Defining effective strategies to integrate multi-sample single-nucleus ATAC-seq datasets via a multimodal-guided approach

BackgroundChromatin accessibility, measured via single-nucleus Assay for Transposase-Accessible Chromatin with sequencing (snATAC-seq), can reveal the underpinnings of transcriptional regulation across heterogeneous cell states. As the number and scale of snATAC-seq datasets increases, we need robust computational pipelines to integrate samples within a dataset and datasets across studies. These integration pipelines should correct cell-state-obfuscating technical effects while conserving underlying biological cell states, as has been shown for single-cell RNA-seq (scRNA-seq) pipelines. However, scRNA-seq integration methods have performed inconsistently on snATAC-seq datasets, potentially due to sparsity and genomic feature differences. ResultsUsing single-nucleus multimodal datasets profiling ATAC and RNA simultaneously, we can measure snATAC-seq integration method performance by comparison to independently integrated snRNA-seq gold standard embeddings and annotations. Here, we benchmark 58 pipelines, incorporating 7 integration methods plus 1 embedding correction method with 5 feature sets. Using our command-line tool, we assessed 5 multimodal datasets at 3 different resolutions using 2 novel metrics to determine the best practices for multi-sample snATAC-seq integration. ATAC features outperformed Gene Activity Score (GAS) features, and embedding correction with Harmony was generally useful. SnapATAC2, PeakVI, and ArchRs iterative Latent Semantic Indexing (LSI) performed well. ConclusionsWe recommend SnapATAC2 + Harmony with pre-defined ENCODE candidate cis-regulatory element (cCRE) features as a first-pass pipeline given its metric performance, generalizability of features, and method resource-efficiency. This and other high-performing pipelines will guide future comprehensive gene regulation maps.