A systematic single-cell analysis of enhancer RNAs across human and mouse

Abstract

Enhancer RNAs (eRNAs) contribute to gene expression specificity, yet their expression patterns at single-cell resolution remain poorly understood. Across 184 329 high-quality single cells from 187 full-length single-cell RNA-seq datasets, we identified 85 977 human and 82 727 mouse eRNAs, which were sparsely detected at the single-cell level (<5% in humans and <8% in mice) but showed high intercellular heterogeneity and contributed substantially to lineage-specific marker repertoires, accounting for up to 33.5% of human and 46.5% of mouse lineage markers. Through systematic analysis of eRNA expression patterns in embryonic development, brain tissues, and cancer, cell-specific eRNAs in different cell types were identified and corresponding regulatory networks were constructed. Our results indicate that eRNAs play important roles in diverse biological processes through modulation of gene expression. Finally, RACE confirmed the full transcription of a colorectal cancer-associated eRNA, whose elevated plasma levels, diagnostic performance, and postoperative reduction support its potential as a noninvasive biomarker for colon cancer detection and monitoring. Collectively, this study provides a comprehensive single-cell landscape of eRNA expression and offers a valuable resource for understanding eRNA-mediated regulatory mechanisms.

Graphical Abstract

Graphical Abstract.

Introduction

Enhancers are noncoding DNA elements that regulate the spatiotemporal expression of associated genes [1]. Many enhancer regions recruit transcription factors (TFs) and RNA polymerase II to produce enhancer RNAs (eRNAs) [2–4]. However, eRNAs are usually expressed at much lower levels than nearby coding genes, with reported expression levels 19–34 times lower than those of proximal mRNAs [5]. Although some eRNAs may reach 70–95 copies per cell [6], their detection remains difficult because they are rapidly degraded by RNA exosomes [7]. Despite their instability, eRNAs have been shown to regulate gene expression through multiple mechanisms [8, 9]. For example, FOXC1e, NRIP1e, and TFF1e interact with the cohesin protein RAD21 in breast cancer cells to promote chromatin looping and oncogene expression [6]; MMP9 eRNA and CCL2 eRNA bind the transcriptional coactivator BRD4 and enhance transcriptional activity at target genes [10]. eRNAs can also be regulated by TFs and RNA-binding proteins; for instance, seRNA-1 is regulated by MyoD and interacts with hnRNPL to enhance Mb expression [11]. These findings suggest that eRNAs participate in diverse regulatory networks.

Because eRNAs are unstable and lowly expressed, standard RNA sequencing (RNA-seq) has limited sensitivity for their detection. eRNAs are commonly profiled using nascent RNA-seq methods such as Global Run-On Sequencing (GRO-seq) and Precision Nuclear Run-On Sequencing, which capture RNA molecules during active transcription by RNA polymerase II [12]. Several active enhancer databases have been built from such data, including HACER [13], RAEdb [14], and eRNAbase [15]. However, these methods require complex library preparation and remain costly, limiting data availability. As a result, conventional RNA-seq data are still widely used to identify and study eRNAs, despite their limitations. For example, Zhang et al. integrated RNA-seq data from 9577 GTEx samples across 54 human tissues and identified 45 411 eRNAs [16]. Han et al. analyzed The Cancer Genome Atlas (TCGA) data and identified approximately 300 000 eRNA loci in super-enhancer regions, leading to the construction of the TCeA database [17]. Another TCGA pan-cancer study identified 9108 cancer-associated eRNAs, most of which showed cancer-type-specific expression, and 5715 eRNAs were associated with at least one clinical feature [18]. These studies revealed the broad relevance of eRNAs in human tissues and cancers, but eRNA expression heterogeneity at single-cell resolution remains poorly characterized.

Single-cell RNA-seq (scRNA-seq) has been widely used to define cell types, marker genes, and cellular heterogeneity. However, only a few studies have examined eRNA expression at the single-cell level [19, 20]. C1 CAGE scRNA-seq captures transcription start sites from the 5′ ends of transcripts and has been used to characterize eRNA activity in EpiSC-like cells [21]. RamDA-seq enables full-length total RNA-seq and has revealed cell-specific eRNA activity during mouse embryonic stem cell differentiation [22]. More recently, 5′-end scRNA-seq approaches such as ReapTEC have enabled simultaneous profiling of gene expression and enhancer activity at single-cell resolution, revealing candidate enhancers and cell type-specific regulatory programs in large-scale CD4⁺ T cell atlases [23]. Nevertheless, these studies are limited in scale and biological scope, and most focus on specific cell types, developmental systems, or specialized sequencing platforms. A broad single-cell view of eRNA expression across tissues and biological contexts is still lacking.

However, the expression level of eRNAs is so low that the droplet-based scRNA-seq platforms such as 10X Genomics have limited sensitivity for detecting eRNAs. Full-length scRNA-seq, known for its higher accuracy and coverage, has been verified to be able to detect a broad range of transcripts, including various types of noncoding RNA (ncRNA), such as long noncoding RNAs (lncRNAs) and circulating RNAs (circRNAs). For example, Luo et al. [24] and Wu et al. [25] systematically analyzed the expression of lncRNAs and circRNAs at the single-cell level using full-length scRNA-seq, respectively, and characterized the diverse expression patterns of the two types of ncRNAs across different cell types, leading us to hypothesize that full-length scRNA-seq data are rich in ncRNA information and could be pivotal for exploring eRNAs at the single-cell level.

In this study, we first evaluated the ability of full-length scRNA-seq to detect eRNA expression by comparing multiple sequencing platforms. We then collected and analyzed large-scale full-length scRNA-seq datasets from human and mouse samples to characterize eRNA expression during embryonic development and in brain tissues and human cancers. We examined eRNA heterogeneity, cell type-specific marker eRNAs, and potential regulatory networks in these biological contexts. Finally, we experimentally validated the complete transcript of a representative marker eRNA in colon cancer (CC) and further investigated its potential clinical application. These findings provide insights into the regulatory mechanisms of eRNAs at single-cell resolution and improve our understanding of the roles of eRNAs in diverse biological processes.

Materials and methods

Collection and construction of human and mouse eRNA sets

The human eRNA set was composed of known eRNAs and the predicted eRNAs constructed from annotated enhancers, whereas the mouse eRNA set was only constructed from annotated enhancers because of the absence of a curated mouse eRNA set (Supplementary Fig. S1). The procedures of eRNA collection and construction were as follows: (i) collection of known human eRNAs: on one hand, the enhancers with known transcriptional signals were obtained from the active enhancer databases including HACER [13] and RAEdb [14]. Then we expanded 3 kb to the left and right from the middle loci of the enhancers [26, 27] to construct known eRNA regions. On the other hand, eRNA regions were downloaded directly from known eRNA databases HeRA [16], eRic [18], as well as from a previously published study [28]. The eRNA collection in this step was defined as known eRNAs. (ii) Construction of predicted eRNAs. Human enhancer regions were obtained from FANTOM, Roadmap, and ENCODE, whereas mouse enhancer regions were obtained from FANTOM and ENCODE. Among these resources, Roadmap- and ENCODE-derived enhancers were defined based on H3K4me1 and H3K27ac signals. To obtain similar number of eRNAs in both species, we only selected the enhancer regions annotated in at least two databases (overlap more than 80%) in human, while the enhancer regions from all databases were combined for mouse. Then the enhancer regions were expanded 3 kb to the left and right from the middle loci too to construct the novel eRNA set. As for human, the known eRNAs collected from step [1] and the novel eRNAs constructed from step [2] were merged to construct the human eRNA set. Finally, the eRNAs regions that overlap more than 80% with each other were merged, and those located near to the transcription start or end site of any gene within 1 kb were removed. The genome coordinates of eRNA regions of human and mouse were converted to GRCh38 (hg38) and GRCm39 (mm39), respectively. Besides, the randomized eRNA sets of human and mouse were constructed by extracting the genome regions from nonoverlap intergenic regions with the same length and the same chromosome distributions of our eRNA sets, respectively.

Definition of super eRNA and typical eRNA regions

The classification of eRNAs was based on their association with super-enhancers, using information sourced from SEdb2.0 [29]. An eRNA was categorized as a super eRNA (SE-eRNA) if the overlap between its sequence and the super-enhancer region was more than 80%. In contrast, the eRNAs that did not meet this criterion were designated as typical eRNAs (TE-eRNAs)

Sequence conservation calculation

The conservation scores for each base in human and mouse were obtained from the UCSC database (http://www.genome.ucsc.edu/) and stored in the hg38.phastCons30way and mm39.phastCons35way files, respectively [30]. This analysis employed methods that display multiple alignments across various species and included measurements of evolutionary conservation using the phastCons methods. We employed the “map” function of the BEDTools software [31] to calculate the average conservation score for each eRNA region.

