WITH AGE, CELLS STOP ANSWERING IN UNISON

A leading idea in the biology of aging is that random molecular damage — mutations in our cells’ DNA and chemical marks gone astray — builds up over a lifetime and gradually upsets how genes are switched on and off. Biologists call it transcriptional deregulation. The puzzle is that this damage is random: each cell accumulates its own, different pattern. How does such scattered damage add up to the recognisable decline of aging?

Looking for the right signal

Genes are controlled by networks of other genes, which coordinate their activity and buffer disturbances. If aging erodes these networks, damage that starts local could spread. But measuring that erosion is hard. Looking at how much individual genes vary from cell to cell gives mixed results: a systematic comparison even found that this “transcriptional noise” does not consistently rise with age. The same group had earlier shown that coordination between genes declines with age overall, but that global measure could not say which genes or which cells were drifting.

Yongjian He, Orr Levy and Julia Hanna Haddad Guez of Bar-Ilan University in Israel, with Xu-Wen Wang of Brigham and Women’s Hospital and Harvard Medical School, asked a different question. If the regulatory networks of old cells grow more different from each other, then switching off the same gene should produce increasingly different effects from one cell to the next. Doing that for real, across many genes and thousands of young and old cells, is not feasible at scale.

A knockout in silico

So they used a deep-learning tool they had developed, DKO. For a given cell type, a neural network learns, from ordinary single-cell data, how the set of genes a cell expresses relates to its full activity profile. Once trained, it can remove one gene from a cell and predict the new profile. The difference between before and after measures that gene’s impact in that cell. Repeated over cells, it shows how consistent the impact is.

First, a sanity check on simulated gene networks deliberately damaged at random. As damage increased, the true effects of exact knockouts became more variable between cells — and DKO, seeing only the undamaged-looking data, predicted the same rise.

Old cells drift apart

Then the real data: human ovaries (women aged 23–29 versus 49–54), human skeletal muscle (15–20 versus 70–75 years), mouse bone marrow and heart (3 versus 24 months), C. elegans worms (day 3 versus day 11) and fruit-fly brains (day 0 versus day 30).

Across 45 comparisons of tissues and cell types, the effect of knocking out the same gene varied more among old cells in 41, significantly so in 40. In four, it varied less. The direction was not universal, the authors note, but the dominant pattern was clear. It held even when the network was trained only on young cells, or only on old ones.

Two hallmarks of aging went the same way. In pancreatic cells from five people, the cells carrying the most of an age-related mutation signature had the most divergent responses, in four of the five; in the fifth, mutation levels were too uniform to tell. Similar results came from heart cells of old mice. And in two models of cellular senescence — the state reached by cells that have stopped proliferating — responses grew more scattered as cells moved from proliferating to senescent, and, within the same group, the most senescent cells strayed the most.

A rulebook coming apart

The authors read this as a progressive loss of regulatory robustness, at the level of single genes and single cells, shared by species as distant as humans and worms. Their proposed picture: mutations accumulate and push cells apart, while senescence may be one consequence of that lost stability.

They are careful about what this is not. The knockouts are model predictions, not measurements in those cells. The links with mutations and senescence are associations; they do not prove what causes what. If random damage really drives the drift, they suggest, treatments aimed at a single aging mechanism may fall short, and preserving the robustness of gene control itself could matter. The next step belongs to the laboratory: testing, gene by gene, whether old cells truly answer out of tune.