A cytoplasmic index for quantifying immune-related A-to-I RNA editing

Abstract

Background

Distinguishing self from non-self is a major challenge for the immune system. Endogenous cytoplasmic double-stranded RNA (dsRNA) can mimic viral RNA and activate immune sensors like MDA5. ADAR1-mediated adenosine-to-inosine editing disrupts base-pairing to suppress immunogenicity of these endogenous structures. Global editing indices are widely used to probe this crucial ADAR1 function. However, they are dominated by nuclear pre‑mRNA edits with limited immune relevance. Here we present the cytoplasmic editing index (CEI) that quantifies editing specifically within inverted Alu repeats in 3′ untranslated regions of mature cytoplasmic transcripts, which potentially form cytosolic dsRNA structures carrying higher immunological risk.

Results

Analyzing over 25,000 RNA-sequencing samples, we demonstrate CEI captures ADAR1p150 activity and outperforms the global editing index in terms of sensitivity and signal-to-noise, enabling sharper tissue-specific profiling, enhanced detection power of infection‑induced editing changes, and stronger association with cancer prognoses. An open-source, cloud-native pipeline delivers end‑to‑end, reproducible analysis at very low cost, supporting immediate, scalable adoption.

Conclusions

CEI provides a refined metric for quantifying immune-relevant RNA editing, revealing previously obscured tissue- and disease-specific editing landscapes. The accompanying open-source, cloud-native pipeline enables broad adoption of high-quality editing analysis across research settings. This approach offers new opportunities for investigating ADAR1's role in immunity, infection, and cancer, with potential applications in biomarker development and therapeutic intervention strategies.

Graphical abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s13059-026-04154-3.

Background

Adenosine-to-inosine (A-to-I) RNA editing is a widespread post-transcriptional modification carried out by adenosine deaminase acting on RNA (ADAR) family of enzymes that convert adenosines to inosines within double-stranded RNA (dsRNA) [1–3]. Among the ADAR enzymes, ADAR1 (ADAR) plays a central role in innate immune regulation. Endogenous dsRNAs that resemble viral RNA may be recognized by cytoplasmic RNA sensors and aberrantly activate antiviral pathways, most notably through the MDA5–MAVS axis and PKR, leading to type I interferon (IFN) response [4–6]. Editing by ADAR1 disrupts the base-pairing of the immunogenic self-dsRNAs to the extent that they are no longer recognized by MDA5, or marks them otherwise, preventing inappropriate activation of the antiviral cellular immune system [7–10]. Aberrant editing has been associated with several autoimmune disorders [11–22], and complete loss of ADAR1 is embryonic-lethal due to MDA5-mediated auto-immunity [7–9, 23–25].

In humans, the vast majority of editing events occur within primate-specific Alu retrotransposons [26–32]. These elements are abundant in primate genomes [33, 34], and frequently form long intramolecular dsRNA structures by pairing of neighboring inverted repeats [29, 35]. Most Alu elements lie in intronic and intergenic regions [34, 36]. Of the small fraction incorporated into mature mRNAs, Alu elements are especially enriched in noncoding regions, mainly the 3′ untranslated regions (3′UTRs). These elements may result in dsRNA structures that engage the cytoplasmic innate immune sensors such as MDA5 [7–9, 35].

Given the essential function of A-to-I editing in preventing inappropriate immune activation, there is growing interest in accurate quantification of editing activity [32, 37–44]. Current quantitative methods remain limited in interpretability and precision. The simplest method — counting edited sites — fails to account for the uneven distribution of Alu element coverage and the typically low editing frequency at individual sites within them. Accordingly, it is highly sensitive to sequencing depth and does not provide a robust estimate for enzymatic activity [45]. More advanced metrics, such as the Alu editing index (AEI) [39], calculate a global score by averaging Alu editing frequencies across the human genome, weighted by expression levels. While these are useful for gauging overall ADAR activity [14, 16, 17, 20, 21, 46–64], they conflate signals from heterogeneous transcripts and subcellular compartments including intronic or nuclear-retained RNAs. The pre-mRNA structures are rarely seen in the cytoplasm [65] and are therefore invisible to the cytoplasmic innate immune sensors. Previous analysis [66] has found that only ~ 0.15% of putative dsRNA structures in pre-mRNA exist in mature mRNAs. Furthermore, the vast majority of the global editing signal is due to a very large number of weakly edited Alu repeats [39, 67], whose structure is largely insensitive to ADAR1 perturbation. Consistently, Sun et al. [68] demonstrated that only 1–2% of cellular dsRNAs are truly immunogenic when unedited, bringing the number of critical MDA5 substrates to hundreds or thousands rather than millions. Consequently, aggregate genome-wide editing indices that are mostly dominated by weakly-edited nuclear structures lack specificity for immune-relevant substrates and may be insensitive to the subtle dynamics of editing that regulate innate immunity.

To address this gap, we introduce a cytoplasmic editing index (CEI): a biologically informed metric that specifically quantifies A-to-I editing events within putative cytoplasmic dsRNAs – clusters of inverted Alu elements residing in 3′UTRs, which are expected to form long intramolecular dsRNA structures capable of activating cytoplasmic immune sensors [69]. Integrating both structural and spatial constraints, this metric provides focused measurement of editing in immunologically relevant substrates. The full repertoire of immunogenic dsRNAs remains undefined. Current evidence [25, 68, 69] suggests that only a fraction of cytoplasmic Alu–Alu duplexes are immunostimulatory, and that additional immunogenic dsRNAs may arise from non‑Alu regions [19, 69, 70]. Nevertheless, we propose using the editing status of Alu elements likely to encounter cytoplasmic dsRNA sensors as a biologically meaningful proxy for ADAR1p150 activity, which correlates with its immunosuppressive function.

We demonstrate that the CEI provides enhanced sensitivity for detecting immunologically relevant editing changes and serves as a valuable tool for investigating RNA editing across diverse biological contexts. Benchmarking across over 25,000 RNA-seq samples from infectious disease and normal tissue datasets, we show that the CEI robustly captures editing dynamics under diverse conditions. To support broad adoption, we provide an accessible, cost-efficient cloud-computing framework enabling scalable and systematic quantification of immunogenic dsRNA editing across diverse experimental settings.

Results

A targeted and scalable index for cytoplasmic A-to-I RNA editing

To develop a cytoplasmic editing index that captures immunologically relevant dsRNA substrates, we first established quantitative criteria for identifying Alu elements with immunogenic potential (Fig. 1a). We defined inverted 3′UTR clusters as regions containing ≥ 2 Alu elements (each > 200 bp) with at least one element oriented in each direction, enabling intramolecular dsRNA formation (Fig. 1d). The selected regions are both retained in cytoplasmic transcripts and capable of forming stable secondary structures detectable by cellular dsRNA sensors.

Fig. 1.

