Semaglutide attenuates a proteomics-based dementia risk signature in older adults with overweight or obesity and cardiovascular disease without diabetes: A post hoc analysis of the SELECT phase 3 trial
Abstract
INTRODUCTION
Plasma proteomics detect multi-pathway biological changes preceding dementia onset. The Dementia SomaSignal Test (dSST) is a validated 25-protein score predicting 5- and 20-year all-cause dementia risk. Preclinical and clinical data suggest glucagon-like peptide-1 receptor agonists may have neuroprotective effects.
METHODS
In a post hoc analysis of the Semaglutide Effects on Heart Disease and Stroke in Patients With Overweight or Obesity (SELECT) trial, adults ≥ 65 years with overweight/obesity and cardiovascular disease without diabetes (n = 2970) were randomized to semaglutide 2.4 mg or placebo. Non-fasted serum samples at baseline and week 104 were analyzed using the dSST.
RESULTS
Semaglutide reduced increases in predicted dementia risk versus placebo: 2.5-fold less increase in 5-year risk (26.0% lower predicted event rate; odds ratio [OR] 0.74, 95% confidence interval [CI] 0.65–0.85) and 1.67-fold less increase in 20-year risk (8.8% lower; OR 0.91, 95% CI 0.88–0.94). It also reduced odds of higher dementia risk classification by 36% (β −0.44; P < 0.001).
DISCUSSION
Semaglutide slowed progression of a validated proteomics-based dementia risk signature.
Highlights
- Semaglutide attenuated a proteomics dementia risk signature in the Semaglutide Effects on Heart Disease and Stroke in Patients With Overweight or Obesity (SELECT) trial.
- Five-year predicted risk increased 2.5-fold less (26% lower modeled dementia rate).
- Twenty-year predicted risk increased 1.67-fold less (8.8% lower modeled dementia rate).
- Proteomics can characterize effects of cardiometabolic interventions on dementia risk.
1 BACKGROUND
Dementia most often develops over decades, arising from neurodegenerative proteinopathies—including amyloid beta (Aβ) and tau in Alzheimer's disease (AD) or alpha synuclein in Parkinson's disease and Lewy body dementia (LBD)— frequently compounded by co-occurring pathophysiological processes such as cerebrovascular disease and neuroinflammation. Biomarkers such as amyloid and phosphorylated tau measured in cerebrospinal fluid (CSF), positron emission tomography, or more recently in plasma, have improved diagnosis and biological detection of AD. However, AD-specific biomarkers do not fully capture non-AD neuropathologies (e.g., cerebrovascular injury) or the cumulative impact of systemic processes contributing to all-cause dementia risk, particularly in late life and in populations with high cardiometabolic comorbidity.1, 2 The 2024 Lancet Commission found that ≈ 45% of dementia risk can be attributed to modifiable risk factors.3 Hence, it remains essential to identify biomarkers that are orthogonal to core AD neuropathology and reflect pathways that can be modified before symptom onset.
Recent proteomic studies have revealed robust molecular signatures across asymptomatic and symptomatic AD and have identified biologically distinct AD subtypes in CSF proteomics, highlighting heterogeneity that is likely to influence treatment response.4-6 In parallel, plasma proteomics has advanced toward clinically scalable prediction of incident dementia using multi-protein panels and machine learning.7
The Dementia SomaSignal Test (dSST) is a 25-protein, machine learning–derived score developed using SomaScan proteomics that predicts individualized 5-year and 20-year risk of all-cause dementia diagnosis.7 The dSST was developed using the SomaScan Assay (SomaLogic), which uses DNA-based binding reagents (modified aptamers) to assess plasma levels of ≈ 5000 proteins with high specificity and limits of detection largely comparable to antibody-based assays.7, 8
The dSST is derived from a panel of 25 circulating proteins spanning multiple interconnected biological domains, including immune-metabolic signaling (CCL18, S100A9, IL18, HAVCR1, AGER, NOTCH1), vascular and extracellular matrix remodelling (SVEP1, ADAMTS13, CILP2, CDCP1, TNFRSF11B), metabolic and endocrine regulation (PTN, PPY, IGFALS, NADK, MAP2K2, ASB9, CHI3L1), neuronal integrity (NPTXR, NEFL, CDON), and cytoskeletal and proteostatic maintenance (TBCA, NDST1, H2AC25, S100A13; Table S1 in supporting information).7
The dSST 20-year model was trained in The Atherosclerosis Risk in Communities (ARIC) Study visit 3 cohort (n = 11,277), a community-based biracial US cohort designed to study atherosclerosis risk factors, with median age 60, mean body mass index (BMI) 27.7 kg/m2, and low prevalence of diabetes (6.3%) at baseline. The model was then further validated in ethnically and geographically independent cohorts including the Baltimore Longitudinal Study of Aging (BLSA, n = 1214) and the Japanese National Institute for Longevity Sciences–Longitudinal Study of Aging (NILS-LSA, n = 340). A 5-year prediction output was further validated in the ARIC visit 5 cohort (n = 4985).
