September 16, 2026
Sometimes a drug does exactly what it was designed to do and still fails.
That uncomfortable reality has played out dramatically over the past several weeks. On September 4, Novartis announced that pelacarsen, an investigational antisense oligonucleotide targeting lipoprotein(a), or Lp(a), had failed to reduce cardiovascular events in the Phase III Lp(a)HORIZON trial. The result was striking because pelacarsen did lower Lp(a) and engaged its intended target. What it did not do, in the overall study population, was significantly reduce the composite risk of cardiovascular death, non-fatal myocardial infarction, non-fatal stroke or urgent coronary revascularization [1].
This was not a small or poorly conceived experiment. Lp(a) is an inherited cardiovascular risk factor affecting approximately one in five people worldwide, and earlier studies using targeted therapy achieved impressive reductions in this cholesterol-carrying protein. In a Phase II trial, pelacarsen reduced Lp(a) concentrations by as much as 80% with one of the dosing regimens evaluated [2]. Lp(a)HORIZON then enrolled more than 8,300 people with established cardiovascular disease to ask the question that ultimately mattered: would lowering Lp(a) translate into fewer cardiovascular events [1]?
The answer, at least from the topline results we have so far, was no.
Only a few weeks earlier, another large cardiovascular trial produced a remarkably similar pattern. ZEUS enrolled more than 6,300 people with atherosclerotic cardiovascular disease, chronic kidney disease and elevated high-sensitivity C-reactive protein (hsCRP). Ziltivekimab, an antibody targeting the inflammatory cytokine interleukin-6 (IL-6), produced the expected biological effects: free IL-6 and hsCRP fell significantly. Yet there was essentially no difference in major adverse cardiovascular events compared with placebo: hazard ratio 0.99, with a 95% confidence interval of 0.88 to 1.11 [3].
Two sophisticated therapies. Two biologically plausible targets. Clear evidence that the intended biology moved. But no corresponding improvement in the primary clinical outcome.
It is tempting to look at results like these and conclude that biomarkers are unreliable, or that we should simply stop trusting surrogate measures and focus on hard clinical outcomes. But this simplistic take misses the more interesting lesson: different measurements answer different questions. Did the drug reach its target? Did the target change? Did the disease biology change? Did the tissue function differently? Did the patient feel better, function better or live longer? Those questions sit along the same biological path, but they are not interchangeable. A positive answer to one question does not guarantee a positive answer to the next.
Lp(a) fell substantially with pelacarsen, and that result is not invalidated because the cardiovascular endpoint was missed. The drug changed exactly what it was designed to change. What the trial tells us is that lowering that particular measure, in that population, with that intervention and over that timeframe, did not translate into the expected reduction in cardiovascular events.
There are many possible reasons. The magnitude or duration of lowering may matter. Modern background cardiovascular therapy may have reduced residual risk in the studied population. Different components of Lp(a) biology may matter more than the concentration we currently measure. The trial population may have lacked the relevant pathophysiology, or may have been too far down the road to cardiovascular disease for Lp(a) lowering to reduce disease burden, with the damage already too advanced. Or the biological model connecting Lp(a) lowering to cardiovascular events may simply be incomplete. These uncertainties are precisely why outcome trials exist.
A surrogate endpoint is a stand-in measurement, like blood pressure or cholesterol levels, used in a clinical trial because it’s easier or faster to measure than the real outcome researchers care about, such as a heart attack or death.
Surrogates rest on the assumption that changing the stand-in will also change the real outcome. They are extraordinarily useful because waiting for mortality, disability or major disease progression in every study would make many trials prohibitively long, large or expensive. The FDA maintains a formal table of surrogate endpoints that have supported both traditional and accelerated drug approvals, reflecting just how embedded these measures are in modern development [4].
But the evidentiary strength behind different surrogates varies considerably. A 2024 JAMA analysis examined 37 surrogate markers used as primary endpoints across 32 non-oncologic (noncancer) chronic diseases. For 22 of those 37 markers (59%), the researchers could not identify a published meta-analysis evaluating the relationship between treatment effects on both the surrogate and a clinical outcome. Even among the markers for which evidence existed, relatively few surrogate-clinical outcome relationships demonstrated high-strength associations [5].
Oncology provides an even starker illustration. Researchers examined 46 cancer indications that had received accelerated approval between 2013 and 2017 and were followed for more than five years. Of the 46 approved drugs, fewer than half (20/46, or 43%) ultimately demonstrated improvement in overall survival or quality of life in confirmatory studies. Ten indications were withdrawn, and seven still lacked a definitive regulatory outcome at the time of analysis [6].
This is not an argument against surrogate endpoints, which in many cases have allowed valuable therapies to reach patients years earlier. It is an argument for greater precision in validating and reporting the clinical outcomes a surrogate has actually mapped to.
A response rate is not overall survival. A molecular change is not restored function. Target engagement is not automatically disease modification.
There is another conclusion we should resist: that the most downstream clinical measure must always be the best measure. Sometimes waiting for the endpoint we care about means waiting until considerable biological damage has already occurred.
A biomarker can reveal that a disease pathway is changing before the patient experiences overt functional decline. It can identify which biological process is abnormal, tell us whether a therapy is reaching the intended tissue, help enrich a trial for patients most likely to respond, and potentially tell us whether an intervention is working while there is still something left to preserve. These are different jobs from proving clinical benefit, not lesser ones. The problem begins when we ask one measure to do all of them.
Clinical trial interpretation becomes particularly difficult when those layers disagree. A primary endpoint may fail while secondary measures move. A molecular marker may show dramatic improvement without a corresponding functional effect (as we saw with the two trials mentioned above). A functional endpoint may remain stable while underlying disease biology continues to deteriorate.
The temptation after a disappointing result is to search the dataset for a more favourable story. That is where scientific discipline matters most: a missed primary endpoint cannot simply be replaced after the fact by a biomarker or secondary outcome that happened to move.
At the same time, the biomarker should not be discarded simply because the trial was negative. It may be telling us something important about where the therapeutic chain broke. Perhaps the target was wrong. Perhaps the marker was too far removed from the disease process that ultimately drives the outcome. Perhaps the intervention came too late. Perhaps we measured the wrong patients. Or perhaps we measured the right biology for the wrong question.
That last possibility deserves much more attention. Before asking whether a biomarker predicts the endpoint, we should probably ask something even more basic: are we actually measuring the disease process we think we are measuring? That is where Part 2 of this series takes over (stay tuned!).
References