Identification of brain region-specific landscape and functions of clustered circRNAs in Alzheimer’s disease using circMeta2

Abstract

Alzheimer’s disease (AD) is an age-related neurodegenerative disorder with regulatory RNAs playing significant roles in its etiology. Circular RNAs (CircRNA) are enriched in human brains and contribute to AD progression. Many circRNA isoforms derived from same gene loci share common back splicing sites, thus often form clusters and work as a group to additively regulate their downstream targets. Unfortunately, the coordinated role of clustered circRNAs is often overlooked in individual circRNA differential expression (DE) analysis. To address these challenges, we develop circMeta2, a computational tool designed to perform DE analysis focused on circRNA clusters, equipped with modules tailored for both a small sample of biological replicates and a large-scale population study. Using circMeta2, we identify brain region-specific circRNA clusters from six distinct brain regions in the ENCODE datasets, as well as brain region-specific alteration of circRNA clusters signatures associated with AD from Mount Sinai brain bank (MSBB) AD study. We also illustrate how AD-associated circRNA clusters within and across different brain regions work coordinately to contribute to AD etiology by impacting miRNA-mediated gene expression and identified key circRNA clusters that associated with AD progression and severity. Our study demonstrates circMeta2 as a highly accuracy and robust tool for analyzing circRNA clusters, offering valuable molecular insights into AD pathology.

circMeta2 performs differential expression analysis for clustered circRNAs. Using this tool, the authors identify brain region-specific circRNA clusters in healthy controls, as well as their key dysregulations associated with Alzheimer’s Disease progression.

Introduction

Alzheimer’s disease (AD) is a prevalent age-related neurodegenerative disorder that accounts for 50%-60% of dementia cases in elderly individuals worldwide¹. Approximately 6.7 million seniors currently live with Alzheimer’s disease in the US, and this number is projected to nearly double to 13.8 million by 2060 barring major medical breakthroughs². Major AD symptoms are characterized by cognitive decline, memory impairment, language difficulties, and compromised problem-solving, severely affecting daily living³. Significant efforts have focused on studying the association of genetic variations with AD severity, identifying numerous genetic AD risk factors⁴. For instance, mutations in APP, PSEN1, and PSEN2 are the major causes of familial AD by dysregulating the formation of amyloid plaques and tangles⁵, whereas mutations in APOE4 and many other genes contribute to sporadic AD by influencing amyloid β aggregation and clearance⁶.

Expanding beyond these well-characterized genetic factors, increasing evidence demonstrates the important roles of non-coding regulatory RNAs in AD, such as microRNAs (miRNAs) and long non-coding RNAs (lncRNAs)^7,8. Particularly, emerging studies have linked AD pathology to a relatively new and less studied class of regulatory RNAs with a circularized structure termed circular RNAs (circRNAs)^9,10. CircRNAs are derived from a unique, non-canonical splicing process known as “back-splicing”, where a downstream splice donor in mRNAs is joined to an upstream splice acceptor to form a closed continuous loop without 5’ and 3’ ends, granting them great stability and resistance to degradation¹¹. Notably, circRNAs show highly tissue-specific expression patterns and are strongly enriched in the mammalian brain compared to other tissues, indicating their critical roles in brain function¹². Due to their relatively long half-lives resulting from resistance to exoribonucleases, circRNAs can accumulate to high levels during the aging process and are proposed to serve as “memory molecules” in the brain^13,14. Mechanistically, accumulating evidence suggests multifaceted roles for circRNAs, including acting as miRNAs sponge¹⁵, interacting with RNA-binding proteins (RBPs)¹⁶, and regulating transcription and alternative splicing process¹⁷.

With the application of sensitive and genome-wide circRNA detection methods from RNA-seq data, it has become increasingly clear that more than 50% of circRNAs are derived from alternative back-splicing (ABS) events, resulting in widespread clusters of circRNAs with common back-splicing junction (BSJ) sites and shared sequence compositions^11,18. CircRNAs within a cluster could potentially exert a cumulative effect in miRNA sponging, acting as competitive endogenous RNAs to absorb and sequester miRNAs, thereby reducing their availability and regulatory impact on silencing mRNA targets. This mechanism can broadly impact target gene expression, enabling a sophisticated and fine-tuned gene regulatory network^12,13. Unfortunately, despite the intriguing biological roles of circRNA clusters, individual circRNA components may not show robust changes and are often overlooked by existing computational tools that focus solely on individual circRNA identification and differential expression (DE) analysis¹⁹. In addition, most of the current DE analysis methods adopt existing algorithms, such as DESeq2 and edgeR, which are specifically designed to identify differential mRNA expression^20–22. An optimized computational framework to identify stage-specific clustered circRNAs from large cohorts of patients is urgently needed.

In this study, we developed a computational tool, circMeta2, to effectively identify DE circRNA clusters compatible with datasets of various levels of sample size. First, we demonstrated circMeta2’s utility by performing a comprehensive analysis of tissue-specific circRNA cluster expression patterns across multiple brain tissues and various organs, successfully identifying unique circRNA clusters and highlighting its versatility in analyzing diverse biological systems. We further focused on its functional capabilities by analyzing circRNA clusters in the frontal cortex and cerebellum using the ENCODE dataset, showcasing its effectiveness detecting functional circRNA clusters on a small sample size of biological replicates. Next, we applied circMeta2 to a large population-scale analysis using the Mount Sinai brain bank (MSBB) AD cohort study, identifying region-specific dysregulated circRNA clusters associated with Alzheimer’s Disease. Our study revealed both the coordinated roles of distinct circRNA clusters and the specific roles of same circRNA clusters across different brain regions in regulating AD-relevant miRNAs and their target genes involved in AD pathology. Furthermore, we demonstrated that critical circRNA clusters are continuously depleted in association with AD progression. Collectively, our study not only presents circMeta2 as a robust and effective computational tool for identifying and quantifying circRNA clusters, but also provided potential molecular mechanisms underlying circRNA clusters’ contribution to AD pathogenic progression.