Cloud-based pipeline and biological rationale for quantifying cytoplasmic versus global A-to-I RNA editing. a Biological rationale of the cytoplasmic editing index. For an RNA molecule to trigger immune activation, it must (1) localize to the cytoplasm and (2) adopt a double-stranded conformation therein. Following transcription and splicing, mRNAs are transported to the cytoplasm. Reversely oriented Alu repeats in the 3′UTRs can fold into intramolecular double-stranded RNA (dsRNA) structures. When these cytoplasmic duplexes are insufficiently edited, they are recognized by cellular innate immune pathways and trigger type-I interferon responses. The cytoplasmic editing index specifically quantifies editing within these immunologically relevant dsRNA substrates. b Cloud-based analysis workflow. RNA-seq samples are (1) retrieved, (2) processed through a modular cloud pipeline, and (3) analyzed for A-to-I editing patterns. The pipeline is designed for high-throughput processing and optimized for efficient deployment across diverse dataset sizes, with compatibility across multiple cloud-computing environments. c Resource optimization and scalable processing. Each pipeline step operates with minimal computational requirements (virtual CPU and RAM allocations shown beside icons), with processing costs of approximately $0.10 per sample (for GCP and AWS platforms). The containerized Nextflow implementation enables portable deployment to any Nextflow-compatible cloud or local computing environment. Complete parameter specifications are detailed in Methods. d Cytoplasmic vs. global editing index computation. The cytoplasmic editing index is calculated exclusively from RNA-seq reads mapped to inverted Alu clusters within 3′UTRs (Methods). These clusters are likely to form long, intramolecular dsRNA structures that remain stable after splicing and nuclear export. By contrast, the global Alu editing index encompasses reads aligned to any genomic Alu element, regardless of orientation, size or genomic context

Applying these criteria genome-wide (hg38), we identified 3,343 Alu elements across 895 genes within structured 3′UTR Alu clusters (see Methods). The cytoplasmic editing index (CEI) is defined as the aggregate A-to-I editing level calculated exclusively across these immunologically relevant Alu elements. For comparison, we computed the global Alu editing index (AEI), which aggregates editing across 1,113,103 genomic Alu elements (merged contiguous elements of any length) irrespective of orientation or transcriptomic context [39]. The computational strategies underlying both indices are summarized in Fig. 1d.

To enable large-scale comparative analysis of cytoplasmic vs. global editing patterns, we developed a cost-efficient, modular, cloud-based computational pipeline that retrieves RNA-seq data, performs preprocessing, alignment, and quantifies expression and editing (Fig. 1b). The input file for this workflow is a list of Sequence Read Archive (SRA) sample accession codes. Resource allocation is optimized for each step, resulting in a per-sample processing cost of ~ US $0.10 (Fig. 1c). We validated our approach by analyzing 36,334 publicly available RNA-seq samples from diverse contexts — infectious disease studies (15,649 samples; Additional file 2: Table S1), GTEx [71] (9,602 samples) and TCGA [72] (11,083 samples) — demonstrating robust scalability across heterogeneous datasets. All pipeline scripts and reference files are open-access to support reproducibility (Methods).

Few Alu clusters contribute to the CEI

First, we characterize the 3,343 Alu elements within 3′UTRs contributing to the CEI, and their genomic architecture. We compared these to the 5,344 elements in uniformly oriented Alu clusters, including singletons (one isolated 3′UTR Alu element) or tandem repeats (Fig. 2a, b). The latter are not expected to form cytoplasmic intramolecular dsRNAs. Furthermore, uniformly oriented clusters are present in ~ five-fold more genes than inverted clusters (Fig. 2c; 4,011 vs. 791 genes). Similarly, analysis of murine B1 and B2 repeat clusters in mouse 3′UTRs show that uniformly oriented clusters dominate and inverted clusters are restricted to far fewer genes (Fig. 2d, e). These results suggest conserved evolutionary constraints against dsRNA-forming configurations in 3′UTRs.

Fig. 2.

Structural constraint on inverted versus tandem Alu clusters in human 3′UTRs. a Schematic of Alu configurations in 3′UTRs. Inverted Alu clusters can fold into intramolecular dsRNA that can be edited by ADAR and can activate cytoplasmic RNA sensors if unedited, whereas tandem Alu clusters lack complementary elements and do not form dsRNA. b Orientation breakdown of Alu elements in human 3′UTRs (3,343 inverted and 5,344 uniformly oriented elements). Each element is assigned to the largest cluster size in which it participates within each orientation (singleton, pair or ≥ 3). Eighty elements appear in both orientation classes (reflecting different transcript isoforms) and are thus counted twice. Uniformly oriented elements predominate and are overwhelmingly singletons, whereas inverted elements, though fewer overall, make the majority of multi-element clusters. c Number of human genes harboring inverted and tandem 3′UTR Alu clusters. There are 104 genes exhibiting both types, each in a different isoform. d Orientation breakdown of B1/B2 elements in mouse 3′UTRs (650 inverted and 3,184 uniformly oriented elements). Three B1 elements and one B2 element appear in both orientation classes (reflecting different transcript isoforms) and are thus counted twice. Uniformly oriented elements predominate, as seen in human (b). e Same as c for mouse B1/B2 clusters (Euler diagram). f Distribution of the largest Alu cluster per gene (cluster found in the transcript whose 3′UTR has the largest number of Alu elements among all gene’s transcripts), stratified by orientation. Right inset: possible tandem and inverted configurations for cluster sizes 2–4, and expected inverted:tandem ratios under a random-orientation model (see Methods). Left inset: observed vs. expected number of genes with inverted Alu clusters. Labels above each bar indicate gene counts per category; empirical inverted:tandem ratios (bottom, sizes 2–4) compare to right inset expectations. Inverted Alu clusters span a broader size range than tandem Alu clusters, reflecting the larger combinatorial space of mixed orientations. Nevertheless, inverted Alu clusters are consistently rarer than expected (panel c), indicating negative selection against dsRNA-forming configurations. g Depletion of inverted clusters relative to the random model. Bars represent the observed proportion of inverted clusters divided by the proportion expected when orientations are assigned at equal probability (for Alu cluster sizes with > 10 genes per orientation; Methods; Additional file 1: Fig. S1). Values < 100% indicate underrepresentation, consistent with selective constraint against dsRNA-forming orientations. The proportion of observed inverted clusters declines with cluster size (χ² test, no multiple testing correction for the three tests; Additional file 1: Fig. S1)

Examining tandem and inverted cluster abundance as a function of cluster size reveals depletion of inverted clusters compared to tandem clusters (Fig. 2f). For size-2 clusters, the number of tandem configurations exceeds that of inverted ones; for sizes ≥ 3, inverted clusters outnumber tandems (Fig. 2f), but not as much as expected in the absence of evolutionary purification (left inset, Fig. 2f). For a size-N cluster, there are $2^N$ orientation combinations, of which $2^N-2$ are mixed-strand configurations associated with inverted clusters, compared to only two homogeneous-orientation ones associated with tandem clusters. Thus, a random, equally probable and uncorrelated assignment of orientations should result in an inverted:tandem ratio of $2^{N-1}-1$ for a cluster of size N (right inset, Fig. 2f). The observed ratio is well below this expectation (left inset, Fig. 2f), indicating strong depletion of dsRNA-forming configurations (Fig. 2g; Additional file 1: Fig. S1). This finding is consistent with previous reports on evolutionary selection against dsRNA-forming arrangements in mature transcripts [66].