Human–mouse eRNA conservation analysis

To directly evaluate cross-species conservation of eRNA loci between human and mouse, reciprocal liftOver analysis was performed using hg38 and mm39 genome assemblies. Human eRNA regions were converted from hg38 to mm39 using the hg38ToMm39.over.chain.gz chain file, and mouse eRNA regions were converted from mm39 to hg38 using the mm39ToHg38.over.chain.gz chain file. Considering the rapid evolutionary turnover of noncoding regulatory elements, minMatch = 0.3 was used as the primary liftOver threshold.

Lifted human eRNA regions were intersected with mouse eRNA regions, and lifted mouse eRNA regions were intersected with human eRNA regions using BEDTools. A human–mouse eRNA pair was defined as conserved when the human eRNA overlapped a mouse eRNA after human-to-mouse liftOver and the corresponding mouse eRNA also overlapped the human eRNA after reciprocal mouse-to-human liftOver. eRNAs included in reciprocal conserved pairs were annotated as conserved, whereas all remaining eRNAs were annotated as nonconserved.

The absolute numbers of conserved and nonconserved eRNAs were calculated separately for human and mouse. To compare the single-cell expression characteristics of conserved and nonconserved eRNAs, we calculated average expression level, expressing-cell fraction, and expression heterogeneity for each eRNA. Differences between conserved and nonconserved eRNAs were assessed using the Wilcoxon rank-sum test.

Full-length scRNA-seq data collection

A large number of full-length scRNA-seq data for various types of tissue and cell line in human and mouse were collected from public databases including the Gene Expression Omnibus (GEO), Single Cell PORTAL, and Single Cell Expression Atlas. The SRA-Toolkit (version 3.0.0) was employed to download all original data. The types of full-length scRNA-seq include SMARTer, Smart-seq, Smart-seq2, Quartz-seq, RamDA-seq, SUPeR-seq, MATQ-seq, and Tang’s method.

scRNA-seq data analysis and integration

The genome annotation files of human (hg38) and the mouse (mm39) were downloaded from the GENCODE database [32]. For the full-length scRNA-seq and bulk RNA-seq, the raw sequencing reads were aligned by Hisat2 (v.2.2.1), and StringTie (v.2.1.4) was used for gene and eRNA expression quantification. As for the 10X Genomics, the STARsolo tool with default parameters was performed to calculate the number of reads per gene and per eRNA.

Low quality cells were filtered out by the perCellQCMetrics and isOutlier algorithms in Scater (v.1.22.0) software with default parameters [33]. In detail, the perCellQCMetrics function was first performed to estimate the level of library size, the number of gene reads and mitochondrial reads in each dataset. Then the isOutlier function based on the median absolute deviation (MAD) was used to filter the outlier cells. Next, the expression profiles were normalized using the SCTransform algorithm in Seurat (v.4.3.0), and batch effect correction and data integration were conducted by Harmony algorithm. Finally, 4000 highly variable genes were selected for downscaling by principal component analysis and cell-type clustering was performed by uniform manifold approximation and projection (UMAP)/t-stochastic neighbor embedding (t-SNE). The cell types were annotated based on the authoritative cell marker genes annotated in CellMarker [34] and PanglaoDB databases [35]. Normal epithelial cells and carcinoma cells with aneuploid rearrangement were determined by CopyKAT R package [36] where the parameter “ngene.chr” was set to 5. The number of reads for each eRNA and each gene was further transformed into reads of exon model per million mapped reads (RPM), which was used as the expression levels of eRNAs and other genes. eRNAs, mRNAs, and lncRNAs with an average RPM ≥ 1 were considered expressed and retained for downstream analyses.

Comparison of eRNA expression detection feasibility between full-length scRNA-seq and other sequencing techniques

GRO-seq technology, known for its ability to detect nascent RNA, was utilized as the benchmark for eRNA expression detection. Then, to compare and evaluate the feasibility of full-length scRNA-seq detecting eRNA expression, we collected the raw sequencing reads of the same cell line with different sequencing platforms including the gold standard GRO-seq. Specifically, the MCF-7 cell line for human and the E14 cell line for mouse were selected, since each of which had available data from multiple sequencing platforms, as detailed in Supplementary Table S1.

We set average RPM ≥ 1 as the threshold for expressed eRNA detection. For consistency, the same RPM threshold was also applied to mRNAs and lncRNAs when defining expressed transcripts. At the single-cell level, a transcript was considered detected in an individual cell if its RPM was ≥1. Then to evaluate the efficacy of different sequencing platforms in detecting eRNA expression, the number of expressed eRNAs detected was used as sensitivity indicator while the relationship between the average expression level and variation was applied as the specificity, where the variation was represented by the MAD. For this comparison, higher specificity was operationally defined as improved detection of lowly expressed eRNAs with low variability. Using GRO-seq as a gold standard, both Pearson and Spearman correlation coefficients of eRNA expression levels detected by various platforms and GRO-seq were calculated to gauge accuracy. In addition, previously published scRNA-seq (C1 CAGE) data were downloaded and analyzed to compare the performance of full-length scRNA-seq methods (Smart-seq2, SMARTer, SUPeR-seq) in mouse embryo cells too (Supplementary Table S1).

Co-expression network construction

For co-expression network construction, we first selected tissue/cell types with sufficient cell numbers and detectable eRNA expression, including 14 tissue/cell types in human and 18 tissue/cell types in mouse. Within each selected tissue/cell type, Spearman correlation coefficients were calculated between eRNAs and other genes based on their single-cell expression profiles. The corresponding P values were calculated using the Student asymptotic test implemented in the WGCNA R package and were adjusted by the false-discovery rate (FDR) method [37]. eRNA–gene pairs with |Spearman correlation coefficient| >0.3 and FDR < 0.05 were retained to construct tissue/cell type-specific co-expression networks. Then we only obtained the co-expressed relationships observed in more than three datasets to build the final eRNA co-expression network in both human and mouse respectively.

Functional annotation of eRNAs

To annotate the functions of eRNAs, the gene ontology enrichment analysis was performed for the co-expressed genes of each eRNA by clusterProfiler R package using FDR < 0.05 as significant level [38]. To integrate the functions of various eRNAs, we quantified the occurrence of each eRNA function. In addition, we employed Fisher’s method to consolidate multiple P values, which involves aggregating P values from numerous independent tests into a single statistic that conforms to a chi-square distribution.

TF–eRNA network construction

The binding regions of TFs across different cell lines were downloaded from the Cistrome database [39]. TF–eRNA relationships were initially defined when a TF binding region overlapped an eRNA region. To define the TF–eRNA regulatory network in a specific biological process, we further required that the TF and eRNA were co-expressed in the corresponding biological process. Specifically, Spearman correlation coefficients were calculated between TF expression and eRNA expression across cells or samples within the same biological process. And the TF–eRNA pairs with |Spearman correlation coefficient| ≥ 0.3 and FDR < 0.05 were considered as co-expressed.

eRNA–target network construction

Enhancer-target relationships were obtained from public databases, including EnhancerAtlas 2.0 (http://singlecelldb.com/index.php) and the ABC Model (https://www.engreitzlab.org/resources/). If the overlap between an enhancer and an eRNA region was greater than 80%, the corresponding enhancer-target relationship was converted into an eRNA–target relationship. To refine the eRNA–target network in a specific biological process, we further evaluated co-expression between eRNAs and candidate target genes within the same biological process. Spearman correlation coefficients were also calculated between eRNA expression and target gene expression across cells or samples. And the eRNA–target pairs with |Spearman correlation coefficient| ≥ 0.3 and FDR < 0.05 were further filtered.

Expression specificity score calculation

To characterize the expression specificity of eRNAs across multiple tissues/cell types, we calculated the specificity score of eRNAs or other genes utilizing the widely used Tau method [40]:

Where N is the number of tissues/cell types and is the average expression value of the ith eRNA or the ith gene in a certain tissue/biological process.

The R package ‘SCEnt’ [41] was utilized to assess the heterogeneity of gene expression within a cell population by Kullback–Leibler Divergence method.