In the ARIC cohort, dementia diagnosis was adjudicated through surveillance of cognitive and functional assessment tests (including the Clinical Dementia Rating [CDR] scale; the Functional Activities Questionnaire [FAQ]; and the delayed word recall task, digit symbol substitution from the Wechsler Adult Intelligence Scale–Revised [WAIS-R], a letter fluency task), as well as telephone screenings, informant ratings, hospital records, and death record reviews. Using these data, dementia was classified based on National Institute on Aging and Alzheimer's Association (NIA-AA) 2018 and Diagnostic and Statistical Manual of Mental Disorder, Fifth Edition (DSM-5) criteria.
The dSST demonstrated prognostic value with areas under the curve (AUC) of 0.68 to 0.70 for 20-year dementia risk (0.75–0.81 when combined with age), and 0.78 for 5-year risk in older adults. The dSST outperformed the apolipoprotein E genotype (AUC 0.60) and established plasma biomarkers (measured using the single molecule array [Simoa]: phosphorylated tau [p-tau]181 [0.63], Aβ42/40 [0.64], neurofilament light chain [0.62], and glial fibrillary acidic protein [0.56] for 20-year dementia risk prediction). The dSST was associated with longitudinal cognitive decline across multiple domains (including the Mini-Mental State Examination [MMSE]), accelerated brain atrophy (particularly in medial-temporal regions), and neuroimaging/neuropathology measures.7 In sex-stratified analyses, the model performance did not differ and the combination of the dSST, age, and sex yielded the same performance compared to the age + dSST model (AUC 0.81).
The dSST predicts individualized 5-year and 20-year probabilities of a dementia diagnosis, which can further stratify individuals into clinically meaningful predefined low, medium–low, medium–high, and high risk of dementia groups, with a 20-year event rate ranging from 3.8% (low) to 35.6% (high) and a 5-year event rate ranging from 1.5% (low) to 20.3% (high).7
The average risk of developing dementia (i.e., the predicted probability of developing dementia based on the mean of the 25 proteins selected for the dSST) within 20 years was 15% (or a dSST score of 15) in the training data set. Four non-overlapping risk strata were classified based on relative risk to the training data set average: > 50% reduced risk (low; dSST score ≤ 7), between 50% reduced risk and average risk (medium–low; dSST score > 7 and ≤ 15), between average risk and 50% elevated risk (medium–high; dSST score > 15 and ≤ 22), and > 50% elevated risk (high; dSST score > 22).
The average risk of developing dementia within 5 years was 6.5% (or a dSST score of 6.5) in the training data set. Four non-overlapping risk strata were classified based on relative risk to the training data set average: > 50% reduced risk (low; dSST score ≤ 3.1), between 50% reduced risk and average risk (medium–low; dSST score > 3.1 and ≤ 6.5), between average risk and 50% elevated risk (medium–high; dSST score > 6.5 and ≤ 9.7), and > 50% elevated risk (high; dSST score > 9.7).7
Cardiometabolic risk factors (obesity, insulin resistance, vascular disease) are consistently associated with later life dementia risk, motivating preventive strategies in these populations. Glucagon-like peptide-1 receptor agonists (GLP-1RAs) such as semaglutide produce substantial weight loss and improve cardiometabolic profiles while also demonstrating pleiotropic anti-inflammatory, immune-modulating, and vascular effects that may plausibly influence neurodegenerative pathways. Early clinical and real-world evidence suggests possible reductions in cognitive decline or dementia risk among GLP-1RA–treated populations, particularly in type 2 diabetes, although causal mechanisms and generalizability remain uncertain.9-11 Recent syntheses emphasize neuroimmune pathways as plausible mediators linking GLP-1 signaling to brain health,12, 13 and preliminary evidence suggests proteomics can identify AD-related biomarkers impacted by GLP-1RAs.14, 15 In addition, biomarker data from the recent phase 3 clinical trials evoke and evoke+ with once-daily oral semaglutide 14 mg in early AD indicate semaglutide reduced CSF concentrations of several relevant biomarkers related to AD pathology, neurodegeneration, and neuroinflammation, but did not translate to clinical efficacy.16 It is yet to be confirmed whether pharmacological interventions can shift long-term dementia risk biology before disease symptomatology. The present study is the first to evaluate whether semaglutide modifies a validated plasma proteomic dementia risk signature using randomized clinical trial data.
