Published on 16 September 2026
Plasma proteomic clocks of cell-type ageing: does astrocyte age predict Alzheimer’s disease?
In brief
Drawing on the plasma proteomics of 60,542 people from three independent resources, the authors build ageing clocks for more than forty cell types, then test their prognostic value. In the UK Biobank, extreme astrocyte ageing is associated with subsequent incident Alzheimer’s disease over fifteen years of follow-up, with a hazard ratio of 12.59 between the most aged and the youngest participants on this marker, and of 5.10 against the normal ageing group. Set against established risk factors in a single model, this marker weighs as much as APOE4 carriage, 5.16 versus 5.30. The work also reports an association between skeletal myocyte ageing and amyotrophic lateral sclerosis. Two reservations dominate the reading. These clocks are learned by penalised regression on proteins correlated with chronological age, which says nothing about what would happen if one intervened on them. And the last author declares co-founder positions in three biotechnology companies, while a patent application related to this work has been filed with Stanford University. The source claims no clinical use.
The context
The question of Alzheimer’s disease risk comes up in the consulting room long before the first objectively measurable memory complaint. A sixty-year-old patient whose parent had the disease asks what can be known, and what can be done. Until now the answer came down to two things: APOE genotype and, in some specialist pathways, amyloid and tau markers. The first is fixed at birth and says nothing about timing. The second are close to the disease itself, and therefore late.
The hypothesis explored here is of a different kind. It posits that tissues do not age at the same pace within one individual, and that this mismatch between organs can be read in the blood. If a plasma protein can be attributed to the cell type that produces it, a biological age can be estimated for each cell type, and one can look at which of them drifts before the disease.
The mechanism
What a cell-type clock measures, and what it does not
| Element | What the study says |
|---|---|
| Cell-type attribution | Single-cell transcriptomics from the Human Protein Atlas, version 24.1 |
| FindingA protein is assigned to a cell type only if the corresponding gene is expressed at least twice as highly in that type as in any other, or exclusively in it. This filter keeps only a minority of the proteins assayed, 16.5% on SomaScan (1,202 of 7,289) and 24.2% on Olink (708 of 2,923). The clock of a cell type is therefore an aggregate of circulating proteins, not a measurement made in the tissue, and the authors themselves point out that transcript levels do not settle protein abundance. | |
| Model construction | Elastic net penalised regression, aggregated over 100 resamples |
| FindingAmong the proteins assigned to a cell type, the model selects those that best predict chronological age, with sex as a covariate and ten-fold cross-validation. It therefore retains what covaries with age, without ranking what causes it. A model is kept only above a correlation threshold and with at least four proteins: 43 models on SomaScan, 48 on Olink. | |
| Two platforms | SomaScan for two cohorts, Olink for the third |
| FindingThe two technologies are not applied to the same participants: SomaScan is used for the GNPC consortium and the NSHD cohort, Olink for the UK Biobank. Agreement between platforms is therefore a replication across cohorts, not a paired comparison on the same samples. It remains a serious argument against an assay artefact, without correcting the limitations inherent to plasma measurement. | |
| What is not measured | The tissue itself |
| FindingNo histological verification and no astrocyte imaging supports the plasma estimate. External validation relies on indirect markers: in the Insight 46 substudy of the NSHD cohort, 483 participants, the astrocyte age gap is associated with plasma phosphorylated tau 217 burden, coefficient 1.08, adjusted P 3.92 × 10−7, and several cell types are associated with the PACC composite cognitive score. The link between a circulating astrocyte clock and the actual state of astrocytes remains an inference. | |
The study at a glance
| Population | |
| 60,542 adults from three independent resources: the UK Biobank (44,458 participants after quality control, aged 40 to 69 at recruitment, Olink platform), the Global Neurodegeneration Proteomics Consortium or GNPC (14,281 participants from 14 cohorts, SomaScan platform) and the British 1946 birth cohort NSHD (1,803 participants, SomaScan platform). The GNPC is not a disease-free population: 7,074 participants are defined as healthy, alongside patients with Alzheimer’s disease (2,761), amyotrophic lateral sclerosis (245), Parkinson’s disease (476), frontotemporal lobar degeneration (199) and mild or subjective cognitive impairment (1,992). The incidence analyses are confined to the UK Biobank. | |
