https://www.sciencedirect.com/science/article/pii/S156816372500217X
This paper was recently promoted by one of its authors which is where I picked it up from. I am not persuaded that methylation clocks have the accuracy to be able to distinguish between noise and small changes in the ageing pathway, but still:
chatGPT(5.5paid):
The uploaded paper is a 2025 Ageing Research Reviews article, “From clock to clock: Therapeutic target discovery for aging and age-related diseases” by Chen et al. It is primarily a review/perspective, but it also contains a small original analysis testing whether genes selected by aging clocks are enriched for known drug targets.
Summary
The paper makes two closely connected arguments.
First, the authors advocate a “dual-purpose” drug-development strategy: rather than finding a drug for one age-related disease in isolation, identify targets that participate both in fundamental aging biology and in one or more specific age-related diseases. The hope is that a successful drug could subsequently be developed across several indications. Rapamycin and metformin are presented as existing examples of compounds that affect broad aging-related pathways while also having disease-specific applications.
The authors propose the hallmarks of aging as one way of finding such targets. A candidate associated with several hallmarks is argued to be more likely to sit at a biologically important convergence point. They use two examples:
- TNIK, associated with six hallmarks including inflammation, senescence, nutrient sensing, genomic instability, stem-cell exhaustion and altered intercellular communication.
- Apelin/APLNR, linked to eight hallmarks including autophagy, inflammation, senescence, telomere attrition and proteostasis.
TNIK is particularly prominent because Insilico Medicine used its AI platform to identify TNIK as a target in idiopathic pulmonary fibrosis and developed the inhibitor INS018_055. The authors report that target discovery to preclinical candidate nomination took about 18 months and cost under $3 million.
The second and more interesting argument is that aging clocks should not merely be regarded as biomarkers; their constituent features may be useful for identifying therapeutic targets.
They review three main clock classes:
Epigenetic clocks — Horvath, Hannum, PhenoAge, GrimAge/GrimAge2, DunedinPACE, pan-mammalian clocks and others. These infer age, mortality risk or aging rate from DNA methylation.
Transcriptomic clocks — clocks based upon bulk or single-cell gene-expression patterns. These are more dynamic and potentially more mechanistically informative, though also noisier.
Multi-omic/AI clocks, especially Precious1GPT/2GPT/3GPT. Precious1GPT combines age prediction with disease-versus-control classification and uses SHAP feature importance to identify genes relevant to both aging and disease.
The paper’s main quantitative analysis
The authors take 11 recently published aging clocks, extract their clock-associated genes, filter these for druggability and safety, and then ask:
Are known drug targets unusually frequent among genes selected by aging clocks?
Their baseline is the druggable genome:
- 5,870 druggable/safe genes
- 990 are known drug targets
- therefore 16.9% of the druggable genome consists of known drug targets.
Across the clock-associated druggable genes, however, the average was about 27% known drug targets. Five of the eleven clocks showed statistically significant enrichment at P < 0.05 and another four had nominal/borderline enrichment at 0.05–0.1.
Examples include:
| Clock | Known targets / druggable clock genes |
|---|---|
| Single-cell transcriptomic | 6/25 = 24% |
| BiT age | 12/40 = 30% |
| GrimAge2 | 41/195 = 21% |
| DunedinPACE | 8/30 = 26.7% |
| Pan-mammalian clock A | 19/58 = 32.8% |
| Pan-mammalian clock B | 36/125 = 28.8% |
| Pan-mammalian clock C | 37/111 = 33.3% |
The pan-mammalian clocks show particularly strong enrichment, several with P < 0.001.
Interestingly, TNIK and APLNR also emerge independently from aging-clock analyses, giving some convergence between the hallmark-based approach and clock-based target identification.
The paper then proposes additional uses for clocks in drug development: patient stratification, disease monitoring, measuring treatment response, and eventually creating a feedback loop in which clock-derived targets are tested in trials and successful interventions in turn validate or improve the clock.
Finally, the authors emphasize an explicitly commercial aspect of the strategy: discovering new dual-purpose compounds early enough that there is substantial patent life remaining, allowing subsequent indication expansion.
