A post hoc analysis of the phase 3 SELECT trial found that once-weekly semaglutide attenuated a 25-protein blood-based dementia risk signature in 2,970 adults aged 65 and older with cardiovascular disease and overweight or obesity. Predicted 5-year dementia event rates were 26.0% lower and 20-year…
Semaglutide slowed the worsening of a machine learning-derived dementia risk signature based on 25 blood proteins in older adults with cardiovascular disease and overweight or obesity, according to a post hoc analysis of the phase 3 SELECT trial. The analysis, published in 2026 in Alzheimer's & Dementia: Diagnosis, Assessment & Disease Monitoring DOI 10.1002/dad2.70432 , was led by first author Mart Jimnez-Mausbach. It examined 2,970 SELECT participants aged 65 years or older who were treated for 104 weeks with once-weekly semaglutide or placebo. In the semaglutide group, the magnitude of increase in the proteomics-predicted 5-year dementia risk was 2.5-fold lower than in the placebo group, corresponding to a 26.0% lower predicted 5-year dementia event rate.
The study was funded by Novo Nordisk, which manufactures semaglutide and participated in study design, data collection, analysis, and manuscript review. Illumina collaborated on the statistical analysis, and Amsterdam UMC was a research collaborator. News-Medical published an assessment of the journal analysis on August 10, 2026; the article was written by Dr. Sanchari Sinha Dutta and reviewed by Lauren Hardaker.
The analysis used the Dementia SomaSignal Test dSST , a machine learning-derived blood plasma risk score that integrates 25 proteins into a predicted probability of dementia over 5-year and 20-year horizons. Because SELECT enrolled people without diabetes, the result suggests that GLP-1 receptor agonist treatment can affect pathways relevant to brain health in a cardiometabolic population independent of glycemic control.
The effect on the longer horizon was smaller but still present: a 1.67-fold lower magnitude of increase in the proteomics-predicted 20-year dementia risk, corresponding to an 8.8% lower predicted 20-year dementia event rate. At week 104, semaglutide was associated with 36% lower odds of being classified into a higher dSST dementia risk category than placebo.
The analysis is a surrogate-signal result, not a clinical one. It does not show fewer dementia diagnoses, slower cognitive decline, or a change in any adjudicated neurological outcome. What it shows is that 104 weeks of a peptide-based metabolic drug moved a molecular readout associated with future dementia risk, and moved it in the direction that would be expected if the drug were protecting brain-relevant biology.
The dSST converts measurements of 25 plasma proteins into a predicted probability of dementia within 5 years and within 20 years. The score is not a diagnostic test. It is a statistical construct: a machine learning model that maps the relative abundance of those proteins onto a forecast of later dementia outcomes, developed and validated on EDTA-plasma samples. Applying it to the SELECT samples turns stored serum into a longitudinal readout of how each participant's dementia-related proteomic profile changed during two years of treatment.
In this post hoc analysis, serum samples collected at baseline and at week 104 from the 2,970 participants aged 65 and older were run through the signature, and the change in predicted risk between the two time points was compared across treatment groups. The randomized contrast is preserved. Because SELECT assigned treatment in a double-blind fashion, the week 104 measurements should not carry the selection biases that typically distort observational biomarker studies, even though the analysis of those measurements came after the trial ended.
The signature carried prognostic signal in this cohort. All 18 participants who experienced dementia-related adverse events during the study had baseline dSST risk categories of high or medium-to-high, and none were low risk at baseline. That internal check matters. It shows that a machine learning score trained elsewhere still ranked participants in the expected order when applied to these stored samples.
The composition of the signature explains why a cardiometabolic drug could move it. 76% of the dementia-associated dSST proteins have established roles in cardiometabolic disease, spanning inflammation, lipid biology, and vascular function. Semaglutide already modifies several of those systems, so a dementia signature enriched for cardiometabolic proteins is a mechanistically plausible target even without a direct effect on amyloid or tau.
The key results:
That last result is the strongest evidence of a weight-independent mechanism. If the effect on the dementia signature were simply a downstream consequence of fat loss, most of the association should have vanished after adjusting for BMI. It did not. The 28% that did vanish is itself informative: it places a floor under how much of the observed change could be secondary to weight reduction, and it implies that weight loss and direct pharmacological effects both contribute.
The divergence between the two horizons is one of the most instructive patterns in the data. A 104-week intervention attenuated the 5-year signature 2.5-fold but the 20-year signature only 1.67-fold. Short-horizon predictions in proteomic risk models tend to be weighted toward proteins that track current physiological state, including inflammatory and vascular markers that can shift within months. Long-horizon predictions are weighted toward cumulative processes that build over decades. A treatment that improves the first set of proteins will compress short-horizon risk quickly, while long-horizon risk, reflecting years of accumulated pathology, will move more slowly.