| Exposure | |
| Biological age estimated for each cell type from high-density plasma proteomics, 7,289 proteins measured on SomaScan and 2,923 on Olink. | |
| Comparator | |
| Internal reference. For each cell type, the gap between predicted and expected age is converted to a z score: extreme ageing corresponds to a score above 2, extreme youthfulness to a score below −2, the rest forming the normal ageing group. No comparison arm in the sense of a trial. | |
| Outcomes | |
| Over fifteen years of follow-up in the UK Biobank: incident Alzheimer’s disease, amyotrophic lateral sclerosis, lung cancer, lymphoma, type 2 diabetes, chronic obstructive pulmonary disease, heart failure and stroke, as well as all-cause mortality (5,649 deaths). A composite polycellular risk score is then validated in the NSHD cohort. | |
| Design | |
| Observational cohorts with statistical learning, Oxford level of evidence 2b. No randomisation, no intervention. The GNPC component is cross-sectional; prospective follow-up comes from the UK Biobank and the NSHD cohort. |
Quality control
| Criterion | Judgement |
|---|---|
| Size and replication | Exceptional |
| FindingThree independent resources and two proteomic platforms. This is the main strength of the work, and it is rare at this density of measurement. | |
| Competing interests | Declared and substantial |
| FindingThe last author declares that he is a co-founder and scientific advisor of Teal Omics, Alkahest and Qinotto, with equity stakes; one co-author is also a co-founder of Teal Omics. A patent application related to this work has been filed by three authors and Stanford University. Two other co-authors declare ties to the pharmaceutical and diagnostics industries. Funding is mainly public and philanthropic, with declared support from Nan Fung Life Sciences, and the authors state that the funders had no role in the conduct of the study or in its publication. | |
| Nature of the inference | Predictive, not causal |
| FindingA model that predicts does not establish that acting on its variables would change the outcome. Nothing in this design supports a therapeutic target. | |
| Precision of the estimates | Unevenly reported |
| FindingThe most striking hazard ratios, 12.59 for astrocytes and Alzheimer’s disease and 12.74 for skeletal myocytes and amyotrophic lateral sclerosis, are given without a confidence interval. The supplementary tables do, however, provide bounds for a different comparison, extreme agers against the rest of the cohort: 4.70, 3.70 to 5.97, for astrocytes and Alzheimer’s disease on 702 cases, and 2.59, 1.58 to 4.23, for skeletal myocytes and amyotrophic lateral sclerosis on 275 cases. | |
| Representativeness | Older, Caucasian population |
| FindingThe authors state that the cohorts were predominantly older and Caucasian, and that validation in younger and more diverse populations is essential; they also specify that they used neither ancestry nor ethnicity in the analyses. The two British cohorts, UK Biobank and NSHD, account on their own for 46,261 of the 60,542 participants; the GNPC, for its part, brings together teams from several countries. | |
The findings
| Outcome | What the source reports |
|---|---|
| Astrocyte ageing and Alzheimer’s disease | 12.59 between extremes, 5.10 against normal |
| FindingAssociation measured on incidence in the UK Biobank, over fifteen years of follow-up, with significant stratification of the survival curves. Conversely, a young astrocyte profile is associated with a hazard ratio of 0.39 against the normal group, which the authors express as a risk reduction of more than 60%. No analysis measures the interval between blood sampling and the diagnosis of Alzheimer’s disease. | |
| Against established risk factors | 5.16 versus 5.30 for APOE4 |
| FindingIn the comparative model, extreme astrocyte ageing gives a hazard ratio of 5.16, 95% CI 4.06 to 6.56, comparable to APOE4 carriage, 5.30, 4.54 to 6.18, and higher than the polygenic risk score, 2.14, 1.92 to 2.39. The two combined, APOE4 carriage and extreme astrocyte ageing, give 11.58, 8.56 to 15.66. The associations are stronger in women than in men. | |
| Combination with genotype | 38.3% cumulative incidence at 15 years |
| FindingFigure observed in APOE4 homozygotes with extreme astrocyte ageing, that is 18 cases among 47 participants, against 12.6% in homozygotes with normal ageing (144 cases among 1,140) and no cases among the 23 homozygotes with a young astrocyte profile. The most exposed subgroup is therefore very small, which should temper the reading of this percentage. | |
| Muscle ageing and ALS | 12.74 between extremes |