Methods

ENCODE and MSBB data

Ribosomal RNA (rRNA)-depleted RNA-seq datasets and matched miRNA-seq datasets were collected the across six brain regions including cerebellum, diencephalon, frontal cortex, occipital lobe, parietal lobe, and temporal lobe, as well as six different organs including spinal cord, heart, liver, lung, ovary, and spleen, each with two biological replicates from ENCODE²³ (Supplementary Data 1). Moreover, rRNA-depleted RNA-seq datasets from 364 human brains were selected from the Mount Sinai Brain Bank (MSBB) cohort²⁴. The age at death of the selected samples ranged from 61 to 108, with a mean of 84.7 ± 9.7, and included 238 females and 126 males primarily of European ancestry. All ages 90 and above were censored to 90+ for HIPAA compliance. Neuropathological assessments were conducted according to the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) protocol and were used for classification of AD severity status: 1 = Normal, 2 = Definite AD, 3 and 4 = Possible AD. RNA-seq data from four Brodmann areas-the frontal pole (BM10), the superior temporal gyrus (BM22), the parahippocampal gyrus (BM36) and the inferior frontal gyrus (BM44) were used for analysis, considering their high vulnerability to Alzheimer’s Disease²⁵.

Identification of circRNA and circRNA clusters

CircRNAs were identified through the widely recognized circRNA prediction method CIRCexplorer2, using rRNA-depleted RNA-seq data²⁶. To ensure confidence in the identified circRNAs, extremely long (>10 kb) and short (<100 bp) predicted circRNAs were removed from further analysis. Only circRNAs with back-splicing junction read counts (BSJ RC) > 2 were retained for the downstream data analysis. CircRNA clusters were defined based on the proximity and type of alternative back-splicing events. Specifically, clusters were formed by grouping a minimum of two circRNAs that either shared a common 5’ back-splicing site (A5BS) or a common 3’ back-splicing site (A3BS). This approach allowed us to categorize two types of circRNA clusters: A5BS clusters or A3BS clusters, where circRNAs share the same 5’ or 3’ back-splicing junction.

Differential expression analysis for individual circRNA for a small sample size of RNA-seq data

circMeta2 accommodates the identified circRNAs in the output format of CIRCexplorer2 to perform DE analysis for individual circRNA for a small sample size of RNA-seq data, e.g., biological replicates for tissue or cell type. Let $Y_{ij}$ denote the read counts for circRNA i in condition j. $Y_{ij}$ can be modeled by either a NB distribution $Y_{ij}~NB\left(s_{ij}*{\mu }_{ij},{\varphi }_{ij}\right)$ or a Poisson distribution as $Y_{ij}~Pois\left(s_{ij}*{\mu }_{ij}\right)$ where $s_{ij}$ adjusts for sequencing depth, ${\mu }_{ij}$ is the mean expression level for the ith circRNA in the jth condition, and ${\varphi }_{ij}$ is the dispersion parameter. The expression level of circRNA is measured by the junction reads that span the back-splicing sites. As the junction read counts in the back splicing sites are generally less over-dispersed compared to mRNA read counts²². To account for this uncertainty for data modeling, we chose a distribution assumption for the junction read counts based on deviance goodness of fit (GOF) test for between NB and Poisson distributions for each circRNA. circMeta2 consider the GOF when selecting the appropriate distribution assumption and statistical test to identify differentially expressed (DE) circRNA on an individual level. Specifically, if the NB distribution shows an overall better GOF, circMeta2 uses edgeR for DE analysis as default²¹. If Poisson distribution shows a better fit, circMeta2 employs an approximated Z-test to compare the Poisson rates between two conditions.

$$z_i=\frac{{\tilde{\mu }}_{i2}-{\tilde{\mu }}_{i1}}{\sqrt{\frac{{\tilde{\mu }}_{i2}}{n_2}+\frac{{\tilde{\mu }}_{i1}}{n_1}}},{\tilde{\mu }}_{ij}={\sum }_{ijl}\frac{Y_{ijl}}{s_l},j=1,2$$

Where $l$ represents different replicates. For low circular reads, a square root transformation is applied to ensure that z-test statistics approach normality. The Z-score is approximated normal, and p-values can be readily computed.

$$z_i=\frac{\sqrt{{\tilde{\mu }}_{i2}}-\sqrt{{\tilde{\mu }}_{i1}}}{\frac{1}{2}\sqrt{\frac{1}{n_2}+\frac{1}{n_1}}}$$

Differential expression analysis on individual circRNA for a population-scale RNA-seq data

Similarly, circMeta2 leverages GOF to select the most appropriate distribution, either a Negative Binomial or Poisson distribution, to model junction read counts within in a generalized linear model (GLM) framework. Depending on the distribution, this results in either a Negative Binomial GLM or a Poisson GLM, with different choices of link functions. Unlike the Z-test used for DE analysis on a small sample size, GLMs can accommodate covariates such as gender and race, which are common in population-scale RNA-seq data. Hypothesis testing focuses on the condition variable (e.g., disease versus normal) using the Wald test to determine the DE circRNAs.

A weighted Z-score approach for performing DE analysis on circRNA cluster

Meta-analysis is a widely adopted statistical technique used to combine the evidence from multiple studies to provide a more accurate estimate of the effect. As the confidence of evidence (e.g., effect size or sample size) varies across different studies, a weighted Z-score approach has been proposed to combine p-values, resulting in a combined p-value to evaluate the overall evidence²⁷. We hereby adopt the idea of meta-analysis to identify DE circRNA clusters by leveraging the effect sizes of each individual circRNA in the cluster, which are obtained from the aforementioned statistical tests either for a small sample size of RNA-seq data or for a population-scale RNA-seq data.

