A post hoc analysis of the SELECT phase 3 trial found semaglutide 2.4 mg slowed progression of a validated 25-protein proteomic signature of dementia risk in 2,970 adults aged 65 years and older with overweight, obesity, and cardiovascular disease but no diabetes. Over 104 weeks, predicted 5-year…
A post hoc analysis of the phase 3 SELECT trial has found that semaglutide 2.4 mg slowed progression of a validated proteomics-based dementia risk signature in adults 65 years and older with overweight or obesity, cardiovascular disease, and no diabetes. The analysis included 2,970 participants and compared changes in a 25-protein dementia risk score between non-fasted serum samples collected at baseline and at week 104. Semaglutide was associated with a 2.5-fold smaller increase in 5-year predicted dementia risk than placebo, a 26.0% lower predicted 5-year dementia event rate OR 0.74, 95% CI 0.65 to 0.85 , and a 36% reduction in the odds of being classified into a higher dementia risk category β -0.44; P < 0.001 .
This is the first evidence from a large phase 3 randomized trial that a GLP-1 receptor agonist attenuates changes in a validated proteomics-based dementia risk score. Semaglutide is a peptide-based GLP-1 receptor agonist, and the finding extends earlier preclinical and clinical signals of GLP-1 receptor agonist neuroprotection into a randomized, placebo-controlled setting, using a molecular readout rather than a clinical diagnosis of dementia. Those earlier signals came from cellular and animal models of neurodegeneration, from observational cohorts, and from small mechanistic studies; none carried the randomization, the sample size, or the two-year placebo comparison of a phase 3 cardiovascular outcomes trial.
The result is not a demonstration that semaglutide prevents dementia. The analysis is post hoc, the outcome is a proteomic surrogate, and the samples were non-fasted. But the size, direction, and statistical strength of the effect justify a close reading of what the study can and cannot show, and of what it implies for a semaglutide research program already reaching well beyond diabetes and obesity.
The analysis is also an example of method. A validated proteomic surrogate made it possible to ask a neurological question of a trial that was never designed to answer one, at a cost limited to the assay itself. Two years is a short interval measured against the decades-long preclinical phase of dementia. A drug effect on a circulating risk signature detectable within that window, in a cohort not selected for cognitive complaints, is the main reason the analysis matters. Randomization is what gives that signal its purchase. An observational study could not exclude the possibility that the protein signature and the drug were independently tied to some third factor, such as overall metabolic health. In a randomized comparison, allocation to semaglutide or placebo could not have been influenced by any participant characteristic, including the proteins measured at baseline. That is also the reason the result must be read as a biological signal rather than a clinical promise.
Plasma proteomics can detect multi-pathway biological changes that precede dementia onset, and the Dementia SomaSignal Test dSST is built on that premise. The dSST is a validated score composed of 25 proteins that predicts 5-year and 20-year all-cause dementia risk. Rather than tracking a single disease pathway, the score integrates proteins across biological systems to produce a risk estimate. A validated score of this kind is calibrated against external cohorts, which is what gives its two prediction horizons meaning, and its 25 components are selected to capture biological drift that precedes clinical disease.
The word "dementia" in the test's name covers the full diagnostic spectrum: Alzheimer's disease, vascular dementia, and mixed pathologies. That breadth matters for this cohort. The participants in this analysis all had established cardiovascular disease, a population in whom vascular injury is a plausible contributor to dementia risk. A score that captured only amyloid-related biology would be a poor fit for such a group; a 25-protein composite spanning inflammation, metabolism, and vascular function is better matched to it.
In this analysis, non-fasted serum samples drawn at baseline and at week 104 were measured on the dSST, and the change in predicted risk over that 104-week interval was compared between randomized groups. For the 5-year prediction horizon, predicted dementia risk rose in both groups over the two years. But the increase was 2.5-fold smaller with semaglutide than with placebo. That corresponds to a 26.0% lower predicted 5-year dementia event rate, with an odds ratio of 0.74 95% CI 0.65 to 0.85 . The 20-year horizon showed the same direction at a smaller absolute magnitude: a 1.67-fold smaller increase in predicted dementia risk and an 8.8% lower predicted 20-year event rate OR 0.91, 95% CI 0.88 to 0.94 .
An odds ratio of 0.74 is not the same quantity as a 26.0% lower event rate. The first is a ratio of odds; the second is a model-derived probability. They agree in direction, which is the point. Both are transformations of the same underlying comparison, and both are generated by the dSST's calibration rather than by observed dementia events. The third endpoint was categorical rather than continuous. The dSST assigns participants to dementia risk categories, and semaglutide reduced the odds of being classified into a higher risk category by 36% , with an effect estimate of β -0.44 and P < 0.001 . The two continuous trajectories and the categorical shift all point the same way: among these participants, the proteomic dementia risk signature progressed more slowly on semaglutide than on placebo.
