config.json

{
 "approved_by_author": "\u231b",
 "citation": "Tyshkovskiy, Alexander, et al. \"Universal transcriptomic hallmarks of mammalian ageing and mortality.\" Nature 654 (2026): 173-188.",
 "citations": 0,
 "citations_date": "2026-08-21",
 "clock_name": "tagemortality",
 "data_type": "transcriptomics",
 "doi": "https://doi.org/10.1038/s41586-026-10542-3",
 "journal": "Nature",
 "last_author": "Vadim N. Gladyshev",
 "model_type": "elastic net regression",
 "n_features": 10487,
 "notes": "Elastic Net over 10,487 mouse-Entrez genes, multispecies multi-tissue, scaleddiff variant. Cohort-relative: predict_age runs the tAge cohort preprocessing on the raw RNA-seq counts itself, so a prediction is a hazard shift against the reference group rather than an absolute risk. Name the cohort's species with a 0/1 column among var_names (mouse, rat, macaque or human; absent or all-zero means mouse) and the samples to centre against with a truthy adata.obs[\"tage_reference_group\"] (absent centres on the whole cohort); at least two samples are needed. The published pipeline's SimpleImputer, mean-only StandardScaler and pass-through SelectKBest are folded into the packaged linear layer, and the imputer medians are carried as reference_values so a gene the sample does not measure contributes its training median. Output is log10(hazard ratio) -- base 10, not the natural log the 'log hazard' unit label usually implies -- and unlike the chronological clocks it is never rescaled by species maximum lifespan, so it is directly comparable across species. Released under the MGB Open Access License 1.0: non-commercial academic research use only.",
 "platform": [
  "RNA-seq"
 ],
 "population": "multiple mammalian species",
 "predicts": [
  "mortality risk"
 ],
 "reference_values": true,
 "research_only": true,
 "species": "multiple species",
 "tissue": [
  "multi-tissue"
 ],
 "training_target": [
  "mortality"
 ],
 "unit": [
  "log hazard"
 ],
 "version": "0.5.0",
 "year": 2026
}