Similar results are obtained when the analysis is restricted to the longest 3′UTR isoform, as well as for analyses of murine B1 and B2 repeat clusters in mouse 3′UTRs (Additional file 1: Fig. S2; Additional file 1: Fig. S3).

We tested whether cluster architecture might be influenced by Alu evolutionary age or subfamily composition, but found only minor differences in sequence divergence and slight deviations from genome-wide proportions (Additional file 1: Fig. S1).

Cytoplasmic index reflects ADAR1p150-dependent immune activation

ADAR1p150, the interferon-inducible cytoplasmic isoform of ADAR1 [2, 73], plays a critical role in preventing cytoplasmic dsRNA from triggering innate immune responses. We thus asked whether the CEI, probing editing in Alu elements within inverted Alu clusters, exhibits elevated sensitivity to editing by ADAR1p150. We compared the CEI to two additional editing metrics: the conventional global Alu editing index (AEI), and a tandem editing index quantifying editing within tandem-oriented Alus (negative control, see Methods). Analysis of knockout [10] and reconstitution [74] datasets revealed that the CEI was selectively sensitive to ADAR1p150 loss or reconstitution compared to the AEI, while the tandem index remained consistently low across all conditions.

As expected, ADAR1 knockout suppressed all editing indices; however, selective loss of ADAR1p150 reduced the CEI more strongly. Similarly, in the reconstitution system, ADAR1 knockdown reduced editing signals in both indices as well. Subsequent ADAR1p150 expression selectively restored the CEI almost to its wild-type levels, while global editing remained diminished (Fig. 3a).

Fig. 3.

Biological perturbations of A-to-I RNA editing reveal differential sensitivity of CEI and AEI. a Impact of ADAR1p150 loss and reconstitution on editing indices. Left [10]: HEK293 cells with ADAR1 and ADAR1p150 knockout (KO) ± IFN-β treatment; right [74]: TBP-ARPE cells with ADAR1 knockdown followed by ADAR1p150 overexpression (each sample cultured and sequenced in triplicate). Loss of ADAR1 activity reduces global editing in both systems, while cytoplasmic editing is selectively impaired by ADAR1p150 absence and restored upon isoform presence, demonstrating that the cytoplasmic index specifically measures ADAR1p150-mediated activity. The tandem index (computed from exclusively tandem Alu elements, see Methods) remains near-zero across all genotypes, validating the assumption that these regions cannot form dsRNA. b ADAR1p150 is required for IFN-β responsiveness. Bars show fold-change in mean editing (IFN-β treated versus untreated) from the ADAR1 KO experiment in panel a (mean ± s.d., n = 3). While IFN treatment enhances the cytoplasmic index in WT cells, this response is abolished in both ADAR1p150 KO and ADAR1 KO cells, confirming isoform specificity. Global and tandem indices show minimal IFN responsiveness in WT cells. c Temporal dynamics of cytokine-induced editing in EndoC-ꞵH1 cells. Two independent time-course datasets (IFN-α, left [75]; mixed cytokines, right [76]) reveal index trajectories over 2–48 h. The cytoplasmic index rises earlier and more dramatically than the global index, while the tandem index remains consistently low and largely unaffected by cytokine treatment. d Fold-induction for datasets shown in panel c (mean ± s.d., n = 5). The CEI consistently exhibits greater dynamic induction compared to the AEI. Apparent fold-changes in the tandem index do not reflect biologically meaningful induction, as the absolute editing levels remain near baseline (panel c). e Species conservation of ADAR1p150 dependence. Left [78], tissue-specific patterns: brain (predominantly ADAR1p110-expressing) shows marked reduction of CEI and AEI upon ADAR1p110 knockout, with a further decrease in the ADAR1p110/ADAR2 double knockout. In contrast, thymus (predominantly ADAR1p150-expressing) maintains high editing across genotypes. Right [79], rescue experiments: editing-deficient mouse embryonic fibroblasts (MEFs; ADAR1⁻/⁻; ADARB1⁻/⁻; Gria2ᴿ/ᴿ) show no editing when expressing RFP controls, modest CEI restoration with ADAR1p110, and complete rescue with ADAR1p150. In both contexts, AEI and tandem indices remain consistently flat. f Subcellular specificity. Analysis of subcellular fractions from HeLa cells [80] reveals highest CEI in the cytoplasm, decreasing progressively in nuclear and chromatin fractions, whereas AEI and tandem indices remain relatively constant across compartments

In addition, the knockout dataset included IFN-β stimulation, enabling assessment of interferon‑induced editing responses across genotypes. In wild-type cells, IFN-β treatment induced a marked increase in the CEI compared to untreated controls, while the AEI showed no significant change and the tandem index remained largely unresponsive. IFN-induced enhancement of cytoplasmic editing was absent in both complete ADAR1 knockout and selective ADAR1p150 knockout cells (Fig. 3b), confirming that cytoplasmic editing induced by interferon specifically requires ADAR1p150.

Time-course analyses in two independent datasets [75, 76] revealed that CEI rose faster and higher than AEI following IFN-α or cytokine treatment, whereas the tandem index remained largely unaffected (Fig. 3c,d). Similarly, the cytoplasmic index exhibits preferential responsiveness to cytokines in human islets [75, 77] (Additional file 1: Fig. S4). The contrasting behavior of inverted versus tandem Alus, despite both being part of cytoplasmic 3′UTRs, demonstrates that structural architecture, not subcellular context alone, determines immune-responsive editing activity.

Analysis of mouse tissues [78] and embryonic fibroblasts [79] demonstrated evolutionary conservation of isoform-specific cytoplasmic editing. In mouse thymus, predominantly expressing ADAR1p150 [78], cytoplasmic editing does not change much upon knockout of ADAR1p110 and ADAR2. Conversely, in mouse brain tissue, primarily expressing ADAR1p110 [78], ADAR1p110 knockout causes marked CEI reduction; combined ADAR1p110/ADAR2 knockout further diminishes this effect. In editing-deficient fibroblasts (ADAR1⁻/⁻; ADARB1⁻/⁻; Gria2ᴿ/ᴿ), cytoplasmic editing was rescued robustly by ADAR1p150, partially by ADAR1p110, and not at all by an RFP control. Across these conditions, the effect on AEI and tandem index is limited, confirming the isoform specificity of cytoplasmic editing (Fig. 3e).

Direct assessment of subcellular localization further supported the cytoplasmic specificity of the CEI. Fractionation experiments in HeLa cells [80] revealed highest CEI in cytoplasmic fractions, intermediate in nuclear fractions, and lowest in chromatin fractions, whereas the AEI and tandem indices presented negligible compartmental variation (Fig. 3f). Thus, CEI demonstrates isoform specificity absent in conventional editing indices, with particular utility for monitoring immune activation and infection responses.

