ADAR1-mediated RNA editing promotes B cell lymphomagenesis

Summary Diffuse large B cell lymphoma (DLBCL) is one of the most common types of aggressive lymphoid malignancies. Here, we explore the contribution of RNA editing to DLBCL pathogenesis. We observed that DNA mutations and RNA editing events are often mutually exclusive, suggesting that tumors can modulate pathway outcomes by altering sequences at either the genomic or the transcriptomic level. RNA editing targets transcripts within known disease-driving pathways such as apoptosis, p53 and NF-κB signaling, as well as the RIG-I-like pathway. In this context, we show that ADAR1-mediated editing within MAVS transcript positively correlates with MAVS protein expression levels, associating with increased interferon/NF-κB signaling and T cell exhaustion. Finally, using targeted RNA base editing tools to restore editing within MAVS 3′UTR in ADAR1-deficient cells, we demonstrate that editing is likely to be causal to an increase in downstream signaling in the absence of activation by canonical nucleic acid receptor sensing.


INTRODUCTION
Diffuse large B cell lymphoma (DLBCL) is one of the most common and aggressive types of lymphoma. Based on gene expression profiling, two major ''cell-of-origin'' subtypes of DLBCL have been identified, referred to as germinal center B cell-like (GCB) and activated B cell-like (ABC), respectively. 1 In recent years, high-throughput genomic studies further identified recurrent genetic alterations within the key diseasedriving genes/pathways in DLBCL, such as BCL6 and MYC, epigenetic modifiers, p53/DNA damage response pathway, B cell receptor (BCR)/NF-kB-, NOTCH-and JAK-STAT-signaling pathways as well as genes related to antigen presentation and immune evasion. [2][3][4][5][6][7] Based on these genetic alterations, several molecular subtypes have been suggested. 4,6,8,9 Nevertheless, the disease heterogeneity of DLBCL cannot be fully explained by the genetic changes, both regarding the biological nature of the cancers and their response to treatment. 10 Although DNA alterations (mutations or more extensive structural changes such as large deletions, duplications and translocations) or gene expression changes can be critical to tumor initiation, maintenance and progression, post-transcriptional RNA modifications together with altered RNA binding protein activity (the 'epitranscriptome') have recently been characterized as contributors to clonal evolution in cancer, through dynamic determination of RNA function and gene expression in response to environmental stimuli. [11][12][13][14][15] Of all epitranscriptomic modifications recently described, RNA editing is the most common and straightforward to catalog and validate. RNA editing occurs by hydrolytic deamination of cytosine (C) to uracil (U) (mediated by the activation-induced cytidine deaminase (AID)/APOBEC family of proteins) or of adenosine (A) to inosine (I) (decoded as guanosine, mediated by the ADAR family of proteins), which cause respectively thymine (T) or guanine (G) base changes in cDNA, as compared to the genomic reference 16 (hence the term ''editing''). Next-generation sequencing approaches have been particularly adept at identifying inconsistencies in base-pair composition between the genomic sequence (DNA) and its corresponding transcribed RNA (cDNA), which are then referred to as putative RNA editing events. Based on In contrast to genomic mutations, which represent the historical archive of insults to the DNA, RNA editing is an active mark: If the editing enzyme is inhibited or lost, editing disappears, and the transcriptome reverts to faithfully mirroring the expressed genome of the cell. However, an unedited transcriptome is also a rich source of dsRNAs, and acute loss of ADAR1 leads to inflammation and cell death in healthy tissue through the accumulation of unedited endogenous dsRNA and consequent activation of corresponding sensing pathways. [20][21][22][23] Hypomorphic mutations in ADAR1 were observed in patients with Aicardi-Goutiè res syndrome, which is characterized by uncontrolled interferon (IFN) expression and secretion 24 and belongs to a group of autoinflammatory disorders referred to as type I interferonopathies. 25 In the context of tumor tissues, it has recently been shown that general loss of ADAR1 (and of editing) causes cell-intrinsic lethality through endogenous dsRNA mediated induction of IFN, a key anti-tumor cytokine, and subsequently to robust inflammation of the tumor microenvironment (with substantial tumor infiltration by CD8 + T cells) and increased response to immune checkpoint inhibitors. [26][27][28] Here, we profile the RNA editing events and assess the role of ADAR1 in immune evasion in DLBCL. We used matched DNA and RNA sequencing datasets derived from DLBCL tumor biopsies to precisely determine the transcripts that are edited in DLBCL and the signaling pathways they represent. We observed that specific pathways associated with disease progression can be altered either by mutation and/or by editing -an unexplored concept in lymphomagenesis, which we then validated using targeted base editing to perturb the specific transcripts (MAVS, encoding mitochondrial antiviral signaling protein), on an editing-null background. We also assess the role of ADAR1-mediated MAVS editing in mediating inflammation through the induction of IFN/NF-kB, and find that a lower level of ADAR1 or MAVS in B cell lymphoma does not correlate with an increase of IFN response as observed for solid tumors (such as lung adenocarcinoma or pancreatic cancer 29 ). Increased level of ADAR1 or MAVS is, however, associated with increased T cell exhaustion. Overall, ADAR1-mediated RNA editing represents an additional mechanism underlying B cell lymphomagenesis, and our findings have implications for ADAR1-centered immunotherapies currently under development.

