Leakage at the Donor Level Reshapes Aging Predictability Across Tissues in Single-Cell Atlases
Year: 2026
Type: article
Abstract: Single-cell RNA sequencing atlases provide a powerful resource for quantifying molecular aging across tissues, but predictive analyses are vulnerable to donor-level leakage when multiple samples derived from the same individual are split across training and test sets. Here, we define AtlasEva as a leakage-aware evaluation framework and use it to systematically audit how donor-level leakage affects age-predictability estimates in the Tabula Muris Senis atlas. AtlasEva organizes three coordinated outputs: a donor-held-out estimate, an intentionally ungrouped diagnostic, and a mouse-level permutation null. Using mouse-level pseudobulks and simple supervised baselines, we show that evaluation design can alter not only absolute performance but also tissue-level biological conclusions. Across 18 evaluable tissue tasks, ungrouped evaluation changed balanced accuracy and reshuffled tissue rankings, with a maximum positive difference of 0.778 and low concordance between grouped and ungrouped rankings. Direct same-mouse reuse was structurally possible in 11 tasks; differences in the other seven reflected fold composition rather than donor reuse. A striking example was bladder, where apparent predictability under ungrouped splitting collapsed under mouse-held-out evaluation. Conversely, several tissues retained age-predictive signal under donor-held-out evaluation, including brain, marrow, lung, and mesenteric adipose tissue. In brain-derived pseudobulks, age-associated signal was linked to disease-associated microglial and interferon programs. These results show that evaluation design—and donor-level leakage when repeated observations permit direct reuse—can be a source of false tissue-level inference in single-cell aging prediction.
Source: IEEE Access
Authors: Zongmin Liu, Alex Lu, Wei Wang, Jianghong Rao, Yanmin Yang
Institutions: Stanford Medicine, Stanford University, Johns Hopkins University
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7202242736