Summary
A bioRxiv preprint introduces EpiZoo, a 2.6-billion-parameter model trained on about 20.9 million single cells from multiple species. Its authors report performance on several single-cell analysis tasks and describe applications in comparing regulatory patterns and prioritising cancer mutations.
A bioRxiv preprint introduces EpiZoo, a model designed to analyse single-cell epigenomic data across species while incorporating information from DNA sequences. Its authors report that the model performs well on several analysis tasks using external datasets excluded from training, and describe uses ranging from comparisons of primate regulatory patterns to prioritising somatic mutations in cancer. The paper is a preprint and has not yet been peer reviewed.
Representing cells across species
Single-cell epigenomics examines regulatory features in individual cells. One important signal is chromatin accessibility: how open or accessible parts of DNA are, which can help reveal where regulatory activity may occur. Comparing these patterns between species can help researchers investigate which regulatory features are conserved and which have changed.
The authors say many existing models rely on genomic coordinates, making them difficult to apply across species, and do not incorporate regulatory information encoded in DNA sequences. EpiZoo is designed to address both issues. It converts high-dimensional single-cell profiles into compact representations the researchers call “cell sentences”. These combine DNA-sequence information with epigenomic context that is independent of sequence and the relative importance of accessible regions.
The model uses a mixture-of-experts transformer architecture and has 2.6 billion parameters in total. It was pretrained on the researchers’ manually curated Omni-scATAC corpus, which contains approximately 20.9 million cells from multiple species. The abstract does not specify the species represented or give a breakdown of the corpus.
Reported tasks and applications
On external datasets that were not included in pretraining, the authors report state-of-the-art performance in feature extraction, cell-type annotation and data imputation. These tasks respectively involve deriving useful data representations, assigning cell identities and estimating missing measurements. The abstract does not provide benchmark scores, so the performance claim is best understood as the authors’ summary of their comparisons.
The researchers also describe using EpiZoo to compare regulatory conservation and divergence during primate evolution. They report that the model can support context-aware prioritisation of cancer-associated somatic mutations and predict cell-type-specific chromatin accessibility from DNA sequences across genomic regions and species.
Together, these aims make EpiZoo a computational approach for connecting DNA sequence with cell-level regulatory measurements across species. Its reported applications concern analysis and prioritisation; the preprint does not describe a clinical test or treatment.