Robust ADAR1 mediated RNA editing in DLBCL
To explore the role of ADARs RNA editing in DLBCL, we first profiled 106 DLBCL samples, using matched genomic (in the form of whole genome or whole exome sequencing) and transcriptomic sequencing data (Table S1). Then, we applied a recently developed machine learning-based method to identify specific RNA editing sites 30 to our dataset ( Figure S1). This classifier, named RED-ML, was built on carefully constructed positive and negative training sets, and can efficiently remove possible sequencing artifacts, misalignments during analysis, as well as genomic mutations 30 (for details also see the material availability and STAR Methods section). In addition, paired DNA sequencing was used to further filter out any DNA mutation present in the dataset, substantially reducing the possibility that RNA editing events called through this pipeline would inadvertently include true DNA mutations. Using these pipelines, we identified 57610 RNA editing sites across all DLBCL samples. The vast majority of the editing sites (97%) identified in our dataset have already been reported to the REDIPORTAL database 31 ( Figure S1C), a rigorously annotated database of RNA editing, supporting the reliability of our RNA editing detection pipeline. Furthermore, we have validated the identified RNA editing sites using matched RNA and genome DNA samples derived from the same tumor biopsies by Sanger sequencing and cloning in selected genes (CARD11 and TP73) ( Figure S2). Consistent with previous reports, the most abundant editing type was A-to-I (A-to-G) (reviewed in Eisenberg and Levanon 2018 16 ) ( Figure S1D), and we further focused on these dominant editing events. The human genome encodes three ADARs: ADAR1 is ubiquitously expressed and, together with ADAR2, are catalytically active, whereas ADAR3 is catalytically inactive and acts as an inhibitor of editing. 32,33 Initial comparison of the expression level of each of the three ADARs encoded in the human genome suggested ADAR1 as the main adenosine editor in DLBCL ( Figure 1A). Furthermore, ADAR1 expression positively correlated with the total number of editing sites per patient (r = 0.43, p < 0.0001; Figure 1B) and 90% A-to-I editing involved Alu repeats within non-coding regions (i.e., introns and 3 0 UTRs) (Figures 1C, and S1D), consistent with previous reports on ADAR1 activity. [34][35][36][37] Moreover, motif analysis of these sites showed preferential usage of U and G, as 5 0 -and -3 0 neighbor bases of the edited A, as known flanking sequence preference of ADAR1 ( Figure 1D). [38][39][40] In addition, the edited transcripts were not biased for transcript length, GC content or expression level ( Figure S3), in line with previous reports. 35,37,[41][42][43] As a control, we performed the same analysis on a previously published dataset from EBV immortalized peripheral blood B cells from 41 healthy individuals (referred to as control B cells), [44][45][46] where only 17,858 RNA editing sites were identified. To allow unbiased comparison between samples/cohorts, we next broadly quantified ADAR1-mediated RNA editing activity by calculating the Alu editing index (AEI). 47 ADAR1 expression showed an even stronger positive correlation with the AEI per DLBCL patient (r = 0.508, p < 0.0001; Figure 1E) as compared to that with the total number of editing sites ( Figure 1B). Using this approach, we further confirmed that DLBCL samples have significantly increased editing compared to control B cells (p < 0.0001, Figure 1F). Similarly, increased level of editing was also observed in another type of B cell lymphoma, follicular lymphoma (FL; n = 20; Figure 1F), suggesting the increase of editing as a more general feature for germinal center related B-cell malignancies (see also ref. 48). Altogether these data demonstrate the presence of robust A-to-I ADAR1 mediated RNA editing in DLBCL with sequence features similar to those identified for other types of cancers. 11,49 Increased expression of ADAR1 in cancer is often assumed to refer to overexpression of ADAR1 p150, the interferon-driven ADAR1 isoform. 50,51 Unexpectedly, we observed that in DLBCL, increased ADAR1 expression is instead closely correlated with increased expression of the ubiquitous isoform of ADAR1,  iScience Article p110. Since p110 is also the nuclear form responsible for the majority of editing, this explains the high level of editing in these tumors ( Figure 1F). Indeed, the higher ratio of ADAR1 p110/p150 is not only significant in DLBCL versus FL or control B cells ( Figure 1G) but is also associated with a late disease stage (Table S2).