Time-specific eRNA identification based on scRNA-seq data

To further determine at which time point that the specifically expressed eRNAs is specific in a certain cell type, we then calculated the specificity score of each eRNA for each different time point in the certain cell type by using the following formula as previously reported [42].

where p_ij is the proportion of cells expressing gene i above a threshold in a certain type of cell at the time point j, and q_ij is the proportion of cells expressing gene i above the same threshold in the same type of cell at all the time points excepting for the time point j. The expression threshold is set as the 75th percentile of the expression level of the gene in all cells at the same time point.

Then, the eRNAs were assigned into the time point with the highest specificity score to make potential time-specific eRNAs set. To further rank the specificity of potential time-specific eRNA at each time point, the expression enrichment score of each eRNA within cluster was further calculated by using the following formula,

where u_ij represents the average expression of the eRNA in a certain cell type at the time point j, and v_ij is the average expression of this eRNA in other cells excepting for the cells at the time point j.

Finally, we identified time-specific eRNAs based on the following two criteria: (i) the eRNA exhibited high expression at the relevant developmental stage, with a log₂ of expression fold change (log₂FC) between a given stage and the average expression across all other stages no less than 1.5 and (ii) the enrichment scores e_ij of the specific eRNAs at this developmental stage should be larger than 2.

eRNA-associated cell communication analysis

Cell communication networks among different cell types were inferred using the CellChat R package with default parameters [43]. CellChat analysis was performed based on the expression of protein-coding ligand and receptor genes, rather than eRNA expression directly. To evaluate potential eRNA-associated cell–cell communication, ligand–receptor interactions inferred by CellChat were subsequently integrated with eRNA-associated regulatory relationships. Specifically, when a ligand or receptor gene in a CellChat-inferred interaction was linked to at least one eRNA according to eRNA–target network, this ligand–receptor interaction was defined as an eRNA-associated cell communication event.

pySCENIC regulon analysis of eRNA-associated TFs

To investigate eRNA-associated TF regulatory activity in tumor cells, pySCENIC analysis was performed using the normalized expression matrix of tumor cells. Gene regulatory networks were inferred using GRNBoost2 with a curated human TF list. Motif enrichment analysis was performed using human cisTarget motif ranking databases, and regulon activity in individual cells was quantified using AUCell.

TFs linked to eRNAs were extracted from the TF–eRNA association network and intersected with pySCENIC-derived regulons. The activity of eRNA-associated TF regulons was compared between aneuploid and diploid tumor cells. Mean regulon activity was calculated for each tumor-cell state, and differential regulon activity was assessed using the Wilcoxon rank-sum test followed by Benjamini–Hochberg correction. Representative eRNA-associated TF regulons with significant differences between aneuploid and diploid cells were visualized using heatmaps and violin plots. A SCENIC-supported TF–eRNA regulatory network was constructed by retaining TF–eRNA pairs whose TFs were supported by pySCENIC regulon inference.

Monocle pseudotime analysis of cisplatin-response tumor cells

To investigate dynamic eRNA expression associated with cisplatin response, pseudotime trajectory analysis was performed using the Monocle R package on the oral squamous cell carcinoma (OSCC) cisplatin-response dataset. Tumor cells were grouped into drug-sensitive, drug-holiday, and drug-resistant states according to the original annotation. The normalized expression matrix and cell metadata were used to construct a Monocle-compatible object.

Ordering genes were selected from variable or differentially expressed genes across cisplatin-response states. Dimensionality reduction and cell ordering were then performed to infer pseudotime. Pseudotime values were projected onto the UMAP embedding for visualization. To identify dynamic eRNAs along the cisplatin-response trajectory, eRNA expression was modeled as a function of pseudotime. Cells were divided into pseudotime bins, and the average expression of dynamic eRNAs in each bin was calculated and scaled for heatmap visualization. Representative dynamic eRNAs were visualized using smoothed trend plots along pseudotime. Differences in pseudotime distribution among cisplatin-response states were assessed using the Wilcoxon rank-sum test.

Rapid amplification of complementary DNA ends

To validate the transcript structure of the predicted eRNA in cancer cells, rapid amplification of cDNA ends (RACE) was performed in the HCT116 human CC cell line. Total RNA was extracted using TRIzol reagent according to the manufacturer’s instructions. The quality and concentration of RNA were assessed before reverse transcription. For 5′ and 3′ RACE, cDNA was synthesized using a RACE kit according to the manufacturer’s protocol, with gene-specific primers designed for the target eRNA (Supplementary Table S2). The 5′ and 3′ ends of the eRNA transcript were amplified by PCR, followed by nested PCR to improve specificity. PCR products were separated by agarose gel electrophoresis, and bands of the expected size were purified and subjected to Sanger sequencing. The sequencing results were aligned to the predicted eRNA genomic region to confirm the amplified transcript fragments.

Analysis of clinical applications of marker eRNAs

The cancer types represented in our scRNA-seq datasets, including BRCA, COAD, READ, GBM (glioblastoma), KICH, KIRC, KIRP, LIHC, PAAD, LUAD, and LUSC, were selected for further analysis and comparison with the corresponding TCGA cohorts (Supplementary Table S3). The RNA-seq BAM files and patient clinical characteristics were downloaded from the TCGA data portal [44], and StringTie (version 2.1.4) was used to align these sequence reads on the human eRNA regions. The identification of marker eRNAs across aneuploid epithelial cells from various cancers was performed using the FindAllMarkers function in Seurat. The abundance profiles of these marker eRNAs, identified by StringTie, were normalized using RPM. The relationships between expression of marker eRNAs and clinical characteristics including age, sex, stage, and survival time were evaluated. The Wilcoxon rank-sum test was used for the measurement data, and the chi-square test was used for the enumeration data. Univariate Cox regression analysis was applied to evaluate the prognostic value of the marker eRNAs by the “survival” R package

Patients

Plasma samples were collected from 94 CC patients and 96 healthy individuals at the First Affiliated Hospital of Ningbo University. None of the patients had received chemotherapy or radiotherapy prior to surgery. In addition, we collected postoperative plasma samples from these patients. These plasma samples were preserved at −80°C. Every participant agreed to take part in this project and this study was approved by the Ethics Committee of the hospital.

Real‐time quantitative polymerase chain reaction

Total RNA was extracted from the samples using the TRIzol kit (Invitrogen, USA). The concentration of the extracted RNA was measured with a Full Wavelength Microplate Reader (DeNovix, USA). The Tiosbio™ Polestar 1st cDNA Synthesis Kit (Tiosbio, Tokyo) was used to reverse transcribe the total RNA into cDNA. Real-time PCR amplification was then conducted using the Applied Biosystems™ QuantStudio™ 3 (Thermo Fisher, USA). The real-time quantitative PCR (qPCR) amplification products were subjected to T-A cloning and sequencing, and the sequencing results were compared with the original sequence segments to ensure consistency. There is no recognized reference for the quantification of RNA molecules in plasma; therefore, an absolute quantification method was used. Then, the real-time qPCR product sequences of eRNA were cloned into the pUC57 plasmid, and the C_q values were determined by real-time qPCR. Linear regression was used to fit the absolute quantification standard curve formula for eRNA, enabling the determination of the absolute copy number of eRNA in plasma samples. The primer sequences for chr4:74621763–74628200 were 5′-TCCAGGACTCGTGGGATTTT-3′ and 5′-CCACCAGGCATTCACCTATTT-3′. The Wilcoxon signed-rank test was applied to compare the absolute copy number.

Statistical analysis

Unless otherwise specified, comparisons between two groups were performed using the Wilcoxon rank-sum test. Comparisons among more than two groups were performed using the Kruskal–Wallis test. Correlation analyses were performed using Spearman correlation unless otherwise stated. P values were adjusted for multiple testing using the Benjamini–Hochberg method where applicable. Statistical significance was indicated as follows: ns, not significant; *P < .05; **P < .01; ***P < .001; ^****P < .0001.