The CEI is defined as a weighted average of editing levels across individual elements, with weights proportional to their expression levels. As such, it is inherently sensitive to both editing efficiency and the relative expression of each element. In principle, a change in CEI could therefore reflect shifts in the expression landscape rather than genuine changes in editing activity. However, because the CEI integrates signals from thousands of regions – each contributing only marginally to the overall index – the influence of expression is likely limited. To empirically assess this, we examined the change in CEI upon IFN induction as reported in Fig. 3a.

We first characterized the contribution of individual elements to the CEI. As expected, in both wild-type (WT) and IFN-treated samples, the individual contribution of each element is very small; > 99.7% of the elements in each condition account for less than 1% of the total index each (Additional file 1: Fig. S5a). Thus, even large changes in the expression of isolated elements are unlikely to substantially affect the CEI. Next, we examined per-element editing levels, which are unaffected by expression variation, and observed that the majority of elements exhibited increased editing upon IFN treatment (Additional file 1: Fig. S5b). Finally, to disentangle the contributions of expression and editing, we constructed two pseudo-indices: one that applies IFN-induced editing levels to the WT expression profile, and another that combines the IFN-induced expression profile with WT editing levels. This decomposition clearly separates the two effects and demonstrates that the observed increase in CEI upon IFN stimulation is predominantly driven by elevated per-region editing levels, with only a minimal contribution from expression changes (Additional file 1: Fig. S5c).

Collectively, these findings establish CEI as an ADAR1p150-sensitive, biologically grounded, and computationally efficient metric that specifically captures editing activity in immunologically relevant dsRNA substrates, a small but critical fraction of all cellular dsRNAs [68, 69], with particular utility for monitoring immune activation and infection responses.

Cytoplasmic editing index provides superior signal, specificity and computational performance

We next evaluated the technical performance of the CEI relative to the widely used global Alu editing index (AEI) [39]. Leveraging our standardized cloud-based computational pipeline (Fig. 1b,c), we quantified both indices across 15,649 publicly available RNA-seq samples from 55 infectious-disease datasets, enabling consistent large-scale benchmarking (Figs. 1b,c; Fig. 5a; Methods). We compared editing signal, signal-to-noise ratio, and mean coverage per editable site to assess the relative advantages of the two indices.

Fig. 5.

Cytoplasmic editing index enables detection of immune-regulated RNA editing changes. a Overview of RNA-seq sample collection and analysis workflow. A total of 15,649 publicly available RNA-seq samples spanning 55 datasets and over 20 distinct infectious diseases were compiled and analyzed using standardized cloud-based pipelines to quantify the CEI and AEI. b Comparison of effect sizes (Cohen's d, mean difference divided by pooled standard deviation) for case–control analyses between CEI and AEI across infectious disease datasets (for all case–control comparisons with ≥ 4 samples per group, ≥ 10 total samples). The green triangle indicates the region where the CEI identified a greater positive effect size than the AEI, demonstrating enhanced biological sensitivity. Points are colored based on statistical significance detected exclusively by CEI, exclusively by AEI, by both indices, or by neither index (Wilcoxon test, paired samples where applicable, see Methods). Case–control status was assigned from sample metadata and pathogen load annotations (Methods; Additional file 2: Table S1; Additional file 2: Tables S3-S4). c Longitudinal CEI analysis of tuberculosis treatment and recovery [81] in human samples. Boxplots show editing index trajectories at several time points after treatment initiation, grouped by recovery time (either week 4 or weeks 8–12). The CEI decreases significantly after recovery, while the AEI shows no significant changes (Wilcoxon rank-sum test, raw p-values denoted under comparisons). d Variability of RNA editing indices across TCGA cancer samples. Each point corresponds to one tissue type, and presents the normalized standard deviation (s.d./mean editing) of the index across samples in primary tumors and normal tissues. Inter-tumor variability of CEI is significantly greater than AEI in primary tumors (Wilcoxon signed-rank test), whereas normal tissues show no significant difference between indices. e Kaplan–Meier survival analysis across TCGA cancer types with ≥ 100 samples. Adjusted P-values (Benjamini–Hochberg correction) represent the optimal separation of samples into two groups based on the editing index (Methods). The two indices showed complementary patterns of survival association, with CEI performing better in some cancer types and AEI in others. f Linear modeling (CEI ~ AEI) across antemortem donors of GTEx for tissues with ≥ 10 samples. Points represent model slopes (± standard error) per tissue, ranked by slope magnitude. Inset shows linear fits for tissues with the highest (stomach) and lowest (muscle skeletal) slopes, highlighting variable tissue-specific relationships between cytoplasmic and global editing indices. g Postmortem vs. antemortem editing index stability across GTEx tissues with ≥ 10 samples in both cohorts. Bars indicate postmortem-to-antemortem mean editing ratios for CEI and AEI, highlighting the reduced susceptibility of CEI to postmortem RNA degradation. Insets show distributions of postmortem and antemortem CEI and AEI values in adipose visceral tissue, exemplifying the robustness of CEI against tissue-degradation artifacts

For both indices, A-to-G mismatches consistently ranked first in frequency (> 99% of samples), with C-to-T second (> 93%; Additional file 1: Fig. S6). The values reported for the two indices were correlated across samples, but CEI consistently outperformed AEI across multiple metrics. CEI delivered a ~ 2.6-fold higher editing signal than AEI (ordinary least squares slope = 2.57; Fig. 4a) and exhibits a 2.15-fold higher signal-to-noise ratio (Fig. 4b), reflecting improved sensitivity and specificity for A-to-I events. This enhanced performance stems partly from CEI's ~ 16-fold greater average coverage per editable site (slope = 16.5; Fig. 4c), reflecting its restriction to mature, cytoplasmic transcripts. These gains were obtained without increasing variability across datasets (Fig. 4d), underscoring CEI’s efficiency and robustness for large-scale RNA editing analysis. Given that ~ 0.3% of the genomic Alu elements are included in the cytoplasmic index, their coverage is ~ 16.5-fold enriched, and their editing ~ 2.15-fold enhanced, we estimate that approximately 11% of the A-to-G mismatches contributing to the AEI are captured in the CEI.

Fig. 4.