| FindingHazard ratio between extremely aged and extremely young skeletal myocytes, reported without an interval. Cardiomyocyte ageing gives 6.59 in the same comparison. The association persists when only cases diagnosed more than three years after blood sampling are retained. On 275 incident cases, the supplementary table gives 2.59, 1.58 to 4.23, for extreme agers against the rest of the cohort: the disease is rare, which makes the precision of the estimate all the more decisive. | |
| Antagonistic pleiotropy of APOE4 | Older astrocytes, younger macrophages |
| FindingAPOE4 carriers show older astrocytes and younger macrophages than APOE3 carriers, with APOE2 carriers showing the reverse profile. The authors read this as antagonistic pleiotropy at the cellular level and relate it to an evolutionary hypothesis: immune vigilance that was advantageous under heavy infectious pressure, at the cost of accelerated brain ageing. It is separately, when discussing astrocyte ageing combined with genotype, that they put forward a lead for understanding why only some APOE4 carriers develop the disease. | |
Critical appraisal
| Domain | Risk of bias |
|---|---|
| Residual confounding | Moderate |
| FindingThe Cox models are adjusted for chronological age and sex. Additional covariates are used only for some diseases, for example glycated haemoglobin, body mass index, smoking and renal function for type 2 diabetes. For Alzheimer’s disease and for amyotrophic lateral sclerosis, no adjustment for comorbidities, inflammatory status or current treatments is reported, although these factors alter the plasma proteome. Two sensitivity analyses taking renal function into account appear in the supplementary tables. | |
| Exposure classification | Low |
| FindingAn objective biological measurement, taken before the event. The risk of differential misclassification is low. | |
| Outcome measurement | Registries, no adjudication |
| FindingIncident diagnoses come from the UK Biobank First Occurrences and Cancer Register fields and, where they exist, from algorithmic criteria, with ICD-10 and ICD-9 codes set out explicitly, F00 and G30 for Alzheimer’s disease. Dementia without specified cause and without documented alternative aetiology is reclassified as Alzheimer’s disease. No clinical adjudication is described, which limits the specificity of the outcome. | |
| Overfitting | Contained but real |
| FindingPenalised regression on several thousand variables is exposed to overfitting. Training and testing are separated by assessment centre in the UK Biobank, and the composite score is validated in a distinct cohort on a distinct platform. External replication greatly limits the risk, without eliminating it. | |
| Participant selection | Volunteer participants |
| FindingLarge volunteer cohorts are healthier than the general population, which shifts absolute values without necessarily distorting the direction of the association. | |
Level of evidence
Oxford level of evidence 2b, observational cohorts with independent replication. Confidence is high in the existence of an association between a so-called astrocyte plasma proteomic profile and subsequent incident Alzheimer’s disease, and in the reproducibility of the measurement framework from one cohort and one platform to another. It is moderate on the size of the effect: the reference estimates come with bounds, but the most striking values rest on comparisons between the extremes of the distribution, sometimes involving a few dozen participants, and with no interval reported. It is low on the biological interpretation of the signal, and nil on any therapeutic consequence. The work suggests a research target; it does not demonstrate that acting on astrocytes would change a patient’s trajectory.
The colleague test
What an experienced colleague would say if you put this study to them in two minutes, between two consultations.
“ Sixty thousand people and three cohorts, fine, that is serious work. But a hazard ratio of twelve between the two ends of the distribution, for a multifactorial disease, mostly tells me that the marker and the disease share a common cause I cannot yet see. And I have nothing to offer my patient: this assay exists only in research. I will keep the idea that organs do not age together, it is a fine one, and I will wait to see whether someone replicates it without the co-founder in the last author position. ”
What this means in practice: nothing changes in tomorrow’s clinic. What changes is how we will understand, a few years from now, why two carriers of the same genotype do not share the same fate.
What you can do with this
- Be ready to answer the patient who has read that Alzheimer’s disease can be predicted years in advance: this is an association measured over fifteen years of follow-up in research cohorts, and the source claims no clinical use.
