All biological data changes with age, and enormous sets of such data can be recorded with comparative ease these days. Any sufficiently large set of data can be processed via suitable machine learning approaches in order to produce clocks that correlate with biological age. Some are better than others, some appear to be more sensitive or less sensitive to certain aspects or processes of aging. At the end of the day, these efforts will likely prove useful, but so far they have yet to result in the ability to reliably and rapidly assess a potential rejuvenation therapy for its ability to slow or reverse aging. A clock will always deliver a number, but since the connection between the clock and underlying processes of aging remains unclear, it also remains unclear as to whether that number will in fact usefully reflect changes in biological age produced by a given therapy.
Complexity is a fundamental feature of biological systems. Omics techniques like lipidomics can simultaneously quantify many thousands of molecules, thereby directly capturing the underlying biological complexity. However, this approach transfers the original biological complexity to the resulting datasets, posing challenges in data reduction and analysis. Aging is a prime example of a process that exhibits complex behaviour across multiple scales of biological organisation. The aging process is characterised by slow, cumulative and detrimental changes that are driven by intrinsic biological stochasticity and mediated through non-linear interactions and feedback within and between these levels of organization (ranging from metabolites, macromolecules, organelles and cells to tissue and organs).
Only collectively and over long timeframes do these changes manifest as the exponential increases in morbidity and mortality that define biological aging, making aging a problem more difficult to study than the aetiologies of specific diseases. But aging’s time dependence can also be exploited to extract key insights into its underlying biology. Here we explore this idea by using data on changes in lipid composition across the lifespan of an organism to construct and test a LipidClock to predict biological age in the nematode Caenorhabditis elegans. The LipidClock consist of a feature transformation via Principal Component Analysis followed by Elastic Net regression and yields a Mean Absolute Error of 1.45 days for wild type animals and 4.13 days when applied to mutant strains with lifespans that are substantially different from that of wild type. Gompertz aging rates predicted by the LipidClock can be used to simulate survival curves that are in agreement with those from lifespan experiments.