A team of Skoltech researchers has shown that the currently available models for testing the effects of cellular reprogramming techniques that make old human cells young again are inadequate to the task. In their recent study in Aging Cell, the researchers call for revised aging clock models that would indicate the uncertainty in biological age estimation. While this could initially lead to a more cautious reappraisal of rejuvenating effects, the resulting new appraisals will have the virtue of being closer to reality.

Aging clocks are models used to study what causes aging and how to reverse it. They estimate the biological age of a person based on the results of blood tests, advanced DNA sequencing, and other studies. The clocks are therefore central to assessing the effects of any agents and manipulations that promise to rejuvenate cells, restoring them to a youthful state.

Among the most widely publicized rejuvenation approaches is cellular reprogramming. The idea is that a harmless virus delivers an RNA fragment containing four genes — known as Yamanaka factors — into the cell. There, molecular machines called ribosomes use the RNA as a blueprint to assemble proteins, which then enter the nucleus of the cell and activate the four corresponding genes in the cell’s DNA. In at least one type of skin cell, this has been shown to reliably induce transformation over the course of several weeks into a stem cell (which can turn into any other kind of cell), accompanied by rejuvenation, as measured by aging clocks.

The rationale behind such rejuvenation involves multiple processes that are intensified by the transition to a stem cell: autophagy cleans out damaged or redundant cellular components. DNA repair fixes certain single-letter errors in the genetic code. Besides that, fewer immune agents causing inflammation are released, and more new cells are produced by cell division.

But while various rejuvenation techniques do seem promising, validating and comparing their efficiency often relies on biological aging clocks. Crudely speaking, by measuring cell age before and after the supposed rejuvenation, and subtracting one from the other, you can know which approach works better, or indeed works at all. But are our aging clocks accurate enough for this?

A study by Skoltech researchers, which came out July 28 in Aging Cell, looked into the uncertainty inherent in all existing aging clocks and concluded that it is way too high to make these models useful for validating rejuvenation effects in cellular reprogramming.

Aging clocks tend to corroborate the effectiveness of cellular reprogramming, churning out near-zero and sometimes negative age estimates post-rejuvenation, but the underlying machine training models have overwhelmingly been trained on regular, unrejuvenated cell data and are therefore — ironically — unfit for gauging the age of old cells supposedly made young again, which are after all an oddity not found in nature.

“I like to compare this to a neural network trained to differentiate between the photos of cats and dogs. Trained on a vast array of data, it might exhibit near-perfect judgment on the matter, but as soon as you feed an artist’s rendition of a cat or dog to it, the system is out of depth. It might still make the correct guess, but such ‘out-of-domain’ input introduces a fair amount of uncertainty. Now, if you go as far as to submit a photo of a sofa or something, the system will still say ‘cat,’ or it will say ‘dog,’ because these are its only outputs. Obviously, the uncertainty in this prediction is so high as to render it meaningless,” said Dmitrii Kriukov, the lead author of the study, a Skoltech PhD student of the Life Sciences program and a research scientist at AIRI (Artificial Intelligence Research Institute).

According to the researcher, the impact of Yamanaka factors and comparable agents on the cell is so unprecedented as to render the rejuvenated cell radically out of domain for current aging clock models. In fact, even some of the pre-rejuvenation inputs turn out to introduce reasonably high uncertainty. For example, if the model has been trained on data from individuals of one ethnicity and is applied to those of another, the uncertainty in age estimation might go from give or take one year to, say, several years. By the same token, the uncertainty for rejuvenated cell age increases so sharply that any conclusions drawn about reprogramming techniques become shaky.

“We therefore propose incorporating this whole notion of uncertainty into aging clock models so that the algorithms would label their own predictions in a way that makes it clear which inferences from them are sound or otherwise,” said the study’s principal investigator, Associate Professor Ekaterina Khrameeva of Skoltech Bio.

She added that the challenges exposed by the team’s research also indicate that entirely new approaches to evaluating biological age are called for: “A conventional aging clock attempts to accurately predict the number of years lived. What you could do instead is build a model predicting the number of years remaining to be lived.” While this might sound unsettling, given enough data, such an algorithm would actually stand a chance of accurately comparing the effects of potential rejuvenating agents.