Two properties of the data are worth weighing. The effect appeared at both prediction horizons, which argues against an artifact of a single scoring window. The relative effect was larger at 5 years than at 20 years, which is what would be expected if semaglutide slowed the rate of change in the signature rather than shifting baseline risk. A drug that slows an ongoing biological drift will show its largest relative effect over the short horizon; the longer window dilutes it. By the same logic, the 20-year odds ratio sits closer to the null because a 104-week treatment window is a smaller fraction of a 20-year risk horizon than of a 5-year one.
The 20-year horizon deserves a separate caution. Two years of follow-up cannot directly observe a 20-year outcome. The longer prediction is an extrapolation from the model's calibration in external cohorts, and it assumes that the relationship between the 25 proteins and long-term dementia risk is stable across time and across populations. That assumption is testable in principle, but not in these data. The design does, however, strengthen the measurement in one respect. The endpoint was the change in predicted risk between baseline and week 104, so each participant served as their own baseline control. Within-person change scores absorb much of the between-person variation in absolute protein concentrations, a useful property when samples are non-fasted. That does not, however, convert a surrogate measurement into a clinical one.
SELECT was a phase 3 randomized trial in which participants received semaglutide 2.4 mg or placebo. The subpopulation analyzed here consisted of adults aged 65 years or older with overweight or obesity, established cardiovascular disease, and no diabetes. The 2,970 participants with paired samples are a subset of the full SELECT cohort, and the dementia analysis was not a prespecified primary or secondary endpoint of the trial.
Post hoc status changes how much weight the result should carry. The comparison retains the central strength of randomization: allocation to semaglutide or placebo could not have been influenced by any characteristic of the participants, including their dementia risk. Randomization balances measured and unmeasured confounders at baseline, and for that reason a post hoc analysis of a randomized trial sits well above an observational biomarker study in the causal hierarchy. But a post hoc analysis planned after the trial was unblinded carries a higher risk of selective reporting, and the reported confidence intervals and P value do not account for the multiple analyses typically run on such a dataset. This is a strong hypothesis-generating signal, not confirmatory evidence.
The subset itself introduces a further question. Participants were included because they had usable serum samples at both time points, and the possibility that sample availability was related to participant characteristics cannot be dismissed from the data as reported here. Any nonrandom selection within a randomized trial can, in principle, disturb the balance that randomization created. The analysis would need to demonstrate that the two arms remained balanced within the analyzed subset for the comparison to retain its full force. The fact that this is a tractable check does not make the result depend on it, but it is a reason to weigh the consistency of the three endpoints more heavily than the exact P value.
The endpoint is the second structural limit. The dSST is a validated prognostic test: it predicts the probability of future dementia from protein levels measured in blood. But a biomarker that predicts risk in an untreated population is not automatically a valid surrogate for treatment effect. Surrogate validity requires evidence that a change in the biomarker under a specific intervention tracks a change in the clinical outcome, and that evidence does not yet exist for the dSST and semaglutide. What the analysis demonstrates is a drug effect on a molecular signature, not a drug effect on dementia diagnoses.
The non-fasted sampling deserves a closer look. Postprandial state can shift the concentration of metabolically regulated plasma proteins, adding noise to the measurement. Noise of that kind usually blunts a signal rather than creates one, which suggests the observed effect may be conservative. There is also a theoretical concern that semaglutide, by reducing food intake, could alter the postprandial state at the time of sampling and thereby introduce a systematic difference between groups that is not a direct effect on the dementia signature. The analysis cannot rule that out, and the direction of any such bias is unknown.
Finally, the population limits the reach of the finding. These were adults 65 years and older with overweight or obesity, cardiovascular disease, and no diabetes. How the result applies to younger adults, to people with diabetes, or to people without established cardiovascular disease cannot be inferred from this analysis.
Dementia prevention is a difficult endpoint for clinical trials because the disease develops over decades and the earliest molecular changes precede symptoms by years. A trial built on adjudicated dementia diagnoses requires a large cohort, very long follow-up, and a high cost per recorded event. Within a two-year window, the number of incident dementia cases in a cohort of this size would be small, which is why the SELECT analysis could not have used a clinical endpoint even if one had been planned. That is precisely where a molecular signature that responds within two years becomes attractive: it can rank candidate interventions before the far more expensive commitment to a clinical-endpoint trial.
The SELECT analysis is an example of that logic applied to an existing trial. The dSST required only non-fasted serum samples at two time points, and those samples were already being collected for the cardiovascular outcomes program. Reanalysis of banked material produced a randomized, placebo-controlled estimate of the drug's effect on a dementia-related molecular signature at essentially no additional trial cost. The analysis contains no information about cognitive status, because the outcome was…
Peptides referenced: Semaglutide, GLP-1.
Related reading: Four peptide leads target Klebsiella DnaA in an in-silico screen, Semaglutide Solid Lipid Nanoparticles Boost Oral Permeability Sixfold, Peptide self-assembly death pathways classified into five themes, GLP-1 Initiation Tied to Fewer Coded GI Symptoms in IBS.