Deep learning enables cross-species annotation and attribution of ageing states in haematopoietic stem and immune cells

Abstract

Mouse single-cell ageing studies provide experimentally controlled age contrasts, but using mouse-labelled data to annotate human ageing states is limited by species, donor and assay effects in sparse transcriptomic and chromatin profiles. We developed a cross-species annotation workflow that treats mouse-to-human prediction as a target-validated domain-adaptation problem. The workflow uses orthologue-aligned features, a residual encoder, an age classifier and a species discriminator trained with two-phase adversarial optimisation, and couples prediction with stability-based gene attribution. In haematopoietic stem cells (HSCs), the model achieved held-out human AUROCs of 0.933 in scRNA-seq and 0.953 in scATAC-seq. In an independent CD8+ T-cell scRNA-seq setting, the held-out human AUROC was 0.941. Ablation analyses indicated that residual connections, ELU activation and two-phase training improved predictive performance and attribution stability. Consensus attributions from DeepLIFT, Integrated Gradients and saliency recovered conserved ageing-associated genes with greater cross-species overlap than differential expression alone. In a COVID-19 convalescent cohort, severe disease in younger adults was associated with a higher fraction of CD8+ cells classified as old-like by the pretrained model. These results support a reproducible framework for testing, interpreting and releasing cross-species single-cell ageing models, while highlighting the need for target-domain validation when mouse labels are transferred to human data.