🧬 Aging Wiki

lehallier-proteomic-clock-2019

Properties1
aliasesLehallier 2019, plasma proteomic clock, proteomic aging clock, SomaScan aging, non-monotonic proteome aging

12 min read

Lehallier Proteomic Clock (2019)

A plasma proteomics-based biological aging clock built from SomaScan aptamer measurements of ~2,925 proteins in human plasma. The Lehallier 2019 Nature Medicine study trained a 373-protein LASSO model to predict chronological age (r=0.93–0.97 across cohorts; MAE ~2.9 years per Fig. 1g, not stated in manuscript text) and made two foundational discoveries: (1) plasma protein profiles change non-monotonically with age, with identifiable protein “waves” cresting at approximately 34, 60, and 78 years of age; and (2) a proteome-derived biological age captures aging-relevant variation distinct from DNA methylation clocks 1. As a non-DNA biomarker, the proteomic clock is mechanistically complementary to DNAm clocks and potentially more accessible for clinical translation.

Identity and Origin

  • Primary citation: Lehallier B et al. (2019) Nature Medicine 25:1843–1850 — doi:10.1038/s41591-019-0673-2
  • Archive status: downloaded; green OA via PMC (PMC7062043); local PDF verified 2026-05-05
  • Citation count: ~888 (as of 2026-05-05; 100th citation percentile in archive)
  • Clock type: First-generation proteomic (chronological-age trained)
  • First/senior authors: Benoit Lehallier, Tony Wyss-Coray (Wyss-Coray lab, Stanford)
  • Measurement platform: SomaScan v4 aptamer-based proteomics (Somalogic), ~2,925 SOMAmer reagents per sample
  • Cohorts: LonGenity (Ashkenazi Jewish longevity cohort), INTERVAL (blood donor cohort, UK), and additional validation cohorts
  • Output: Predicted chronological age in years (derived from 373-protein plasma signature)

Key Finding: Non-Monotonic Protein Waves

The most conceptually important contribution of Lehallier 2019 is the identification of three waves of protein abundance change that do not follow a simple linear age trajectory. Rather than proteins monotonically increasing or decreasing throughout life, the study identified distinct temporal patterns 1:

Wave / crest ageApproximate ageBiological interpretation (per paper)
Wave 1~34 yearsDownregulation of extracellular matrix (ECM) and structural pathway proteins; enrichment in ECM-related GO/KEGG pathways
Wave 2~60 yearsHormonal activity, binding functions, and blood pathways; CVD-associated proteins enriched; proteins most strongly linked to cognitive and physical function decline
Wave 3~78 yearsBlood pathways and bone morphogenetic protein (BMP) signaling; MMP12 and CHRDL1 prominent; proteins linked to Alzheimer’s disease and Down syndrome proteomes

This non-linear structure implies that protein-level aging is not a uniform progressive process — distinct biological transitions occur at mid-life, late middle age, and old age. Interventions effective at one transition may be ineffective at others. The finding also explains why cross-sectional studies of “aging biomarkers” conflate potentially distinct aging phases.

Important caveat: The three-wave result is based on cross-sectional (not longitudinal) plasma protein measurements. Whether the same individual passes through each wave sequentially, or whether the waves reflect cohort-specific effects or population heterogeneity, requires longitudinal data to confirm. needs-replication

Training Details

ParameterValue
Primary cohortsINTERVAL: n=3,301 (2 subcohorts, age 18–76y); LonGenity: n=962 used (1,030 enrolled minus 68 excluded for missing data, age 61–95y)
Total training + test samplesn=4,263 (INTERVAL + LonGenity only); 171 additional samples from 4 independent cohorts used for robustness assessment only
Training / test splitTwo-thirds (n=2,817) training; one-third (n=1,446) validation
Protein platformSomaScan: 5,284 aptamers (LonGenity), 4,034 aptamers (INTERVAL); 2,925 aptamers with identical SeqId in both cohorts used for analysis
Feature selectionLASSO regression (glmnet, alpha=1, 10-fold cross-validation) on 2,925 proteins → 373-protein predictor
Age range covered18–95 years (INTERVAL 18–76; LonGenity 61–95; independent cohorts 21–107)
Held-out test performancer=0.93–0.97 vs chronological age across cohorts (discovery n=2,817: r=0.97; 4 independent cohorts n=171: r=0.93); MAE ~2.9 years needs-replication — value not stated in PMC manuscript text; appears in main Fig. 1g