What is genuinely novel?
I would separate the novelty into three levels.
1. The strongest novelty: using aging-clock features as a target-discovery dataset
The idea that aging clocks contain biologically informative features is not new, and the paper explicitly cites previous attempts to derive targets from them.
What is more novel is the systematic question:
If you take genes selected independently by different aging clocks, are they disproportionately already validated pharmaceutical targets?
The answer is apparently yes: about 27% versus 16.9% in the background druggable genome.
That is probably the most substantive contribution of the paper.
It turns aging-clock construction on its head:
conventional use:
molecular features → predict aging
versus
proposed use:
molecular features that predict aging → prioritize biological interventions.
That is conceptually useful.
2. Combining disease prediction and aging prediction to discover “dual-purpose” targets
Precious1GPT is particularly interesting conceptually because it is not simply asking which genes predict age. It first learns an aging representation and then incorporates disease-control classification.
That makes the underlying question something closer to:
Which molecular features encode both aging and disease?
rather than merely:
Which molecular features correlate with chronological age?
This is much closer to a therapeutic-target problem.
3. A drug-development strategy rather than merely a biological-aging strategy
The paper also links:
aging clocks → target identification → druggability filtering → disease indication → clinical trials → clock-based pharmacodynamic measurement → indication expansion.
Figure 1 explicitly presents this as an alternative commercial drug-development pipeline rather than just a geroscience research methodology.
That synthesis is useful even though none of its individual components is completely new.
Critique
There is a potentially important idea here, but I think the paper substantially overstates what the enrichment analysis demonstrates.
1. Prediction is not causation
This is the central problem.
An aging clock is optimized to find features that predict age or an age-related phenotype.
That does not mean those features cause aging.
For example, a gene may change because of:
- inflammation,
- changes in immune-cell composition,
- declining renal function,
- hormonal changes,
- medication exposure,
- accumulated disease,
and therefore be an excellent predictor of age while being a poor intervention point.
The causal possibilities are:
Target → aging
but also:
aging → target
or:
third process → both aging and target.
Feature importance, regression coefficients or SHAP values cannot distinguish these.
Consequently, the paper’s key transformation—
clock feature → therapeutic target
—requires considerably more causal evidence than is provided.
2. Recovering known drug targets is not necessarily evidence that the clock identifies aging mechanisms
This is the biggest methodological weakness of their quantitative result.
The comparison is essentially:
16.9% known targets in the druggable genome
versus
~27% among druggable clock genes.
That is interesting, but there are many possible reasons unrelated to causal aging biology.
Genes that are:
- highly expressed,
- extensively studied,
- implicated in inflammation,
- extracellular or membrane-associated,
- involved in common diseases,
- represented well on methylation arrays,
are both more likely to become drug targets and more likely to appear in aging datasets.
Thus there may be a strong annotation/study-intensity bias.
The analysis would have been much stronger if the authors had constructed matched controls for:
- expression level,
- gene length,
- number of publications,
- disease-association count,
- protein class,
- tissue expression,
- GWAS evidence,
- network connectivity,
- methylation-array coverage.
Without such matching, the enrichment does not demonstrate that the aging-clock information itself is responsible for the enrichment.
3. There is a possible circularity in the filtering procedure
They first select clock features and then apply druggability and safety filters before testing enrichment for known drug targets.
Those filters are not independent of existing pharmacological knowledge.
If a protein already belongs to a well-understood druggable protein family—kinases, GPCRs, ion channels, proteases, etc.—it is simultaneously:
- more likely to pass a druggability filter, and
- more likely already to be a known drug target.
That can inflate the apparent enrichment.
A better test would compare clock-associated genes against matched equally druggable genes and ideally predict future targets that were not known when the clock was built.
That latter prospective test would be particularly persuasive.
4. Different clocks are being treated as more comparable than they really are
The models have radically different structures.
Some select:
- CpG sites,
- individual genes,
- expression features,
- SHAP-important genes,
- cross-species methylation loci.
The authors then map these heterogeneous representations onto gene lists and compare them.
