One blood draw clocks the biological age of 40-plus cell types, Stanford team shows in 60,542 people

2026-08-19

Mapping plasma proteins to cell types via the Human Protein Atlas, Stanford researchers trained elastic net aging clocks for 40-plus cell types in 60,542 people; extremely aged skeletal myocytes carried a 12.7-fold higher 15-year ALS risk.

What problem this solves

Aging is asynchronous: within one person, different cells and organs can sit at very different biological ages. Measuring aging at cellular resolution has required harvesting tissue for single-cell transcriptomics, which is invasive and impossible to scale across living populations. Plasma proteomics previously topped out at organ-level clocks, so resolution stopped at the organ.

This Nature Medicine paper asks whether a single blood draw can push the measurement down to cell-type level, and whether "one cell type is unusually old" predicts disease. The corresponding author is Stanford's Tony Wyss-Coray, whose group pioneered plasma proteomic aging clocks; the sample spans 60,542 people across three independent cohorts.

Method

Three steps, with the attribution problem solved first:

• Using Human Protein Atlas single-cell transcriptomics, 1,202 of 7,289 SomaScan-measurable plasma proteins (16.5%) were assigned to specific cell types, requiring expression at least twofold higher in that type than in any other. On the Olink platform, 708 of 2,923 proteins (24.2%) mapped.
• For each cell type, an elastic net regression (glmnet in R, averaged over 100 bootstrap replicates, sex as covariate) was trained on its专属 protein subset to predict chronological age. The quality gate was r ≥ 0.25 in training, r ≥ 0.15 in test, at least 4 protein features; 43 models survived for SomaScan and 48 for Olink (including 14 lineage-level models such as a lymphoid lineage pooling B, T and NK cells).
• Each person gets an age gap per cell type: predicted biological age minus the average prediction for their chronological age, z-scored, with >2 marked extremely old and < -2 extremely young.

The design's strongest feature is cross-platform, cross-cohort validation: GNPC (14,281, SomaScan), NSHD (1,803, SomaScan) and UK Biobank (44,458, Olink) replicate one another, so conclusions do not hinge on one assay technology.

Results

Even among healthy people the distribution is uneven: in GNPC, 35.4% had no extreme cellular age gaps, 24.4% had exactly one extremely old cell type, and 1.5% had ten or more. UKB numbers are close (26.1% / 22.7% / 2.8%).

Disease prediction is the headline, UKB 15-year follow-up, adjusted for age and sex:

The skeletal myocyte-ALS link survives excluding cases diagnosed within 5 years of blood draw, so it is not reverse causation from subclinical disease. Among APOE4 homozygotes, extreme astrocyte aging came with 38.3% 15-year cumulative Alzheimer's incidence versus 12.6% for normal astrocytes; none of the 23 APOE4 homozygotes with youthful astrocytes developed AD. APOE4 carriers show older astrocytes but younger macrophages, with APOE2 the inverse, matching antagonistic-pleiotropy expectations.

Cumulative burden is dose-responsive: 90% 15-year survival with no extreme cell types versus 34% with more than 20. A polycellular aging risk score (PARS, Cox model on extreme-aging indicators plus sex) split the top and bottom 5% cleanly in training and test, and transferred to the SomaScan-based NSHD cohort.

Why it matters

This is the first demonstration of cellular-resolution aging measurement at scale in living humans, from one blood draw. For AI4Science practitioners it is a clean case study: single-cell atlases as priors, routine clinical assays as input, elastic net as the model. No deep learning anywhere; the bottleneck fell to data-attribution design. On the prevention side, cell-specific early warning is the appeal, with the skeletal myocyte clock flagging ALS risk years before any existing screen (though clinical use is far off). Commercially, organ-level plasma clocks already support companies; pushing to cell-level resolution is a direct upgrade of the same playbook.

A cold shower on magnitude: these HRs compare extreme groups, base event rates are low (ALS hit 13 of 1,342 in the extreme group), and individual-level risk prediction remains distant.

Limitations

Author-stated: cell-type coverage is limited to the Human Protein Atlas catalog, leaving some specialized populations out; transcript levels do not always track protein abundance, and proteins assigned to cardiomyocytes (FABP3) or microglia (HAVCR1) may be released more broadly in disease; cohorts skew old and Caucasian, so younger and more diverse validation is needed.

Three further concerns from reading it:

• The r ≥ 0.25 gate is lenient; several cell clocks have weak age prediction, and every downstream disease association rides on that moderate signal.
• Protein-to-cell mapping is inferential, and in disease the cellular origin of proteins may shift, at which point the age gap no longer cleanly means "this cell type aged".
• The respiratory-aging and lung-cancer link adjusted for pack-years, but residual confounding by smoking cannot be excluded; the authors stick to "58% higher hazard than smoking alone".

Data and code availability statements sit in the online methods; no public repository link appears in the text.

Terms

• Aging clock: a regression model predicting biological age from molecular features; prediction minus chronological age gives the age gap.
• Elastic net: linear regression with mixed L1/L2 regularization, suited to protein data with many correlated features.
• SomaScan / Olink: the two dominant plasma proteomics platforms, aptamer-based and proximity-extension-based respectively.
• Age gap: an individual's predicted cell-type biological age minus the same-age average, z-scored with ±2 marking extremes.
• APOE4: the strongest genetic risk allele for Alzheimer's disease.
• Hazard ratio (HR): the ratio of event rates between two groups per unit time.

Source

• Paper

What people are saying

• Nature Medicine: Plasma proteomic signatures predict human disease EricTopol

All paper explainers