Cross-Species Ageing DANN

Mouse cohorts are the workhorse of experimental ageing biology: animals can be aged under controlled conditions, so every cell carries an unambiguous young-or-old label. Human single-cell data has no such luxury — chronological age is entangled with donor genetics, health history, and batch, and the profiles themselves are sparse. Carrying mouse age labels over to human cells fails for a specific technical reason: species, donor, and assay differences dominate the ageing signal the classifier is supposed to learn.

This model reframes that failure as a domain-adaptation problem with target-domain validation. It was built by the Nerlov lab in the MRC Molecular Haematology Unit at the MRC Weatherall Institute of Molecular Medicine, University of Oxford, with collaborators at Imperial College London and the University of Cambridge. Mouse and human measurements are projected into a shared orthologue-aligned feature space, and a residual encoder trains against two heads at once — an age classifier and a species discriminator — so the embedding keeps ageing information while shedding species identity.

It is deliberately narrow. Unlike broad single-cell foundation models such as scGPT and Geneformer, which learn general-purpose cell representations for many downstream tasks, this is a single-task binary young/old classifier for haematopoietic stem and immune cells. What distinguishes it is the final experiment: the frozen pretrained checkpoint was applied zero-shot to an external COVID-19 convalescent cohort, where severe disease in younger adults was associated with a higher fraction of CD8+ cells scored as old-like.

#Key Features

• Two-phase adversarial training: The age classifier and species discriminator are optimised in two successive adversarial phases rather than a single joint pass, which ablations link to both higher accuracy and more stable attributions.
• Orthologue-aligned cross-species features: Mouse and human profiles are mapped onto a shared orthologue feature list, letting a model trained on labelled mouse cells score human cells directly.
• Transcriptome and chromatin: The same workflow was fitted to scRNA-seq and scATAC-seq profiles of haematopoietic stem cells, and separately to CD8+ T cells.
• Stability-based gene attribution: Consensus across DeepLIFT, Integrated Gradients, and saliency recovers conserved ageing-associated genes with greater cross-species overlap than differential expression alone.
• Fixed-checkpoint zero-shot scoring: Rather than refitting per dataset, the released checkpoint is applied unchanged to a new clinical cohort.

#Technical Details

The network is a domain-adversarial neural network (DANN): a residual encoder feeds a low-dimensional embedding to an age-classification head and, adversarially, to a species-discrimination head. The released reference configuration uses 256-unit hidden layers, a 64-dimensional embedding, ELU activations, dropout of 0.3, a learning rate of 1e-4, batch size 64, and unit weighting on the domain loss; checkpoints ship as Keras/TensorFlow files. On held-out human data the model reached AUROCs of 0.933 for HSC scRNA-seq, 0.953 for HSC scATAC-seq, and 0.941 in an independent CD8+ T-cell scRNA-seq setting. Ablations attribute gains to the residual connections, the ELU activations, and the two-phase schedule. Parameter counts are not reported; the network is small by contemporary standards, and its substance lies in the training procedure and the validation discipline rather than scale.

#Applications

The direct use is annotating ageing state in human haematopoietic and immune single-cell data where age labels are unreliable or absent, which lets groups with well-controlled mouse cohorts extract more from them. Because the model emits a per-cell old-like score, it also supports population-level readouts: the fraction of a donor's CD8+ compartment scored as old-like becomes a candidate measure of immune ageing that can be tested against clinical outcomes, as in the COVID-19 convalescent analysis. The MIT-licensed reproducibility package ships checkpoints, feature lists, and evaluation scripts, so the pipeline can be re-run or retrained; work outside haematopoietic and immune lineages would require refitting.

#Impact

The contribution is methodological as much as predictive: it argues that cross-species single-cell transfer should be judged on held-out target-domain data rather than on source-domain accuracy or embedding-mixing metrics, and it releases the artefacts needed to check that claim. Several caveats belong alongside the results. The work is a bioRxiv preprint that has not completed peer review, and the model repository notes that access to the weights may be restricted to reviewers and editors during that period. The age label is binary, so the output is a young/old call, not a biological age clock. And the COVID-19 finding is observational — an association between old-like CD8+ fraction and disease severity, not evidence that the ageing state caused the outcome.