But a CpG associated with a nearby gene does not necessarily regulate that gene; enhancer-promoter relationships can be long-range, and methylation may simply be a marker rather than a regulator.
So:
CpG → nearest/mapped gene → drug target
is often a fairly weak biological inference.
That issue is especially important because several of their strongest enrichments come from epigenetic clocks.
5. Chronological-age clocks may enrich for consequences of aging rather than drivers
This distinction deserves more emphasis than the paper gives it.
Suppose expression of gene X steadily increases because damaged cells accumulate with age.
X may become an extremely strong clock feature.
Blocking X might:
- slow aging;
- have no effect on aging;
- remove an adaptive response and actually worsen aging.
The clock alone gives essentially no way of distinguishing these possibilities.
A useful target-discovery system therefore needs to combine clocks with causal information such as:
genetics/Mendelian randomization + perturbation experiments + longitudinal trajectories + intervention response + tissue-specific biology.
The paper mentions integration with target-identification systems but does not fully solve this problem.
6. “Multiple hallmarks” is not necessarily evidence for a superior target
The TNIK and apelin examples sound persuasive because the proteins can be connected to six or eight hallmarks.
But the hallmarks are highly interconnected.
A regulator of inflammation, for example, can readily generate secondary links to:
senescence → stem-cell dysfunction → altered communication → proteostasis → nutrient signalling.
Therefore, counting the number of hallmarks with which a protein is associated risks producing a connectivity/popularity score rather than identifying an upstream causal mechanism.
A highly pleiotropic target may also be less attractive pharmacologically because perturbing it may cause effects across many physiological systems.
Rapamycin itself illustrates that tradeoff quite nicely in the paper.
7. The paper mixes a scientific hypothesis with an Insilico Medicine commercial case study
This deserves explicit attention.
All authors are affiliated with Insilico Medicine, and the conflict declaration states that Insilico is a commercial generative-AI company possessing several hundred patents, patent applications and commercial software.
Several flagship examples in the review are also Insilico technologies or programmes:
- PandaOmics
- Chemistry42
- PreciousGPT
- TNIK
- INS018_055.
That does not invalidate the work, but it means that claims such as the superiority of the AI-powered development pathway, dramatic cost/time reductions and the value of clock-derived target discovery should ideally be independently replicated.
Figure 1 in particular should be read partly as a strategic/commercial model, not as a controlled empirical comparison of AI versus conventional drug discovery.
8. The paper’s strongest evidence is retrospective
They essentially ask:
Do clock genes contain targets we already know are drugs?
A much stronger demonstration would be:
- construct the aging clock;
- freeze the model and target predictions;
- nominate previously unrecognized targets;
- perturb them experimentally;
- demonstrate reversal of multiple independent measures of aging;
- show improvement in functional outcomes or lifespan.
The crucial validation is therefore prospective, not retrospective.
Until then I would describe the paper’s result as evidence that aging clocks contain pharmacologically interesting biology, rather than evidence that they can reliably discover anti-aging targets.
One particularly important conceptual issue
I think the paper could have made a much stronger distinction between:
a biomarker of aging
and
a control point of aging.
Those are very different things.
An analogy would be a clock on a car dashboard. Engine temperature may predict when an engine is failing, but attacking the temperature sensor is not necessarily a way to repair the engine.
The valuable clock features for therapeutics are therefore not necessarily those with the largest predictive coefficients. They are the subset for which there is evidence that perturbing the feature changes the underlying aging trajectory.
That distinction is the major gap between the paper’s attractive computational idea and a convincing therapeutic-discovery methodology.
Overall assessment
I would rate the paper as conceptually interesting and useful, but stronger as a drug-discovery perspective than as evidence that aging clocks already constitute a validated target-discovery technology.
Its most interesting finding is the ~27% versus 16.9% enrichment of known drug targets among clock-associated druggable genes. That is enough to justify further investigation.
But the decisive next experiment is not another clock. It would be a prospective perturbational study showing that targets selected uniquely by the clock actually modify aging phenotypes better than equally druggable targets selected without the clock.
If that worked, the central claim of this paper would become considerably more compelling.