Predicting age from fibroblast gene expression
Neonatal dermal fibroblasts (blue: nuclei; green: tubulin; red: actin)

Neonatal dermal fibroblasts (blue: nuclei; green: tubulin; red: actin)

As we have discussed here before, the ability to predict a person’s age by making some biological measurement would be of immense value in the field of biogerontology: a reliable biomarker of age would enable us to monitor the rate of aging in real time, dramatically accelerating progress in identifying lifespan-enhancing therapeutics.

A great deal of effort has been directed toward identifying such biomarkers, and the most promising candidates proposed thus far as quite diverse, ranging from methylation clocks to calculations based on hospital records. Each has its own advantages, but also disadvantages, and despite recent progress there is still a great deal of room for improvement. The field continues to await a simple, inexpensive, non-invasive measurement that can be performed repeatedly on the same person, enabling both individual monitoring and longitudinal studies within populations.

A promising step in the right direction is exemplified by a study published last week. a collaboration led by researchers at the Salk Institute reported that gene expression in a specific type of skin cells can predict the donor’s age with striking accuracy, with an error just a few percent of the maximum human lifespan.

As their starting material, the authors obtained dermal fibroblasts from 133 healthy people ranging in age from 1 to 94 years of age. Fibroblasts are an ideal choice for studies of this kind in large part because they normally divide very slowly, making it more likely that individual cells will hold a record of age-related damage. They’re also relatively easy to obtain, although the procedure involved is much more invasive (and painful) than a blood draw.

After culturing the fibroblasts for a period of time, the authors obtained gene expression data for each sample by RNA sequencing (RNA-seq), a next-generation technique for comprehensively measuring RNA levels in a sample of cells. They then used the data to train a machine learning system to predict the donor’s chronological age from gene expression levels, and compared its performance against previously reported efforts to use gene expression as a biomarker of age.

The trained system could assign ages to donors with impressive accuracy, at a median absolute error of just 4 years, far better than previously published models. Importantly, the system assigned higher-than-actual ages to cells from subjects with Hutchison-Gilford Progeria Syndroms (HGPS), whose symptoms mimic accelerated aging—a feat that none of the competitor models were able to match.

It remains an open question whether the error (i.e., the difference between predicted and chronological age) is noise, or instead reflective of real differences among the subjects. People age heterogeneously, so it could be that two people with the same chronological age have quite different physiological ages.

In future research, health outcome data could be used to test this idea (probably focusing on a narrow slice of the age range in order to control study costs): Do 65-year-olds assigned an age of 70 enjoy 10 years less healthy life than 65-year-olds assigned an age of 60?

If so—and especially if the costs of next-generation sequencing continue to decrease with time—an approach like this could be adapted into a kind of ‘molecular physical exam,’ providing patients with vital information about their individual rates of aging, which could in turn inform their decisions about lifestyle, healthcare, and long-term planning. Applied to model organisms, the same strategy could be used to monitor the effectiveness of anti-aging therapies much more rapidly than is possible when using lifespan as a study endpoint.