In the context of DE analysis for circRNA clusters, each individual circRNA in the cluster may have different expression levels and different DE directions across conditions. Instead of grouping all junction reads from all circRNAs in a circRNA cluster for performing DE analysis directly, we leverage the meta-analysis approach, considering the heterogeneity of individual circRNAs in both expression levels and DE directions within the cluster. First, DE analysis is conducted independently for each circRNA component, using statistical approaches as mentioned previously, either for a small sample size of biological replicates or population-scale RNA-seq data. Next, for each circRNA cluster, we adopted a weighted Z-score scheme integrated with Stouffer’s method²⁸ to obtain a combined p-value from individual circRNAs. P-value for each circRNA is first converted into Z-scores using the standard normal cumulative density function. We further adopt two weights to weight the Z-scores. The first directional weight $w_d$ indicates the alignment of direction between the cluster and individual circRNA within the cluster. CircRNAs sharing the same DE direction as the circRNA cluster are upweighted, while those in the opposite direction are downweighed. Without loss of generality, circRNAs aligning in DE direction with the cluster receive a directional weight $w_d$ of 2 and those in opposite directions receive a weight of 0.5. The second expressional weight $w_e$ considers the expression level of each individual circRNA within the cluster. $w_e$ is essentially the proportion of log-transformed read counts of each circRNA among all log-transformed read counts of all circRNAs in the cluster. Therefore, the overall weight of each Z-score for each circRNA, considering both expression level and DE direction, is the product of the directional weight $w_d$ and expressional weight $w_e$. Subsequently, we apply Stouffer’s Z-score method to calculate a weighted Z-score for the circRNA cluster, defined as:

$$Z=\frac{{\sum }_{i=1}^k{w}_d{w}_e{Z}_i}{\sqrt{{\sum }_{i=1}^k{\left(w_d{w}_e\right)}^2}}$$

A combined p-value is obtained assuming the weighted Z-score follows the standard normal distribution. This p-value is used to assess the statistical significance of DE circRNA cluster. To account for multiple tests, we apply the Benjamin-Hochberg method to adjust the p-values for controlling the false discovery rate²⁹.

Analyzing circRNA clusters in MSBB AD cohort study

DE analysis for individual circRNA is performed using Poisson GLM, where circRNA expression is treated as the outcome and other personal traits as the covariates, such as gender and age. The covariate of interest is three AD stages: Definite AD, Possible AD, and Normal. Two comparisons Definite AD vs Normal and Possible AD vs Normal are performed separately with the Normal stage as the reference level. Log2 fold change and p-value are reported for each circRNA in the DE analysis. CircRNA clusters are further identified on the proximity and type of alternative back-splicing events. DE circRNA cluster are performed by employing the weighted Stouffer’s method to amalgamate p-values from individual circRNAs in each cluster, considering both the expression direction and expression level.

Predicting biological functions for circRNA clusters

To predict the biological functions of circRNA clusters in miRNA sponging, we first selected the shortest circRNA from each circRNA cluster and extracted its exon sequence from the Circbank database as the common sequence of circRNA cluster³⁰. Using Targetscan_70³¹, we predicted miRNA binding sites for circRNA cluster based on the defined common sequence. We downloaded small RNA-seq data for the respective biological system from ENCODE database and quantified miRNA expression using miRge2.0³². Only miRNAs expressed in the corresponding samples were considered as circRNA clusters’ targets and used for functional prediction. TargetScanHuman database was used to find potential mRNA targets of the identified miRNAs, with targets having weighted context scores above the 80^th percentile considered as top targets³³. Comparisons of expression changes for miRNA top targets and randomly selected, non-target genes were performed using Student’s t-test to identify significant change of miRNA activity (p-value < 0.05). Gene Ontology analysis for the miRNA target mRNAs with consistent upregulation or downregulation as circRNA clusters was performed by PANTHER³⁴.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Results

A robust and comprehensive computational framework for circRNA cluster DE analysis

More than 50% of circRNAs are clustered due to alternative back splicing (ABS), forming groups of circRNAs that share one back splicing junction (BSJ) site (referred to as circRNA clusters hereafter) (Fig. 1A)¹¹. The circRNAs within each cluster share common sequences and can potentially play additive roles in regulating the availability of miRNAs and RNA-binding proteins (RBPs)¹¹. However, most current circRNA differential expression (DE) analytic tools and algorithms focus on individual circRNAs, which could overlook the significant additive gene regulatory roles of clustered circRNAs if individual circRNAs within clusters show insignificant changes. To address this challenge, we developed circMeta2, a computational tool specifically designed for circRNA DE analysis at the cluster level (Fig. 1). As illustrated in the flowchart, circMeta2 is designed to work with various output formats from several major algorithms for circRNA identification and quantification, including CIRI2, find_circ and CIRCexplorer2 (Fig. 1A)^26,35,36. This flexible input compatibility maximizes the utility of circMeta2 for easy and seamless workflow integration. Clustered circRNAs with either alternative 3’ back splicing (A3BS, circRNAs sharing 3’ splicing sites) or alternative 5’ back splicing (A5BS, circRNAs sharing 5’ splicing sites) can be identified by circMeta2 based on shared BSJ sites (Fig. 1A). Importantly, circMeta2 incorporates either statistical tests embedded in circMeta for small sample size RNA-seq data or a GLM framework for large population-scale RNA-seq data²², which enables controlling for confounding variables such as age and gender in the DE analysis of individual circRNAs (Fig. 1B). This design makes circMeta2 versatile in handling various sample sizes. circMeta2 further adopts a weighted Stouffer’s method to combine the p-values of individual circRNAs into a single combined p-value for their residing circRNA clusters, reflecting both basal expression level and DE direction (Fig. 1B).

Fig. 1. Overview of the circMeta2 framework.

A Illustration of circRNA clusters formed by alternative back splicing and their additive roles in miRNA sequestration. B Algorithms used for differential expression analysis of circRNA clusters in datasets with a small number of samples and in larger population-level studies.