A landscape for RNA editing in DLBCL
To begin to characterize the role of ADAR1-mediated editing in DLBCL, we concentrated on A-to-I RNA editing sites that can be assigned to a gene locus and that were enriched in DLBCL patient samples. For this, we first excluded all editing occurring in non-genic regions and then cataloged editing sites occurring in more than 10% of DLBCL patient samples (but which were absent from the transcriptomes of control B cell samples). This analysis yielded 4605 unique editing sites present in 834 genes (Tables S3A and S3B). Furthermore, pathway analysis using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database suggested that these edited genes were for pathways of relevance to DLBCL (like the apoptosis/p53 signaling pathway or the NF-kB signaling pathway - Figure 2A) but also for pathways that have not been associated with DLBCL (like the RIG-I-like Receptor (RLR) signaling pathway; Figure 2A). iScience Article Delving deeper into the genes that define these pathways, we assessed how many of these were mutated at the DNA level or were found edited at the RNA level. We observed that genes that belong to the apoptosis/p53 pathway (like ATM, XIAP and MDM4), or genes which belong to the NF-kB pathway (such as PLCG2 and CARD11) or finally genes within the RLR pathway (which culminates in IFN signaling in response to viral infection) were often found either edited or mutated ( Figures 2B-2D). For example, TP53 was mutated in $30% of the patients (red bars) but only occasionally edited (blue bars), whereas ATM was heavily edited but infrequently mutated ( Figure 2B). Another example in the apoptosis/p53 signaling pathway is MDM4 (encoding a regulator of p53 activity - Figure 2B), which was edited in almost all samples, but not mutated. In the RLR pathway, DDX3X was only mutated but not edited in DLBCL, whereas MAVS was never mutated but heavily edited ( Figure 2C). Similarly, in the NF-kB pathway, MYD88 was only mutated, whereas other genes such as CARD11 and PLCG2 were more often edited (Figure 2D). Finally, altered editing frequencies did not arise from differences in mRNA expression levels for the genes of interest, between DLBCL samples and control (non-malignant) B cells ( Figure S3). These results highlight that RNA editing occurs within transcripts belonging to pathways necessary for DLBCL survival and maintenance, and within those pathways and with few exceptions, editing and mutation appear to be mutually exclusive (Table S4, Figure S4). This aspect suggests that functionally, RNA editing may represent a mechanism complementary to that of DNA mutation.