Results

eRNA expression can be effectively detected with full-length scRNA-seq technology

A comprehensive human eRNA atlas was constructed by integrating eRNA datasets from HACER [13], RAEdb [14], HeRA [16] eRic [18], and a previously published study [28]. All genomic coordinates were standardized to hg38. eRNAs with more than 80% overlap were merged, and regions located within 1 kb of gene transcription start or end sites were removed. This process yielded 60 970 known human eRNAs. To further expand the atlas, enhancer regions from FANTOM [5], ENCODE [45], and Roadmap [46] were incorporated. Only enhancer regions supported by at least two databases were retained, extended by ±3 kb from the enhancer midpoint, and merged when overlap exceeded 80%. This approach identified an additional 51 903 eRNAs, resulting in a final set of 93 347 human eRNAs (Supplementary Table S4). Using a similar strategy, 104 656 eRNAs were identified in mouse (Supplementary Table S5). A higher proportion of human eRNAs overlapped super-enhancer regions compared with mouse eRNAs, with SE-associated eRNAs accounting for 60.2% of human eRNAs and 31.3% of mouse eRNAs (Fig. 1a). Most eRNAs in both species were 6–9 kb in length (Fig. 1b). Consistent with greater evolutionary divergence of enhancer-associated transcripts, eRNAs showed lower sequence conservation than lncRNAs and mRNAs in both species (Fig. 1c).

Figure 1.

Information about eRNA atlas and feasibility assessment of full-length RNA-seq to detect eRNA expression. (A) The numbers of SE_eRNAs and TE_eRNAs analyzed in human and mouse. The green bar represents the eRNAs located on the super-enhancer regions and purple bar represents the eRNAs located in typical enhancer regions. (B) The density plots for the distributions of eRNA lengths in human and mouse respectively. (C) The boxplots compare the sequence conservation among mRNAs, lncRNAs, and eRNAs. ^**** means P < .0001 by Wilcoxon rank-sum test. (D) The number of eRNAs detected by different sequencing platforms in the same cell line in human and mouse respectively. (E) Scatter plots show the relationships between MAD and the median of eRNA expression detected by different sequencing platforms in human. The x- and y-axes represent the median and MAD value of the eRNA expression, respectively. (F and G) The correlation of eRNA expression detected by different sequencing platforms with that detected by GRO-seq in the same cell line in human (F) and mouse (G). The color strength of heatmap represents the value of Pearson correlation coefficient. The filled color aside represents the type of sequencing platform, the legends of which are the same as in (D). (H) The correlation of eRNA expression levels detected by SMARTer and GRO-seq in human MCF-7 cell line. The axis represents the mean of eRNAs expression. (I) The correlation of eRNA expression detected by GRO-seq and other types of full-length scRNA-seq including SMARTer-seq, Smart-seq2, Tang’s methods in mouse E14 cell line. The axis represents the median of eRNAs expression. (J) The percentage of expressed eRNA detected by different types of scRNA-seq in mouse embryo cells based on our eRNA atlas and random set.

To evaluate whether full-length scRNA-seq can detect eRNAs, we compared eRNA detection across sequencing platforms in the human MCF-7 cell line. GRO-seq, used as the gold standard, detected 75 066 eRNAs, corresponding to 80.4% of all annotated human eRNAs (n = 93 347; Fig. 1D). SMARTer scRNA-seq detected 45 816 eRNAs (49.1%), total RNA-seq detected 44 820 eRNAs (48.0%), and poly (A)+ RNA-seq detected 38 127 eRNAs (40.8%). In contrast, 10X Genomics scRNA-seq detected only 1329 eRNAs (1.4%) (Fig. 1d). A similar pattern was observed in mouse E14 cells. Among 104 656 annotated mouse eRNAs, GRO-seq detected 42 725 eRNAs (40.8%), SMARTer scRNA-seq detected 31 495 eRNAs (30.1%), total RNA-seq detected 31 001 eRNAs (29.6%), poly (A)+ RNA-seq detected 21 196 eRNAs (20.3%), Tang’s method detected 18 886 eRNAs (18.0%), Smart-seq2 detected 17 571 eRNAs (16.8%), and 10X Genomics scRNA-seq detected 1 047 eRNAs (1.0%) (Fig. 1d). These results indicate that full-length scRNA-seq detects substantially more eRNAs than droplet-based 10X Genomics scRNA-seq.

To further assess detection specificity, we examined the variance and median expression levels of eRNAs across platforms, under the assumption that higher specificity is reflected by improved detection of lowly expressed eRNAs with low variance. We observed that highly expressed eRNAs generally exhibited lower variability (Fig. 1e, Supplementary Fig. S2). Consistent with expectations, lowly expressed eRNAs with low variance were more frequently detected by GRO-seq and total RNA-seq, likely due to their greater sequencing depth (Fig. 1e, Supplementary Fig. S2). Full-length scRNA-seq and poly (A) RNA-seq demonstrated comparable specificity, whereas 10X Genomics scRNA-seq showed the lowest performance (Fig. 1e, Supplementary Fig. S2).

To evaluate quantification accuracy, we calculated Pearson correlations between eRNA expression measured by each platform and GRO-seq. In human cells, total RNA-seq showed the highest concordance with GRO-seq, followed by poly (A)+ RNA-seq and SMARTer scRNA-seq, with correlation coefficients of 0.54 and 0.48, respectively (Fig. 1f,h, Supplementary Fig. S3a). These correlations were higher than random controls (r < 0.001; Supplementary Fig. S4a). In contrast, 10X Genomics scRNA-seq showed the weakest correlation (r = 0.096; Supplementary Fig. S3a). Similar results were observed in mouse E14 cells (Fig. 1g,i, Supplementary Figs S3b and S4b). These findings suggest that full-length scRNA-seq quantifies eRNA expression with accuracy comparable to poly (A)+ RNA-seq, which has been widely used in bulk eRNA studies [17, 18, 28, 47].

Finally, we compared full-length scRNA-seq with C1 CAGE, a single-cell method designed to capture transcription start sites and enhancer activity. SMARTer scRNA-seq detected a proportion of eRNAs comparable to C1 CAGE (47.9% vs. 46.1%), and other full-length scRNA-seq platforms also outperformed random sets (3.5%–7%) (Fig. 1j). Pearson correlation analysis further showed that eRNA expression profiles from SMARTer, Smart-seq2, and SUPeR-seq were significantly correlated with those from C1 CAGE (r > 0.5, P < .05; Supplementary Figs S3c and S4c). These results support the use of full-length scRNA-seq for detecting low-abundance eRNAs at single-cell resolution. Recent methods such as STORM-seq further illustrate the potential of full-length, ribo-reduced single-cell sequencing for profiling transient eRNA expression [48]. Collectively, these findings indicate that full-length scRNA-seq provides high sensitivity for eRNA detection and offers a robust foundation for characterizing eRNA-mediated regulatory landscapes.

The expression characteristics of eRNAs at the single-cell level

To systematically analyze single-cell eRNA expression, we collected publicly available full-length scRNA-seq datasets from GEO, Single Cell PORTAL, and Single Cell Expression Atlas. In total, 187 studies covering 44 human and mouse tissues or cell types were included (Fig. 2a, Supplementary Table S6). These datasets contained 207 096 cells, including 23.7% human and 76.3% mouse cells. After quality control, 44 605 human cells and 139 724 mouse cells were retained. Using an overall RPM ≥ 1 across all expression profiles within each species as the filtering criterion, we identified 85 977 (92.1%) human eRNAs and 82 727 (79.0%) mouse eRNAs.

Figure 2.

The expression patterns of eRNAs across different cell types. (A) Numbers of cells collected from each tissue and sequencing platform in human and mouse datasets. Bar colors indicate sequencing platforms. (B) UMAP visualization of four major human cell lineages: epithelial, immune, progenitor, and stromal cells. (C) Numbers of detected eRNAs (RPM ≥ 1) across the four major human cell lineages. (D) Proportions of annotated eRNAs, lncRNAs, and mRNAs expressed per cell across human cell lineages. (E) Distributions of expression heterogeneity for eRNAs, mRNAs, and lncRNAs across human cell lineages, measured by the coefficient of variation. (F) Violin plots comparing the average log₂ fold change (avg_log2FC) of mRNA, eRNA, and lncRNA markers across human lineages. (G) Stacked bar plot showing the numbers of lineage markers, subdivided into mRNA, eRNA, and lncRNA markers. (H) UMAP visualization of human epithelial cells subclustered into epithelial subtypes. (I) Dot plot showing the top three eRNA markers for each epithelial subtype. Dot size represents the percentage of cells expressing each marker, and color indicates the average expression level. (J) UMAP FeaturePlot showing the expression pattern of eRNA61378 across human epithelial subtypes. High-expression cells are shown in red.