Identification and functional characterization of brain region-specific circRNA clusters in the human frontal cortex versus cerebellum

To demonstrate the effectiveness of circMeta2 in identifying and functionally characterizing DE circRNA clusters from RNA-seq with a small sample size, we identify and compare circRNA profiles identified from RNA-seq data with two biological replicates in human frontal cortex and cerebellum from ENCODE³⁷ respectively. Genome-wide circRNAs were first identified by CIRCexplorer2 and then used for comparison between the frontal cortex and cerebellum. As a result, 50,720 unique circRNAs were identified in total, with 4008 (7.9%) circRNAs specifically identified in the cerebellum and 30,373 (59.9%) in the frontal cortex (Fig. 2A). These findings suggest that the number of individual circRNAs expressed in the human frontal cortex is higher than in the cerebellum. Using circMeta2, we identified 12,274 and 11,694 of circRNA clusters in frontal cortex and cerebellum respectively, with 11,687 of these clusters expressed in both brain regions. Among all the human cortical and cerebellar circRNA clusters, circMeta2 found the majority (60.85%) of circRNA clusters were composed of 2–3 individual circRNAs with either A5BS or A3BS (Fig. 2B). From 12,281 total circRNA clusters, DE analysis by circMeta2 using the small sample size setting (Methods) discovered 3909 cortical-enriched and 2,856 cerebellar-enriched clusters, respectively (FDR < 0.05, Fig. 2C).

Fig. 2. Identification of brain region-specific circRNA clusters and their functional roles in miRNA regulation.

A Percentage of circRNAs commonly and specifically expressed in frontal cortex and cerebellum. B Numbers of circRNA clusters composed of different number of individual circRNAs in frontal cortex (blue) and cerebellum (orange). CircRNA clusters derived from A3BS and A5BS were counted separately. C Volcano plot showing 3,909 frontal cortical-enriched (red) and 2856 cerebellum-enriched (blue) circRNA clusters. D Schematic representation of expression changes in individual circRNAs within the PROM1-4323-A3BS cluster, with red arrows indicating upregulation and asterisk denoting significance. Strong binding affinity is shown between the cluster’s common sequence and miR-138-5p. E Cumulative plot indicates log₂FC of miR-138-5p’s top 10% target genes is significantly lower than log₂FC of random selected non-target genes (p-value = 0.049). F Gene Ontology (GO) analysis for miR-138-5p’s target genes upregulated in the frontal cortex compared to cerebellum.

Among the 3909 clusters with higher expression profiles in the frontal cortex than in the cerebellum (log₂ Fold Changes (log₂FC) > 0), the expression of circRNA cluster PROM1-4323-A3BS differed drastically between the frontal cortex and cerebellum (FDR = 3.17 × 10⁻⁸, log₂FC = 8.24, Fig. 2D). A close inspection of circRNA cluster PROM1-4323-A3BS revealed six unique circRNAs within this cluster. However, four out of six were not found to be significantly enriched in the frontal cortex by individual circRNA DE analyses, indicating the importance of functionally investigating clustered circRNAs as groups (Fig. 2D). Using a well-established algorithm, TargetScan, which predicts mRNA targets of miRNA based on conserved seed sequence³¹, we found that all six individual circRNAs derived from circRNA cluster PROM1-4323-A3BS contained a common sequence element harboring two miR-138-5p binding sites that could potentially interfere with the availability and silencing activities of miR-138-5p specifically in the cortex. The predicted miRNA binding sites were further confirmed by the high binding affinity between the cluster common sequence and miR-138-5p using STarMir (Fig. 2D)³⁸.

Consistent with the high expression of circRNA cluster PROM1-4323-A3BS, which could soak and limit the availability and silencing activity of miR-138-5p, its activities were indeed significantly deactivated, evidenced by a higher log₂FC of miR-138-5p top 10% target genes than randomly selected non-target genes, suggesting a global activation of these target mRNAs (Fig. 2E, p-value = 0.049). We then focused on upregulated miR-138-5p target genes evaluated by mRNA DE analysis, which could be regulated by the PROM1-4323-A3BS/miR-138-5p axis. Gene Ontology (GO) analysis revealed these genes are highly related to brain functions, such as nervous system development, generation of neurons, brain development, regulation of neuron migration, and synaptic vesicle cycle (Fig. 2F)³⁴. These findings are consistent with previous reports that miR-138-5p is critical in neuron differentiation and memory^39,40. One of the miR-138-5p target, WW and C2 containing 1 (WWC1) is significantly upregulated in AD and has been reported to play an important role in memory function^40,41. Our analysis demonstrates that circMeta2 is a practically useful tool for circRNA cluster DE analysis in biological experimental settings with a limited number of replicates. To the best of our knowledge, this study also systematically surveys the circRNA cluster landscape in different human brain regions for the first time, revealing their important roles in modulating the availability of regulatory mRNAs.

Applying circMeta2 to identify tissue-specific circRNAs and circRNA clusters

In addition to the comparison between the frontal cortex and cerebellum, we extended our analysis by applying circMeta2 to a wide range collection of tissues and organs. Specifically, circMeta2 was applied to seven major organs, including the brain, spinal cord, heart, liver, lung, ovary, and spleen. Each organ, except for the brain, has two biological replicates of total RNA-seq samples. Within the brain, we further evaluated six distinct brain regions including the frontal cortex, cerebellum, diencephalon, occipital lobe, parietal lobe, and temporal lobe, each with two biological replicates (Supplementary Data 1). By calculating the distances among the seven tissues based on their circRNA expression profiles, we found that six brain regions were more closely related to each other than to the spinal cord, consistent with the difference between central nervous system and peripheral nervous system (Fig. 3A). Tissue-specific circRNA clusters were identified among seven tissues (FDR < 0.05 and a Fold Change > 1) by treating one tissue as a group and other tissues as the other group: cerebellum (1994 clusters), diencephalon (712 clusters), frontal cortex (839 clusters), occipital lobe (806 clusters), parietal lobe (642 clusters), spinal cord (1035 clusters), and temporal lobe (1017 clusters) (Fig. 3B). These findings illustrated circMeta2’s capacity to effectively identify circRNA clusters uniquely expressed in different tissues within the nervous system. To further investigate, we selected frontal cortex as a representative of the brain organ and compared it to five non-CNS organs. Using circMeta2, organ-specific DE circRNA cluster analysis was conducted by treating each organ as a group and using the remaining organs as the other group. Consequently, we identified organ-specific circRNA clusters for brain (2436 clusters), heart (751 clusters), liver (980 clusters), lung (569 clusters), ovary (554 clusters), and spleen (701 clusters) (Supplementary Fig. 1). We also used other brain regions to represent the brain organ and demonstrated the expression of organ-specific circRNA clusters (Supplementary Fig. 2). Collectively, these results confirmed circMeta2’s effectiveness in identifying both tissue- and organ-specific circRNA clusters.