The Semaglutide Effects on Heart Disease and Stroke in Patients With Overweight or Obesity (SELECT) trial was a large randomized cardiovascular outcomes trial in participants with overweight/obesity and established cardiovascular disease without diabetes comparing once-weekly subcutaneous semaglutide 2.4 mg versus placebo.17 This setting provides an opportunity to test whether GLP-1RA treatment modifies circulating proteomic signatures associated with long-horizon dementia risk in an older, high-risk cardiometabolic population.
RESEARCH IN CONTEXT
-
Systematic review: We reviewed literature on plasma proteomics for dementia risk prediction, SomaScan-based signatures, glucagon-like peptide-1 receptor agonists (GLP-1RA), dementia risk, and cognition using PubMed and related sources. No randomized controlled trial has previously examined GLP-1RA effects on a proteomic dementia risk signature in a cardiometabolic population.
-
Interpretation: In the Semaglutide Effects on Heart Disease and Stroke in Patients With Overweight or Obesity (SELECT) randomized cardiometabolic outcomes trial, semaglutide attenuated longitudinal worsening of a validated proteomic dementia risk signature in older adults with overweight/obesity and cardiovascular disease without diabetes (n = 2970), the first evidence of pharmacological effect on a multi-pathway dementia risk signature in a cardiometabolic population.
-
Future directions: Prospective studies should evaluate whether treatment-induced proteomic risk reductions translate to slower cognitive decline or lower dementia incidence in individuals with cardiometabolic disease.
2 METHODS
2.1 Study design and population
This post hoc analysis used data from the SELECT trial (NCT03574597), a randomized, double-blind, placebo-controlled cardiovascular outcomes trial enrolling 17,604 adults aged ≥ 45 years with overweight or obesity (BMI ≥ 27 kg/m2) and established atherosclerotic cardiovascular disease, defined as one or more of the following: previous myocardial infarction, stroke, or symptomatic peripheral artery disease. Exclusion criteria included glycated hemoglobin (HbA1c) of ≥ 48 mmol/mol (≥ 6.5%); history of type 1 or 2 diabetes; presence of end-stage kidney disease; or previous myocardial infarction, stroke, hospitalization for unstable angina pectoris, or transient ischemic attack within 60 days of screening; or New York Heart Association class IV heart failure.
Participants were randomized in a 1:1 ratio to once-weekly subcutaneous semaglutide, titrated to a target dose of 2.4 mg over 16 weeks, or matched placebo, in addition to standard of care. The SELECT trial design, inclusion/exclusion criteria, and primary cardiovascular outcomes have been described in detail elsewhere.17 All participants provided written informed consent, consistent with the Declaration of Helsinki.
Semaglutide 2.4 mg was superior to placebo in reducing the incidence of death from cardiovascular causes, non-fatal myocardial infarction, or non-fatal stroke at a mean follow-up of 39.8 months (hazard ratio 0.80; 95% confidence interval [CI] 0.72–0.90; P < 0.001).17
For the present analysis, we included participants aged ≥ 65 years at baseline with available non-fasted serum samples at baseline and week 104 (n = 2970; Figure 1). The age threshold was prespecified for this analysis to enrich for individuals at higher short- to intermediate-term dementia risk and to align with the calibration range of the proteomic risk model. Serum samples in SELECT were pre-specified for collection at weeks 0, 20, and 104; the week 104 timepoint was selected for this analysis as the latest pre-specified sampling point with stored samples available for proteomic profiling. All analyses were conducted according to the randomized treatment assignment (intention-to-treat principle).