To further annotate these eRNAs from a cross-species perspective, we mapped eRNA loci between hg38 and mm39. In total, 24 682 (26.4%) human eRNAs and 26 585 (25.4%) mouse eRNAs were classified as conserved (Supplementary Fig. S5a). At the single-cell level, conserved eRNAs showed lower detection frequency and lower average expression than nonconserved eRNAs in both species, but displayed higher expression heterogeneity, measured by log₁₀ (coefficient of variation + 1) of expression (Supplementary Fig. S5b–d). These results suggest that conserved eRNAs are generally more sparsely detected and more heterogeneous across single cells.

Based on canonical marker genes, human cells were classified into four major lineages: epithelial, immune, progenitor, and stromal cells (Fig. 2b). Mouse cells were classified into five lineages: endothelial, epithelial, immune, neural, and stromal cells (Supplementary Fig. S6a). We next examined eRNA expression across major cell lineages. The epithelial cells showed the highest eRNA diversity, with 79 517 (85.2%, Fig. 2c) and 92 614 (88.5%, Supplementary Fig. S6b) detected eRNAs in human and mouse, respectively. eRNA expression patterns also differed between species. In human cells, eRNAs were generally lowly expressed, whereas in mouse neural, epithelial, and stromal lineages, eRNA expression reached levels comparable to or higher than lncRNAs (Supplementary Fig. S6c). At the single-cell level, only a small fraction of annotated eRNAs was detected in individual cells (<5% in humans and <8% in mouse, Fig. 2d, Supplementary Fig. S6d). In contrast, mRNAs showed robust coverage, peaking at nearly 30% in human epithelial cells (Fig. 2d) and 40% in mouse neural and epithelial cells (Supplementary Fig. S6d). Across lineages, eRNAs showed higher intercellular heterogeneity than mRNAs and lncRNAs, especially in the mouse immune lineage (Fig. 2e, Supplementary Fig. S6e). This is consistent with previous reports that burst-like eRNA transcription from super-enhancers can rapidly activate inflammatory genes in immune responses [49]. Together, these results indicate that eRNAs are low-abundance, sparsely detected, heterogeneous, and highly context-dependent transcripts at the single-cell level.

We next evaluated whether eRNAs could serve as cell-lineage markers. In epithelial cells, eRNA markers exhibited significantly higher avg_log2FC values than both mRNA and lncRNA markers, indicating stronger epithelial-lineage specificity. In stromal cells, eRNA markers showed lower avg_log2FC values than mRNA markers but were comparable to lncRNA markers. By contrast, no significant differences among the three transcript classes were observed in immune or progenitor cells (Fig. 2f). In mouse cells, eRNA markers differed significantly from lncRNA markers across all examined cell types, suggesting more prominent transcript-class-dependent marker patterns in mouse datasets (Supplementary Fig. S6f). Although mRNAs remained the most abundant marker class, eRNAs accounted for a considerable fraction of lineage markers, including 33.5% of markers in human epithelial cells and 33.3% in human stromal cells (Fig. 2g), as well as 46.5% in mouse epithelial cells and 46.3% in mouse endothelial cells (Supplementary Fig. S6g).

Given that human epithelial cells displayed abundant eRNA markers and pronounced heterogeneity, we further subclustered this lineage into six distinct epithelial subtypes (Fig. 2h). Characterization of the marker composition across these subtypes revealed that eRNAs constitute a substantial portion of the cell-state-specific transcripts. Most notably, in proliferating epithelial cells, eRNAs contributed the most marker genes, outnumbering mRNAs by nearly two-fold (4935 eRNA markers compared with 2602 mRNA markers and 189 lncRNA markers; Supplementary Fig. S6h). Top candidate eRNA markers displayed subpopulation-associated expression patterns across epithelial cell subsets (Fig. 2i). Their relative marker strength varied across epithelial subpopulations and was not consistently higher than that of mRNA or lncRNA markers (Supplementary Fig. S6i). As a representative example, eRNA61378, a top-ranked marker for the proliferating subtype, exhibited highly localized expression restricted to this subpopulation on the UMAP plot (Fig. 2j). Together, these findings indicate that eRNAs can serve as effective markers to distinguish both major cell lineages and distinct cellular states.

eRNAs show dynamic regulation during embryo development and are strongly activated in early stages

A significant body of studies have explored the gene heterogeneity in embryonic development at the single-cell level [50, 49], but the dynamics and stage-specific regulation of eRNAs remain unclear. To address this, we analyzed 15 full-length scRNA-seq datasets, including 5 human and 10 mouse datasets, comprising 5744 cells across 17 embryonic stages (Fig. 3a). eRNA expression was enriched at early stages, especially oocyte, zygote, two-cell, and four-cell stages, in both human and mouse (Supplementary Fig. S7a and b). UMAP analysis based on eRNA expression separated embryonic cells by developmental stage in both species, similar to mRNA-based clustering (Fig. 3b–e). Correlation analysis further showed that eRNA expression patterns were more stage-specific and more closely aligned with developmental progression than lncRNAs and mRNAs (Fig. 3f and g).

Figure 3.

The dynamic expression pattern of eRNAs in embryo development. (A) The number of cells collected at different developmental stages. Human and mouse cells are distinguished by squares and circles, respectively. The size of the dot indicates the number of cells. The color of the dot represents the types of embryo stage. (B–E) The UMAP plots show the categorization of embryo cells clustering based on eRNAs and mRNAs expression profiles in human (B and C) and mouse (D and E), respectively. The developmental stages of embryo cells are color-coded. (F and G) The Pearson correlation coefficient heatmaps of eRNA (left), mRNA (middle), and lncRNA (right) expression among different development stages in human (F) and mouse (G), respectively. (H) The number of specifically expressed eRNAs and mRNAs at different developmental time points. (I) Violin and box plots showing the distribution of eRNA expression levels across embryonic developmental stages. eRNA expression was measured as log₁₀ (RPM + 1), and the black line indicates the median expression trend across stages. Statistical significance was assessed using the Kruskal–Wallis test. (J) The number of TFs regulated by stage-specific eRNAs in different developmental stages. (K) Sankey plots showing selected recurrent TFs across developmental stages in human embryos. The width of each flow represents the number of linked eRNAs associated with each TF at the corresponding stage. (L) The heatmap of specific eRNAs in different developmental stages and corresponding main functions in human. (M) An example for the dynamic expression pattern of regulatory pair for TE/ICM stages in mouse. (N) Genome accessibility profiles show the ATAC-seq signals at genomic loci bound by Cux2 and its co-expressed eRNAs across developmental stages.

We next identified stage-specific eRNAs using enrichment score and log₂ fold-change thresholds. Across different thresholds, the distribution of stage-specific eRNA numbers was highly consistent between human and mouse (pcc = 0.97), with the totipotent stage showing the largest number of stage-specific eRNAs (Supplementary Fig. S7a). Using an enrichment score >2 and log₂FC >1.5, we identified 562, 193, 43, and 77 stage-specific eRNAs in human across the totipotent stage (TS), TE/ICM, EPI/PrE, and IE, respectively (Fig. 3h, Supplementary Table S7). In mouse, 510, 204, 52, 144, and 90 stage-specific eRNAs were found in the corresponding stages (Fig. 3h, Supplementary Table S7). The number of stage-specific eRNAs exceeded that of mRNAs, suggesting that eRNAs provide additional regulatory information during embryogenesis (Fig. 3h). Stage-specific eRNA expression decreased across development in both species, indicating higher eRNA activity during early embryogenesis (Supplementary Fig. S7c and d). SE-eRNAs also showed higher expression levels and detection frequencies than TE-eRNAs across human and mouse developmental stages (Supplementary Fig. S8a–d). These patterns were consistent with previous reports of elevated eRNA activity during early development [51]. Bulk RNA-seq data from mouse embryos further supported this trend, showing the highest eRNA expression at the two-cell stage followed by a general decline during subsequent development (Fig. 3i) [52, 53]. We further examined the conservation status of stage-specific eRNAs. In human embryos, 24.1% (211/875) of stage-specific eRNAs were conserved, whereas only 12.9% (129/1000) were conserved in mouse embryos. Across developmental stages, the proportion of conserved stage-specific eRNAs ranged from 9.3% to 32.0% in human and from 7.8% to 15.7% in mouse, with the highest conserved fraction observed at the human totipotent stage (Supplementary Fig. S8e). This suggests that many stage-specific eRNA programs are species-specific.