Cytoplasmic editing index demonstrates superior signal, specificity, and efficiency over global editing metrics. a–c Scatter-density plots compare the CEI and AEI in terms of a) the editing signal (A-to-G index, b) signal-to-noise ratio (SNR), and c) mean coverage per editable site. Points are colored by sample density; dashed black lines mark x = y, and red lines are ordinary-least-squares fits through the origin, giving slopes of 2.57, 2.15 and 16.5, respectively. SNR is the A-to-G index divided by the maximum non-A-to-G mismatch index (four samples with infinite SNR were excluded). Average coverage is computed per editable site (Methods). d Dataset-level variability. Each point is the standard deviation of the index values across all samples within a single dataset normalized by its mean editing level, restricted to samples with noise < 0.3 in both indices (dataset sizes 22–1,299 samples, Additional file 2: Table S2). No significant difference is found between indices (Wilcoxon signed-rank). e Empirical cumulative distribution of editing noise (maximum non-A-to-G mismatch index per sample). Although the two curves largely overlap, the global index has a slightly larger fraction of samples under the threshold of 0.3 (vertical dashed line), consistent with slightly lower background mismatch variability. Inset: noise density histogram. f, g, Computational efficiency. Bars show the mean ± s.d. wall-clock runtime per sample (f) and peak resident-set memory per sample (g), both recorded with the GNU time utility. Both metrics are substantially lower for the CEI, underscoring its performance advantage

The distribution of the noise per sample is similar for both indices, with a slight advantage to the global index (Fig. 4e). In addition, the cytoplasmic index remained robust to sub-optimal mapping conditions (Additional file 1: Fig. S6). Finally, CEI required substantially lower compute time and memory usage than AEI (Fig. 4f,g), highlighting its efficiency. These advantages are recapitulated in 9,602 Genotype-Tissue Expression [71] (GTEx) samples (Additional file 1: Fig. S7).

The cytoplasmic editing index increases sensitivity to editing alteration in infection

We next compared use of CEI and AEI to detect immune-related editing alterations at the population level across diverse infections. Using our uniformly processed compilation of 15,649 RNA-seq samples spanning more than 20 pathogens and associated controls, we computed CEI and AEI for each sample (Fig. 5a), and searched for infection-associated editing changes (Fig. 5b).

Across infections, CEI consistently showed larger absolute effect sizes and greater sensitivity, achieving statistical significance in 46 comparisons, compared to only 28 significant changes in AEI (Fig. 5b). This advantage was especially pronounced for editing increases during active infection, consistent with CEI's sensitivity to ADAR1p150 upregulation, and was most evident for viral infections and respiratory conditions (Additional file 1: Fig. S8). In a longitudinal tuberculosis cohort [81], CEI captured a progressive decrease in editing after treatment and recovery (Fig. 5c), whereas AEI showed no significant changes, highlighting CEI’s capacity to track dynamic, immune-driven regulation of RNA editing.

Recent studies have highlighted the importance of the ADAR protein in the development of autoimmune and autoinflammatory diseases, most notably in Aicardi–Goutières syndrome [13] (AGS) and type 1 diabetes [21]. These findings might lead one to expect elevated CEI values in such conditions. However, the role of ADAR-mediated editing has primarily been demonstrated during the initial, critical stages of disease onset, and may not be accurately captured in bulk tissue profiles obtained at later time points.

To investigate whether altered editing patterns could still be detected in autoimmune and autoinflammatory diseases, we analyzed RNA-seq data from a diverse set of conditions, including multiple sclerosis (MS), Crohn’s disease, atopic dermatitis, psoriasis, Sjögren's syndrome, and ulcerative colitis [82–86]. In these datasets, we observed no significant differences in the overall editing index (Additional file 1: Fig. S9). These findings likely reflect the relatively advanced disease stages represented in the cohorts examined, where multiple feedback loops and compensatory pathways have already been activated. Moreover, clinical tissue samples are often composed of heterogeneous cell populations, such as whole blood or lesional tissue with infiltrating immune cells, which may dilute or obscure disease-relevant editing changes occurring in specific cell subsets.

Nevertheless, it is worth noting that our previous work did identify a general reduction in RNA editing levels in patients with atopic dermatitis and psoriasis [14, 20]. This suggests that disease- and context-specific effects on editing can still occur, even when global CEI measurements appear unchanged.

Improved detection of RNA editing variability in cancer and across normal tissues using the CEI

We next examined whether the enhanced sensitivity of CEI extended beyond infectious diseases. We first looked at variability and patient stratification in cancer samples and controls from the TCGA dataset. CEI captured greater editing variability than AEI across primary tumors but not normal tissues, underscoring its capacity to detect heterogeneity in tumor-associated editing changes (Fig. 5d). Furthermore, Kaplan–Meier survival analyses showed that CEI and AEI provided better patient stratification in different cancer types across TCGA, highlighting complementary strengths and underscoring the clinical value of using both metrics (Fig. 5e). CEI showed stronger signal and lower noise relative to AEI in TCGA tumors, and the indices differed in the cancer types they most strongly distinguished, reflecting complementary strengths (Additional file 1: Fig. S10).

To assess the robustness of CEI across tissues and donor cohorts, we analyzed normal samples from both antemortem (living samples: organ donor, surgical) and postmortem donors, taken from the GTEx [71] dataset. The editing index is affected by postmortem processes, such as hypoxia, cellular metabolism and apoptosis [39, 60], making this distinction critical for evaluating tissue-specific measurements. The correlation between CEI and AEI varied substantially across tissues, revealing that different tissues exhibit specific immune-relevant editing landscapes (Fig. 5f; Additional file 1: Fig. S11). As a notable exception, cortex-derived brain tissues uniquely exhibited CEI values lower than the AEI ones, potentially reflecting distinct editing dynamics driven by high expression of ADAR2 and ADAR3 in this neural tissue [47, 87] (Additional file 1: Fig. S11). Importantly, CEI was less susceptible to postmortem changes (Fig. 5g; Additional file 1: Fig. S11). These findings underscore the stability of CEI across donor cohorts and support its use for quantifying RNA editing in both high-quality and partially degraded tissues.

Discussion

Quantifying immunologically relevant A-to-I RNA editing represents a critical challenge in understanding innate immune regulation, as existing metrics fail to capture the biological specificity required for mechanistic insights. The cytoplasmic editing index (CEI) provides a biologically informed complement to existing RNA editing metrics by targeting the subset of editing events most likely to influence cytoplasmic dsRNA sensing.

The qualitative difference between CEI and global editing metrics is manifested in CEI's selective ADAR1p150 dependence. Notably, comparison with recently mapped immunogenic dsRNA regions revealed substantial overlap with CEI targets: 41% overall (559/1,355), increasing to 76% (542/710) for 3′UTR regions and 81% (111/137) for those with the highest cytoplasmic editing levels [68]. Consistently, the significant depletion of inverted Alu clusters relative to random expectations provides an evolutionary validation for CEI's targeting strategy, indicating a strong selective pressure against dsRNA-forming arrangements, consistent with previous findings [66]. This evolutionary constraint, shown here for human and mouse, confirms that the repeats targeted by CEI targets are enriched in functionally critical editing substrates.

The specific double‑stranded RNAs that trigger innate immune activation remain poorly characterized. Potential contributors to the cytoplasmic pool of immunogenic dsRNAs include cis‑natural antisense pairs, Alu‑derived structures not represented in RefSeq transcripts but present in the cytoplasm due to intron retention or 3′UTR extension, and dsRNAs formed from other repetitive elements such as endogenous retroviruses (ERVs) [19, 60, 69]. The CEI does not account for these largely uncharacterized putative substrates of ADAR1p150 or related RNA‑processing enzymes. Moreover, current evidence suggests that some cytoplasmic Alu–Alu duplexes included in the CEI are immunogenically inert [25, 69, 70]. Given that the complete repertoire of immune‑relevant dsRNAs is still unknown, the CEI framework deliberately focuses on a defined subset of ADAR1p150 editing targets to yield a measurable indicator of enzyme activity, while acknowledging that not all of this activity is directly linked to immune modulation.

