Abstract
Single-cell ATAC-seq (scATAC-seq) data provides critical insights into cellular heterogeneity and transcriptional regulation. However, the analysis of this data is frequently hampered by its inherent sparsity, near-binarization, and ultra-high dimensionality. We present SCARP (Single-Cell ATAC-seq Analysis via Network Refinement with Peaks Location Information), a novel network diffusion-based method designed to address these challenges. SCARP is the first scATAC-seq analysis tool to utilize genomic peak location information, modeling the prior probability of co-accessibility as a function of genomic distance. By constructing a bipartite network of cells and peaks and applying a network refinement diffusion process, SCARP generates high-quality low-dimensional embeddings. Our experiments demonstrate that SCARP achieves superior cell clustering, identifies rare biologically significant subpopulations, and uncovers co-accessibility relationships that provide new insights into transcriptional regulation and disease-related pathways.
Introduction
The Single-cell Assay for Transposase-Accessible Chromatin using sequencing (scATAC-seq) has revolutionized our ability to study the epigenetic landscape at a single-cell resolution. By identifying regions of open chromatin, researchers can infer which regulatory elements are active in specific cell types. Despite its potential, scATAC-seq data is notoriously difficult to analyze due to three primary factors:
- High Sparsity: The "curse of missingness" means that many accessible regions are not captured during sequencing.
- Near-Binarization: The data typically exists as a matrix of 1s and 0s, offering limited signal depth compared to transcriptomic data.
- Ultra-high Dimensionality: Experiments often track over 100,000 genomic regions (peaks) across thousands of cells.
Existing computational methods often treat genomic peaks as independent features, ignoring a fundamental biological principle: genomic regions located near one another on a chromosome are more likely to be co-accessible and share regulatory relationships. SCARP was developed to bridge this gap by incorporating spatial genomic context into the analysis pipeline.
Methodology
The SCARP framework utilizes a bipartite network where both cells and genomic peaks are represented as nodes. The method is built upon several key innovations:
Genomic Distance Prior
Unlike traditional methods, SCARP calculates "prior weights" between adjacent peaks based on their physical distance on the genome. By modeling co-accessibility as a decreasing function of genomic distance, SCARP effectively encodes the biological likelihood that neighboring regions interact. This spatial prior significantly improves the characterization of peak relationships.
Network Refinement (NR) Diffusion
To combat data sparsity, SCARP employs a Network Refinement process based on a degree-normalized version of the Katz index. This diffusion process allows information to flow across the network, effectively "filling in" missing data by aggregating signals from a node’s neighbors. This denoising step reduces the influence of "promiscuous" peaks (those open across all cell types) while amplifying the signal of rare, cell-type-specific regions.
Computational Efficiency and Embedding
To manage the massive dimensionality of whole-genome data, SCARP processes each chromosome in parallel. Following the diffusion process, the method uses Uncentred Principal Component Analysis (UCPCA) to generate low-dimensional embeddings for both cells and peaks. These embeddings serve as the foundation for downstream tasks such as clustering and regulatory network inference.
Results and Discussion
Benchmarking and Clustering Performance
SCARP was evaluated against nine benchmarking datasets, including Forebrain, Leukemia, and Splenocyte. In every instance, SCARP outperformed state-of-the-art tools like cisTopic and scAND, consistently achieving the highest scores in Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI). This indicates that SCARP produces more accurate and biologically meaningful cell groupings.
Discovery of Rare Cell Subpopulations
In an analysis of a 10X Multiome PBMC dataset, SCARP identified a previously overlooked subpopulation of CD14 monocytes. These cells were characterized by high differentiation activity, a finding that was subsequently validated using independent scRNA-seq data from the same cells. This demonstrates SCARP's sensitivity in detecting subtle cellular heterogeneity that other methods might miss.
Insights into Transcriptional Regulation
Beyond clustering, SCARP’s peak embeddings allow for the calculation of co-accessibility scores between genomic regions, predicting cis-regulatory interactions such as enhancer-promoter loops. In a study of melanoma progression (SOX10 knockdown), SCARP identified HIF1A as a critical factor interacting with the master regulator MITF. This interaction provides a clearer picture of how hypoxia-related pathways drive cancer progression. Furthermore, predicted interactions were validated against Promoter-capture Hi-C (PCHi-C) data, showing a highly significant correlation (Student’s t-test $P=1.31 \times 10^{-230}$).
Conclusion
SCARP represents a significant advancement in scATAC-seq analysis by integrating genomic spatial information with network diffusion. By treating the genome not as a collection of independent points but as a structured physical entity, SCARP provides a more robust and biologically accurate representation of single-cell chromatin accessibility. Future work will focus on extending this framework to multi-omic integration, combining refined chromatin maps with transcriptomic and proteomic data to provide a holistic view of cellular regulation.
Summary updated