Fig. 3. Tissue-specific circRNA cluster analysis in ENCODE dataset.

A Hierarchical tree of ENCODE data based on circRNA expression profiles. B Heatmap of normalized expression levels of tissue-specific circRNA clusters with FDR < 0.05 and Fold Change > 1. C Venn diagrams comparing individual circRNA and circRNA cluster from ENCODE, validation, and synthetic data for frontal cortex tissue.

To validate the identified circRNAs and circRNA clusters, we used frontal cortex as an example and adopted another non-ENCODE related frontal cortex validation dataset, which consists of two total RNA-seq data⁴² to assess the reproducibility of the findings at both individual and cluster levels. Using CIRCexplorer2, we identified 46,712 individual circRNAs from the ENCODE frontal cortex dataset and 15,252 from the validation dataset, with 10,532 common circRNAs confirmed by the validation dataset (Fig. 3C1). To evaluate the significance of this overlap, we generated a synthetic background dataset by randomly sampling circRNAs from six different brain regions and the spinal cord within the ENCODE dataset, matching the number of circRNAs identified in the ENCODE frontal cortex. We then compared these circRNAs to those identified in the validation dataset (Fig. 3C2). The common circRNAs identified using the ENCODE frontal cortex were more numerous than those identified using the synthetic background dataset. Fisher’s exact test, considering numbers of common and uncommon circRNAs in both ENCODE frontal cortex and synthetic background datasets, revealed that the overlap between ENCODE frontal cortex and the validation was statistically significant, with an odds ratio of 1.57 and a -log10(p) value > 16. Additionally, circMeta2 identified 18,885 expressed circRNA clusters from ENCODE frontal cortex and 10,249 clusters from the validation dataset (expression level > 0.5), with 7,409 common circRNA clusters (Fig. 3C3). The number of common circRNA clusters was also statistically significant with an odds ratio of 1.28 and a -log10(p) value > 16 based on Fisher’s exact test (Fig. 3C4). These results validated circMeta2’s reliability and applicability in identifying tissue-specific circRNAs and circRNA clusters.

Evaluation of brain region-specific landscape and functions of clustered circRNAs in a large cohort of AD samples using circMeta2

Recent studies have established the critical contribution of individual circRNAs to AD pathogenesis^9,10. However, the systematic evaluation of clustered circRNA as groups simultaneously contributing to AD progression has not been studied. Here, we applied circMeta2 to RNA-seq data from the MSBB study²⁴, which contains hundreds of bulk RNA-seq profiled from four different cortex regions (Fig. 4A). Using CIRCexplorer2, we identified 24,373, 20,810, 20,436, and 23,599 individually expressed circRNAs in BM10, BM22, BM36 and BM44, respectively. Among these circRNAs, 17,195 were identified in all four brain regions, while each cortex region exhibited region-specific circRNAs (1998 in BM10, 507 in BM22, 458 in BM36, and 1693 in BM44), underscoring both common and region-specific roles of circRNAs (Fig. 4B). Compared with other regions, 1,998 circRNAs were specifically expressed in BM10, a region with unique cognitive and executive functions, including risk and decision-making and working memory⁴³. These findings indicate the distinct roles of these circRNAs in the frontal pole. According to shared BSJ sites, circMeta2 identified 10,285, 8578, 8412, and 9835 circRNA clusters in BM10, BM22, BM36, and BM44, respectively. Importantly, more than 55% of circRNAs formed circRNA clusters in all four cortex regions (Supplementary Fig. 3). Consistent with the feature in ENCODE data from the cortex and cerebellum, most circRNA clusters were composed of two or three individual circRNAs (Fig. 4C). Using the Poisson GLM followed by weighted Stouffer’s method integrated in circMeta2, we identified hundreds of significantly AD-dysregulated circRNA clusters in each brain region (FDR < 0.05) (Fig. 4D).

Fig. 4. Dysregulation of circRNA clusters in various cortex regions of Alzheimer’s Disease.

A 3D brain map highlighting locations of four brain regions used in this study: BM10 (Frontal Pole), BM22 (Superior Temporal), BM36 (Parahippocampal), and BM44 (Pars Opercularis). B Venn diagram illustrating the overlap of circRNAs identified across the four cortex regions. C Bar plot depicting the number of A5BS circRNA clusters composed of different numbers of individual circRNAs. D Volcano plots showing dysregulated circRNA clusters in Alzheimer’s Disease identified in each cortex region.

One highly expressed circRNA cluster in the BM10 region was circRNA cluster MAN2A1-11161-A5BS (chr5-109091029), which showed aberrant and upregulation in the AD BM10 region (FDR = 5.4 × 10⁻³, log₂FC = 0.83). circRNA cluster MAN2A1-11161-A5BS was composed of three individual circRNAs, yet none of them exhibited significant upregulation by individual circRNA DE analysis, implying that their functional significance was neglected by DE analysis solely on individual circRNA (Fig. 5A, B). Functional prediction suggested that the common sequence shared among all three individual circRNAs within circRNA cluster MAN2A1-11161-A5BS possesses two potential binding sites for miR-9-5p (Fig. 5A), whose downregulation has been tightly linked to AD pathology^44,45. Consistent with the upregulation of MAN2A1-11161-A5BS in AD, we also confirmed that the activity of miR-9-5p is significantly repressed by comparing its target genes with randomly selected miR-9-5p non-target genes (Fig. 5C, p-value = 1.4 × 10⁻⁷). GO analysis revealed AD-upregulated miR-9-5p target genes are closely related to brain functions that potentially influence AD pathology, including chemical synaptic transmission, trans-synaptic signaling, CNS development, neurogenesis, and synapse organization (Fig. 5D). Specifically, several AD risk genes, such as GSK3B, PTEN, and STMN, were upregulated in AD and involved in multiple GO terms (Fig. 5E)⁴⁶.

