuinely useful job for a biomarker, and it is a different job from being a treatment target. A number can be excellent at the first and useless at the second.
Why did raising HDL cholesterol not work when low HDL predicts heart disease so reliably?
The most likely explanation is that low HDL cholesterol is a marker of other things — insulin resistance, high triglycerides, a worse metabolic picture overall — rather than a cause of atherosclerosis in its own right. The genetic evidence points the same way: people born with naturally higher HDL cholesterol do not have fewer heart attacks.
Is niacin harmful? Should anyone be taking it?
HPS2-THRIVE tested extended-release niacin combined with laropiprant at 2 g daily, on top of statin therapy, and found no reduction in vascular events alongside significant excesses of several serious adverse events. That is evidence about that formulation at that dose in that population. It is not a statement about dietary niacin, which is an essential vitamin, and it is not advice about any prescription you are currently taking — that is a conversation for the doctor who prescribed it.
Did Lp(a)HORIZON prove that lipoprotein(a) doesn't matter?
No. It showed that this drug, at this dose, in people who already had established cardiovascular disease, did not reduce events over 4.7 years despite lowering Lp(a) substantially. Lp(a) remains an independent predictor of cardiovascular risk with supporting genetic evidence. Several other Lp(a)-lowering drugs are in ongoing outcome trials, and some are testing earlier-stage patients.
How do I apply this to a supplement that claims to improve a biomarker?
Ask whether anyone has run a randomised trial of that supplement with hard outcomes, and what it found. Most supplement evidence stops at the biomarker, which is exactly the stage at which torcetrapib, dalcetrapib, evacetrapib, niacin, B vitamins, pemafibrate and pelacarsen all still looked promising.
References
- [1]Cho L, Nicholls SJ, Frndak S, et al. Pelacarsen and Cardiovascular Outcomes in Patients with High Lipoprotein(a). N Engl J Med. 2026.
- [2]Echt DS, Liebson PR, Mitchell LB, et al. Mortality and morbidity in patients receiving encainide, flecainide, or placebo. The Cardiac Arrhythmia Suppression Trial. N Engl J Med. 1991;324(12):781–788.
- [3]Fleming TR, DeMets DL. Surrogate end points in clinical trials: are we being misled? Ann Intern Med. 1996;125(7):605–613.
- [4]Voight BF, Peloso GM, Orho-Melander M, et al. Plasma HDL cholesterol and risk of myocardial infarction: a mendelian randomisation study. Lancet. 2012;380(9841):572–580.
- [5]Barter PJ, Caulfield M, Eriksson M, et al. Effects of torcetrapib in patients at high risk for coronary events. N Engl J Med. 2007;357(21):2109–2122.
- [6]Schwartz GG, Olsson AG, Abt M, et al. Effects of dalcetrapib in patients with a recent acute coronary syndrome. N Engl J Med. 2012;367(22):2089–2099.
- [7]Lincoff AM, Nicholls SJ, Riesmeyer JS, et al. Evacetrapib and cardiovascular outcomes in high-risk vascular disease. N Engl J Med. 2017;376(20):1933–1942.
- [8]HPS3/TIMI55–REVEAL Collaborative Group. Effects of anacetrapib in patients with atherosclerotic vascular disease. N Engl J Med. 2017;377(13):1217–1227.
- [9]AIM-HIGH Investigators. Niacin in patients with low HDL cholesterol levels receiving intensive statin therapy. N Engl J Med. 2011;365(24):2255–2267.
- [10]HPS2-THRIVE Collaborative Group. Effects of extended-release niacin with laropiprant in high-risk patients. N Engl J Med. 2014;371(3):203–212.
- [11]Lonn E, Yusuf S, Arnold MJ, et al. Homocysteine lowering with folic acid and B vitamins in vascular disease (HOPE-2). N Engl J Med. 2006;354(15):1567–1577.
- [12]Das Pradhan A, Glynn RJ, Fruchart JC, et al. Triglyceride lowering with pemafibrate to reduce cardiovascular risk (PROMINENT). N Engl J Med. 2022;387(21):1923–1934.
- [13]Ference BA, Kastelein JJP, Ray KK, et al. Association of triglyceride-lowering LPL variants and LDL-C-lowering LDLR variants with risk of coronary heart disease. JAMA. 2019;321(4):364–373.
- [14]Tibau A, Romano A, Liu ITT, et al. Assessing outcomes emerging after conversion to regular approval for cancer drug indications granted accelerated approval, 1992–2021. J Natl Cancer Inst. 2025;117(10):2103–2111.
- [15]Cholesterol Treatment Trialists' (CTT) Collaboration. Efficacy and safety of more intensive lowering of LDL cholesterol: a meta-analysis of data from 170,000 participants in 26 randomised trials. Lancet. 2010;376(9753):1670–1681.
