ComicGTN infers disease-associated rare cell states from single-cell multiomic data using DNA sequence-augmented graph transformer networks.

Single-cell multiomic technology makes it possible to profile both the transcriptome and epigenome within individual cells. The detection of rare cell states from such data is crucial for exploring novel disease biomarkers with clinical potential. However, existing methods typically focus on the gene and peak counts while ignoring the underlying genomic sequence at accessible sites. Here, we propose ComicGTN, an innovative computational framework that integrates single-cell multiomic data with DNA sequence information via enhanced graph transformer networks to accurately identify rare cell clusters. ComicGTN consistently outperforms nine state-of-the-art methods in identifying rare cells across multiple complex scenarios. In mouse breast cancer data, ComicGTN detects several functionally distinct immune subpopulations. In cerebral cortex of epilepsy patients, it uncovers oligodendrocytes undergoing particular differentiation states. For polycystic kidney disease patient-derived organoids, ComicGTN elaborates pathological pathways in a captured rare glomerular subgroup. Overall, ComicGTN accelerates the localization of disease-associated rare cell populations, facilitating the derivation of clinical insights in development and disease progression.