The CEI is a weighted average signal across numerous Alu clusters and may therefore be influenced by variations in their relative expression, even when the editing rate at individual elements remains constant. Such shifts can arise from changes in the expression of the host gene or from alterations in isoform composition. In particular, alternative polyadenylation, frequent in cancer and immune activation, can lead to 3′UTR shortening, resulting in the loss of Alu‑containing regions or the removal of one member of an editable Alu pair, thereby impacting the CEI. Our analyses indicate that these transcriptomic effects are largely averaged out in the aggregated CEI signal, and the dominant determinant of CEI variation is ADAR1p150 enzymatic activity. Nonetheless, coordinated 3′UTR remodeling, such as widespread shortening, could disrupt this relationship. Future work integrating isoform‑resolved quantification and RNA secondary‑structure profiling may help disentangle changes in editing efficiency from changes in substrate availability.

Our scalable, cost-effective computational pipeline delivers substantial technical advantages, and is particularly suitable for large-scale or heterogeneous datasets. Per-sample processing costs are reduced to ~ $0.10 while maintaining reproducibility through containerized, cloud-native workflows. This proved crucial for our analysis of > 25,000 samples across diverse contexts, demonstrating that CEI consistently outperformed AEI with higher editing signal, improved signal-to-noise ratio and higher coverage per editable site, without increased variability. These advantages translated into enhanced biological detection across infectious diseases, complementary patient stratification for cancer contexts, and reduced susceptibility to postmortem artifacts.

Importantly, our implementation of the CEI is highly accessible and enables widespread adoption. The user is required to supply just a list of SRA accession codes, and all steps from data download to index calculation are fully automated.

Conclusions

CEI's enhanced sensitivity suggests a possible clinical utility for monitoring immune status in infectious diseases and autoimmune conditions where ADAR1 dysfunction drives pathology. The improved cancer stratification hints at potential prognostic applications, though prospective validation would be required. By targeting functionally meaningful substrates rather than aggregate editing, CEI provides a powerful analytical framework for dissecting how A-to-I editing shapes immune regulation in health and disease, complementing existing global metrics.

Methods

Terminology

Alu element

A single Alu element > 200 bp as defined in the RepeatMasker database. It may or may not overlap other Alu elements.

Alu cluster

The set of all Alu elements located within a 3′UTR of a curated RefSeq transcript. Alu elements that overlap the 3′UTR partially and exceed beyond its limits are trimmed to include only the subsequence overlapping with the 3′UTR. Two distinct 3′UTR isoforms may define two distinct Alu clusters.

Inverted Alu cluster

An Alu cluster containing at least one element of each orientation.

Tandem Alu cluster

An Alu cluster that is not an inverted Alu cluster, where all Alu elements have the same orientation.

Infectious disease data collection and annotation

Fifty-five publicly available infection-related RNA-seq datasets, comprising 15,741 RNA-seq samples, were obtained from the NCBI Sequence Read Archive (SRA) [88]. Datasets were annotated via Sequence Read Archive (SRA) metadata, with additional curation from Gene Expression Omnibus (GEO) [89] SOFT files when metadata were incomplete. The datasets are listed in Additional file 2: Table S1. Accession numbers, metadata, and per-sample annotations are summarized in Additional file 2: Table S4.

Each study was classified as case–control or alternative design (e.g., longitudinal, treatment). For longitudinal datasets, two time-points were selected as case and control based on study specifications, as detailed in Additional file 2: Table S1. The time-point with the highest pathogen load was chosen as the case group whenever possible, unless a different selection criterion was noted in the primary publication.

For multi-strain pathogens, strains were analyzed separately only if each strain group contained at least four samples. Otherwise, strains were combined into a single group. Datasets containing multiple libraries from the same donor were treated as independent observations; no donor-level random-effect correction was performed. Case–control comparisons are detailed in Additional file 2: Table S3. Pathogen classification by clinical manifestation and pathogen type is listed in Additional file 2: Table S6.

Of 15,741 samples in this cohort, 92 failed to run due to various issues, most commonly because of mismatched paired-end lengths (e.g., 50 bp + 150 bp) or missing FASTQ files. The failed samples and a detailed breakdown of failure reasons are provided in Additional file 2: Table S5.

ADAR1 perturbation data collection

In addition, we analyzed 147 publicly available RNA-seq samples from datasets profiling genetic or cytokine perturbations of RNA editing regulators in human and mouse along with cell fractions [10, 74–80]. These datasets included experimental conditions such as ADAR knockdown or reconstitution, as well as interferon stimulation time courses along with sequencing of cell fractionation protocol. Accession numbers, metadata, and sample groupings for these experiments are provided in Additional file 2: Table S7.

Autoimmune diseases data collection

We analyzed 384 RNA-seq samples from 4 independent publicly available datasets encompassing 9 autoimmune conditions in human tissues. Accession numbers and disease-control groupings for these datasets are provided in Additional file 2: Table S8.

GTEx data collection and annotation

Raw RNA-seq data for 9,602 samples from the Genotype-Tissue Expression v10 (GTEx) [71] project were obtained via dbGaP [90] (phs000424). Only normal samples sequenced with paired-end, unstranded, short-read protocols with available FASTQ files via SRA were included. Surgical and organ-donor samples were classified as antemortem (tissues collected from living donors), whereas postmortem samples originated from deceased donors. Accession numbers, metadata, and sample group assignments for all analyzed GTEx samples are provided in Additional file 2: Table S9.

TCGA data collection and analysis

Aligned RNA-seq data from The Cancer Genome Atlas [72] (TCGA) were obtained from the Genomic Data Commons (GDC) [91] via dbGaP (phs000178). We excluded multi-mapped and unmapped reads, to retain only uniquely mapped primary alignments compatible with our reference annotations, using the command samtools view -h -b -q 255 -F 2308 -f 2. Of the initial 11,083 samples, 452 samples were excluded due to the absence of reads after filtering. Accession numbers, metadata, filtering outcomes, and sample group assignments for all analyzed TCGA samples are provided in Additional file 2: Table S10. Detailed workflow adaptations for TCGA data are documented and available on GitHub.

Cloud-based computational pipeline and data processing

All computational analyses used our cloud-optimized Nextflow pipeline engineered for cost-efficient processing of high-throughput RNA-seq data. The workflow runs on Google Cloud Platform (GCP) Batch and Amazon Web Services (AWS) Fargate, but is portable to other providers owing to its modular design. All steps are executed using version-pinned Docker images to ensure reproducibility across analyses. Stage-specific resource allocation reduced average end-to-end cost to ~ $0.10 per sample (Fig. 1c).

