Abstract
INTRODUCTION: Chromatin accessibility profiling is an important tool for understanding gene regulation and cellular function. While public repositories house nearly 10,000 scATAC-seq experiments, unifying this data for meaningful analysis remains challenging. Existing tools struggle with the scale and complexity of scATAC-seq datasets, limiting tasks like clustering, cell-type annotation, and reference mapping. A promising solution is using foundation models adapted to specific tasks via transfer learning. While transfer learning has been applied to scRNA-seq, its potential for scATAC-seq remains underexplored. METHODS: We introduce Atacformer, a transformer-based foundation model for scATAC-seq data analysis. Unlike other models that only produce cell-level representations, Atacformer generates embeddings for individual cis-regulatory elements. Pre-trained on a large atlas of scATAC-seq experiments, Atacformer learns robust representations of genomic regulatory regions for downstream use. After pretraining, the model is fine-tuned for cell-type prediction and batch correction. We also integrated Atacformer with RNA-seq data to build a Contrastive RNA-ATAC Fine Tuning (CRAFT) model capable of cross-modal alignment and RNA imputation from ATAC data. RESULTS: Atacformer matches or exceeds leading scATAC-seq clustering tools in adjusted rand index and runtime, with fine-tuned models achieving top performance across datasets. It processes raw fragment files end-to-end 80% faster than existing tools while preserving biological structure. Fine-tuned on bulk BED files, it recovers cell type and assay labels with >80% accuracy. We show how the Atacformer architecture produces contextualized embeddings of individual genomic regions, which we use to identify unannotated, cell-type-specific promoter elements directly from chromatin accessibility data.