- Keep in mind the idea that is useful in a memory clinic: ageing is not synchronous across organs, which helps explain why chronological age predicts individual risk so poorly.
- Have a concrete argument to hand on the unequal susceptibility of APOE4 carriers, a question families often ask.
- Keep the reflex of asking who stands to gain from the result: here, a patent application and companies exist, and that is information the reader must have.
- Watch, over the coming years, for independent prospective validation before any clinical translation.
Frequently asked questions
Is a plasma proteomic ageing clock test available for patients?
The measurements are performed on research platforms, SomaScan and Olink, within established cohorts. The source describes no test intended for routine practice and claims no clinical use.
Does a hazard ratio of 12.59 mean my patient’s Alzheimer’s risk is twelve times higher?
No. It is a comparison between the two extremes of the marker’s distribution within a cohort, a cell-type age gap above two standard deviations on one side, below minus two on the other. Against the normal ageing group, the hazard ratio falls to 5.10. Absolute risk depends on age, genotype and length of follow-up, and the highest cumulative incidence figure reported by the authors, 38.3%, rests on 18 cases among 47 participants.
Do astrocytes cause Alzheimer’s disease?
The design does not allow that claim. The model selects proteins associated with age, then observes their predictive value. An unmeasured common cause would explain the association just as well, and the authors themselves present their result as a stratification biomarker, not as an established mechanism.
Does this study say anything about preventing Alzheimer’s disease?
No intervention analysis appears in this paper. The silence of the source on this point allows no conclusion either for or against the efficacy of any preventive measure.
Annotated bibliography
Source study. Ding DY, Bot VA, Chen KL, Groves JW, Pálovics R, Masuda D, Farinas A, Oh HS, Wagner V, Lu N, The Global Neurodegeneration Proteomics Consortium (GNPC), Cruchaga C, Isakova A, Schott JM, Wyss-Coray T. Plasma proteomic signatures of cellular aging predict human disease. Nature Medicine, 2026; 32: 2060-2072. Received 21 April 2025, accepted 4 May 2026, published online 15 June 2026, June 2026 issue. DOI 10.1038/s41591-026-04446-y. PMID 42297981. Funding. Stanford Alzheimer’s Disease Research Center and National Institute on Aging (P50AG047366, P30AG066515, AG072255), Milky Way Research Foundation, Nan Fung Life Sciences, Knight Initiative for Brain Resilience, NOMIS Foundation, AHA-Allen Brain Health and Cognitive Impairment Cross-Network Collaborative Grants (23BHCICG1188316), MAC3 Impact Philanthropies, Alzheimer’s Research UK (ARUK-PG2014-1946 and ARUK-PG2017-1946), Alzheimer’s Association (SG-666374-UK birth cohort), British Heart Foundation and UK Dementia Research Institute. The authors declare that the funders had no role in study design, data collection and analysis, the decision to publish or preparation of the manuscript. Competing interests. T. Wyss-Coray and H. Se-Hwee Oh are co-founders and scientific advisors of Teal Omics Inc. and hold equity stakes; T. Wyss-Coray is also a co-founder and scientific advisor of Alkahest Inc. and Qinotto Inc., with equity stakes. C. Cruchaga declares research support from GSK and Eisai, sits on the scientific advisory board of Circular Genomics, in which he owns stock, and sits on the scientific advisory board of ADmit. J. M. Schott declares research funding and the supply of PET tracer by AVID Radiopharmaceuticals, a subsidiary of Eli Lilly, and by Alliance Medical, consultancy for Roche, Eli Lilly, Biogen, MSD, GE Healthcare, Alamar Biosciences and Receptive Bio, and the role of Chief Medical Officer of Alzheimer’s Research UK. The other authors declare no competing interests. T. Wyss-Coray, D. Y. Ding, V. A. Bot and Stanford University have filed a patent application related to this work.
Supplementary material. The supplementary material of the publication was consulted: Supplementary Figures 1 to 11, the Nature Portfolio reporting summary and Supplementary Tables 1 to 12, of which Tables 8 to 10 gather the hazard ratios by cell type in the UK Biobank with their 95% confidence intervals.