Samples were processed and stored at three different sites within the trial infrastructure. To ensure analytical consistency, 15 participants with discordant storage sites between baseline and week 104 samples were excluded. After applying these quality control filters, 2970 participants with complete paired samples at both timepoints were included in the analysis.
2.2 Proteomic assessment and dementia risk model
Circulating proteomic profiles were measured using the SomaScan aptamer-based platform version 4.1 (SomaLogic) for the high-throughput quantification of ≈ 7000 proteins.
Predicted dementia risk was estimated using the dSST, a previously developed and validated 25-protein, machine learning–derived plasma signature trained to predict individualized 5- and 20-year risk of all-cause dementia diagnosis.7, 18
In SELECT, the dSST algorithm was applied without modification to non-fasted serum samples collected at baseline and week 104. Serum samples were frozen at −80°C until analysis. The model outputs continuous predicted probabilities for 5- and 20-year dementia risk, as well as assignment to ordered risk categories (low, medium–low, medium–high, high) based on previously defined thresholds corresponding to increasing observed event rates in validation cohorts (Figure 2).7 Log10-transformed protein measurements were centered and scaled using means and standard deviations from the ARIC visit 3 training data sets.
Because the dementia risk prediction model (dSST) was originally developed using SomaScan 5K ethylenediaminetetraacetic acid (EDTA) plasma data and subsequently validated in independent EDTA plasma SomaScan 7K datasets (7), we evaluated potential matrix effects arising from the use of serum samples in SELECT. Matrix comparability was assessed using paired EDTA plasma and serum samples from ≈ 1000 individuals in the Covance cohort,19 assayed on the SomaScan 5 and 7K platforms for plasma and serum, respectively. Concordance between matrices was quantified using the Lin concordance correlation coefficient (CCC). Across the 25 proteins included in the dSST model, the median CCC between EDTA plasma and serum was 0.808, indicating good agreement at the analyte level. Importantly, concordance of the resulting model predictions was high, with a CCC of 0.959 between plasma- and serum-derived dSST scores. These results support cross-matrix robustness of both individual features and the integrated model output.
2.3 Outcomes
The outcomes of this post hoc analysis, comparing semaglutide versus placebo, were change in predicted 5-year dementia risk from baseline to week 104; change in predicted 20-year dementia risk from baseline to week 104; and shift in ordered dSST risk categories at week 104.
As an exploratory analysis, baseline dSST risk categories were examined among participants who experienced dementia-related adverse events during the trial, defined using the narrow-scope standardized MedDRA query (SMQ) for dementia. These events were not adjudicated and were used solely for descriptive enrichment analyses.
2.4 Statistical analysis
All analyses were conducted in the subset of participants aged ≥ 65 years with available non-fasted serum samples at baseline and week 104 (n = 2970) and compared semaglutide versus placebo according to randomized assignment.
The resulting regression coefficient (β1) was exponentiated to obtain odds ratios (ORs) with 95% CIs, reflecting the relative difference in risk progression between treatment groups. To minimize the influence of extreme predicted values, survival probabilities were truncated to a minimum of 10−6 or a maximum of 1–10−6 prior to logit transformation. For the 20-year risk model, no predictions exceeded these bounds. For the 5-year risk model, 179 (6.03%) predictions at baseline and 107 (3.6%) at the 2-year visit were below the lower threshold and therefore were truncated to 1×10−6.
This model adjusts for baseline risk category as a covariate, with the treatment effect coefficient subsequently exponentiated to yield an OR. Model estimates are reported as regression coefficients (β), standard errors (SE), P values, and corresponding ORs with 95% CIs.
Given the post hoc and hypothesis-generating nature of these analyses, no adjustment for multiplicity was prespecified. All statistical tests were two sided with P < 0.05 considered statistically significant. Missing week 104 samples were assumed to be missing completely at random, and no imputation was performed. Participants without paired baseline and week 104 samples were excluded from the analysis.