ADAR1 editing of MAVS is correlated with increased MAVS protein levels and increased downstream signaling
Narrowing down from transcriptome-level editing (specified by the AEI) to gene-level editing in the context of pathways affected ( Figures 2B-2D), the RLR pathway was identified as a potentially unrecognized candidate disease-driving pathway for DLBCL. In the current view of RLR pathway signaling, the signal consists of the deposition of viral RNA within the cytoplasm, which is sensed by the cytoplasmic RNA sensors RIG-I (DDX58) or MDA5 (IFIH1). Once these sensors bind the viral dsRNA, they interact with the MAVS protein on the mitochondrial membrane and nucleate the formation of MAVS polymers. The switch from soluble MAVS to self-propagating helical MAVS fibers (reminiscent of amyloid and prions) is thought to activate and amplify downstream signaling 52 through cytosolic protein kinases of the IKK family, which in turn activate the transcription factors NF-kB and IRF3. Activated NF-kB and IRF3 subsequently translocate to the nucleus, where they induce the expression of type I IFN and other inflammatory molecules ( Figure S5A and ref. 21 ). Because sensing through the RLR pathway culminates in NF-kB and type I IFN signalings (important for tumor cell survival or inflammation in the tumor microenviroment), we further investigated a possible role of ADAR1-mediated RNA editing in the activation of this pathway.
In DLBCL samples and within the RLR pathway, MAVS is not mutated, but it is highly edited ( Figure 2C). Of interest, MAVS editing is positively correlated with increased transcript levels ( Figure 3A). To assess whether MAVS protein levels were affected by editing, we performed mass spectrometry-based proteomic analyses on a small subset of DLBCL samples (n = 14, noted as red dots in expression data in Figure 3A). Within these randomly selected samples, we could identify a positive correlation of ADAR1 and MAVS protein levels ( Figure 3B). These data together demonstrate that ADAR1-mediated RNA editing is positively correlated with an increase in MAVS mRNA and also in MAVS protein levels in the absence of additional, exogenous signaling (e.g., viral infections associated with DLBCL, including HBV and EBV - Figure S6). We further showed that MAVS expression, both at the transcript and the protein level, is positively correlated with the ISG/NF-kB gene expression scores ( Figures 3C and 3D). We validated these findings in two independent DLBCL cohorts (validation cohorts 1 and 2, Figures 3E and 3F), as well as the FL samples (validation cohort 3; Figure 3G); in contrast, a negative correlation was observed in a lung adenocarcinoma cohort ( Figure 3H). Notably, unlike DDX58 (RIG-I) or IFIH1 (MDA5), MAVS itself is not an ISG (ref. 53 and Figure S7), and overexpression of MAVS is sufficient to trigger pathway activation. 53,54 We, therefore, hypothesize that robust ADAR1-mediated editing of the MAVS transcript observed in DLBCL patients ( Figure 2B) would provide an alternative, and RLR-independent way of signaling toward ISG/NF-kB pathway activation ( Figure S5B). iScience Article ''hyper-edited'' regions surrounded by relatively unedited stretches ( Figure S8). Although many editing sites are present in both DLBCL samples and control B cells, most editing events within clusters 3 and 4 are specific to DLBCL ( Figure S8B and Table S3A) and are also present within a DLBCL-derived cell line RC-K8 ( Figures S8B-S8D). We thus proceeded to knockout ADAR1 from RC-K8 cells using the CRISPR/ Cas9 technology ( Figure S9). This substantially reduced editing ( Figure S9E) but did not abolish it: all three targeted clones we obtained show deletions in the proximity of the sgRNA sites resulting in loss of exons 3 and 4, at the cDNA and protein level ( Figures S9B-S9F); this deleted version of ADAR1 retains residual RNA editing activity. Despite this, total RNA editing, quantified as AEI, is substantially reduced (especially in clone 18, which we then used in restoration experiments) ( Figure S9E). Comparing RNA editing sites detected from RC-K8 wildtype versus knockout RNA-seq data, we could confirm the absence or reduction of all editing sites within MAVS -demonstrating that they are ADAR1-dependent ( Figure S8C). Finally, we also validated a few sites using RT-PCR and Sanger sequencing ( Figure S8D).
