AI Insight
Researchers developed a deep learning framework that uses mouse aging data to predict aging states in human blood stem cells and immune cells, overcoming the challenge that human aging data is scarce while mouse data is abundant but species-specific. The system achieved over 93% accuracy in classifying young versus old human cells across multiple cell types and data modalities, and identified conserved aging features across species more effectively than traditional methods. When applied to COVID-19 patients, the model revealed that younger adults with severe disease had more "old-like" immune cells, suggesting accelerated immune aging.
Why it matters
This tool could accelerate human aging research by leveraging abundant mouse experimental data, potentially revealing conserved mechanisms of immune aging and identifying therapeutic targets. The framework's application to COVID-19 suggests it may help stratify disease risk and understand why some younger patients experience severe outcomes.
Understand the Science
⚠️ Preprint – Noch nicht peer-reviewed
Dieser Artikel wurde noch nicht von unabhängigen Experten begutachtet. Die Ergebnisse sind vorläufig und sollten mit Vorsicht interpretiert werden.
Cross-species transfer is a central bottleneck in single-cell ageing research: mouse studies provide experimentally tractable age labels, whereas human datasets are scarcer and strongly shaped by species-specific transcriptomic and chromatin programs. Here we present an interpretable resource for cross-species ageing-state annotation that combines a residual encoder, two-phase domain-adversarial optimisation and stability-based attribution. In haematopoietic stem cells (HSCs), the framework achieved human-test AUROCs of 0.933 in scRNA-seq and 0.953 in scATAC-seq. In an external CD8+ T-cell scRNA-seq setting, the held-out human AUROC was 0.941. Attribution across DeepLIFT, Integrated Gradients and Saliency identified reproducible consensus genes and showed greater cross-species conservation than differential expression alone. Applied to a COVID-19 convalescent cohort, the pretrained CD8+ model associated severe disease with a higher fraction of old-like CD8+ cells in younger adults. Together, the framework provides a reproducible and interpretable platform for cross-species annotation, conserved-feature discovery and translational immune-state stratification. Ageing is difficult to study directly in humans, particularly in rare stem and immune compartments for which repeated sampling across the lifespan is impractical. By contrast, murine ageing datasets are abundant and experimentally tractable, but biological differences between species create a substantial transfer problem. In haematopoietic stem cells (HSCs), ageing is associated with impaired regenerative capacity, altered differentiation output and platelet-myeloid bias, making this compartment a demanding but biologically informative testbed for cross-species annotation (refs. 1-3). Most cross-species single-cell analyses remain descriptive: they compare differential expression, pathway shifts or manifold structure, but they rarely provide transferable annotation of human ageing state from mouse-labelled data. A useful resource should therefore do more than benchmark classifiers. It should annotate unlabelled human cells, expose conserved ageing features and make the basis of those predictions inspectable enough for biological follow-up. Domain-adversarial neural networks offer a natural starting point because they learn task-relevant signal while suppressing domain-specific variation (ref. 4). In sparse high-dimensional single-cell data, however, standard gradient-reversal training can be unstable when the objectives of age discrimination and species confusion compete within the same encoder. We therefore focused on an optimisation scheme that stabilises domain-adversarial learning in this setting. Here we present a cross-species framework for ageing-state annotation that combines a residual multilayer perceptron encoder, two-phase domain-adversarial optimisation and stability-based attribution. We evaluate the framework in HSCs using both scRNA-seq and scATAC-seq, validate transfer in an independent CD8+ T-cell scRNA-seq setting, compare attribution with differential expression and then apply the pretrained CD8+ model to a COVID-19 convalescent cohort to assess translational utility.