Plasma protein intensities from DIA-MS in 350 Russian adults supported chronological-age estimation (MAE about 8.9 years in random cross-validation, rising to 10.56-11.87 years when whole MS batches were held out), but acquisition order was itself strongly age-associated, so the findings support only a proof-of-concept and not a batch-independent model, clinical tool, or biologically validated ageing clock.
Many plasma proteins were associated with chronological age: 148 of 222 protein groups at nominal p < 0.05, and 98 after Bonferroni correction.
The strongest age associations involved fibrinogen beta chain, SERPINA3, ITIH4, complement C5, and ceruloplasmin.
In internal random-split analyses, a 50-feature Gradient Boosting strategy estimated age with a mean absolute error (MAE) of about 9 years.
Holding out complete MS batches increased prediction error, and these batch-held-out results are the primary estimate of generalization.
Reducing the number of proteins selected within each training fold increased error, showing a trade-off between assay complexity and accuracy.
Acquisition order and MS batch were confounded with age and sex, and acquisition order alone predicted age.
The authors conclude the results support only a proof-of-concept for plasma-proteomic prediction of age within this dataset.
This study measured proteins in blood plasma from 350 adults in Moscow, aged about 20 to 99, to see whether protein levels could be used to estimate a person's age. Many proteins changed with age: 98 of 222 remained significantly linked to age even after strict statistical correction. The strongest links were with proteins involved in clotting, inflammation, and immune function, such as fibrinogen beta chain, SERPINA3, ITIH4, complement C5, and ceruloplasmin. A machine-learning model using 50 proteins estimated age to within roughly 9 years on average when tested with standard random splits of the data. However, the samples had been run on the mass spectrometer in an order that was related to age, so the instrument run order alone could predict age to within about 4 years. When whole instrument batches were held out to test more realistic generalization, the error rose to about 10.6-11.9 years. A smaller 8-protein version had higher error (about 11.8 years in random splits). This research suggests that plasma proteins carry age-related information, but the results are tied to this dataset and its design. The authors state that this is only a proof-of-concept and not a clinical tool or a validated biological ageing clock. Independent cohorts with better-balanced sample processing would be needed to confirm the findings.
Arbatskiy M, Balandin D, Novikova S, Vavilov N, Maiorov V, Churov A. (2026). Age-Associated Plasma Protein Signatures and Proteomic Age Estimation by Data-Independent Acquisition Mass Spectrometry in a Russian Cross-Sectional Cohort.. International journal of molecular sciences. https://doi.org/10.3390/ijms27198707