What AI sees in aging data
As people age, their blood composition, gene activity, tissue condition and other measurements change. Machine learning can look for combinations of these features in large datasets and compare them with participants’ ages or health. Some models are called ‘aging clocks’: they calculate a measure from a set of biomarkers, or measurable signs of biological processes [2].
Biological age is not a single quantity that can be read precisely from the body, like a date of birth in a passport. Different clocks use different data and can produce different results. An estimate reflects the particular model and sample; it does not tell a person how long they will live and is not, on its own, a diagnosis [2].
An example: blood proteins and organ age
In a study published in Nature in 2023, researchers used machine learning and data on proteins in blood plasma to estimate age-related signatures of 11 organs and systems. The models were tested across several groups comprising 5,676 adults in total. For some participants, the estimate for one organ differed from those for others [3].
Certain protein signatures were associated with diseases and mortality risk at the level of the groups studied. This is a useful clue for further work, but it does not prove that those proteins caused the disease. The study also did not show that changing an organ’s calculated age would extend an individual’s life [3].
From a molecule to a hypothesis, not a ready treatment
AI is also useful at an earlier stage of research. For example, the AlphaFold system predicts a protein’s three-dimensional structure from its sequence. Understanding a protein’s shape can help researchers investigate how it works and generate new hypotheses for biomedical research, including research into aging processes [4].
A predicted structure, however, remains a model: its accuracy depends on the case, and important details need experimental verification. Even when an idea is supported in cells or animals, that does not mean it is safe and effective in people. Separate research stages lie between a laboratory finding and a clinical outcome [4].
What a study in people can show
An instructive example is CALERIE, a randomized study involving 220 adults without obesity. One group was assigned to a calorie-restriction intervention, while the other ate without that restriction. In a subsequent analysis of blood samples from 197 participants, researchers assessed changes in several measures based on DNA methylation — chemical marks associated with the regulation of gene activity [5].
Over two years, one measure, DunedinPACE, changed in a direction the authors interpreted as a slower measured pace of aging; two other measures, PhenoAge and GrimAge, did not change significantly. This is a biomarker research result, not proof of longer life: the analysis did not establish an effect on lifespan. The example shows why a favorable result for one measure cannot automatically be treated as a final clinical outcome [5].
Why data quality and independent validation matter
A model learns from the data it is given. If a sample underrepresents certain age groups or people of different backgrounds, the result may work less well for them. That is why it matters to test a model on independent data and in different groups, and to assess how accurately and reproducibly the biomarker itself is measured. For example, repeated analysis of the same sample should give comparable results. Testing in other laboratories helps determine whether a difference reflects biological processes or features of the measurement method. It also matters whether the association with an outcome remains after accounting for participants’ ages [2].
A model’s intended task matters too: finding patterns in a research study is different from using a result in a clinical decision. In their joint principles for AI use in drug development, the U.S. Food and Drug Administration and the European Medicines Agency emphasize the context of use, data governance and risk-based assessment of model performance [6].
What evidence is needed to claim longer life
To establish that an intervention benefits people, it is not enough to show that a blood test or an algorithm’s result changed afterward. Research needs an appropriate comparison group, safety assessment and outcomes that matter to people, such as disease, the ability to perform daily activities and survival. Some answers require long-term follow-up [2] [5].
Randomization helps researchers compare groups and reduce the influence of differences between participants, but study size, duration and reproducibility matter too. A biomarker may help researchers track changes before such outcomes occur, but first it must be shown to reliably reflect the specific clinical outcome being assessed [2].
The takeaway
AI is already expanding the tools available to aging research: it helps researchers detect signals in data, build models and select hypotheses to test. But a prediction is the beginning of an investigation, not a guarantee of longevity. The most convincing findings bring together high-quality data, independent validation and clinical studies that measure what really matters in people’s lives.
Sources
- World Health Organization — Healthy ageing and functional ability
- Moqri et al., Nature Medicine — Validation of biomarkers of aging
- Oh et al., Nature — Organ aging signatures in the plasma proteome track health and disease
- Jumper et al., Nature — Highly accurate protein structure prediction with AlphaFold
- Waziry et al., Nature Aging — Effect of long-term caloric restriction on DNA methylation measures of biological aging in healthy adults from the CALERIE trial
- U.S. Food and Drug Administration and European Medicines Agency — Guiding Principles of Good AI Practice in Drug Development