Scientists have unveiled a new artificial intelligence–derived blood test that distills the molecular noise of cellular aging into a single number, one that appears to forecast who will die, who will fall ill, and how quickly the body accumulates the senescent cells that many researchers now consider a driving force behind growing old. The biomarker, called the SASP Score, was described in the journal Aging Cell and built from plasma protein data on tens of thousands of middle-aged and older adults in the UK Biobank. Unlike earlier attempts to quantify the biology of aging with linear statistics, the new score is generated by a semi-supervised deep learning architecture that can capture the tangled, non-linear relationships among dozens of secreted inflammatory proteins, and it can be ported across different laboratory measurement platforms, a technical hurdle that has long frustrated the field.
The biological target of the score is the senescence-associated secretory phenotype, or SASP, a cocktail of cytokines, chemokines, growth factors, proteases, and other molecules shed by senescent cells. Cellular senescence is a state of permanent cell-cycle arrest that is generally protective against cancer in the short term but becomes increasingly problematic with age. Senescent cells do not simply sit quietly; they remain metabolically active and progressively remodel their secretome, flooding surrounding tissue with inflammatory and tissue-remodeling signals. Through this secretory output, a relatively small population of arrested cells can exert outsized effects on neighboring and even distant tissues, promoting chronic inflammation, tissue dysfunction, and the multicolored tapestry of age-related disease, from heart failure to dementia. Because these molecules circulate in the blood, they offer an accessible window onto a burden that would otherwise require tissue biopsies to assess.
To build the score, the research team first curated a panel of 38 plasma proteins that are consistently dysregulated when cells of different types are pushed into senescence by different triggers. The list was assembled by cross-referencing databases such as CellAge, the SenNet Consortium’s markers, and the SASP Atlas, together with manual curation of the literature, and it includes well-known inflammatory messengers such as interleukin-6, CXCL8, CCL2, GDF15, and soluble tumor necrosis factor receptors. The team then turned to the UK Biobank Pharma Proteomics Project, which measured 2,923 plasma proteins in more than 54,000 participants using the Olink Explore 3072 proximity extension assay. After excluding three proteins with excessive missingness and imputing the remainder, the researchers worked with proteomic data from 50,997 participants, splitting them into a training set of 43,358 individuals and a held-out test set of 7,639.
The model at the heart of the SASP Score is a Guided AutoEncoder with Transformer, or GAET. Classic autoencoders compress high-dimensional input into a low-dimensional latent representation and then reconstruct the original data, a strategy that excels at nonlinear dimensionality reduction but can drift toward biologically meaningless encodings. The GAET adds an auxiliary guidance head that predicts chronological age from the latent representation during training, steering the model toward aging-relevant protein structure. Crucially, chronological age is used only as a training signal; once the model is trained or fine-tuned, the score itself is computed purely from protein measurements. The Transformer backbone, borrowed from the architecture that powers modern language models, encodes both the quantitative protein values and the textual identity of each protein through a dual-embedding mechanism, allowing the network to learn inter-protein dependencies while retaining the semantic anchors needed to transfer the model to other cohorts and assays.
The results in the UK Biobank test sample were striking. The SASP Score correlated strongly with chronological age (Spearman rho = 0.70) and with established composite measures of biological aging, including PhenoAge (rho = 0.66), BioAge (rho = 0.71), the proteomic aging clock (rho = 0.67), and the healthspan proteomic score (rho = −0.63). It also tracked weaker but consistent signals across physical and cognitive traits: higher scores went hand in hand with higher waist-to-hip ratio, higher body mass index, elevated systolic blood pressure, greater frailty on a 49-item index, shorter leukocyte telomeres, weaker grip strength, slower walking pace, and slower reaction times. Median scores were significantly higher in men, in adults aged 60 and over, in people with high waist-to-hip ratios, in former or current smokers, and in those with frailty, hypertension, or hypercholesterolemia.
Most consequentially, the score predicted what happened next. Over a mean follow-up of 13.3 years, each one-standard-deviation increase in the SASP Score was associated with a 41 percent higher risk of death (adjusted hazard ratio 1.41, 95 percent confidence interval 1.25 to 1.60), after accounting for age, sex, ethnicity, education, material deprivation, smoking, alcohol status, systolic blood pressure, and body mass index. The score was also significantly associated with the incidence of multiple age-related conditions, with particularly robust signals for chronic kidney disease, heart failure, lung cancer, and neurological disorders including delirium, dementia, and Parkinson’s disease. When the researchers stratified participants using a cutoff set at the 90th percentile of scores among healthy individuals, those above the threshold showed visibly worse Kaplan–Meier survival curves for mortality and several major chronic diseases.