Given the crucial role of TFs in embryonic development, we next examined their regulatory relationships with stage-specific eRNAs. In human, the TF network included 5430 regulatory relationships involving 672 eRNAs and 359 TFs. Among these, 46.2% (166/359) of TFs exhibited stage-specific activity, predominantly at the totipotent stage (TS) (64.5%, 107/166) (Fig. 3j, Supplementary Table S8). Some TFs exhibited regulatory roles across multiple developmental stages. In human, 125 TFs were active in two stages, with 84.0% (105/125) acting in consecutive stages. Similarly, 74.2% (46/62) of TFs active across three stages showed consecutive activity (Supplementary Fig. S9a). Several TFs, including CEBPA, POU5F1, NR2F2, and TEAD4, were active throughout human embryonic development (Fig. 3k) [51, 54–57]. Functional annotation of the corresponding stage-specific eRNA-associated programs suggested enrichment in biological processes such as hematopoiesis and pattern specification (Fig. 3l, Supplementary Table S9). In mouse embryos, the TF–eRNA network contained 6111 regulatory relationships involving 746 eRNAs and 335 TFs. Among them, 207 TFs acted at a single stage, with most acting at the earliest stage (86.0%, 178/207) (Fig. 3j, Supplementary Table S8). For TFs active across multiple stages, 60.5% (52/86) of those active across two stages and 48.5% (16/33) of those active across three stages showed consecutive-stage activity, suggesting a tendency toward continuous regulatory control of eRNAs during development (Supplementary Fig. S9a). Several TFs, including CREBBP, GATA4, MYC, NANOG, POU5F1, TET1, and SOX2, were active throughout mouse embryonic development (Supplementary Fig. S9b). Notably, POU5F1 was shared between human and mouse, consistent with its central role in pluripotency regulation [51, 58–62]. Similar functional patterns were observed for mouse stage-specific eRNA-associated programs, including enrichment in hematopoietic processes (Supplementary Fig. S9c, Supplementary Table S9) [63].

We further constructed stage-specific TF–eRNA subnetworks for each developmental stage (Supplementary Fig. S10a and b). To link these interactions with downstream genes, we integrated TF–eRNA pairs with eRNA–target associations and generated representative TF–eRNA–target triplets (Supplementary Fig. S10c and d). These representative triplets in both human and mouse embryos showed stage-associated expression dynamics, further supporting the temporal specificity of TF–eRNA–target regulatory relationships. As an example, Cux2, a TF involved in neurogenesis and cell fate regulation [64], was predicted to regulate eRNA chr12:72792542–72801803 (Fig. 3m). ATAC-seq data from mouse embryonic stages showed that Cux2 binding signals increased together with this eRNA, suggesting coordinated activation of the TF–eRNA pair (Fig. 3n). These results highlight stage-specific TF–eRNA regulatory dynamics during embryonic development.

Specific expression patterns of eRNAs in various brain cell types

The regulatory roles of noncoding RNAs, especially lncRNAs, have been widely studied in the brain [65, 66]. eRNAs have also been implicated in neuronal activation, such as eRNAs at the Fos locus that respond to neuronal or cAMP stimulation [67]. However, their single-cell expression patterns in brain cells remain unclear. To address this, we integrated 30 full-length scRNA-seq datasets from human and mouse, including 6 human and 24 mouse datasets, and retained 60 206 high-quality brain cells after filtering, including 56 623 mouse cells and 3583 human cells.

Mouse cells were classified into 11 clusters, with major populations including GABAergic neurons (GABA), glutamatergic excitatory neurons (GLUT), and microglia (Fig. 4a). Neurons constituted the largest fraction (52.6%), including GABA (25.5%), GLUT (22.4%), and medium spiny neurons (MSN, 4.6%), while microglia represented the most abundant non-neuronal population (25.4%). We then quantified eRNA expression at the single-cell level across these brain cell subtypes. The number of detected eRNAs per cell varied markedly among subtypes. Vascular leptomeningeal cells, oligodendrocyte precursor cells, oligodendrocytes, and pericytes showed relatively higher eRNA detection, whereas microglia and neural stem cells showed lower detection levels (Supplementary Fig. S11a). Compared with mRNAs and lncRNAs, eRNAs were detected at lower frequencies but showed clear subtype-dependent variation (Supplementary Fig. S11b).

Figure 4.

Cell specific expression pattern of eRNAs in brain. (A) A total of 56 623 high-quality mouse brain cells are projected by t-SNE plot. The distinct colors in the illustration represent brain cell types, while the presence of contour curves delineates the boundaries among these clusters. In the circular plot, the external axis illustrates the log scale of the total number of each cell type. The circular plot includes four color-coded tracks, from outer to inner, they represent cluster (mirroring the central t-SNE’s color scheme), project ID, expression levels of marker eRNAs, and the count of expressed eRNAs, respectively. The legends denote the marker eRNAs (left) and project ID (right) of each cell type. The bottom corner insets show t plots colored by marker eRNAs (left) and the number of expressed eRNAs. (B) The bar plots from left to right show the number of cells, expressed eRNAs, expressed eRNAs in single cells, expression levels of eRNAs, the number of marker eRNAs, the number of marker mRNAs, and the number of marker lncRNAs of each cell type. The cell types belonging to neurons are framed in pink while non-neurons are in green. (C) Volcano plot of differentially expressed eRNAs in the Alzheimer’s disease mouse model. Red dots indicate eRNAs that are significantly upregulated in the disease model, while blue dots represent those that are significantly downregulated. (D) UMAP visualization of subclustered microglia, highlighting the eRNA26986 + microglia state. (E) Dot plot showing the expression of representative microglia-related genes in eRNA26986 + microglia and other microglia. Dot size represents the percentage of cells expressing each gene, and color indicates the average expression level. (F) The bar plots show the proportions of each cell type in human adult, fetal, and organoid. (G) The cell distribution of human adult, fetal, and organoid brain colored by cell types. (H) The heterogeneity scores of eRNAs in different status of brain.

We then identified cell-type-specific marker eRNAs. Although mRNAs had a higher average number of markers, eRNAs showed greater variability across cell types, whereas lncRNAs had the fewest markers (Fig. 4b). Brain marker eRNAs were mainly nonconserved across most brain cell subtypes, except for GABAergic neurons and pericytes (Supplementary Fig. S11c). Although TE-eRNAs were more abundant in the mouse background eRNA set, brain marker eRNAs showed a relative increase in SE-eRNA contribution, with SE-eRNAs predominating in several glial, progenitor, and endothelial populations (Supplementary Fig. S11c).

The oligodendrocyte precursor cell cluster, despite having the fewest cells, exhibited the highest eRNA expression levels (Fig. 4b). Oligodendrocyte precursor cells are essential for myelin sheath development and maintenance [68]. Similarly, vascular leptomeningeal cells, despite their relatively low abundance, exhibited the highest number of marker eRNAs. These observations indicate that eRNA activity is not simply determined by cell number, but instead varies across brain cell subtypes in a cell-state-dependent manner.

Cell-type-specific eRNAs may also provide clues for disease-associated regulatory states. For example, a microglia-specific eRNA (chr13:68028916–68035522, corresponding to eRNA26986) was co-expressed with Trem2, a known risk gene for Alzheimer’s disease, suggesting a potential association with neuroinflammatory states. We further examined this observation using data from an Alzheimer’s disease mouse model (GSE288170) [69], in which this eRNA was significantly upregulated compared with controls (Fig. 4c). In addition, several Alzheimer’s disease-associated eRNA loci, including one located on chromosome 1, showed significant upregulation in the Alzheimer’s disease mouse model (Fig. 4c). We further subclustered microglia to test whether marker eRNAs could define finer cell states. A microglial subcluster with enriched expression of eRNA26986 also showed higher expression of microglia-related genes, including Trem2 [70], Apoe [71], and P2ry12, and was defined as an eRNA26986⁺ microglia state (Fig. 4d and e). These results suggest that brain cell-type-specific eRNAs may help capture finer regulatory states within brain subtypes.

