Semaglutide, proteomics, and a lower 5-year dementia risk signal

A question that won’t leave you once you see it

Picture a routine checkup: blood drawn in the morning, no scanners, no spinal taps, no long series of cognitive tests. Then imagine the lab returns a number that estimates a person’s risk of dementia years and years later. Can that kind of prediction be real—or is it just a science-fiction promise?

A new post hoc analysis in Alzheimer’s & Dementia links semaglutide (a GLP-1 receptor agonist) to a measurable reduction in a proteomics-based dementia risk signature in older adults with overweight/obesity and cardiovascular disease without diabetes. The headline number—about 26% lower modeled 5-year dementia event rate vs placebo—sounds exciting. But the real story is how proteomics turns the messy chemistry of illness into something computationally trackable over time. (pmc.ncbi.nlm.nih.gov)

Let’s walk through the moving parts: what semaglutide is, what the “dementia risk signature” actually measures, what the SELECT trial contributed, and how to interpret results that are predicted rather than clinically diagnosed.

Semaglutide in one paragraph: a hormone mimicker with system-wide effects

Semaglutide is a GLP-1 receptor agonist—a drug designed to mimic a gut hormone called glucagon-like peptide-1 (GLP-1). GLP-1 helps regulate appetite and blood sugar by signaling through the GLP-1 receptor in multiple tissues, not only the pancreas. In large trials, semaglutide has been used at higher doses for weight management in people with obesity and related conditions. (accessdata.fda.gov)

Why would a weight-and-metabolism drug matter for dementia biology? Because dementia risk is strongly intertwined with the body’s long-term “system stress”: vascular injury, inflammation, metabolic dysfunction, and changes at the blood–brain barrier. GLP-1 receptor agonists are being studied for exactly those kinds of upstream pathways in addition to glucose and weight effects. ()

Proteomics: measuring the body’s protein “language”

Proteomics is the study of proteins—the workhorses that carry out most cellular functions. In blood, proteins reflect a mixture of organ activity: liver, immune system, vascular cells, and more. Measuring proteins can give a snapshot of biology that blood sugar alone can miss.

This study uses an aptamer-based proteomics platform (commonly described through the SomaScan approach). Instead of using antibodies for every target, SOMAmer reagents (a kind of modified DNA aptamer) bind to specific proteins, and the assay outputs a quantitative signal tied to protein abundance. (somalogic.com)

High-throughput proteomics can measure thousands of protein features in one sample. That’s where the next concept enters: a risk score is not “one protein.” It’s usually a learned pattern across many.

The key tool: dSST, a 25-protein dementia risk score

The analysis relies on the Dementia SomaSignal Test (dSST), a plasma proteomic predictor that estimates 5-year and 20-year all-cause dementia risk using a 25-protein score. A “score” here means a single computed number derived from those proteins—built with machine learning so that the combination correlates with later dementia incidence in population cohorts. (med.nagoya-u.ac.jp)

What makes dSST interesting is that it tries to capture multi-pathway risk, not just Alzheimer’s disease (AD) neuropathology. Dementia is often shaped by overlapping processes—vascular disease, neuroinflammation, and other systemic contributors—so a proteomics pattern is a natural way to summarize that overlap.

A practical version of the question many people search is: Can a blood test really predict dementia risk years in advance? dSST was developed and validated to predict long-horizon dementia risk based on measured blood proteins, using datasets like ARIC and others described in dSST materials. ()

One nuance worth keeping in mind: prediction models can be accurate about risk categories while still being imperfect about what an individual will experience. Risk scores are probabilities, not destinies.

How SELECT created a biology time-course (without diagnosing dementia)

The parent trial is SELECT (Semaglutide Effects on Heart Disease and Stroke in Patients With Overweight or Obesity). SELECT studied semaglutide 2.4 mg once weekly vs placebo in adults with overweight/obesity and established cardiovascular disease but without diabetes—a group where cardiometabolic risk and dementia risk often overlap.

From the dementia-risk analysis perspective, the important design detail is timing and sampling: researchers analyzed non-fasted serum at baseline and at week 104 (about two years) and applied the dSST model to see whether semaglutide changed the predicted dementia risk trajectory. (ncbi.nlm.nih.gov)

So this wasn’t “everyone gets dementia imaging.” It was “everyone gets blood-based proteomics, and the model translates biology into a dementia-risk estimate.”

