{
  "abstract": "OBJECTIVE: Single-cell foundation models such as scGPT learn rich representations of cellular identity, yet the biological programs encoded in their internal activations remain opaque. We investigate whether sparse autoencoders (SAEs), a mechanistic interpretability technique from AI safety research, can decompose these representations into sparse, biologically interpretable features. METHODS: We extract residual-stream activations from all 12 transformer layers of a pre-trained scGPT model processing 1000 human immune cells from the Tabula Sapiens atlas. We train SAEs with dictionary size M=2,048 at multiple sparsity levels (λ∈{1,3,10}) and evaluate recovered features using cell-type classification (AUROC), gene set enrichment (Fisher's exact test, FDR <0.05), and comparison with PCA baselines. RESULTS: >0.76. Later-layer SAE features recover biologically coherent programs aligned with annotated cell types, with 64% of alive features receiving significant gene set annotations at layer 11 (λ=3). We observe a sparsity-dead-feature trade-off: at λ=10, up to 66% of dictionary elements become inactive. CONCLUSION: Mechanistic interpretability methods developed for large language models transfer productively to biological foundation models, but require domain-specific calibration. SAEs provide a principled approach to understanding what single-cell foundation models learn about cellular identity, with potential applications in model auditing and biological discovery.",
  "abstract_shown": true,
  "affiliations": [
    "Department of Computer Science, University of Tübingen, Tübingen, Germany. Electronic address: ihor.kendiukhov@student.uni-tuebingen.de"
  ],
  "article_type": null,
  "arxiv": null,
  "author_total": 1,
  "authors": [
    {
      "affs": [
        0
      ],
      "name": "Ihor Kendiukhov",
      "orcid": "0000-0001-5342-1499"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/9020npexefbg0.bib",
  "cited_by": 17,
  "doi": "10.1016/j.jbi.2026.105056",
  "fulltext_url": null,
  "id": "9020npexefbg0",
  "issns": [
    "1532-0464",
    "1532-0480"
  ],
  "landing_url": "https://doi.org/10.1016/j.jbi.2026.105056",
  "language": "en",
  "licence": "CC-BY",
  "licence_label": "CC BY",
  "licence_source": "journal",
  "licence_url": "https://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": null,
  "native_subjects": [
    [
      "openalex",
      "1312|Single-cell and spatial transcriptomics"
    ]
  ],
  "openalex": "W7161630786",
  "pdf_url": "https://www.sciencedirect.com/science/article/pii/S1532046426000808/pdf",
  "pmcid": null,
  "pmid": "42155660",
  "published": "2026-05-19",
  "publisher": "Elsevier BV",
  "ris": "https://research.draconitic.com/paper/9020npexefbg0.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T13:26:46Z",
      "native_id": "W7161630786",
      "source": "openalex",
      "url": "https://doi.org/10.1016/j.jbi.2026.105056"
    }
  ],
  "subjects": [
    {
      "field": "biochemistry",
      "origin": "openalex:1312|Single-cell and spatial transcriptomics",
      "subfield": "molecular-biology"
    }
  ],
  "title": "Sparse autoencoders reveal interpretable cell-type programs in single-cell foundation model representations",
  "title_text": "Sparse autoencoders reveal interpretable cell-type programs in single-cell foundation model representations",
  "type": "peer-reviewed",
  "updated_at": "2026-11-05T13:20:46Z",
  "url": "https://research.draconitic.com/paper/9020npexefbg0",
  "venue": "Journal of Biomedical Informatics",
  "versions": []
}