To investigate the role of eRNAs in brain development, we analyzed two human brain datasets [72, 73] comprising adult brain, fetal brain, and organoid tissues. The cell composition varied significantly among the three states: adult brains were predominantly composed of GLUT, microglia, astrocytes, and oligodendrocytes, while fetal brains consisted mainly of neuronal cells and some neural progenitor cells. Organoids were dominated by neural progenitor cells (Fig. 4f and g). eRNA heterogeneity was higher in fetal and organoid cells than in adult cells, suggesting more dynamic eRNA activation and repression in these developmental states (Fig. 4h). These findings further support the context-dependent and development-associated heterogeneity of eRNA expression in brain cells.

eRNAs exhibit pronounced heterogeneity and potential oncogenic roles in human cancers

Previous studies have shown that eRNAs can act as regulatory elements and potential therapeutic targets in cancer [6, 74, 75]. However, pan-cancer eRNA heterogeneity at single-cell resolution remains unclear. We analyzed 9149 cells from 11 human cancer datasets. After quality control, 8863 cells were retained and grouped into 16 clusters, mainly epithelial cells (56.5%, 5005/8863) and immune cells (28.8%, 2551/8863), including macrophages (77.4%), B cells (15.4%), and T cells (7.3%) (Fig. 5a). Consistently, eRNAs also showed stronger cell-type-restricted expression than lncRNAs and mRNAs across tumor-associated cell populations (Supplementary Fig. S12a). We next summarized cell composition and eRNA expression features across cancer types, revealing marked differences in tumor-cell abundance, cell-type composition, the number of expressed eRNAs, and average eRNA expression levels among cancers (Fig. 5b). To further investigate eRNAs in tumor heterogeneity, we focused on epithelial cells, among which 2941 of 4967 cells were classified as aneuploid by CopyKAT (Supplementary Fig. S12b). In aneuploid tumor cells, BC-HER2 showed the highest number of expressed eRNAs, whereas myxoid liposarcoma showed the fewest, indicating substantial cancer-type-dependent variation in eRNA activity (Fig. 5b). We further examined the cancer-type specificity of eRNA expression. Compared with lncRNAs and mRNAs, eRNAs showed a much stronger cancer-restricted expression pattern, with nearly half of expressed eRNAs detected in only one cancer type and most eRNAs detected in fewer than five cancer types (Fig. 5c). In contrast, only 3.2% of lncRNAs and 1.8% of mRNAs were specific to one cancer type. For example, eRNA chr10:5485675–5492022 was expressed in two cancer types, including three breast cancer subtypes, with the highest expression in BC-HER2. This eRNA overlaps NET1e, a previously validated breast cancer-associated oncogenic eRNA [18]. These results suggest that eRNAs exhibit pronounced cell-state and cancer-type heterogeneity at single-cell resolution.

Figure 5.

The cell heterogeneity of eRNAs in human cancer. (A) The t-SNE plot of all 8863 cells from human cancer datasets, colored by annotated cell types. (B) A multi-dimensional analysis of cellular and eRNA characteristics for each cancer type. The panels, arranged from left to right, are the cell count, the proportion of each different tumor-cell type, both the count and mean expression value of eRNAs within each individual cell type, the number of expressed eRNAs, and the average expression value of expressed eRNAs, respectively. In the rightmost panel, the dot size represents the number of expressed eRNAs, while the color intensity reflects their mean expression value. (C) The proportions of cancer-specific expression of eRNA, lncRNA, and mRNA, colored by the number of cancer type. (D) The expression distribution of eRNAs in epithelial cells for each cancer type. The molecular subtypes of BC are framed by pink. (E) Left: the number of marker eRNAs in each different cancer type. Middle: the expression level of top three marker eRNAs in each cancer type. Right: the enriched gene ontology (GO) terms of marker eRNAs for each cancer type. The color of bars and dots indicates the type of cancer. (F) The specific enriched gene ontology terms of each cancer type. (G) The genomic structure of eRNA chr14:54443897–54450392 and CDKN3. (H) eRNA-mediated cell–cell communication among different cell types for each cancer. The size of each pie indicates the number of eRNAs involved in these interactions. The proportion within each pie reflects the percentage of communication. The color of the bottom circle corresponds to the cell type. (I) Top: the distributions of the number of expressed eRNAs in each individual cell in different drug resistance states. Middle: the distributions of the average expression values of expressed eRNAs in each individual cell in different drug resistance states. Bottom: the counts of marker eRNAs (bottom) in different drug resistance groups.

The number of expressed eRNAs varied across cancer types and was independent of the number of cells in each cancer (Wilcoxon rank-sum test, P = .525). We next compared eRNA expression between aneuploid and diploid epithelial cells. eRNA expression was higher in aneuploid than diploid cells in most cancer types (77.8%, 7/9), except glioblastoma and OSCC (Fig. 5d). In breast cancer, epithelial cells from high-grade triple-negative breast cancer (TNBC) showed higher eRNA expression than those from low-grade subtypes in both diploid and aneuploid states (Fig. 5d). TCGA-BRCA data analysis further showed higher eRNA expression in patients with poor prognosis than in those with good prognosis (P < .001, Wilcoxon rank-sum test, Supplementary Fig. S12c). For example, eRNA chr3:104519945–104526347 was highly expressed in TNBC and was linked to 108 genes, including APOE, CSF1R, CX3CR1, SPP1, TNF, and TREM2, which are associated with breast cancer progression [76–79].

We further identified tumor-cell marker eRNAs for each cancer type. Both the number and co-expression-based annotations of these marker eRNAs differed among cancers (Fig. 5e, Supplementary Table S10). For example, CC marker eRNAs were linked to DNA damage response, neuroblastoma marker eRNAs to ncRNA processing, and BC-luminal B marker eRNAs to RNA splicing. Some cancer-specific eRNAs were also associated with immune-related genes (Fig. 5f). For instance, TNBC-specific eRNA chr3:104516499–104524037 was co-expressed with 11 HLA family genes. Another HCC-specific eRNA, chr14:54443897–54450392, may regulate CDKN3, a gene implicated in cell cycle regulation, apoptosis, and tumor–immune evasion [80, 81] (Fig. 5g). These results suggest that cancer-specific eRNAs may reflect tumor-type-specific regulatory states.

To explore eRNA-associated cell–cell communication, we integrated eRNA co-expression networks with CellChat-inferred ligand–receptor interactions. In most cancers, eRNA-associated interactions mainly occurred within epithelial cells (Fig. 5h). In BC-TNBC, however, most eRNA-associated communications occurred between immune cells and other cell types (73.6%, 1181/1605), suggesting that eRNAs may be linked to tumor–immune interactions in this subtype.

We next examined eRNAs in drug resistance using the OSCC cisplatin-response dataset (GSE117872), which includes drug-holiday, drug-sensitive, and drug-resistant cells [82]. Drug-resistant cells showed a greater number of expressed eRNAs and higher eRNA expression levels than the other groups (P < .001, Wilcoxon rank-sum test, Fig. 5i). TF–eRNA network analysis identified two marker eRNAs, chr14:34502365–34510596 and chr17:64676912–64684451, as potential targets of SOX9, a TF associated with cancer drug resistance [83]. Another eRNA, chr11:65908577–65914942, was linked to resistance-associated genes including CFL1, EIF1AD, and FOSL1 [84–86].

To further characterize eRNA-associated regulatory programs in tumor cells, we performed pySCENIC analysis of eRNA-linked TFs. In total, 200 eRNA-associated TF regulons were identified, including 51 regulons with higher activity in aneuploid tumor cells and 149 regulons with higher activity in diploid tumor cells (Supplementary Fig. S13a and b). Among them, RAD21 showed increased regulon activity in aneuploid tumor cells, whereas several TFs, including SIN3A, IRF1, TCF4, RELB, BCL3, MXI1, SOX2, RELA, and PITX1, showed higher regulon activity in diploid tumor cells. These results suggest that eRNA-associated TF regulons differ between aneuploid and diploid tumor-cell states (Supplementary Fig. S13c and d). In the OSCC cisplatin-response dataset, Monocle pseudotime analysis revealed a trajectory associated with drug-response states. Sensitive cells were mainly distributed at the early pseudotime stage, with a median pseudotime of 0.020, whereas drug-holiday cells occupied an intermediate state, with a median pseudotime of 2.35, and resistant cells were enriched at the late pseudotime stage, with a median pseudotime of 7.60 (Supplementary Fig. S13e and f). Along this trajectory, 5969 eRNAs were significantly associated with pseudotime, including 1906 positively correlated and 4063 negatively correlated eRNAs (Supplementary Fig. S13g). These dynamic eRNAs formed distinct temporal expression modules, suggesting that eRNA activity changes during the transition from cisplatin-sensitive to resistant states.

