IMMClock

A person's T cell compartment ages in two ways at once. Naïve cells give way to memory and effector populations, changing what the compartment is made of, while the cells that stay in a given state accumulate transcriptional changes of their own. A clock fitted to pooled PBMC expression reads only the sum, so a donor whose composition has shifted looks identical to one whose individual cells have aged.

IMMClock (IMMune Cell cLock) separates them by construction. Rather than one clock for blood, it fits an independent age model for each immune cell type, and within T cells for each differentiation state, then combines the intrinsic and compositional signals into an integrated clock. Because the fit is per cell type, it scores a single cell as readily as a person: averaging one cell type's per-cell predictions recovers the person-level estimate. It was developed at the Cancer Data Science Laboratory of the National Cancer Institute and posted as a preprint in November 2024.

Earlier single-cell aging clocks were fitted on small bespoke cohorts, and DNA methylation clocks remain hard to run at single-cell resolution because whole-epigenome profiling of individual cells is sparse. IMMClock instead trains on a cohort two orders of magnitude larger than prior single-cell clocks and then runs unchanged on new data. Where Cross-Species Ageing DANN scores immune cells as young or old, IMMClock returns a continuous age.

#Key Features

• One clock per cell type: Separate models for CD8⁺ T cells, CD4⁺ T cells, NK cells, B cells and monocytes, so an estimate is never averaged across populations that age at different rates.
• Intrinsic and compositional clocks combined: Four variants per T cell type — naïve and central memory, effector memory, subtype abundance, and an integrated model — separate aging within a cell state from shifts between states.
• Applied without fitting to the target data: Trained once on the reference cohort, then run on external scRNA-seq cohorts, on bulk whole-blood RNA-seq, and on perturb-seq cells that have no chronological age to fit against.
• Gene-level weights: Elastic-net coefficients are projected from principal-component space back into gene space, giving a signed per-gene contribution per clock.

#Technical Details

Expression is z-scored log-CPM and chronological age is Box-Cox transformed as the regression target. Genes are filtered to those significantly age-associated (Bonferroni-corrected p < 0.01), then balanced between the top 50% of positive and negative correlates; PCA retains 80% of the variance, and a scikit-learn elastic net is fitted over the resulting components, with selection and PCA computed on training data only. Training uses OneK1K: roughly 1.3 million PBMCs from 982 healthy donors aged 19 to 97, as cell-type pseudo-bulk per person. Under repeated stratified five-fold cross-validation the integrated clock reaches person-level Spearman correlations of 0.80 for CD4⁺ T cells, 0.68 for CD8⁺ T cells and 0.56 for NK cells; B cells (0.44) and monocytes (0.28) were dropped. On six independent PBMC single-cell datasets covering 823 CD8⁺ and CD4⁺ T cell samples, correlations range from 0.56 to 0.88. Each external run restricts the fit to genes shared with the target panel, so the clock is a training recipe over a fixed corpus rather than one universal coefficient vector.

#Applications

The clock turns a transcriptome into an immune-age readout for cohort and screen analysis. In the Framingham Heart Study — bulk whole-blood RNA-seq from 2,691 individuals — predicted immune age tracks chronological age (ρ = 0.56–0.59), rises with C-reactive protein, and is elevated in people with a history of cancer, heart disease, stroke or type 2 diabetes, after adjusting for age and sex. At single-cell resolution, higher predicted age accompanies senescence and exhaustion signatures, and tumour-infiltrating CD8⁺ lymphocytes score older than matched circulating T cells from the same patients. Across a CRISPRa perturb-seq screen of 70 genes in primary human T cells, mean immune age per perturbation correlates at ρ = −0.85 with the functional activation score, and immune age measured at rest predicts activation after restimulation.

#Impact

IMMClock's contribution is a per-cell immune-age measurement trained at cohort scale and validated by transfer, which lets it reach settings — bulk cohorts, perturbation screens — where no age label exists to fit against. The link it draws between immune age and T cell activation is the result most likely to be built on, since it suggests screening for perturbations that rejuvenate and activate T cells at once. Several caveats are load-bearing. The work remains a bioRxiv preprint, and neither code nor a fitted model has been released: the code availability statement promises a Zenodo deposit after publication, and none has appeared. Training is cross-sectional, limiting its use for tracking an individual over time, and the training donors come from a single Australian cohort, with ancestry breadth entering only through external validation.