To test whether ADAR1-mediated editing of MAVS 3 0 UTR is a direct and RLR-independent mechanism of activating canonical MAVS signaling, we used a recently described targeted base editing system 55 to reintroduce editing specifically onto the MAVS 3 0 UTR in ADAR1-''deficient'' RC-K8 cells. This system uses an engineered mutant of the DNA repair protein O 6 -alkylguanine-DNA alkyltransferase, called SNAPtag, which is able to form a covalent bond with O 6 -benzylguanine (BG), 56 conjugated to the deaminase domain of ADAR1 carrying the E406Q mutation typical of the hyperactive variants 57 Figure 4A). Second, we designed two chemically stabilized bisbenzylguanine(BB)-modified guide RNAs (BB-gRNAs or gMAVSs) to target simultaneously clusters 3 and 4 within MAVS at multiple sites and restore ADAR1-mediated RNA editing in these DLBCL-specific regions ( Figure S10). As a control, we used the same gRNAs before the addition of BB group (NH-gRNAs or gCTRLs), leading to a gRNA able to create a dsRNA but not to recruit SA1Q-GFP iScience Article allowing us to discriminate between SA1Q-GFP and unspecific editing ( Figures 4A, and S10). Using these strategies, we observed specific gMAVS-dependent editing at the targeted location via RT-PCR (Figure S11A). However, within the targeted transcript, editing was not necessarily site-specific. Indeed, RNA editing analysis from RNA-seq showed restoration of editing in several sites along the 3 0 UTR of MAVS, altogether leading to a higher editing index for this region in the presence of gMAVSs compared to gCTRLs (Figures 4B and 4C). Thus, selective loading of the enzymes onto a transcript can broadly affect editing within a region flanking the targeted sites. Unexpectedly, we also observed a considerable difference in the editing measured from RNA-seq data or RT-PCR and Sanger sequencing ( Figures 4C, S11A), iScience Article suggesting that the guide-dependent editing on the 3 0 UTR of MAVS is possibly higher than what measured from RNA-seq (resulting in reads that are unmappable, as previously suggested 58 ). Finally, we analyzed the ability of gMAVSs to generate off-target editing events. Despite the underestimation of editing in MAVS by RNA-seq, we found the highest number of statistically significant editing sites within its 3 0 UTR ( Figure 4D, Supplementary File 1). In addition, this editing leads to a small but measurable increase in MAVS transcript levels ( Figure S11B), recapitulating the correlation seen in Figure 3A. Altogether, the restoration of editing with gMAVSs was reasonably specific, on par with the best previously published data using this system. 55 To then assess our hypothesis that restoration of MAVS editing could lead to increased downstream signaling (as predicted by correlative data in Figure 3), we calculated the ISG and NF-kB scores based on the transcriptomic data from the RC-K8 ADAR KO cells transfected with SA1Q-GFP and gMAVSs or gCTRLs. This analysis demonstrated that targeted RNA base editing in MAVS could restore (or increase) levels of ISGs and of NF-kB responsive genes, which were reduced (or unaffected) in the ADAR1 knockout cells (Figures 4E and 4F). In line with our hypothesis, expression of MAVS protein was reduced in RC-K8 ADAR1 KO cells ( Figure S11D), whereas MAVS protein levels increased on targeted editing, leading to the increase in ISG and NF-kB scores ( Figures 4F and 4G). However, we also detected many off-target editing events in the presence of gMAVS ( Figure S11C). To evaluate if off-target editing could affect ISG and NF-kB signaling, we identified the top 100 transcripts containing the highest number of differentially edited editing sites ( Figure 4D) and only a minority of those (5%) are ISGs or NF-kB genes ( Figure 4D), strengthening the notion that the increased ISG/NF-kB scores most likely resulted specifically from the restoration of editing on MAVS.
Overall, these data demonstrate that targeted editing of the untranslated region (the 3 0 UTR) of a key gene within a pathway may be sufficient to modulate the output of that pathway. More specifically, these results support the hypothesis that ADAR1-mediated MAVS editing can lead to increased MAVS protein levels and to increased, non-canonical downstream signaling, especially through NF-kB, a key oncogenic signaling pathway in DLBCL pathogenesis. These findings are in line with a recent publication showing that MAVS protein level is affected by ADAR3 through inhibition of ADAR1 editing on MAVS 3 0 UTR. 59 Furthermore, these data demonstrate that ADAR1-mediated MAVS editing also leads to increased ISGs expression in B cell malignancies, which is an unusual observation, given that reduction or loss of ADAR1 in most other studies on solid tumors is associated with activation of inflammatory cascades, as indicated by the ISG signatures, rather than their silencing 21,22,26,27,29,60 and Figure S9G-I, S12A-B). Altogether, these data reveal an alternative, and RLR-independent way of signaling toward ISG/NF-kB via RNA editing.