The headline findings: slower predicted risk growth with semaglutide

Here’s what the post hoc analysis reported when comparing semaglutide vs placebo:

1) 5-year predicted dementia risk

Semaglutide was associated with a 2.5-fold less increase in the model’s 5-year dementia risk versus placebo—described as a 26.0% lower predicted event rate.

The analysis also reports an odds ratio (OR) of 0.74 with a 95% confidence interval (CI) from 0.65 to 0.85. In plain language, an OR below 1 indicates lower odds of the modeled event/risk category relative to placebo. The confidence interval gives the plausible range of that effect size. ()

2) 20-year predicted dementia risk

For longer horizon modeling, semaglutide showed a smaller—but still reported—signal: 8.8% lower modeled dementia event rate over 20 years, with OR 0.91 (95% CI 0.88–0.94). ()

This pattern—stronger short-horizon modeled change than long-horizon modeled change—is common in biology-based prediction studies. Longer-horizon outcomes are influenced by many future exposures, and a two-year intervention window may only move some of the upstream signals.

3) Risk classification shifts

Beyond continuous risk estimates, dSST also groups people into risk categories. The post hoc analysis reports 36% lower odds of being classified as higher risk, with P < 0.001 and a reported effect parameter (β = −0.44). ()

Odds ratio, explained without the math pain

An odds ratio (OR) compares the odds of an event in one group to the odds in another. If OR = 0.74, the odds in the semaglutide group are about 26% lower than placebo on the odds scale.

Two details matter:

• This is about modeled dementia risk, not a confirmed dementia diagnosis during SELECT.
• Post hoc proteomics analyses can be sensitive to modeling choices. That doesn’t make them useless—it means they’re a hypothesis-generating bridge toward clinically testable outcomes.

Why this is promising, and why it still isn’t a dementia “prevention” claim

The most exciting part of the story is the method: instead of waiting years for clinical dementia diagnoses, the study asks whether semaglutide alters a biological signature associated with dementia risk.

That aligns with the broader dementia-prevention thinking that risk is substantially shaped by modifiable factors across the life course. The 2024 Lancet Commission report estimated that roughly 45% of dementia risk is attributable to modifiable risk factors. This doesn’t mean any single drug can prevent dementia—it means upstream interventions may shift downstream probability. (chronicdisease.org)

Still, there are hard limits:

1. Post hoc ≠ primary outcome. The SELECT trial wasn’t designed to measure dementia incidence.
2. Predicted risk ≠ observed diagnosis. A dSST score can move because proteins shift, but that doesn’t guarantee long-term clinical benefit.
3. dSST measures “all-cause dementia risk.” That’s broader than Alzheimer’s disease alone.

A cautious way to interpret the results is: semaglutide may be attenuating some systemic pathways captured by the 25-protein dSST signature. That’s a step toward biological plausibility, and it’s exactly what translational research tries to do.

The bigger technical lesson: risk signatures can quantify drug effects on biology

Even if you never care about semaglutide specifically, the technical takeaway is bigger.

This analysis shows how a proteomics risk model can act like a “biological dashboard.” In a cardiovascular outcomes trial, the primary focus is clinical events. Here, researchers reuse stored blood samples to ask whether semaglutide changes a dementia-related proteomic pattern.

That kind of design matters because dementia is slow. If blood-based multi-protein signatures reliably track meaningful change, they can reduce the time needed to test prevention hypotheses.

And yes, the tricky parts remain—model stability, population differences, and the perennial question of whether biomarker shifts translate into clinical outcomes. But this is the pathway researchers keep walking: from proteins, to predicted risk, to tests of whether those predictions hold in real-world endpoints.

Bottom line

Semaglutide appears to reduce the modeled growth of dSST-based dementia risk over two years of treatment in older adults with overweight/obesity and cardiovascular disease without diabetes, with reported estimates such as 26% lower predicted 5-year dementia event rate vs placebo. ()

It’s a compelling biological signal—one that illustrates how proteomics and machine-learned protein signatures can turn cardiometabolic interventions into measurable, dementia-relevant pathway changes. The study doesn’t prove dementia prevention, but it does sharpen the target for what future dementia-focused trials need to test.