Background: Single-cell Assay for Transposase Accessible Chromatin using sequencing (scATAC-seq) measures genome-wide chromatin accessibility for the discovery of cell-type specific regulatory networks. ScATAC-seq combined with single-cell RNA sequencing (scRNA-seq) offers important avenues for ongoing research, such as novel cell-type specific activation of enhancer and transcription factor binding sites as well as chromatin changes specific to cell states. On the other hand, scATAC-seq data is known to be challenging to interpret due to its high number of zeros as well as the heterogeneity derived from different protocols. Because of the stochastic lack of marker gene activities, cell type identification by scATAC-seq remains difficult even at a cluster level.
Results: In this study, we exploit reference knowledge obtained from external scATAC-seq or scRNA-seq datasets to define existing cell types and uncover the genomic regions which drive cell-type specific gene regulation. To investigate the robustness of existing cell-typing methods, we collected 7 scATAC-seq datasets targeting mouse brain for a meta-analytic comparison of neuronal cell-type annotation, including a reference atlas generated by the BRAIN Initiative Cell Census Network (BICCN). By comparing the area under the receiver operating characteristics curves (AUROCs) for the three major cell types (inhibitory, excitatory, and non-neuronal cells), cell-typing performance by single markers is found to be highly variable even for known marker genes due to study-specific biases. However, the signal aggregation of a large and redundant marker gene set, optimized via multiple scRNA-seq data, achieves the highest cell-typing performances among 5 existing marker gene sets, from the individual cell to cluster level. That gene set also shows a high consistency with the cluster-specific genes from inhibitory subtypes in two well-annotated datasets, suggesting applicability to rare cell types. Next, we demonstrate a comprehensive assessment of scATAC-seq cell typing using exhaustive combinations of the marker gene sets with supervised learning methods including machine learning classifiers and joint clustering methods. Our results show that the combinations using robust marker gene sets systematically ranked at the top, not only with model based prediction using a large reference data but also with a simple summation of expression strengths across markers. To demonstrate the utility of this robust cell typing approach, we trained a deep neural network to predict chromatin accessibility in each subtype using only DNA sequence. Through model interpretation methods, we identify key motifs enriched about robust gene sets for each neuronal subtype.
Conclusions: Through the meta-analytic evaluation of scATAC-seq cell-typing methods, we develop a novel method set to exploit the BICCN reference atlas. Our study strongly supports the value of robust marker gene selection as a feature selection tool and cross-dataset comparison between scATAC-seq datasets to improve alignment of scATAC-seq to known biology. With this novel, high quality epigenetic data, genomic analysis of regulatory regions can reveal sequence motifs that drive cell type-specific regulatory programs.
Optimal marker gene selection for cell type discrimination in single cell analyses