In DLBCL, MAVS expression is associated with T cell exhaustion within a relatively uninflamed tumor microenvironment
Recent work has shown that loss of ADAR1 from tumors inflames the tumor microenvironment (TME) through ISG upregulation, consequently leading to substantial infiltration by CD8 + T cells. 27 Here, we have shown that MAVS expression (modulated at least in part through ADAR1-mediated MAVS editing) is positively correlated with the ISG/NF-kB scores in DLBCL (Figures 3C and 3D). Therefore, a change in MAVS expression (because of ADAR1 editing) within the tumor seems to be a strong corollary of ISG/NF-kB signaling and presumably of downstream interferon production and tumor survival signaling, respectively. We then asked how MAVS expression would correlate with an influx of cells within the TME consistent with an increase in inflammation. We used the TIMER tool 61 to predict the composition of the tumor-infiltrating immune cells between tumors that expressed high levels of MAVS (top 50%) versus those that showed low MAVS expression (bottom 50%). Within this analysis, we did not find substantial changes, especially T cells in TME composition relating to the expression of MAVS ( Figure S13), in contrast to what has been reported for solid tumors 27 and consistent with our observation here, i.e., in B cell lymphomas, expression of MAVS positively correlated with ISG/NF-kB scores ( Figures 3C-3F). However, when T cell functional features from the DLBCL tumors were further evaluated, we found that high expression of MAVS is significantly correlated with increased T cell exhaustion signature in DLBCLs ( Figure 5A and 5B). Similar trends were observed in the two independent DLBCL cohorts ( Figures 5C and 5D), but a negative correlation was noted in lung cancers ( Figure 5E). Similar results were obtained when comparing the ADAR1 high-and low-expressing tumors (Figures 5F-5J). Taken together, our data suggest that MAVS expression (or ADAR1-mediated MAVS editing) in DLBCL (as opposed to other solid tumors such as lung cancer) is positively correlated with inflammation cascade and exhaustion status of the tumor-infiltrating T cells. In addition, our data imply that DLBCL tumors with high MAVS expression (through increased editing) might be associated with a worse prognosis, and this is certainly a trend in our data with a significantly higher score of the international prognostic index (IPI) in MAVS high versus low group (Table S5). iScience Article Beyond this, our data suggest that different types of tumors can benefit from general ADAR1 inhibition (or perhaps specific inhibition of MAVS editing) via different mechanisms, i.e., either through an increase in inflammation within the TME (for solid tumors) or through a decrease in T cell exhaustion (for B cell lymphoma). Thus, MAVS editing can potentially be a better biomarker than ISG scores, 26 in the context of selecting tumors for therapy with ADAR1 inhibitors.

DISCUSSION
Aberrant inflammation is an emerging hallmark of tumorigenesis. In that regard, pattern recognition receptors (PRRs) which are essential mediators of the host immune response, are also critical elements affecting tumor pathobiology, with the Toll-like receptor (TLR) or Nod-like receptor (NLR) pathways critically involved in cancer. 62 A third, the RLR pathway, has yet to be evaluated in the context of cancer.
Signaling through the RLR pathway is tightly regulated, with activation usually through the sensing of viral nucleic acids (through sensors such as RIG-I, MDA5 or even TLR3) and suppression through a set of negative regulators, one of which is the RNA editing enzyme ADAR1. Under homeostatic conditions, aberrant activation of the RLR pathway is prevented by editing endogenous dsRNA by ADAR1. Consequently, loss of ADAR1 results in the induction of IFN, a key anti-tumour cytokine, and subsequently to robust inflammation of the tumor microenvironment (substantial tumor infiltration by CD8 + T cells) and increased response to tumor restriction by checkpoint inhibitors. 27 Although this is the case for three types of solid tumors studied (melanoma, lung adenocarcinoma and acinar pancreatic carcinoma, 26,27,29,60 we discovered that in DLBCL, one of the most common and aggressive types of B cell lymphomas, reduction or loss of ADAR1 was irrelevant to key proximal measures of inflammation within the tumor microenvironment ( Figure 5). Thus, the regulation of the RLR pathway in B cells, which are part of the adaptive immune system, and their malignant counterparts, is most likely different from other types of non-immune cells. This statement is further iScience Article supported by our analyses in DLBCL derived cell lines (like RC-K8) which are capable of responding to stimulation through the RLR pathway ( Figure S12C) and also by recent reports of an alternative pathway to MDA5/MAVS that can rescue the inflammatory phenotype mediated by loss of ADAR1 in mature immune cells 63 (and which could be preferentially expressed in B cells or other cells of the immune system). It is also worth noting that while loss of MAVS is well tolerated in solid tumor derived cell lines (e.g. HCC366 26 ), it has a moderate anti-proliferation effect on B cells ( Figure S14, and ref. 64 ), again favoring the notion that the regulation of the RLR pathway in B cell derived tumors, is distinct.