Fig. 5. Upregulated circRNA cluster in Alzheimer’s Disease and its molecular function.

A Schematic illustration of individual circRNA in cluster MAN2A1-11161-A5BS, showing miR-9-5p binding sites in the common sequence. Expression change in AD is indicated by red arrows (upregulated) and blue arrow (downregulated). B Expression levels of individual circRNAs from the MAN2A1-11161-A5BS cluster in normal control and AD. C Boxplot illustrating that the log₂FC of miR-9-5p target genes is significantly higher than that of randomly selected non-target genes (p-value = 1.4 × 10⁻⁷). D GO analysis of miR-9-5p targets upregulated in AD BM10 region. E AD risk genes targeted by miR-9-5p, upregulated in the AD BM10 region are involved in multiple GO terms. F GO terms for genes targeted by the top 10 dysregulated circRNA clusters via miRNA soaking in four different cortex regions. Each row represents a cluster while each column indicates a specific GO term. Commonly affected GO terms across all four cortex regions are listed.

A similar strategy was used to predict biological processes that regulated by the top 10 dysregulated circRNA clusters from each cortex region (Fig. 5F). Interestingly, multiple GO terms were commonly affected in all four cortex regions by their top dysregulated circRNAs, including synapse organization, vesicle mediated transport in synapse and regulation of neuron projection development (Fig. 5F), all of which potentially contribute to AD pathogenesis.

Coordinated roles of circRNA clusters in AD pathogenesis via miRNA sponging

In addition to the additive effects of different circRNAs from a specific cluster, it is plausible that different circRNA clusters may also soak the same miRNA to regulate its activity in a given brain region. For example, two circRNA clusters in the BM10 region, namely ARMC8-9535-A5BS and WDR17-10690-A3BS, both upregulated in AD, could coordinately modulate the same miRNA according to their shared miRNA binding sites (Fig. 6A, B). As shown in Fig. 6C, individual circRNAs ARMC8-6121 (chr3-137940767:137964025), ARMC8-9453 (chr3-137940767:137942575), and ARMC8-14822 (chr3-137940767:137947852) account for 40.9%, 27.6%, and 31.6% expression of circRNA cluster ARMC8-9535-A5BS (chr3-137940767) in the normal control group, respectively. In comparison, the expression of all three individual circRNA components in this cluster was increased, leading to a significantly increased overall expression of circRNA cluster ARMC8-9535-A5BS (Fig. 6C). Similarly, an upregulation of cluster WDR17-10690-A3BS was also recognized in the AD BM10 region (Fig. 6D).

Fig. 6. Coordination of circRNA clusters in BM10 region in AD pathogenesis.

A Illustration of two AD-upregulated circRNA clusters from BM10 brain region that coordinate to sequester miR-144-3p. Red arrow indicates upregulation of circRNA, and asterisk denote significant expression change. B Schematic representation of structure and components of ARMC8-9535-A5BS and WDR17-10690-A3BS clusters. C, D Expression levels of individual circRNAs from ARMC8-9535-A5BS cluster (C) and WDR17-10690-A3BS cluster (D) in normal control and AD patients. E Boxplot showing that the log₂FC of miR-144-3p target genes is significantly higher than random selected non-targets genes (p-value = 4.14 × 10⁻¹¹). F Gene Ontology analysis of miR-144-3p targets, upregulated in BM10 region of AD.

Interestingly, the common sequences of both circRNA clusters are predicted to bind to 27 miRNAs (Supplementary Data 2). Among them, miR-144-3p could be simultaneously soaked by both clusters, suggesting that the upregulation of ARMC8-9535-A5BS and WDR17-10690-A3BS works coordinately to repress the activity of miR-144-3p in AD. This hypothesis is also supported by the global upregulation of miR-144-3p target genes compared with randomly selected miR-144-3p non-targeted genes (Fig. 6E, p-value = 4.14 × 10⁻¹¹). GO analysis revealed that miR-144-3p target genes are enriched in multiple neuronal functions such as synapse organization, neuron differentiation, and trans-synaptic signaling, potentially contributing to AD pathology (Fig. 6F). Our analysis demonstrates the coordinated roles of distinctive circRNA clusters in the same brain regions to influence a common set of miRNAs, suggesting a sophisticated and fine-tuned network composed of circRNA clusters, miRNAs and mRNAs contributing to AD pathogenesis.

Besides the coordinated roles of different circRNA clusters from the same cortex region, we also explored whether circRNA clusters in different cortex regions could work together to promote AD pathogenesis through the regulation of common AD-related miRNAs. We found that the downregulation of circRNA cluster ARHGAP26-11281-A5BS (chr5-142434003) in BM44 and ARHGAP26-10077-A5BS (chr5-142416760) in BM22 could both release miR-140-5p and enhance its activity in AD via their cluster common sequences, suggesting that different circRNA clusters could regulate the same downstream miRNA in different brain regions (Fig. 7A). Interestingly, miR-140-5p has not only been shown to be aberrantly upregulated in the AD brain but also to help ameliorate AD phenotype when silenced^47,48. A significant binding affinity between the common sequences of both clusters and miR-140-5p was confirmed using STarMir (Fig. 7A)³⁸.

Fig. 7. Brain region-specific functions of circRNA clusters in AD pathogenesis.