3 RESULTS
3.1 Study sample
The analysis included a subset of 2970 SELECT participants with available non-fasted serum samples at baseline and week 104. Baseline clinical characteristics were balanced between treatment groups as expected from randomization (Table 1).
| Characteristic |
Placebo (n = 1588) |
Semaglutide 2.4 mg (n = 1382) |
|---|---|---|
| Age [years], mean (SD) | 69.8 (3.9) | 69.7 (3.8) |
| Female sex, no. (%) | 414 (26.1) | 400 (28.9) |
| Race, no. (%) | ||
| White | 1467 (92.4) | 1270 (91.9) |
| Asian | 79 (5.0) | 62 (4.5) |
| Black or African American | 23 (1.4) | 22 (1.6) |
| Other/not reported | 19 (1.2) | 28 (2.0) |
| Body mass index [kg/m2], mean (SD) | 32.9 (4.5) | 32.9 (4.7) |
| Cardiovascular disease inclusion criteria, no. (%) | ||
| Myocardial infarction only | 981 (61.8) | 875 (63.3) |
| Stroke only | 302 (19.0) | 243 (17.6) |
| Peripheral artery disease only | 79 (5.0) | 75 (5.4) |
| Two or more inclusion criteria | 187 (11.8) | 163 (11.8) |
| Other | 39 (2.5) | 26 (1.9) |
| Glycated hemoglobin [%], no. (%) | ||
| < 5.7 | 516 (32.5) | 436 (31.5) |
| 5.7 to < 6.5 | 1072 (67.5) | 946 (68.5) |
| eGFR [ml/min/1.73 m2], mean (SD) | 74.7 (15.1) | 74.5 (15.9) |
| High-sensitivity CRP [mg/L], median (IQR) | 1.6 (0.8−3.5) | 1.6 (0.8−3.6) |
- Abbreviations: CRP, C-reactive protein; eGFR, estimated glomerular filtration rate; IQR, interquartile range; SD, standard deviation.
3.2 Semaglutide slowed the progression of predicted 5-year dementia risk
Over 104 weeks of follow-up, predicted 5-year dementia risk increased in both treatment groups, consistent with aging in an older, high-risk population. However, the magnitude of increase was significantly attenuated among participants receiving semaglutide compared to placebo (Figure 3).
Specifically, the 5-year proteomics-predicted dementia risk increased 2.5-fold less in the semaglutide group over 104 weeks, corresponding to a 26.0% lower predicted event rate (OR 0.74; 95% CI 0.65–0.85). This effect represents a significant slowing of proteomic risk progression in this older cardiometabolic population.
3.3 Semaglutide slowed the progression of predicted 20-year dementia risk
A similar, albeit more modest, pattern was observed for long-horizon risk estimates (Figure 3). The 20-year risk increased 1.67-fold less, corresponding to an 8.8% lower predicted event rate (OR 0.91; 95% CI 0.88–0.94).
The more modest relative effect on 20-year risk compared to 5-year risk is consistent with the longer time horizon and reduced short-term sensitivity of long-range risk calibration, as reported in the dSST validation studies.7 This pattern may reflect greater immediate effects on pathways driving near-term vulnerability, while long-horizon risk also encompasses accumulated lifetime exposures less amenable to short-term intervention.
To assess whether the observed effect was mediated by weight reduction, we performed a sensitivity analysis adjusting for change in BMI from baseline to week 104. In this adjusted model, the treatment effect coefficient was attenuated from β −0.092 to β −0.066 (P < 0.001), corresponding to a reduction of 28% in the estimated effect size. This attenuation indicates that 72% of the treatment-associated difference in 20-year dementia risk persisted after accounting for BMI change, suggesting that mechanisms beyond weight loss contribute to the observed proteomic signature modification.
3.4 Semaglutide shifted the distribution of proteomic risk categories
Ordinal regression analysis demonstrated that semaglutide treatment was associated with a significant downward shift in the distribution of dSST risk categories at week 104 (Figure 4). Semaglutide was associated with 36% lower odds of being classified into a higher dementia risk category compared to placebo (β −0.44; SE 0.078; P < 0.001), corresponding to an OR of 0.64 (95% CI 0.55–0.75).
This effect reflected a general shift toward lower or more stable risk categories rather than reclassification of a small subset of individuals, suggesting a broad impact on the underlying biological processes captured by the dSST proteomic signature.