Focusing on the RLR pathway, a potential pathway in the context of cancer in general and DLBCL in particular, we find that MAVS, the central RLR adaptor molecule, while not mutated, is robustly edited ( Figure 2C), with editing correlated with increased gene and protein expression levels and increased downstream signaling (Figure 3). But why would MAVS translation be promoted when the transcript is better edited?
The MAVS 3 0 UTR contains six Alu repeats, 4 of which are direct repeats. Direct Alu repeats have been correlated with transcript localization to the cytoplasm followed by increased translation, 65 whereas inverted repeats are correlated with increased nuclear retention. It is possible that the balance toward direct Alu repeats encoded in the MAVS 3 0 UTR it the molecular cause for the observed increase in translation.
MAVS is known to aggregate in response to ligand (viral RNA) bound RIG-I (DDX58) or MDA5 (IFIH1); however, viral infection did not correlate with MAVS (or ADAR1) levels in DLBCL: even though some of the samples showed evidence of viral infection (illustrated by HBV + or EBV + status), the editing rate (as defined by the AEI index) was not altered ( Figure S6). We hypothesized that MAVS aggregation on protein overexpression subsequent to RNA-editing of the MAVS 3 0 UTR (but in the absence of a canonical upstream ligand) would be sufficient to increase downstream signaling. To directly assess this we used specific RNA baseediting technologies to target RNA editing to the MAVS 3 0 UTR on an ADAR1 deficient background, revealing an alternative, RLR-independent way of signaling toward ISG/NF-kB via RNA editing. This is, also beyond the context of DLBCL, a demonstration that RNA editing of single or several sites within a 3 0 UTR can activate a downstream signaling pathway.
We hypothesize that the outcome of the RLR signaling pathway is modulated not only by MAVS editing but also by editing of additional components, including additional RLR components. Overall, we propose that specific patterns of editing in DLBCL (represented by the DLBCL cell line RC-K8) lead to activation to the NF-kB branch of the RLR pathway supporting tumor survival signaling. Finally, MAVS editing/signaling is not associated with the number of T cells infiltrated in the TME but is associated with the T cell exhaustion status, suggesting a possible specific immune evasion mechanism for DLBCL.
Overall, the experiments reported here demonstrate specific functions for ADAR1 in DLBCL that are distinct from those described for solid tumors and have implications for ADAR-centered immunotherapies currently under development.

Limitations of the study
The main limitation of this study concerns the technical characteristics of targeted RNA-base editing approaches. Despite the fact that the SNAP-ADAR system we employed has been shown to be the most specific, 55 all RNA base editing technologies developed thus far show abundant (if low level) off-target editing. 66 This is especially the case in two-component systems like the one employed here. Thus, we cannot completely exclude the possibility that the output of targeted editing on MAVS (namely ISG/NF-kB activation), is at least in part the result of off-target editing. Aggregating ISG and NF-kB scores mitigates this concern as these ''scores'' evaluate pathway activation as a whole, thus minimizing the effect of changes in a handful of genes that display off-target editing.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following:

DECLARATION OF INTERESTS
The authors declare no competing interests. iScience 26, 106864, June 16, 2023 iScience Article (Table S1). Informed consent was obtained from all patients, and the institutional review boards approved the study at Tianjin Medical University Cancer Hospital, Uppsala University and the Karolinska Institutet.