A Illustration of AD downregulated circRNA cluster ARHGAP26-11281-A5BS and its common function to sponge miR-140-5p in brain regions BM44 and BM22. Blue arrows indicate downregulation of circRNA, with asterisk denoting significant expression change. B Schematic representation of structure and components of ARHGAP26-11281-A5BS cluster, including two miR-140-5p binding sites in its common sequence. C Expression levels of individual circRNAs from ARHGAP26-11281-A5BS cluster in normal control and AD patients. D Boxplot showing that the log₂FC of miR-140-5p target genes is significantly lower than that of random selected non-targets genes (p-value = 1.22 × 10⁻⁴). E Scatter plot showing common (orange), BM22 region-specific (light green) and BM44 region-specific (light blue) GO terms that regulated by ARHGAP26-11281-A5BS cluster.

CircRNA cluster ARHGAP26-11281-A5BS (chr5:142434003) is composed of four individual circRNAs, ARHGAP26-969 (chr5-142434003:142513670), ARHGAP26-1341 (chr5-142434003:142437312), ARHGAP26-1342 (chr5-142434003:142526946) and ARHGAP26-15800 (chr5-142434003:142500712) (Fig. 7B). Each accounted for 16.8%, 49.6%, 30.4% and 3.2% of the total expression of the circRNA cluster in the normal control. The cumulative level of common sequence among all circRNAs in circRNA cluster ARHGAP26-11281-A5BS was significantly downregulated in AD, which is predicted to coordinately release their commonly bound miR-140-5p and enhance its activity (Fig. 7C). Similarly, circRNA cluster ARHGAP26-10077-A5BS was also significantly downregulated in AD, resulting in the release of miR-140-5p from its common sequence, which also harbors the miR-140-5p binding site (Supplementary Fig. 4A, B). Importantly, by comparing the expression change (log₂FC) of miR-140-5p targets versus random non-target genes, we confirmed that the activity of miR-140-5p was truly enhanced in both BM44 and BM22 cortex regions (Fig. 7D, p-value = 1.22 x 10^-4, Supplementary Fig. 4C, p-value = 1.08 × 10−⁶). Specifically, 61 and 51 miR-140-5p targets were significantly downregulated in BM22 and BM44 and were used to compare the functional impact of miR-140-5p release in these two brain regions by GO analysis (Fig. 7E). Interestingly, numerous overlapping biological processes were commonly affected in BM22 and BM44 regions, including nervous system development and central nervous system development (Fig. 7E), indicating that circRNA could play important and common roles across different brain regions. Collectively, our study suggests that different circRNA clusters from either the same or different cortex regions could work coordinately to sponge miRNA and regulate gene networks that involved in AD pathogenesis.

Functional roles of circRNA clusters in promoting AD progression

Recent studies have identified the progressive dysregulation of multiple circRNAs that correlated with AD severity using the MSBB dataset, shedding light on their functional impact in disease progression and potential roles as biomarkers^9,10. To investigate the roles of circRNA clusters in different AD stages, we first sub-grouped AD samples into normal healthy donors (NP.1 = 1), possible AD (NP.1 = 3 or 4), and definite AD (NP.1 = 2) groups based on neuropathology categories as measured by the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) and identified specific circRNA clusters associated with AD progression based on their continuous accumulation or depletion along the severity scale (Fig. 8A and Supplementary Data 3)⁴⁹. One example of our findings was circRNA cluster ASAP1-12100-A5BS in BM36, which exhibited consistent depletion along with increased AD severity (Fig. 8B, FDR = 2.59 × 10^-4).

Fig. 8. Continuous dysregulation of circRNA clusters with Alzheimer’s Disease (AD) progression.

A. Schematic illustrations of AD brain in condition of normal control, possible AD, and definite AD (Created in BioRender. Zhao, F. (2024) BioRender.com/f01j596). B Expression level of circRNA cluster ASAP1-12100-A5BS in BM36 brain region across normal control, possible AD and definite AD. C Log2FC comparison showing significant differences between miR-455-3p target and randomly selected non-target genes for possible AD vs. normal control (p-value = 0.0010, left) and definite AD vs. normal control (p-value = 4.05 × 10⁻⁶, right). D Bubble plot show GO analysis of miR-455-3p target genes, downregulated in definite AD. Bubble size represents the number of genes per GO term and Z-score indicates the likelihood of a decrease or increase in biological function. E Heatmap displaying log₁₀FDR of top ranked AD risk genes across multiple GO terms that related to brain function and AD pathogenesis with FDR = 1 (blue) indicating gene absence in specific GO terms.

The common sequence of ASAP1-12100-A5BS contains a binding site for miR-455-3p, which has been shown to be aberrantly upregulated in AD-affected brains^50,51. The activity change of miR-455-3p was assessed by comparing the log₂FC of its target genes and randomly selected non-target genes in Possible AD and Definite AD groups, respectively. Consistent with the progressive downregulation of circRNA cluster ASAP1-12100-A5BS, the activity of miR-455-3p was continuously enhanced in both Possible AD (p-value = 0.001) and Definite AD (p-value = 4.05 × 10⁻⁶) (Fig. 8C). To further ascertain the statistical significance of miRNA activity change between Possible AD and Definite AD, the Wilcoxon rank-sum test was applied to compare the differences in log₂FC between Possible AD vs. Normal and Definite AD vs. Normal (p-value < 2.2x10^–16)⁵², and the result strongly supports that downregulation of ASAP1-12100-A5BS would release miR-455-3p, which significantly contributes to the progression of AD.

To investigate the molecular mechanism of the ASAP1-12100-A5BS/miR-455-3p pathway in promoting AD progression, GO analysis was performed for miR-455-3p targets that are progressively downregulated in AD (Fig. 8D). Multiple biological processes and pathways that could be disrupted in AD were identified, including neuronal development, synaptic signaling and nerve system development (Fig. 8D). A closer look also suggests that multiple AD risk genes are progressively downregulated in AD and involved in these biological pathways, such as SNCA, APOE, and PSEN1 (Fig. 8E). Our study identified a significant correlation between the dysregulation of circRNA clusters and the progression of Alzheimer’s disease. Specifically, we noted that the depletion of the circRNA cluster ASAP1-12100-A5BS correlates with increased activity of miR-455-3p in AD-affected brains. This finding highlights a potential regulatory interaction that may impact disease progression.