Shapley Additive exPlanations analysis, a technique borrowed from machine-learning interpretability research, revealed which proteins carried the most weight in the score. The top five contributors were GDF-15, CXCL-9, PGF, TNFRSF11B, and CCL-20, a lineup that the authors interpret as highlighting mitochondrial function, immune signaling, and apoptosis control as central threads of the senescence secretome. Importantly, the composite score consistently outperformed any single protein in predicting mortality and nearly every age-related disease examined, with diabetes the lone exception. The researchers attribute this advantage to the model’s capacity to capture nonlinear interactions among proteins, which they argue better reflects the genuine complexity of the secretome than the one-marker-at-a-time approach used in most prior studies, an approach that also inflates the risk of spurious statistical associations through repeated testing.
The external validation came from the MEDEX study, a randomized clinical trial of 585 adults aged 65 to 84 that compared mindfulness-based stress reduction, multimodal exercise, and their combination in older people with subjective cognitive complaints. The team fine-tuned the pre-trained GAET model on MEDEX proteomic data, which had been generated on a completely different platform, a Luminex multiplex immunoassay, using conservative settings to avoid overfitting. In this healthier and age-restricted sample, the score correlated modestly with age (rho = 0.34) and comorbidity burden (rho = 0.16). The longitudinal results were arguably the most provocative finding of the study: over 18 months, the SASP Score rose significantly in participants who did not exercise, but showed no significant increase in those assigned to the exercise arm. Exercise did not lower the score, but it appeared to halt its age-related climb, a pattern the authors describe as consistent with a senomorphic, senescence-stabilizing effect of physical activity.
The authors are careful to position the SASP Score as a measure of one specific hallmark of aging rather than a general biological age clock, and they note that it can be used alongside proteomic aging clocks and organ-specific scores to give complementary views of an individual’s aging biology. They also acknowledge limitations: there is no consensus on a universal set of SASP markers, plasma protein levels may carry nonspecific inflammatory or tissue-damage signals, and the precise cellular sources of the circulating proteins remain unknown. Disease classification in the UK Biobank also lacks granularity for factors such as tumor subtype and stage. Even so, the ability to compute a senescence biomarker across different proteomic platforms, and to detect its stabilization in response to a non-pharmacological intervention, positions the score as a practical tool for observational studies and, perhaps more importantly, for clinical trials of senolytic and senomorphic drugs, where a sensitive, transferable readout of cellular senescence burden has been a long-standing unmet need.
Subject of Research: A deep learning–based composite blood biomarker of cellular senescence burden for predicting mortality and age-related health outcomes
Article Title: A Deep‐Learning Based Biomarker of Systemic Cellular Senescence Burden to Predict Mortality and Health Outcomes
Article References: Zhao, S., Kuo, C.-L., Lenze, E. J., Wetherell, J. L., Haynes, L., El‐Ahmad, P., Fortinsky, R., Kuchel, G., Harris, T., & Diniz, B. S. (2026). A Deep‐Learning Based Biomarker of Systemic Cellular Senescence Burden to Predict Mortality and Health Outcomes. Aging Cell, 25(10), Article e70737. https://doi.org/10.1111/acel.70737
Image Credits: AI Generated
DOI: 10.1111/acel.70737
Keywords: cellular senescence, SASP, deep learning, biomarker, UK Biobank, proteomics, aging, mortality, exercise, Transformer, autoencoder, geroscience
Cite Scienmag News
APA
MLA
Chicago
Beatrice Stafford. (October 1, 2026). AI Turns Blood Proteins Into a Senescence Score That Predicts Death and Disease. Scienmag. https://scienmag.com/ai-turns-blood-proteins-into-a-senescence-score-that-predicts-death-and-disease/
Beatrice Stafford. “AI Turns Blood Proteins Into a Senescence Score That Predicts Death and Disease.” Scienmag, 1 October 2026, https://scienmag.com/ai-turns-blood-proteins-into-a-senescence-score-that-predicts-death-and-disease/. Accessed 1 October 2026.
Beatrice Stafford. “AI Turns Blood Proteins Into a Senescence Score That Predicts Death and Disease.” Scienmag. October 1, 2026. https://scienmag.com/ai-turns-blood-proteins-into-a-senescence-score-that-predicts-death-and-disease/
Copy citation
Download RIS
Tags: Agingaging and disease risk predictionaging biomarkersAI-driven blood testautoencoderbiobank aging researchbiomarkercellular aging biomarkersCellular senescencedeep learningdeep learning for agingExerciseGeroscienceinnovative aging measurement techniquesmortalitynon-linear aging analysisplasma protein biomarkersProteomicsSASPSASP and agingsenescence score predictionsenescent cell secretory phenotypeTransformerUK Biobank