Cell lines
RC-K8 cells (DSMZ, Germany) were cultured at 37 C, 5% CO 2 , in RPMI with L-glutamine and sodium bicarbonate (Sigma), supplemented with 15% fetal bovine serum (FBS, PAN Biotech) and Penicillin/ Streptomycin (Sigma). The RC-K8 cell line was established from a patient with GBC-like lymphoma, described at the time as histiocytic lymphoma (terminal, refractory stage), 70 and later was assigned to the GCB subtype of DLBCL. 71 However, more recent studies suggested constitutive NF-kB signaling in RC-K8 cells, a feature that is usually associated with the ABC subtype. 72 Therefore, we performed both DNA and RNA sequencing on this cell line, and the data suggested that RC-K8 indeed be assigned to the ABC subtype. 64 A549 cells (DKFZ, Germany) and HEK293T (ATCC) were cultured at 37 C, 5% CO 2 , in high glucose DMEM (Sigma), supplemented with 10% FBS and Penicillin/Streptomycin. All cell lines were regularly tested for mycoplasma contamination in the Multiplexion facility (DKFZ, Heidelberg, Germany).

METHODS DETAILS
Processing of DNA sequencing DNA and RNA sequencing data were obtained from a previous publication. 5 Briefly, for DNA sequencing, BWA (v0.5.9) was subsequently used to align reads to the human reference genome (hg19) 73 and VarScan (v2.3.9) was used to detect substitutions and insertions and deletions using the defined parameters.

Processing of RNAseq and calling of RNA editing sites
The RNA sequencing reads were analyzed and mapped to the hg19 reference genome using SOAP2. 74 The aligned sequencing files (BAM) were further used as input for RED-ML, an effective RNA editing detection method based on machine learning. 30 The pipeline extracts candidate RNA editing sites and their corresponding features, with optimal filtering if DNA mutation information is available, then apply a logistic regression classifier to detect authentic RNA editing sites with an associated confidence score. Briefly, three classes of features were used: 1) basic read features, including the number of supporting reads of a candidate site and the putative editing frequency; 2) related to possible sequencing artifacts and misalignments, including mapping qualities of the supporting reads, the relative position of the candidate site in the mapped reads (usually 5bp away from read end as in 75 indication of strand bias, whether the candidate site falls into simple repeat regions; 3) based on known properties of RNA editing, such as the editing type (whether it is A-to-I), whether the candidate site is in an Alu region, and its sequence context. 30 In total, 28 features (Table S4 in ref. 30) for every possible editing site were extracted to remove potential errors during cDNA synthesis, sequencing and mapping or contamination from genomic variants. Finally, this pipeline is concordant with more commonly used algorithms (e.g. REDITools. [76][77][78] Alu editing index (AEI) and MAVS 3 0 UTR editing index After removing the adapter sequence in the raw data and low-quality reads, HISAT was used to align reads to the human reference genome (hg38). The read length of DLBCL and FL samples were cut to 100 bp to avoid potential bias of read length on AEI index. AEI index in each tumor sample was then calculated using the published pipeline. 47 MAVS 3 0 UTR editing index ( Figure 4B) was calculated as the ratio of 'number of reads containing an A-to-G editing' and 'number of total reads of each A' in the region chr20:3848102-3855844 (hg19).
Calculating RNA editing at the gene level In order to calculate the editing frequency at the gene level, all the RED-ML derived editing sites were filtered to keep only reads reporting A-to-G changes. Next, RED-ML derived edit sites were reported in less than two samples, or sites within intergenic regions were filtered out. Starting from this filtered subset of editing sites, we calculated "gene-level" editing as follows: for each gene, we first aggregated all reads covering at least one of the editing sites reported for that gene, with base quality higher than 20 and iScience Article iScience Article (gMAVS or gCTRL) for the variant calling. All positions must be decently covered (min. 10 reads per pull) in both groups, and RNA editing site candidates were considered those that showed at least 1 read carrying an A-to-G variation.

QUANTIFICATION AND STATISTICAL ANALYSIS
Quantification and statistical analysis in this article were performed in R or prism (GraphPad) All of the statistical details for each experiment can be found in the figure legends. The p values in Figures 1F and 1G, 5A, 5C-F5H-5J were calculated using Mann-Whitney U test. The p value in Figure