Discussion

Our study introduces circMeta2, a computational tool specifically designed for circRNA cluster analysis, which, to the best of our knowledge, is among the first of its kind. By applying circMeta2 to ENCODE and MSBB RNA-seq data, we demonstrated its effectiveness in identifying differentially expressed circRNA clusters from RNA-seq data of small sample size and large population-level RNA-seq data. Additionally, we identified both individual and coordinated roles of circRNA clusters in AD pathogenesis, either within brain regions or across multiple brain regions, as well as the functional roles of circRNA clusters across different stages of AD progression.

It has been increasing recognized that circRNAs are critical in normal brain development¹², functions⁵³, and aging⁵⁴. Their genome-wide dysregulation has been linked to various neurological and neurodegenerative disorders^9,55,56. However, the alterations in disease-associated circRNA clusters have not been rigorously explored. Understanding the dynamic landscape of circRNA clusters in various disease state could be key to uncovering underlying disease etiologies. With the growing recognition of circRNAs’ roles in biological processes and human diseases, numerous computational methods have been developed for their analysis, including circRNA identification, quantification, alternative splicing, and functional prediction⁵⁷. However, the lack of methods specifically designed for circRNA differential expression (DE) limits our understanding of their biological functions. Our group have previously developed circMeta for performing DE analysis of circRNAs for small sample studies, which utilized a Poisson-based Z-test to fit the small numbers and sparse distribution of circRNA back-splicing junction read counts²². Building on this, our newly developed computational tool, circMeta2, stands out as an innovative approach for circRNA expression analysis in two major aspects. First, circMeta2 is equipped with two specialized statistical models tailored for different dataset sizes. For datasets with a small number of replicates, circMeta2 adopts the Poisson-based Z-test from circMeta. This approach is advantageous because it effectively handles the challenges posed by limited data size, which is a common scenario in circRNA research. When dealing with large population-level datasets, circMeta2 transitions to using a GLM algorithm. This model is well-suited for analyzing complex datasets that include multiple covariates, allowing researchers to adjust for various factors such as age, gender, and other variables, providing a more comprehensive and nuanced analysis. This dual-model approach of circMeta2 ensures that it is not only versatile across different experimental scales but also maintains high accuracy and robustness in circRNA DE analysis.

The other standout feature of circMeta2 is its innovative approach for DE analysis at the cluster level by employing a meta-analysis approach by combing p values from individual circRNAs within the cluster in a weighted way. The core of this approach lies in its ability to prioritize certain circRNAs within a cluster based on external knowledge, including read counts and regulatory directions. By assigning varying degrees of importance to individual circRNAs, circMeta2 tailors the DE analysis to emphasize more reliable and relevant data points. P-values for each circRNA are converted to Z-scores, which are then weighted according to a predefined scheme that accounts for the read counts and the regulatory direction. This weighting process is designed to enhance the impact of circRNAs with substantial read counts and aligning regulatory directions, while reducing the influence of less relevant ones. The adjusted p-values are then aggregated using Stouffer’s method, combining the multiple tests within a cluster into a single, interpretable statistic, which follows a standard normal distribution, allowing for a robust determination of combined significance.

Although mounting studies have shown the potential biological function of circRNAs in sponging miRNAs or RBPs, one major concern is that most circRNAs show lower stoichiometry, undermining their potential role in titrating other molecules⁵⁸. Alternative back splicing, a commonly occurring event covering a large proportion of circRNAs^11,18,26, allows clustered circRNAs to enhance their soaking capacity through the additive effect, enabling them to bind and sequester miRNAs more effectively, thereby reducing their availability to target mRNAs. Our recent publication has shown that common sequence of circ-cluster ARHGEF28 could potentially bind with miR-454-3p and contribute to oligodendroglia differentiation, indicating the role of circ-cluster in brain function¹¹. In addition, circRNAs from the same cluster are under common regulation both transcriptionally and by the common flanking intron, yet the specific flanking intron provides specific regulation for each circRNA component, leading to a more sophisticated regulatory mechanism²⁶. Although individual circRNA analyses have been conducted in the MSBB dataset⁹, our development of circMeta2 enhances understanding of how the dysregulation of circRNAs at the cluster level contributes to AD pathogenesis. This tool reveals many DE circRNAs that were not detected in previous individual analyses. Furthermore, our study also proposes coordinated roles of multiple circRNA clusters for miRNA sponging, further offering an additional mechanism for the additive effect to strengthen the function of circRNAs as competing endogenous RNAs.

Accumulated in mature neuron and brains¹², circRNAs are well-accepted to play important regulatory roles in neurodevelopment processes and multiple neurodegenerative disorders, including AD^59,60. Recent studies also reported candidate circRNAs whose expression changes are significantly associated with AD severity in multiple brain regions¹⁰, suggesting circRNAs as promising AD biomarkers and therapeutic targets^61–63. However, most of these studies focus on individual circRNAs while leaving circRNA clusters neglected. Our study illustrates that critical functions of clustered circRNAs could also contribute to AD disease progression via their common sequences and similar function to individual circRNAs. Given the strong association between key circRNA clusters and AD progression, AD-associated circRNA clusters might be promising markers for disease progression or even early diagnosis.

Author contributions

B.Y. and L.C. contributed to the conception and design of this study. F.Z. and Y.L. contributed to the primary data analysis, interpretation of analysis results, and presentation of data. F.Z. contributed to the development of the circMeta2 R package. F.Z. and L.C. contributed to the methodology development and drafting. Y.L. and B.Y. provided substantial and critical revisions. L.C. and B.Y. provided resources and guidance throughout the research process. All authors approved the final version of the manuscript for submission.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Ibrahim Javed and Benjamin Bessieres.

Data availability

All Raw data used in this study is publicly available at ENCODE database (Experiment IDs available in supplementary Data 1) and Synapse via SynID: syn3159438. CircRNA clusters differential analysis results for ENCODE and MSBB data is available in Supplementary Data 4.

Code availability

The package circMeta2 is available on GitHub via the link https://github.com/lichen-lab/circMeta2.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-024-07060-1.