Performance Characteristics

Accuracy

The 373-protein predictor achieves r=0.93–0.97 vs chronological age across cohorts (r=0.97 in discovery dataset n=2,817; r=0.93 in 4 independent cohorts n=171). A 9-protein reduced model also predicted age with good accuracy (Extended Data Fig. 3c). MAE ~2.9 years is reported in main Fig. 1g but not stated in the manuscript text; the Horvath MAE comparison (cited elsewhere as ~3.6 yr) is not made in this paper and is sourced from elsewhere 1. unsourced — Horvath MAE comparison requires separate citation.

Mortality and morbidity prediction

Proteomic age acceleration is associated with higher all-cause mortality and worse healthspan outcomes in validation cohorts. The effect sizes for mortality prediction are not as well-characterized as for the DNAm mortality clocks (GrimAge, PhenoAge), but a directional signal exists. needs-replication — specific HR from proteomics predictor vs mortality needs larger-cohort confirmation.

Sex differences

The study identified 895 out of 1,379 age-associated proteins (65%; q<0.05) that also differed significantly between sexes — “2/3 of proteins changing with age also change with sex” (Discussion). This emphasizes that sex-stratified proteomics are necessary for accurate aging characterization. Pooled-sex proteomics aging models conflate sex-specific trajectories. The paper developed a sex-independent clock to address this. Key sex-differentiated proteins include CGA FSHB (follicle-stimulating hormone subunits) and CGA CGB (human chorionic gonadotropin subunit) 1.

Young adults: early drivers of aging biology

Among the most-cited translational findings: several proteins that change early in the proteomic aging trajectory (around the 34-year wave) are related to growth factors, extracellular matrix, and musculoskeletal biology. This suggests that aspects of molecular aging begin earlier than clinical aging phenotypes manifest — consistent with the emerging view that “late-life aging” has molecular precursors in mid-life.

Comparing Proteomic vs DNAm Clocks

FeatureLehallier ProteomicDNAm Clocks (Horvath, GrimAge)
ModalityPlasma proteins (SomaScan)Blood DNA methylation (Illumina array)
Sample typePlasma/serumWhole blood or PBMC
Tissue-specificPlasma-onlyBlood-centric (Horvath is cross-tissue)
Feature count373 proteins71–513 CpGs
Mechanistic interpretabilityHigher (proteins are functionally annotated)Lower (CpGs may not mark causal genes)
Correlation with DNAm ageModerate (r~0.5–0.7)N/A
Intervention responsivenessPartially studiedSee individual clock pages

The proteomic and DNAm clocks are not interchangeable — they capture overlapping but distinct biological signals. The low-moderate correlation between proteomic and DNAm age implies they measure different dimensions of biological aging. needs-replication

Intervention-Responsive Evidence

Partial positive signal: parabiosis and young blood factors (preclinical)

Wyss-Coray lab research (same group as Lehallier 2019) has demonstrated that young plasma infusion in aged mice reverses age-related protein profiles in blood and brain. These are preclinical findings and not direct tests of the Lehallier proteomic predictor specifically. See stem-cell-exhaustion and circulating factor literature for context. needs-human-replication — young blood / plasma proteins interventions are preclinical only.

Partial positive signal: exercise (contested, requires verification)

Exercise training has been shown to alter some of the proteins included in proteomic aging scores (GDF-15, growth factors, inflammatory mediators). Whether exercise meaningfully shifts the 373-protein Lehallier predictor has not been tested in a dedicated RCT as of 2026-05-05. unsourced

Null signal: most pharmacological interventions untested

No pharmacological intervention has been tested against the Lehallier proteomics predictor in a powered RCT. This is an important gap — proteomic clock endpoints are not yet established in any Phase 2 trial. Compare to DunedinPACE (CALERIE-2) and Horvath (multiple RCTs). long-term-unknown

Field successor: Argentieri 2024 ProtAge (UK Biobank)

A larger, methodologically updated proteomic aging clock — Argentieri 2024 ProtAge — was developed in 45,441 UK Biobank participants using the Olink Explore 3072 platform (2,897 plasma proteins) 2. The 204-protein ProtAge predictor achieves r=0.94 vs chronological age in UKB and was validated in Chinese (n=3,977; r=0.92) and Finnish (n=1,990; r=0.94) cohorts. ProtAge predicts the incidence of 18 major chronic diseases plus all-cause mortality, and is associated with telomere length, frailty index, and cognitive performance. ProtAge has now largely supplanted the Lehallier 2019 SomaScan-based clock as the canonical proteomic aging predictor in the field, owing to its larger training cohort, multi-ethnic validation, and Olink-platform availability (lower cost, faster turnaround than SomaScan v4). The Lehallier 2019 paper retains historical importance for the non-monotonic-wave finding and as the conceptual prototype.

