Skip to main content
Back to timeline
bioRxivSource publication:

An aging clock gives one age; scCody splits disease displacement into aging-aligned and off-trajectory parts

Synopsis

Supervised aging clocks return a single age estimate and cannot tell whether a disease-associated change follows healthy aging or departs from it. scCody learns age-conditioned population dynamics from cross-sectional single-cell cohorts using population-level optimal transport with an endpoint-consistency regularizer, reconstructs populations at unseen ages, and recovers age-associated differential expression more accurately than the compared continuous-time generative models. Against the learned healthy-aging reference, selected neurons in Alzheimer's disease show substantial aging-aligned remodeling, whereas COVID-19 shows weaker aging alignment and more pronounced orthogonal immune remodeling.

Source-provided article image: scCody distinguishes aging-aligned from divergent disease-associated cellular remodeling
Figure 1 · bioRxiv · Page 5

Interpretation

Disease-associated cellular change can now be split by direction: one part along healthy aging, another perpendicular to it. Single-cell aging clocks previously compressed molecular change into one predicted age, so a positive age gap did not imply accelerated healthy aging and a near-zero gap did not exclude substantial disease dysregulation. In Alzheimer's disease, aging-aligned displacement in excitatory neurons reached 2.7-fold the ten-year healthy-aging reference; in COVID-19 aging-aligned displacement was weaker across all examined cell types and globally, while orthogonal displacement rose everywhere except platelets.

The approach first learns a healthy-aging population trajectory, then projects disease displacement onto it. Cross-sectional cohorts sample different individuals at different ages, so observed differences mix aging with donor-specific variation, and cells have no known correspondence across ages. The model was trained on the AIDA healthy cohort of 625 donors and about 1.27 million peripheral blood mononuclear cells, with all cells from donors aged 21, 31, 41, 51, 61 and 68 years held out as external unseen ages.

The orthogonal part is further divided into shared, between-patient and within-patient layers. Disease-associated change includes shared responses plus heterogeneity between and within individuals, which a single displacement magnitude cannot separate. In COVID-19 the between-patient standard deviation of the interferon-alpha/beta program was 6.23-, 4.47- and 5.53-fold the healthy reference in the Ren, COMBAT and Stephenson cohorts, while within-patient dispersion rose only 1.12- to 1.30-fold.

Perspective

The demonstrated scope is a healthy peripheral-blood immune-aging trajectory and displacement decomposition on Alzheimer's disease brain single-nucleus data and three COVID-19 peripheral-blood cohorts; the authors propose the reference could help evaluate interventions targeting aging but report no validation experiment.

The authors note that trajectory-based age inference remains less accurate than directly supervised aging clocks and that population-level optimal-transport estimates are affected by finite sampling; some aging-program trajectories are non-monotonic at older ages, which the authors partly attribute to each cohort's age distribution and sparse support for older donors in the training set.