The parent SELECT trial was a randomized, double-blind, placebo-controlled study of once-weekly semaglutide in 17,604 obese or overweight adults aged 45 years or older with established atherosclerotic cardiovascular disease and without diabetes. The target semaglutide dose was 2.4 mg, titrated over 16 weeks. SELECT was designed for cardiovascular outcomes. This post hoc analysis took the 2,970 participants aged 65 years or above, measured their dSST-predicted dementia risk from baseline serum samples, treated them for 104 weeks, and measured the signature again.
What the design can and cannot show is defined by the trial's original purpose. SELECT was not optimized for patient selection or data collection around cognitive health. It did not collect standardized cognitive assessments, and the dementia-related adverse events recorded during the study, 18 in total, were not systematically adjudicated. The proteomic analysis was post hoc, and no prespecified multiple-testing adjustment was applied to these endpoints. That matters because post hoc analyses multiply the number of comparisons made on a single dataset. Without adjustment, some results will look significant by chance. The effect sizes here are large and internally consistent, which argues against pure noise, but the concern cannot be dismissed.
The outcome measured here is a biomarker-based predicted risk score, not a standardized cognitive test result and not a confirmed dementia diagnosis. The analysis therefore establishes that semaglutide shifted a proteomic dementia-risk surrogate favorably over two years in this population. It does not establish that the shift predicts fewer dementia diagnoses or slower cognitive decline in individual patients. That distinction places the finding firmly in hypothesis-generating territory.
A technical caveat deserves emphasis. The dSST was developed and validated using EDTA-plasma samples, whereas the SELECT analysis used serum. Serum and plasma differ in preparation: clotting consumes some proteins, activates others, and exposes samples to proteolytic activity that plasma collection avoids. The authors report high concordance between the sample types, and the fact that baseline risk categories ordered the 18 dementia-related adverse events correctly suggests the measurements retained prognostic information. Still, formal cross-validation of the dSST in serum across independent cohorts would strengthen confidence in any attempt to replicate this result.
Semaglutide is a peptide-based glucagon-like peptide-1 GLP-1 receptor agonist . Its canonical actions are incretin-mediated: glucose-dependent insulin secretion, suppressed glucagon release, delayed gastric emptying, and reduced appetite. But the receptor system is more distributed than the metabolic story implies. GLP-1 receptors are expressed in the central nervous system, including the hypothalamus, brainstem, and hippocampus, regions that regulate appetite and cognition. A synthetic peptide of this class does not necessarily need high blood-brain barrier penetration to reach those circuits. Signaling through vagal afferents and access to circumventricular organs, where the barrier is fenestrated, are both routes by which peripheral GLP-1 receptor activation can influence the brain.
In the vasculature, GLP-1 receptor activation improves endothelial function, reduces expression of adhesion molecules, and lowers circulating inflammatory markers. Those are the same biological systems heavily represented in the dSST protein panel. Chronic low-grade inflammation, insulin resistance, endothelial dysfunction, and atherosclerosis contribute to both cardiovascular disease and neurodegeneration. A drug that improves those systems could plausibly move a dementia risk signature without engaging amyloid or tau directly.
The contrast between the two prediction horizons fits that biology. The 5-year signature moved more than the 20-year signature, consistent with short-horizon risk tracking modifiable inflammation and vascular function while long-horizon risk reflects decades of accumulated pathology. The persistence of roughly 72% of the 20-year association after BMI adjustment reinforces the interpretation that weight loss contributes but does not fully explain the effect.
Adjusting for BMI is a blunt instrument, and the 28% attenuation figure should be read with that in mind. BMI captures only part of the metabolic consequences of obesity; it says nothing about fat distribution, ectopic fat, or the inflammatory state of adipose tissue. If BMI captured the full weight-related pathway, 28% would be an exact estimate of the weight-mediated component. It almost certainly does not. The residual 72% is better described as an association that persists after adjustment for a single, crude measure of adiposity. The honest summary is the one the authors give: semaglutide may influence dementia-related biology through metabolic, inflammatory, vascular, and potentially neurological pathways, and this analysis was not designed to apportion causality among them.
Peptide Atlas's registry lists 668 registered clinical trials for semaglutide on file, with 10 currently listed as recruiting. Among trials with phase designations on file, the breakdown is Phase 2: 4, Phase 4: 4, Phase 3: 1. The database also indexes 197 PubMed papers on the drug. The dementia analysis arrives within a much broader expansion of semaglutide research beyond diabetes and obesity:
Peptides referenced: Semaglutide, Glucagon, GLP-1.
Related reading: Real-world tirzepatide: 68% persist, 55% adherent, 10.5% weight loss, STEP 12: Semaglutide 2.4 mg Cuts Weight in Chinese Adults by 9.9 Points, Why Weight Returns After GLP-1 Agonist Therapy Is Stopped, UK first in Europe to approve Eli Lilly oral GLP-1 orforglipron.