King's College LondonDec 18 2024
Researchers at the Institute of Psychiatry, Psychology & Neuroscience (IoPPN) at King's College London have conducted a comprehensive study to evaluate artificial intelligence based aging clocks, which predict health and lifespan using data from blood.
The researchers trained and tested 17 machine learning algorithms using data on markers in the blood from over 225,000 UK Biobank participants, aged 40 to 69 years when they were recruited. They investigated how well different metabolomic aging clocks predict lifespan and how robustly these clocks were associated with measures of health and aging.
A person's metabolomic age, their "MileAge", is a measure of how old their body seems to be on the inside based on markers in the blood called metabolites. Metabolites are small molecules that are produced during the process of metabolism, for example when food is broken down into energy. The difference between a person's metabolite-predicted age and their chronological age, termed MileAge delta, indicates whether their biological aging is accelerated or decelerated.
The study was published in Science Advances and is the first to comprehensively compare different machine learning algorithms on their ability to develop biological aging clocks using metabolite data, leveraging one of the largest datasets globally. It was funded by the National Institute for Health and Care Research (NIHR) Maudsley Biomedical Research Centre (BRC) and used data from the UK Biobank.
Individuals with accelerated aging (i.e., with a metabolite-predicted age older than their chronological age) were, on average, frailer, more likely to have a chronic illness, rated their health worse, and had a higher mortality risk. They also had shorter telomeres ('caps' at the end of chromosomes), which are a marker of cellular aging and linked with age-related diseases such as atherosclerosis. However, decelerated biological aging (with a metabolite-predicted age younger than chronological age) was only weakly linked with good health.
Aging clocks could help spot early signs of declining health, enabling preventative strategies and interventions before disease onset. They may also allow people to proactively track their health, make better lifestyle choices, and take steps to stay healthy for longer.
Metabolomic aging clocks have the potential to provide insights into who might be at greater risk of developing health problems later in life. Unlike chronological age, which cannot be changed, our biological age is potentially modifiable. These clocks provide a proxy measure of biological age for biomedical and health research, which could help shape lifestyle choices taken by individuals and inform preventative strategies implemented by health services. Our study evaluated a broad range of machine learning approaches for developing aging clocks, showing that non-linear algorithms perform best at capturing aging signals."
Dr. Julian Mutz, King's Prize Research Fellow at the IoPPN and lead author of the study
Professor Cathryn Lewis, Professor of Genetic Epidemiology & Statistics, Co-Deputy Lead of the Trials, Genomics and Prediction theme at the NIHR Maudsley BRC, and senior author of the study, said: "There is substantial interest in developing aging clocks that accurately assess our biological age. Powerful big data analytics can play a critical role in advancing these tools. This study is an important milestone in establishing the potential of biological aging clocks and their ability to inform health choices."
The researchers found that a metabolomic clock developed using a specific machine learning algorithm, called Cubist rule-based regression, was most strongly associated with most health and aging markers. They also found that algorithms which can model non-linear relationships between metabolites and age generally performed best at capturing biological signal informative of health and lifespan.
King's College London
Journal reference:
Mutz, J., et al. (2024). Metabolomic age (MileAge) predicts health and life span: A comparison of multiple machine learning algorithms. Science Advances. doi.org/10.1126/sciadv.adp3743.