Inflammation-Linked Aging Signals in Frozen Single-Cell Foundation Models: Donor-Aware Detection and Robustness Testing

Single-cell foundation models offer a possible route to studying aging biology in latent space, but apparent age effects can be distorted by donor identity and cell-composition differences. We developed a donor-aware interpretability workflow to test whether frozen scGPT and Geneformer representations contain biologically coherent aging signals in age-labeled human single-cell datasets. The workflow combined donor-held-out age decoding, sparse autoencoder feature discovery, cross-model pathway matching, targeted latent-space interventions, donor-bootstrap confidence intervals, and progressively stricter confound controls. Across five datasets, frozen representations contained detectable age information, with best donor-aware balanced accuracy of 0.384. Sparse autoencoders identified 132 donor-aware robust features, and cross-model pathway matching produced 193 paired features, with the strongest convergence in inflammation and NF-kappaB-related programs. The clearest intervention signal was observed in a global Geneformer inflammatory branch, where old-versus-random and old-versus-young contrasts remained positive after split expansion and donor-threshold tightening up to 400 donors. In a monocyte-restricted analysis, both scGPT and Geneformer also showed positive old-versus-random responses in one cohort. The strongest global signal weakened under stricter controls. Fully composition-matched forward-pass reruns yielded 0 of 4 full strict replications. These results indicate that frozen single-cell foundation models do capture biologically plausible aging-related structure, especially around inflammatory programs, but also that donor-aware and composition-aware stress tests are necessary before interpreting such signals as robust mechanisms.