Single-cell foundation models identify shared and divergent transcriptomic signatures of aging across invertebrates and mammals.

Cross-species aging clocks have so far stayed within mammals. We asked whether a single model can predict age from single cells of fly, worm, mouse, and human, species separated by roughly 800 million years of evolution. We fine-tuned scGPT and Geneformer on 1.3 million such cells, using one parameter set per architecture. Both models classified cells as young, middle, or old, reaching 78.6% and 80.5% accuracy. SHAP attribution then showed where the two architectures agreed and where they diverged. scGPT ranked ribosomal proteins highest in every species, with RPL12 first throughout. Geneformer did the same in mouse and human but favored signaling, chromatin, and ubiquitin-ligase genes in fly and worm. Hiding all 51 ribosomal-protein genes at inference cost 22-28 percentage points of accuracy, against 0.3 points for size-matched random gene sets. Bootstrap resampling, five-fold cross-validation, and comparison with six published aging studies left the rankings largely unchanged. Age-associated signal is therefore accessible to a pooled cross-species model, but which genes carry it depends on how the model encodes expression.