Positive signal: exercise (MyoGlu 12-week supervised exercise; Sayed 2025)

A 12-week supervised exercise intervention in 26 men (MyoGlu trial; NCT01803568) reduced the ProtAge gap (ProtAgeGap) by an equivalent of approximately 10 months 3. Most of the 204 proteins remained stable; specific proteins including CLEC14A changed with exercise and tracked improved insulin sensitivity. Transcriptomic data from muscle and fat tissue corroborated PI3K-Akt and MAPK signaling involvement. This is the first interventional demonstration that a proteomic aging clock can be modestly reversed by exercise in humans, and it shifts the field’s view of proteomic aging from largely-static to lifestyle-modifiable. Note: applies to Argentieri ProtAge specifically, not the Lehallier 2019 373-protein SomaScan predictor.

Frailty-trained successor: ProtFI (Garst 2026)

A frailty-trained proteomic biomarker — ProtFI — was published in 2026 (Garst et al., Cell Rep Methods) using UK Biobank Olink proteomics + ¹H-NMR metabolomics in n=40,696 participants 4. ProtFI is an Elastic-Net model that uses a minimal set of proteins to predict the Rockwood frailty index; it outperforms established aging biomarkers in predicting incident cardiovascular disease, handgrip strength, and self-rated health, in internal UKB validation and in two external Dutch cohorts (n=995, n=500). A companion ProtMort clock predicts mortality. This represents another generational shift in proteomic aging clocks: training on frailty (a multi-system functional readout) rather than chronological age yields a more clinically relevant signal. See frailty-index.

Organ-specific proteomic clocks (Goeminne 2025)

A 2025 Cell Metabolism analysis (Goeminne et al.) extended the proteomic-age framework to organ-specific aging by deriving plasma-protein-based age models for individual organ systems and showing that diseases manifest as accelerated aging of specific organismal systems 5. This work emphasizes that “biological age” is not unitary — different organs age at different rates within an individual, and plasma proteomics can resolve this heterogeneity in a way that DNAm-based clocks (typically blood-only) cannot.

SomaScan Platform Limitations

The Lehallier clock requires SomaScan v4 — a proprietary, high-cost aptamer-based proteomics platform. This creates barriers for:

  1. Clinical translation — Current SomaScan v4 cost (~$500–1,500/sample) is prohibitive for routine clinical use
  2. Reproducibility — Platform-specific aptamer reagents and binding characteristics can vary between lots and platform versions (SomaScan v4 vs v4.1 vs SomaScan 5k); cross-version compatibility of the trained model is not guaranteed
  3. Regulatory path — No FDA-cleared version of this test exists; clinical implementation would require independent validation in a CLIA-certified context

Alternative platforms (Olink proximity extension assay, LC-MS/MS proteomics) do not directly replicate SomaScan measurements, limiting transferability. no-mechanism

The Non-Monotonic Wave Finding: Open Questions

  1. Cross-sectional vs longitudinal — The waves are identified from cross-sectional data. Longitudinal validation in individuals measured at multiple ages is needed to confirm that individuals transition through all three waves. needs-replication
  2. Cohort effects — The specific protein levels may reflect birth-cohort exposures (e.g., early-life infection, nutrition, pollutants) rather than pure biological aging processes. contradictory-evidence
  3. Causal interpretation — Which proteins in the wave signatures drive aging vs are driven by aging is unknown for most of the 373 features. Mendelian randomization or perturbational validation is needed.
  4. Intervention targets — The wave structure implies there may be different optimal intervention strategies at different life stages — but no interventional data is available.

