Harvard professor Vadim Gladyshev proposes screening drugs for aging by their effects on biological age and the state of the organism; in his group’s preprint, five of ten selected compounds extended lifespan in old mice
Harvard professor Vadim Gladyshev proposes screening drugs for aging by their effects on biological age and the state of the organism; in his group’s preprint, five of ten selected compounds extended lifespan in old mice
On October 3 at the ARDD conference on aging and drug discovery, Gladyshev presented three conclusions about disease, drugs, and AI. According to Gladyshev, the preprint by Anastasia Shindyapina and Alexander Tyshkovskiy is under peer review.
A mouse lives for two to three years; some whales live for more than two hundred. Gladyshev’s group studies gene activity to understand how nature produces this difference. The researchers call the changes associated with long life longevity signatures, and in 2023 they described these signatures in 41 mammalian species. The signatures of long-lived species resemble one another, as do those of interventions that extend mouse lifespan, such as calorie restriction and rapamycin. The two sets, however, show no relationship to each other. Gladyshev suggests that laboratory interventions act on individual molecular pathways, whereas nature alters many processes at once.
Because lifespan experiments are expensive, the researchers first screened compounds using these signatures. The Connectivity Map database records how more than four thousand compounds change gene activity in cells. The compounds with the closest matches were tested in human liver cells, and 25 compounds were given to young male mice for a month, with gene activity measured in the liver and kidneys. The nine compounds with the highest combined scores across the three tests were then given to mice aged 24–25 months. The tenth was KU0063794, which had been identified through the signatures and tested in mice in 2023.
Four inhibitors of proteins that regulate cell growth (KU0063794, AZD8055, LY294002, selumetinib) and the natural anti-inflammatory compound celastrol extended lifespan in old mice. LY294002 and AZD8055 extended lifespan only in males. The authors describe the proportion of successful compounds as unusually high, but report that a second series of experiments with other compounds is proving substantially less successful.
Aging clocks estimate a tissue’s biological age using models that predict age and mortality risk from gene activity. In mice, mortality-risk clocks distinguish, on average, between interventions that shorten lifespan and those that extend it. According to Gladyshev, they help researchers select compounds before conducting lifespan experiments. AI agents applied these clocks to publicly available data and identified more than 500 interventions that reduce biological age. One of them, ouabain, slowed the progression of age-related frailty in old mice.
In rodent disease models, clocks estimate diseased tissue to be older than healthy tissue, with liver cancer as the exception. Gladyshev extends this interpretation to nearly all age-related diseases, with cancer as a possible exception. An organ that ages earlier than others interacts with the rest of the body: a second heart transplanted into a mouse shifts toward the host’s age. The organ is therefore both older and altered, a state Gladyshev calls disease. “For me, there are no diseases. There is only aging. Different systems, subsystems, and components age,” he says, adding that he does not mean this literally.
Gladyshev proposes adding a second approach to the conventional strategy of finding a target and developing a drug: “move from searching for targets to searching for biological states.” This would involve assessing, for example, how a drug changes biological age. His reasoning is that aging is a systemic process with no single primary cause. “There are no drivers of aging. Aging happens because we live.” He says language models generally use “the human interpretation of biology, rather than biology itself.” Training models on the underlying data therefore requires “data factories” that generate paired measurements of different types across species, over time, and after interventions.