Initialization: a one-time initialization step retrieves all reference annotations and generates index files required for expression quantification, alignment and RNA quantification. The resources are then deposited in a bucket of the user’s choice. This step is done once per computing platform and organism.

Per-sample processing: this includes per-sample tasks — read retrieval, quality control, expression quantification, alignment to the genome, and RNA-editing quantification. For each sample, these steps are executed sequentially in discrete containers on resource-optimized virtual machines. Samples are processed in parallel, enabling thousands of concurrent workflows and simplifying quota management and monitoring. The workflow accepts a list of SRA accession IDs (and an NGC file, for dbGaP data) and produces sample-level outputs ready for downstream analysis.

Retrieving all reference annotations: reference annotations and expression files are downloaded from UCSC Genome Browser [92] (http://hgdownload.soe.ucsc.edu/goldenPath; https://hgdownload.soe.ucsc.edu/gbdb) canonical data sources and processed using standard genomics tools (awk, bedtools [93] and bigBedToBed [92]). Specific resource URLs are detailed in GitHub.

Generation of index files required for expression quantification: Salmon [94] (version 1.10.2) index was generated with RefSeq [92] (latest available version as of early 2024) as the reference transcriptome. The entire genome was used as the decoy sequence as per Salmon’s comprehensive index generation protocol.

Generation of index files required for alignment: for each read-length group we built a length-matched STAR [95] index (--sjdbOverhang = R, where R ∈ {50, 75, 100, 125, 150}).

Generation of index files required for RNA editing quantification: for each regions file, we built the appropriate JSD index using the RNAEditingIndexer. This index is used during the editing quantification step to avoid per-sample re-generation and reduce resource usage.

Resource deposit: resources are transferred to a bucket supplied by the user via AWS CLI or gcloud CLI, according to the cloud provider.

Read retrieval and quality control: raw sequencing reads were retrieved using NCBI SRA Toolkit (v3.0.7), and preprocessed with fastp [96] (v0.23.4) for quality filtering, with non-default options -q 25 -u 20 -e 30 --dont_eval_duplication -G -A. For each dataset (library), we extracted X, the average read length per mate as annotated in SRA. We defined L as the element in the set of standard lengths (L ∈ {50, 75, 100, 125, 150}), that is closest to X. If X ≥ L-3, reads shorter than L-3 bp were discarded using the flag --length_required L-3. Reads longer than L were trimmed to this size from the 3’ using --max_len1 L. Else (X < L-3), reads shorter than X-3 bp were discarded using the flag --length_required X-3, and reads longer than X were trimmed to this size from the 3’ using --max_len1 X. In both cases, the alignment index for this data was set to L. In some cases, SRA annotations were found to be inconsistent with read data and were manually curated. All read lengths decisions for both X and L are detailed in Additional file 2: Table S4 at sample level. Studies containing multiple read-length groups (e.g., 50 bp and 100 bp libraries) were split, and each group of specific read-length was processed independently. GTEx samples were processed using X = L = 75 bp as per their documented sequencing protocol.

Expression quantification: transcript expression levels were quantified using Salmon [94] (version 1.10.2) with automatic library detection (-l A) and transcript-to-gene mapping (-g) using a relevant resource file.

Alignment to the genome: unique alignments were generated with STAR [95] (v2.7.10b), with the relevant index and the following parameters:

--alignSJoverhangMin 8 --alignIntronMax 1000000 --alignMatesGapMax 600000 --outFilterMultimapNmax 1 --outFilterMatchNminOverLread 0.95. The index length was chosen to match preprocessed read lengths as previously detailed. For indices robustness testing, we selected 1,000 random samples and realigned them with the non-AEI-optimal default STAR parameter --outFilterMatchNminOverLread 0.66 while keeping other flags unchanged. Comparison of noise is presented in Additional file 1: Fig. S6b.

RNA-editing quantification: RNA editing levels within Alu regions (AEI, CEI, tandem index) were quantified from STAR alignments using RNAEditingIndexer [39] (v1.0) with RefSeq curated data as reference. Common variants from dbSNP151 [97] (hg38) for human samples or dbSNP142 (mm10) for mouse samples were filtered to exclude known polymorphisms from the analysis. RNA editing levels of known sites of human [98] and mouse [99] are quantified via pipeline using a custom-built program based on samtools [100] mpileup file format.

Resource usage monitoring: runtime and peak memory usage were measured using the GNU time utility and aggregated across samples to benchmark computational efficiency. Results are presented in Fig. 4f and g.

The complete Nextflow workflow, pipeline initialization and documentation are available at https://github.com/a2iEditing/CEI [101]. All known sites and CEI annotations, including pre-computed inverted Alu cluster definitions for human and mouse reference genomes (hg38 and mm10 accordingly), are provided in the repository.

Cytoplasmic index regions

The cytoplasmic editing index is calculated over inverted Alu clusters. These were identified by scanning 3′UTR exons for ≥ 2 Alu elements (each exon separately), each spanning > 200 bp, with at least one Alu element in each of the two strand orientations. Alu elements ≤ 200 bp were excluded, even when positioned inside a candidate cluster. For mouse regions, B1 and B2 families were analyzed separately, with clusters considered inverted when at least one element within a family showed inverted orientation relative to other elements in the same family. Mouse repeats were required to span > 120 bp.

Elements extending beyond annotated 3′UTR boundaries were trimmed to retain only the sequence overlapping the 3′UTR. Transcriptomic annotations were obtained from UCSC Genome Browser [92] (hg38 and mm10 RefSeq Curated, February 2021). Adjacent or overlapping Alu elements were merged with pybedtools [102] (v0.9.1, wrapping bedtools [93] v2.30). The lists of clusters and associated elements are available in Additional file 2: Table S11, Additional file 2: Table S12 and the project’s GitHub repository.

Tandem index regions

Tandem index regions are Alu elements that belong to tandem Alu clusters, do not belong to any inverted Alu cluster, and in addition do not have an inverted Alu element of any length (even ≤ 200 bp) occurring anywhere in the same 3′UTR, to ensure structural independence from potential dsRNA-forming configurations. The same transcript annotations (hg38 and mm10 RefSeq Curated, February 2021) and 3′UTR-trimming rules were applied. Adjacent elements were merged with pybedtools [102] (v0.9.1; bedtools v2.30) using the same parameters as for inverted clusters. For mouse samples, tandem regions were defined separately for B1 and B2 families. The resulting tandem set represents non-dsRNA-forming repetitive element clusters and serves as a control for evaluating orientation-specific editing activity.

RNA editing index calculations

RNA editing levels were quantified using RNAEditingIndexer [39] (v1.0). Two editing indices were computed for all samples, with a third index calculated for specific analyses:

• Global editing index (AEI): quantified editing levels across all annotated Alu elements genome-wide using default hg38 settings for human samples or mm10 settings for mouse samples (B1 and B2 elements).