Limitations and Open Critiques

  1. Cross-sectional training — Like most first-generation clocks, the predictor was trained to predict chronological age cross-sectionally, not to measure longitudinal change. It cannot directly be used as a “pace of aging” measure (unlike DunedinPACE). See dunedinpace-2022.
  2. SomaScan cost and accessibility — High cost limits widespread research and clinical use.
  3. Mortality HR not well characterized — Unlike GrimAge and PhenoAge, mortality HR is not the primary training objective; mortality prediction data are limited.
  4. No MR causal evidence — Genetic instruments for proteomic aging predictor level have not been analyzed in Mendelian randomization studies. The causal-vs-biomarker question is entirely open.
  5. Cross-tissue limitation — Plasma proteins reflect secreted/leaked products from all tissues; the proteomic clock has no single tissue of origin but also cannot identify which organ system is driving age acceleration in an individual.

Therapeutic Implications

The identification of specific proteins in each wave — particularly ECM proteins and structural pathway proteins that shift at ~34 years, and BMP-related proteins prominent at ~78 years — suggests potential therapeutic windows. The paper highlights GDF15 as one of the most prominently age-associated proteins across waves. The paper itself does not claim therapeutic targets; discussion of GDF11 or VEGF-C as rejuvenating factors is from separate Wyss-Coray lab work (refs 6–10 in this paper) and is not a finding of Lehallier 2019 specifically. needs-human-replication unsourced — GDF11/VEGF-C therapeutic claims require separate citations

External Resources

  • SomaScan platform: Somalogic (somalogic.com)
  • Aging Atlas (Open Aging Atlas): https://ngdc.cncb.ac.cn/aging/ — multi-omics aging data including some proteomic datasets
  • Human Protein Atlas: expression and aging context for individual proteins

Cross-references

Footnotes

Footnotes

  1. doi:10.1038/s41591-019-0673-2 · n=4,263 (INTERVAL n=3,301 + LonGenity n=962; age 18–95 yr); 171 additional subjects in 4 independent cohorts · observational cross-sectional · LASSO on 2,925 SomaScan proteins → 373-protein predictor; r=0.93–0.97 vs chronological age across cohorts · model: human plasma · Nat Med 2019 25:1843–1850 · PDF: PMC7062043 (verified 2026-05-05) · key finding: non-monotonic protein waves cresting at ages 34, 60, 78 yr identified by DE-SWAN; 895/1,379 age-associated proteins also sex-associated ↩ ↩2 ↩3 ↩4

  2. doi:10.1038/s41591-024-03164-7 · Argentieri MA et al. · Nat Med 2024 · ProtAge UK Biobank · n=45,441 UKB primary; validation in CKB n=3,977 (r=0.92), THL Finland n=1,990 (r=0.94) · observational cross-sectional + longitudinal mortality follow-up · 204-protein Olink Explore 3072 LASSO predictor; r=0.94 vs chronological age in UKB; predicts incidence of 18 major chronic diseases + all-cause mortality + multimorbidity; associated with telomere length, frailty index, reaction time · model: humans, plasma · pmid:39117878 ↩

  3. doi:10.1038/s41514-025-00318-w · Sayed N et al. · NPJ Aging 2025 · Reversal of proteomic aging with exercise · UKB observational n=45,438 + 12-week supervised exercise sub-study (MyoGlu; NCT01803568) n=26 men · ProtAgeGap reduced by ~10-month equivalent over 12 weeks; CLEC14A and other proteins changed with exercise and tracked insulin-sensitivity improvement; PI3K-Akt and MAPK signaling implicated by muscle/fat transcriptomics · pmid:41449222 ↩

  4. doi:10.1016/j.crmeth.2026.101405 · Garst S et al. · Cell Rep Methods 2026 · ProtFI · UK Biobank Olink proteomics + ¹H-NMR metabolomics; n=40,696 · Elastic-Net model trained on Rockwood frailty index; companion ProtMort trained on all-cause mortality · ProtFI outperforms established aging biomarkers for incident CVD, handgrip strength, self-rated health · external validation: 2 Dutch cohorts (n=995, n=500) · pmid:41966686 ↩

  5. doi:10.1016/j.cmet.2024.10.005 · Goeminne LJE et al. · Cell Metab 2025 (Jan 7) · Plasma protein-based organ-specific aging and mortality models · UK Biobank Olink proteomics · derives organ-system-specific aging clocks from plasma proteins; diseases mapped as accelerated aging of specific organismal systems · pmid: see PMC; verifies organ-specificity that single-tissue DNAm clocks lack ↩

Graph View

Backlinks