3.5 Exploratory analysis of dementia-related adverse events
Among participants aged ≥ 65 years who experienced dementia-related adverse events during SELECT (SMQ Dementia, narrow scope; n = 18), baseline dSST risk categories were medium–high in 7 individuals (39%) and high in 11 individuals (61%). No participants with dementia-related adverse events were classified as low risk at baseline.
Although event numbers were small and dementia was not systematically adjudicated, this enrichment is directionally consistent with the intended risk stratification properties of the dSST and suggests the proteomic signature may identify individuals at elevated risk for near-term cognitive events even within a selected trial population. These findings should be interpreted with caution given the reliance on adverse event reporting, which likely under-ascertains true dementia incidence.
4 DISCUSSION
In a post hoc analysis of serum proteins in a subset of the SELECT participants aged ≥ 65 years with overweight/obesity and established cardiovascular disease without diabetes (n = 2970), semaglutide attenuated longitudinal worsening of a validated 25-protein dSST over 104 weeks. The effect was evident for both 5- and 20-year predicted dementia risk (26% and 8.8% lower modeled 5- and 20-year dementia event rates, respectively), and semaglutide shifted the distribution of categorical risk levels downward versus placebo. These findings suggest that semaglutide may modify systemic proteomic signatures associated with long-term dementia risk in a cardiometabolic population.
Proteomics has transitioned from discovery to increasingly reproducible, multi-cohort signatures that capture the multifactorial biology of dementia and AD.
The dSST was developed specifically to address a gap in dementia risk prediction, enabling the scalable quantification of risk across the multi-decade preclinical/prodromal window and across mixed etiologies contributing to all-cause dementia.7 Importantly, the dSST proteins capture systemic inflammatory, metabolic, vascular, and glial processes in addition to neuronal and synaptic biology. As such, the composite dSST may be sensitive to broader physiological perturbations beyond AD-specific central nervous system pathology alone.
Among the 25 dementia-associated proteins in the dSST, 19 (76%) have also shown established roles in cardiometabolic disease, including vascular and extracellular matrix remodeling, inflammatory signaling, and metabolic and endocrine regulation. This highlights that this protein signature may have the potential, at a biological level, to reflect the benefit of semaglutide on the shared mechanistic axis linking cardiometabolic dysfunction to dementia risk (Table S1).
Our SELECT findings suggest that semaglutide, a pharmacologic intervention with established cardiometabolic benefits, can attenuate the expected age-related increase in a dementia risk proteomic signature. GLP-1RAs may influence dementia-related biologies through multiple mechanisms, such as improved metabolic control and adiposity reduction, reduction of systemic inflammation, improved endothelial function and vascular health, and central effects on neuroinflammation and synaptic/neuronal injury.12, 13 Real-world studies in type 2 diabetes populations have reported associations between GLP-1RA exposure and lower clinical AD diagnosis and all-cause dementia incidence, though confounding and channeling bias remain concerns.9, 10 Controlled clinical evidence is still emerging, and mechanistic biomarkers can help triangulate plausibility.
The larger relative effect on 5-year predicted risk compared to 20-year predicted risk in our analysis could reflect: (1) greater dynamic range/short-term sensitivity of the 5-year calibration among older adults (as reported in dSST validation), (2) partial reversibility of near-term inflammatory/vascular risk processes, or (3) statistical scaling differences in risk calibration at different horizons.7 This pattern is consistent with the idea that cardiometabolic interventions may have more immediate effects on pathways driving near-term vulnerability, while long-horizon risk also reflects accumulated lifetime exposures.
The observed attenuation of the treatment effect by ≈ 28% when adjusting for BMI change indicates that weight loss accounts for a meaningful but minor portion of the signal. The remaining effect may reflect direct neuroprotective, anti-inflammatory, immune-modulating, or vascular mechanisms independent of adiposity reduction, consistent with the pleiotropic properties of GLP-1 receptor agonism.
