Age of individual mouse cells can be estimated from gene activity

Age of individual mouse cells appears to be less straightforward than the age of the animal itself. In the kidney of an old mouse, for example, some cells had gene-expression profiles resembling those normally seen in substantially younger animals. Conversely, cells from young mice could sometimes resemble an older age group.

That does not mean one cell was literally born months before another. The new method, called scMLEAge, estimates a transcriptomic age: how closely the pattern of genes being used by an individual cell resembles cells from animals in a particular age group.

This makes it possible to ask a more interesting question than simply how old an organism is: do all of its cells age together?

One mouse, multiple cellular “ages”

The study used Tabula Muris Senis, a large single-cell RNA-sequencing atlas of mouse aging. The original resource contains more than 350,000 cells from 23 organs and six age groups spanning 1 to 30 months.

The researchers built models for cell types from the bladder, bone marrow, brain, heart, kidney, limb muscle, liver and lung. Each cell type received its own clock. T cells were compared with T cells, kidney tubular cells with the same kidney cell type, and so on.

That distinction matters because aging does not necessarily leave the same molecular signature in a muscle stem cell as it does in a kidney cell.

Kidney proximal-tubule cells provided one particularly clear example. The model separated age groups reasonably well, but it also exposed substantial variation within them. Some cells from 30-month-old mice were assigned a profile closer to the 18-month group. At the other extreme, some cells from three-month-old mice also resembled the 18-month state.

This does not establish that those cells were biologically younger or older by exactly that number of months. It shows that cells sharing the same chronological age can occupy markedly different transcriptional states.

A similar pattern appeared in skeletal-muscle satellite cells, the stem cells involved in muscle regeneration. Their model reached an R² of about 0.81 in both training and testing. Older cells did not simply collapse into a single uniform group; instead, their expression patterns formed an age-related gradient.

Reading age from RNA

Single-cell RNA sequencing records which genes are being used in each cell and approximately how strongly they are expressed.

scMLEAge first learns from cells collected from mice of known chronological ages. For each cell type, it builds an expected pattern of gene-expression counts for the available age groups. When a new cell is examined, the algorithm calculates which age group makes its observed RNA counts most likely.

Rather than relying only on conventional linear regression, the framework models raw count data probabilistically. That is useful for single-cell sequencing because each cell contains relatively few measured RNA molecules and many genes register zero counts, producing highly sparse data.

Performance varied substantially across cell types, with R² values ranging from roughly 0.17 to 0.95. In other words, there is no single clock that performs equally well everywhere. In benchmark tests involving muscle, kidney and lung cells, scMLEAge outperformed ElasticNet and Lasso in most of the tested cell types.

The approach also highlighted genes whose activity tracked with predicted cellular age. These included genes involved in immune responses, ribosomal function, extracellular matrix biology and tissue remodeling. In muscle satellite cells, for example, COL6A1 expression declined with age, while kidney models highlighted age-associated changes involving PCK1 and CD74.

What does the “age” of a cell really mean?

There is an important conceptual limitation. The true biological age of an individual cell is not known in advance. The model therefore has to be trained using the chronological age of the mouse from which each cell was collected. It can then identify cells whose expression profiles depart from what is typical for that chronological age.

For that reason, scMLEAge is better viewed as a way of positioning cells along a transcriptomic aging trajectory than as a stopwatch that reveals a cell’s literal biological age.

The current implementation also predicts discrete age groups rather than a continuous age and was tested only on male mouse data. Its performance in female animals, independent datasets and especially human tissues remains to be established.

Even so, the level at which aging can be examined is changing. A tissue no longer has to be reduced to one average molecular age. Researchers can begin asking which cells inside it look older than their neighbors and which retain a comparatively youthful transcriptional state.

An organism has one date of birth. Its cells, judging by their gene activity, may not all move through aging in synchrony.