To explore the clinical significance of marker eRNAs, we analyzed their expression profiles and associations with clinical characteristics (age, gender, stage, and survival time) using TCGA data. Among 23 357 marker eRNAs, 13 615 (58.3%) were linked to at least one clinical characteristic, and 5060 (37.2%) were associated with prognostic outcomes (Fig. 6a, Supplementary Table S11). We further validated the marker eRNA chr4:74621763–74628200 in HCT116 cells using RACE experiments. Specific primers successfully captured complete sequences from 5′ to 3′ ends. Agarose gel electrophoresis revealed amplified fragments, including a 942-bp 3′ RACE product and a 198-bp 5′ RACE product (Fig. 6B). Spliced sequences matched perfectly with the original, confirming full transcription (Fig. 6C). To assess its diagnostic potential, we examined chr4:74 621763–74628200 expression in plasma samples from CC patients and healthy individuals. Sanger sequencing confirmed the specificity of the real-time qPCR amplicon, and a pUC57-eRNA plasmid standard curve supported reliable absolute quantification (R² = 0.995, P < .001; Supplementary Fig. S14a and b). The eRNA’s absolute copy number was significantly higher in CC patients than in healthy controls (4 551 vs. 2919, P < .001, Wilcoxon rank-sum test, Fig. 6d). Diagnostic efficacy was further confirmed with an area under the curve of 0.73 (Fig. 6e). Postoperative CC patients showed a significant reduction in eRNA levels compared to preoperative patients (4 551 vs. 3132, P = .007, Wilcoxon signed-rank test, Fig. 6f). These results suggest that eRNA chr4:74621763–74628200 is a promising biomarker for CC diagnosis and monitoring, potentially aiding early detection and treatment evaluation.

Figure 6.

The clinical role of eRNAs in cancer. (A) The number of prognostic related marker eRNAs in each cancer type. (B) Top: from left to right: the product of reference nucleic acids, the product of 5′ RACE cloning, the product of intermediate sequences, the product of 3′ RACE cloning. Bottom: the gene structure diagrams show the position of eRNA chr4:74621763–74628200 and adjacent genes. (C) The comparison result between the original sequence and the spliced product sequence by BLAST. These are represented horizontal bars colored by alignment scores. (D) Comparison of absolute eRNA copy number in plasma of healthy individuals and CC patients. (E) ROC curve for distinguishing healthy individuals and CC patients using eRNA chr4:74621763–74628200. (F) Comparison of absolute copy number of eRNA chr4:74621763–74628200 in plasma of preoperative and postoperative patients.

Discussion

Emerging evidence supports the notion that eRNAs play a crucial role in gene transcription regulation in multiple biological processes. Through comparison, we discovered that full-length scRNA-seq is able to detect eRNA expression and the performance is on par with poly(A) RNA-seq and C1 CAGE scRNA-seq. Because poly(A) RNA-seq has been widely used to identify eRNAs and C1 CAGE has demonstrated eRNA detection at single-cell resolution, these comparisons support full-length scRNA-seq as a useful resource for systematic single-cell eRNA profiling [17, 18, 21, 27, 28]. These findings provide insights into eRNA regulatory mechanisms at single-cell resolution and improve our understanding of the roles of eRNAs in diverse biological processes.

By integrating 187 full-length scRNA-seq datasets from human and mouse, we systematically characterized the single-cell expression landscape of eRNAs across embryonic development, brain tissues, and cancer. A major feature of eRNAs was their sparse detection and high intercellular heterogeneity compared with mRNAs and lncRNAs, indicating that eRNA transcription is highly context-dependent rather than broadly ubiquitous (Fig. 2d and e, Supplementary Fig. S6d and e). Despite their low abundance, eRNAs contributed substantially to cell-type- and cell-state-specific marker sets, suggesting that they may provide additional regulatory information beyond conventional mRNA-based annotation.

Our developmental analysis revealed that eRNA activity is particularly prominent during early embryogenesis. Stage-specific eRNAs were enriched in early developmental stages and were associated with stage-dependent TF–eRNA regulatory relationships (Fig. 3h and i, Supplementary Fig. S7c and d). By integrating TF–eRNA and eRNA–target associations, we further identified candidate TF–eRNA–target regulatory programs with stage-associated expression dynamics (Supplementary Fig. S10a–d). These findings suggest that eRNAs may participate in the fine-tuning of developmental gene regulatory programs.

In brain tissues, eRNAs exhibited subtype-dependent expression patterns; brain marker eRNAs were largely nonconserved and showed a relative increase in SE-derived eRNAs compared with the background eRNA set, particularly in glial, progenitor and endothelial populations, suggesting that brain eRNA signatures may reflect species-specific and cell-state-specific regulatory programs (Supplementary Fig. S11c). Notably, marker eRNAs also helped identify finer regulatory states, such as the eRNA26986-positive microglia state associated with microglia-related genes including Trem2, Apoe, and P2ry12 (Fig. 4c–e). These observations suggest that eRNAs may provide useful markers for resolving cellular heterogeneity in neural tissues, although their functional roles in brain development and disease require further mechanistic investigation.

In cancer, eRNA expression was highly heterogeneous across cancer types and tumor-cell states. eRNAs showed stronger cancer-restricted expression patterns than lncRNAs and mRNAs, and aneuploid tumor cells generally exhibited higher eRNA expression than diploid epithelial cells in most cancer types (Fig. 5c and d). Cancer-specific marker eRNAs were associated with diverse biological processes, including immune-related pathways, and eRNA-associated cell–cell communication analysis suggested potential links between eRNAs and tumor–immune interactions, particularly in BC-TNBC (Fig. 5e and h) [87, 88]. In the OSCC cisplatin-response dataset, drug-resistant cells showed increased eRNA expression and distinct eRNA dynamics along the pseudotime trajectory, supporting a potential association between eRNA activity and drug-response state transitions (Fig. 5i, Supplementary Fig. S13e–g).

Finally, we explored the clinical relevance of cancer marker eRNAs. A subset of marker eRNAs was associated with clinical characteristics and prognosis across TCGA cancer cohorts (Fig. 6a). We further validated the transcription of chr4:74621763–74628200 in CC cells using RACE and confirmed its detectability in plasma samples. Its elevated level in CC patients and postoperative reduction suggest that this eRNA may serve as a potential biomarker for diagnosis and postoperative monitoring (Fig. 6b–f). Nevertheless, larger independent cohorts and functional experiments are needed to confirm its clinical utility and biological role.

This study has several limitations. First, because full-length scRNA-seq primarily captures relatively stable or polyadenylated transcripts, some unstable or nonpolyadenylated eRNAs may not be detected. Second, many regulatory relationships were inferred from co-expression, enhancer-target annotation and TF binding information, and therefore require experimental perturbation for validation. Third, the public datasets used in this study varied in sequencing depth, platform, and biological context, which may influence eRNA detection sensitivity. Despite these limitations, our study provides a comprehensive single-cell resource of eRNA expression across human and mouse and offers candidate regulatory programs for future studies of eRNA-mediated regulation in development, neural biology, and cancer.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

This work was supported by the National Natural Science Foundation of China [grant number 32470722]; Ningbo Major Research and Development Plan Project [grant numbers 2023Z226, 2023Z171]; Ningbo Medical and Health Brand Discipline [grant number PPXK2018-02]; the Medical and Health Research Project of Zhejiang Province [grant number 2025KY222]; the Ningbo Municipal Natural Science Foundation [grant number 2021J124]; Zhejiang Key Laboratory of Pathophysiology [grant number 201812]; K. C. Wong Magna Fund in Ningbo University; and the Zhejiang Province Vanguard Goose-Leading Initiative [grant number 2025C01114]. Funding for open access charge: National Natural Science Foundation of China [grant number 32470722].

Ethics approval and consent to participate

The investigation protocol granted approval from the Ethics Committee of Ningbo University (approval no. 2019022501) and was performed in adherence to the ethical principles outlined in the Declaration of Helsinki for medical research including human beings. Before being included in this study, all individuals were required to provide written informed consent.

Data availability

The data that support the findings of this study are openly available in the GEO (https://www.ncbi.nlm.nih.gov/geo/), Single Cell PORTAL (https://singlecell.broadinstitute.org/single_cell), and Single Cell Expression Atlas (https://www.ebi.ac.uk/gxa/sc/home). Publicly available RNA-seq and genotype data were obtained from TCGA through the Genomic Data Commons Data Portal (https://portal.gdc.cancer.gov/). All other data supporting the findings of this study are available within the article and its Supplementary Data.