• Cytoplasmic editing index (CEI): quantified editing within inverted Alu clusters (or inverted B1/B2 elements for mouse) as defined above, using the -rb flag. This metric restricts analysis to cytoplasm-localized, dsRNA-forming clusters and serves as the primary biologically-informed alternative to the AEI.

• Tandem index: calculated for control analyses (Fig. 3) using non-inverted 3′UTR Alu elements via the -rb flag as a structural control.

• All indices used mismatch profiles across specified regions, masking common SNPs (dbSNP151 for hg38; dbSNP142 for mm10). RNAEditingIndexer incorporates gene expression estimates to guide strand assignment and transcript disambiguation during editing quantification (for algorithmic details, see Ref. [39]).

• RNA editing metrics for all samples included in this study are available in Additional file 2: Table S8 and Additional file 2: Tables S13–S16.

Inverted cluster count model

To compute the expected ratio of the numbers of inverted and tandem clusters, we noted that of all 2ᴺ possible allocations of orientation to elements in a cluster with N Alu elements, only two configurations result in a tandem cluster — either all forward or all reverse. Assuming each element is independently oriented, with equal probability for forward or reverse orientation, one expects an inverted-to-tandem ratio:

$$\frac{2^N-2}{2}=2^{N-1}-1$$

Calculation of average coverage per editable site

Mean coverage per editable site is calculated as the total read depth at all A and T positions in the reference sequence, divided by the number of such positions. We excluded contributions from mismatches other than A-to-G and T-to-C. Both forward strand A positions and reverse strand T positions are included since A-to-I editing can occur on either strand depending on transcript orientation.

Statistical analysis of infectious diseases cohort

Case–control comparisons in infection datasets were assessed with Wilcoxon rank-sum tests on noise-filtered data (maximum non-A-to-G mismatch < 0.3); Wilcoxon signed-rank tests were applied when study design allowed. Comparisons were considered only when filtered case and control groups included ≥ 4 samples each and ≥ 10 samples in total. If a dataset included multiple tissues or read lengths, each tissue and read length were compared separately. Technical replicates were merged after noise filtering by averaging over editing measures across all replicates. For paired comparisons, unpaired samples (due to experimental design or filtering) were excluded.

Calculation of standard deviation for ratio of means

Standard deviations for ratios of group means were calculated using error propagation. For each group, the sample mean and standard deviation were computed using the unbiased estimator (denominator n-1). We estimated the coefficient of variation as $CV=\frac{s}{\underset{\_}{x}}$ for each group.

For the ratio $R=\frac{{\underset{\_}{x}}_a}{{\underset{\_}{x}}_b}$, the variance was calculated using the delta method, assuming independence between groups. The standard deviation of the ratio is:

$$SD(R)=\sqrt{Var(R)}=R\times \sqrt{\frac{s_a^2}{{{\overline{x}}_a}^2{n}_a}+\frac{s_b^2}{{{\overline{x}}_b}^2{n}_b}}.$$

Linear modeling and comparative analysis of GTEx cohort

To compare the two indices using linear modeling, we first filtered out noisy samples (maximum non-A-to-G mismatch index ≥ 0.3 for either AEI or CEI). Donor replicates within tissues were merged by averaging over editing measures across all replicates. The filtered and merged cohort included 9,097 samples.

Ordinary least squares regression (CEI ~ AEI, no constant term) was applied to each tissue type with ≥ 10 samples, to find the slopes. To compare antemortem and postmortem cohorts, we repeated the analysis for each cohort separately, requiring ≥ 10 samples in both cohorts.

Survival analysis and comparative analysis of TCGA cohort

Following noise filtering and sample merging as described above, the TCGA cohort included 10,488 samples.

For each TCGA cancer type with ≥ 100 samples, we split the data based on index values to perform survival analysis. Samples were ranked by editing index value and divided into high and low groups using nine different thresholds at 10% increments (10th percentile, 20th percentile, …, 90th percentile, see Additional file 2: Table S17). For each threshold, Kaplan–Meier [103] survival analysis was performed in R v4.5.0 using the survival package (v3.8–3) (survfit()) and log-rank testing (survdiff()). P-values were adjusted for multiple testing (for all cancer types and thresholds, CEI and AEI separately) using the Benjamini–Hochberg procedure to control the false discovery rate, and the threshold yielding the lowest log-rank P-value was chosen. This approach ensures fair comparison between indices by allowing each to be evaluated at its optimal discrimination threshold, while controlling for the multiple comparisons inherent in the threshold optimization procedure.

Statistics

All statistical tests presented in the article are two-sided.

Authors’ contributions

EYL and EE conceived the study and supervised the work. EYL, EE and RCF developed the CEI methodology. IT designed and implemented the cloud platform workflow and conducted the sample analysis. Resource initialization and data analysis were performed by RCF. IT and RCF annotated infectious disease. HK collected and curated infectious disease datasets and contributed to the initial analyses. SHR designed and implemented the known-sites editing quantification program and developed the Docker container. EYL, EE and RCF wrote the manuscript with feedback from IT. All authors read and approved the final manuscript.

Funding

This material is based upon work supported by the Google Cloud Research Credits program with the award GCP19980904. E.Y.L. was supported by the Israel Science Foundation grant 2637/23 and by NIDDK (DK133442). E.E was supported by the Israel Science Foundation grant 867/25.

Data availability

All code, pipeline definitions, and configuration files are released under the open-source GNU General Public License v3.0 (GPL-3.0) and are publicly available in GitHub at https://github.com/a2iEditing/CEI [101] and in Zenodo at https://doi.org/10.5281/zenodo.19158873 [104]. Custom Docker images are archived at https://hub.docker.com/u/levanonlab [105]. Raw RNA-seq data were downloaded from the NCBI Sequence Read Archive (SRA) [88]; accession numbers are provided in Additional file 2: Table S4, Additional file 2: Table S7 and Additional file 2: Table S8. Controlled-access data from TCGA [72] (dbGaP accession phs000178) and GTEx [71] (dbGaP accession phs000424) were obtained under approved data-use agreements; accession numbers are provided in Additional file 2: Table S9 and Additional file 2: Table S10.

All computational steps were executed in version-pinned Docker containers orchestrated by Nextflow [106] (v25.03.1-edge) to ensure end-to-end reproducibility. Statistical analyses were conducted in R v4.5.0 using tidyr v1.3.1, dplyr v1.1.4, purrr v1.0.4, data.table v1.17.6 and stringr v1.5.1. Publication-ready figures were generated in R using ggplot2 v3.5.2, cowplot v1.1.3 and patchwork v1.3.1. Venn diagrams were generated with DeepVenn [107]. Illustrative figures were created with BioRender.com.

Data generated or analyzed during this study are included in this published article and its additional files. 

Declarations

Ethics approval and consent to participate

Public RNA-seq data were downloaded from the NCBI Sequence Read Archive (SRA) [88]. Controlled-access data from TCGA [72] (dbGaP accession phs000178) and GTEx [71] (dbGaP accession phs000424) were obtained under approved data-use agreements.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.