The science of aging clocks
What the field knows, and what FALCONAge implements from it
Every number FALCONAge returns comes from somebody’s paper. This is the record of which paper, what it actually says, and where the published description and the circulating implementation disagree, which they do in eleven documented cases.
It covers what an aging score is, the chemistry and biology behind each input layer, every algorithm family with its equations, the exact transformations and constants published clocks use, the output types and their units, the figures the field reads, and the benchmarking machinery used to judge a clock. It runs to the state of the field in 2026.
Read it as the substrate rather than as the manual. Architecture says which file computes what; this says why that is the right computation, and it is the document to check before trusting a number. Where a section describes something FALCONAge does not implement, it says so, the catalogue here is deliberately wider than v1.0, because the point of writing the field down is to know what is missing.
1. Definitions and framing
1.1 The three quantities
- Chronological age (CA) - time since birth. Years for humans, months or days for rodents, hours for C. elegans.
- Biological age (BA) - the output of a model that maps molecular or clinical features to a number on an age-like scale. In practice BA is whatever a fitted model predicts; its meaning comes entirely from what the model was trained on.
- Age acceleration (AA) - the residual. Two conventions, and they are not interchangeable:
- Absolute age acceleration (AAA), also called delta:
Δ = BA − CA. This is what ComputAgeBench tests (ComputAge benchmarking.py:355). - Relative age acceleration (RAA): residual of
lm(BA ~ CA)in the analysis sample. This removes the systematic regression-to-the-mean bias that all penalised regressions carry. Used by the TwinGene/UKB pipeline asBA_res = residuals(lm(BA ~ age))(Biological-Aging-and-PRS README). - A third convention, used by spatial single-cell clocks, replaces CA with the per-animal per-cell-type mean prediction, so acceleration is measured within an animal rather than against the calendar (SpatialAgingClock deploy.py
get_age_acceleration).
- Absolute age acceleration (AAA), also called delta:
- Pace of aging - a rate, not a level. Units are biological years per chronological year. DunedinPACE and DunedinPoAm are the two published examples; both are centred near 1.0.
1.2 Clock generations
This taxonomy governs everything downstream, what the score means, whether it responds to interventions, and how it must be validated.
| Generation | Training target | Examples | Behaviour |
|---|---|---|---|
| First | Chronological age | Horvath 2013, Hannum, Lin, Zhang EN/BLUP, AltumAge, PedBE, SkinAndBlood | High age correlation, weak or absent mortality signal. Zhang BLUP has RMSE ≈ 2.0 y yet its acceleration no longer predicts mortality, the “oracle paradox”. |
| Second | Mortality or a composite phenotype | PhenoAge, DNAmPhenoAge, GrimAge, GrimAge2, CpGPTGrimAge3, SystemsAge, OMICmAge | Lower CA accuracy, much stronger mortality/morbidity association. |
| Third (“pace”) | Longitudinal rate of change across many organ systems | DunedinPoAm38, DunedinPACE | Rate, responds fastest to interventions. |
| Causal | CA, but with features filtered by Mendelian-randomisation evidence of causality | YingCausAge, YingDamAge, YingAdaptAge | Separates damaging from adaptive/protective methylation. |
| Mitotic | Stem-cell division count / replicative history | epiTOC1, epiTOC2, epiTOC3, stemTOC, HypoClock, epiCMIT-hyper/hypo, RepliTali, MiAge | Cancer risk, proliferative history. |
| System- and organ-specific | Mortality or CA, restricted to organ-enriched features | SystemsAge (11 systems), organAging (24 organ protein panels), MEAT (muscle), CorticalClock (brain), scImmuAging (5 PBMC cell types) | Reveals within-person heterogeneity. |
| Foundation-model | Self-supervised pretraining, then fine-tuned | CpGPT, methylGPT | Cross-platform imputation, SOTA mortality prediction. |
1.3 Biomarker validation taxonomy (Biomarkers of Aging Consortium)
Moqri et al. classify aging biomarkers as predictive (associates with future outcomes), responsive (moves under an intervention), and biological (has a mechanistic, causal claim). A clock that is only predictive is a risk score; only responsive is a surrogate endpoint; a biological biomarker needs causal evidence. The consortium’s Biolearn library and the 2024–25 open competition (chronological age, mortality, multimorbidity; >$200k prizes) operationalise this. Any new tool should record which of the three claims each clock supports.
1.4 Why this matters commercially and scientifically
The geroscience hypothesis holds that aging is the dominant risk factor common to cancer, cardiovascular disease, T2D, Alzheimer’s and Parkinson’s, so slowing aging delays all of them. A validated surrogate endpoint shortens trials that would otherwise need decades of mortality follow-up. TAME (metformin) has raised ~$11M of $40M over seven years with a six-year observation window; a responsive clock is the alternative (computational_aging_course/ml/aging_clocks.md).
Butler’s criteria for a usable aging biomarker: changes with age; predicts mortality better than CA alone; predicts early-stage age-related disease; minimally invasive. Moskalev adds: sensitive early, predictable in a practical time window, robust.
References
The bibliography moved to its own page, where it is numbered once and cited by number from every page of this site. It reads the same in the PDF, where it is bound as the last chapter.
Citations in this document are the bracketed numbers, for example [5] for Horvath 2013. The numbering is global, so a source keeps its number wherever it appears.