In a preprint published in bioRxiv, Prof. Vadim Gladyshev and a team of researchers have used an artificial intelligence-based system to discover a wide variety of potential interventions, including a drug that significantly improves biomarkers of frailty in mice.
Repurposing previous data
Previous research efforts have created a massive dataset in the form of the Gene Expression Omnibus (GEO), which contains the results of a great many experiments related to potentially disease-modifying drugs, many of which are tissue-specific [1]. These researchers refer to this dataset as a “massive missed opportunity” in aging research, because the vast majority of the experiments in the GEO were unrelated to aging and their data was never investigated in that context.
However, investigating all of that data by hand is practically impossible. These researchers note that modern LLMs can “autonomously generate hypotheses, execute complex analytical pipelines, synthesize findings across multiple data sources, and identify patterns that human researchers might overlook.” Combining that ability with the latest generation of clocks, including causality-based clocks such as AdaptAge, CausAge, and DamAge [2], may yield insights that would have simply gotten lost in the noise.
To that end, these researchers created ClockBase Agent, which uses over two million human and murine samples, including both RNA sequencing and epigenetic measurements, and 40 separate aging clocks. Unlike previous efforts in this area, which used simpler AI systems to simply link compounds to improvements in aging biomarkers, ClockBase is built to exhibit real agentic behavior: it uses an LLM to generate hypotheses about this data, then verifies these hypotheses with more in-depth examinations of both the raw data and the literature from which the data was derived.
Much of the data agrees with existing databases
Unsurprisingly, the clocks showed their natures rapidly. The researchers found that first-generation clocks, which were simply meant to estimate chronological age, were strongly correlated with each other, while healthspan-based clocks such as GrimAge were indeed correlated with healthspan and had data clusters accordingly.
Of a total of 43,529 interventions, which included genetics, diseases, pharmacology, and environment, the researchers’ AI model identified 5,756 that were statistically likely to have age-modifying effects. One was the knocking out of IFR4, which is essential in immune cell differentiation, and another was the knockout of Mettl3, which methylates RNA.
The expression of Bach2, which keeps T cells quiescent, was also associated with reduced aging, as was the overexpression of miR-155, a result that the AI gave an extraordinarily low p-value (2.69 * 10^-10), reflecting very high confidence, and the researchers found surprising due to miR-155’s pro-inflammatory effects. On the other hand, the disruption of hedgehog signaling, which is required for tissue homeostasis, and the knockout of H3K9 methyltransferases substantially increased aging; the latter result is wholly unsurprising due to H3K9’s effects on methylation. Most of its results agreed with the existing GeneAge database, and the few that did not could mostly be explained by the negative, age-increasing effects of knocking out “anti-longevity” genes such as Mtor.
The AI agreed wth the consensus that rapamycin and metformin reduce biological aging. It also found that ouabain, a little-known but established senolytic, also substantially reduces aging according to these clocks, as does the dyslipidemia drug fenofibrate. The immunomoulator Serpina3n was strongly linked to reduced aging, while the immune activator 3M-052 accelerated it. Many of the drugs the model identified are already approved by the FDA; unfortunately, it found that nearly two-thirds of the drugs it identified accelerate aging rather than slow it down. Only five of its results were found in the existing DrugAge database, which agreed with the direction of all five.
This model also found that environmental causes led to biological effects. A combination of mechanical overload, which may reflect exercise practices such as resistance training, along with senolytic administration was substantially associated with reductions in age. Hypoxia, the ischemia-reperfusion injury associated with heart attacks and their treatment, infection with viruses, and some metabolic disorders also accelerated age. Exposing embryos to high-intensity light sources accelerates their aging as well.
Overall, the researchers found that their agent found a substantial amount of both corroborating information and potentially actionable new information, stating that it “reveals a substantial set of new intervention candidates for aging research.” While the AI did make a handful of mistakes in its generation of hypotheses, such as being tripped up on clock age versus chronological age and some issues relating to control groups and treatment groups in complex experiments, its overall results provide an immense potential starting point for further work.
Verifying the AI’s data
The researchers took a crucial step to determine if their model was accurate: they used ouabain, the senolytic that the AI identified as being age-decelerating, in their own experiment with standard, 20-month-old, Black 6 mice. They followed the same protocol as the ouabain experiment that the AI had used to generate its conclusion.
In this experiment, the treatment group was far healthier than the control group after three months of intermittent ouabain exposure. This included metrics of frailty, cognitive ability, and fur condition. Their hearts functioned better, as did the microglia in some but not all brain regions. In total, the AI model had correctly identified ouabain as a potential age-modifying drug.
Of course, this was a murine result published in a preprint paper, and ouabain and many of the other interventions will have to go through further experiments and clinical trials before they can be confirmed as treatments and applied to human beings. The AI’s occasional flaws in reasoning mean that, despite the tremendous advances in this field over the past couple of years, it still cannot be fully relied upon to yield perfectly accurate information. However, it is clearly an invaluable tool in giving researchers critical clues that they would probably never have found without it.
Literature
[1] Edgar, R., Domrachev, M., & Lash, A. E. (2002). Gene Expression Omnibus: NCBI gene expression and hybridization array data repository. Nucleic acids research, 30(1), 207-210.
[2] Ying, K., Paulson, S., Reinhard, J., de Lima Camillo, L. P., Träuble, J., Jokiel, S., … & Biomarkers of Aging Consortium. (2024). An open competition for biomarkers of aging. bioRxiv.
View the article at lifespan.io
