Researchers have developed a new transcriptomic aging clock that incorporates information about biological pathways. With this approach, the clock predicts not only age but also how the state of various genetic pathways changes with age, providing insight into the transcriptional changes involved in aging.
Time for a new kind of clock
In recent years, the use of molecular clocks to estimate age has become a staple of longevity research. The basic idea is to measure certain molecular features and use machine learning to correlate them with age. Aging clocks have been developed based on transcriptomic, proteomic, and metabolic data, but the most popular are probably epigenetic clocks, which estimate age based on DNA methylation patterns.
By providing accurate, reliable biomarkers of aging, these clocks have given researchers a relatively easy and straightforward method to distinguish between biological and chronological age. They can be used to measure the efficacy of therapeutic interventions. However, while these clocks provide an effective tool for estimating age, they don’t offer any insight into the processes behind aging. The seem to be reliable predictors, but they are poor guides for interpretation.
A change of design
With this in mind, a team of researchers in Germany developed an aging clock that incorporates knowledge about biological pathways into its design. They accomplished this by restricting how the neurons in an artificial neural network can connect to each other. Normally, the neurons can connect to each other freely, meaning that the network can take on any shape. This is part of the reason these networks are such powerful machine learning tools, but it’s also part of the reason why they are effectively black boxes.
By limiting the connections, the team could guide the flow of information through the network. They constructed it to reflect the pathways in 50 highly conserved and refined gene sets in the ‘Hallmark’ pathway collection. They then trained this model for age prediction using transcriptomic data from nearly 900 skin samples.
Peering into the ticking machine
The resulting model was able to predict age with an average error of 4.7 years, and it included information about the state of the different pathways. Using this, the team could evaluate how aging affected each biological pathway. For example, they report that neurons in the p53- and TNFa/NF-kB-signaling pathways were most strongly activated by age, indicating that these pathways play an important role in aging.
The team then validated their model by carrying out various in silico knockdowns and showing that they recapitulated observed changes in aging. For example, decreasing SIRT1 in the network increased the predicted age. Similar patterns held when they replicated more complex transcriptomic signatures, such as the changes involved in Hutchinson–Gilford progeria syndrome or photo-aging. The model also captured the effect of positive interventions, such as caloric restriction.
The network could do more than just replicate known results. When the researchers simulated knocking down each gene in the network, they turned up some new candidates alongside the expected culprits. For example, they found that knocking down the HK2 gene increased aging, and simulated overexpression led to rejuvenation. HK2 expression has been reported to decrease with age, but so far, the gene hasn’t been the subject of intensive longevity research. The discovery of such candidates highlights the value of building a model that accurately reflects the mechanisms underlying aging rather than simply predicting their outcome.
The development of ‘age clocks’, machine learning models predicting age from biological data, has been a major milestone in the search for reliable markers of biological age and has since become an invaluable tool in aging research. However, beyond their unquestionable utility, current clocks offer little insight into the molecular biological processes driving aging, and their inner workings often remain non-transparent. Here we propose a new type of age clock, one that couples predictivity with interpretability of the underlying biology, achieved through the incorporation of prior knowledge into the model design. The clock, an artificial neural network constructed according to well-described biological pathways, allows the prediction of age from gene expression data of skin tissue with high accuracy, while at the same time capturing and revealing aging states of the pathways driving the prediction. The model recapitulates known associations of aging gene knockdowns in simulation experiments and demonstrates its utility in deciphering the main pathways by which accelerated aging conditions such as Hutchinson–Gilford progeria syndrome, as well as pro-longevity interventions like caloric restriction, exert their effects.
Conclusion
As the researchers themselves write, this clock doesn’t perform better than existing aging clocks and might even be a little less accurate. However, it arguably offers significantly more value because it incorporates biological information rather than being optimized entirely for prediction. This makes it possible to interpret the state of the network and examine changes in different components that reflect known biological pathways. As such, it serves as a tool not only to predict the effect of longevity interventions but to understand them and thus to interrogate the processes behind aging.
Literature
Write a comment: