ATACompass enables cross-species cell identity mapping for scATAC-seq without gene annotations.

Single-cell assay for transposase-accessible chromatin with sequencing (scATAC-seq) provides insights into transcriptional regulation, but there remain challenges in cell identity annotation due to data sparsity and limited gene regulatory annotations, particularly in non-model organisms. To address these limitations, we present ATACompass, a sequence-based framework for cell annotation that directly encodes chromatin accessibility peak sequences using the DNA foundation model HyenaDNA and large language models. By incorporating controlled sequence mutations for data augmentation, ATACompass achieves performance competitive with or superior to existing state-of-the-art methods in intra-species tasks. Critically, ATACompass enables zero-shot cross-species cell annotation by restricting input data to conserved sequences shared between species, allowing non-model organism data to be annotated without species-specific prior knowledge. Furthermore, the integration of datasets from representative evolutionary lineages enhances predictive accuracy, enabling comprehensive annotation across species. In summary, ATACompass provides a gene annotation-independent framework for studying cell types and regulatory landscapes, especially in non-model organisms.