Our results indicate that semaglutide modifies a validated proteomic signature that predicts dementia risk across cohorts.7 If proteomic risk signatures are to be used as pharmacodynamic tools or surrogate-like readouts, the critical next step is demonstrating that treatment-induced changes correlate with cognitive endpoints or adjudicated dementia outcomes, which were not assessed in the SELECT trial. Nonetheless, in the era of prevention trials and pragmatic risk stratification, a scalable blood-based signature that is both prognostic and modifiable could be valuable for trial enrichment (identifying higher risk older cardiometabolic individuals for prevention studies), mechanistic understanding (tracking pathway-level response to interventions that act on systemic metabolism/inflammation/vascular biology), or combination biomarker frameworks (integrating proteomic risk signatures with AD-specific plasma biomarkers such as p-tau217 to distinguish AD biology from broader dementia susceptibility).
Key strengths include the randomized, placebo-controlled trial context; longitudinal sampling over 104 weeks with measurements at baseline and week 104; a sizeable older subgroup (n = 2970); and use of a proteomic signature developed and validated across independent, diverse cohorts with demonstrated long-horizon prognostic value and associations with cognitive decline, brain atrophy, and neuropathology measures.7 Notably, the ARIC training cohort shares several characteristics with SELECT—including high BMI, low prevalence of diabetes, and enrichment for other cardiovascular risk factors—supporting the biological relevance of the dSST in a cardiometabolic population.
Limitations include the post hoc design, lack of prespecified multiplicity control for proteomic endpoints, and reliance on predicted risk rather than adjudicated dementia/cognitive outcomes within SELECT. Dementia-related adverse events captured via MedDRA SMQs are likely under-ascertained and do not substitute for systematic cognitive assessment. A further limitation is the use of serum rather than plasma for proteomic profiling. The dSST was developed and validated in EDTA plasma cohorts, and while aptamer-based assays can be applied to serum, matrix-related differences in protein release during coagulation may alter the absolute levels of certain analytes and potentially affect dSST calibration. Finally, the dSST derivation and validation cohorts (ARIC, BLSA and NILS LSA) were community-based, did not exclude diabetes, and were not enriched for overweight or obesity (even though ARIC had a mean BMI of 27.7). These differences may affect the calibration of absolute dSST risk thresholds in the SELECT population but should not undermine within-trial between-arm comparisons, in which systematic biases apply equally across treatment groups.
Priority next steps include mediation analyses to separate direct versus indirect effects; cross-biomarker concordance with AD-specific plasma biomarkers (i.e., p-tau217) to map where dSST is most sensitive; subtype interpretation connecting signature shifts to proteomic subtype biology described in recent literature;4, 5 and finally, prospective cognitive outcomes to test whether these findings translate to slowed cognitive decline or reduced dementia incidence.
5 CONCLUSIONS
Among participants aged ≥ 65 years with overweight/obesity and established cardiovascular disease without diabetes in the SELECT phase 3 cardiovascular outcomes trial (n = 2970), subcutaneous semaglutide 2.4 mg once weekly slowed the progression of validated proteomics-based 5- and 20-year dementia risk signatures over 104 weeks and reduced the odds of being classified into higher risk categories. These findings support the potential utility of circulating proteomic risk signatures as sensitive, pathway-integrative tools for dementia risk stratification and for evaluating biological effects of cardiometabolic interventions on dementia-related processes. Whether proteomic risk reductions translate to cognitive or dementia outcomes requires prospective evaluation.
ACKNOWLEDGMENTS
We thank the participants, investigators, and trial staff involved in the SELECT clinical trial. Editorial assistance was provided by Jim Wood, OPEN Health Communications. This study was funded by Novo Nordisk A/S. The funder, along with the academic collaborators, was responsible for the study design and contributed to data collection, analysis, preparation, and review of the manuscript in collaboration with the authors.
CONFLICT OF INTEREST STATEMENT
M. Jiménez-Mausbach and J. C. Refsgaard are employees of and minor shareholders in Novo Nordisk A/S. C. Paterson is an employee and shareholder of Illumina. B. M. Tijms is coinventor on a patent of CSF proteomic subtypes (published under patent no. US2022196683A1, owner VUmc Foundation). She receives research funding from ZonMw (VIDI grant #09150171910068), the European Research Council (DecipherAD, #101171721), and the Amsterdam Cohort Hub; all payments were made to her institution. She has received consulting fees from Novo Nordisk and Roche, and payment for lectures/presentations from Sanofi and Novo Nordisk, all paid to her institution. . Author disclosures are available in the Supporting Information.
CONSENT STATEMENT
All human subjects provided informed consent.