Identification of the stress granule transcriptome via RNA-editing in single cells and in vivo

Summary Stress granules are phase-separated assemblies formed around RNAs. So far, the techniques available to identify these RNAs are not suitable for single cells and small tissues displaying cell heterogeneity. Here, we used TRIBE (target of RNA-binding proteins identified by editing) to profile stress granule RNAs. We used an RNA-binding protein (FMR1) fused to the catalytic domain of an RNA-editing enzyme (ADAR), which coalesces into stress granules upon oxidative stress. RNAs colocalized with this fusion are edited, producing mutations that are detectable by VASA sequencing. Using single-molecule FISH, we validated that this purification-free method can reliably identify stress granule RNAs in bulk and single S2 cells and in Drosophila neurons. Similar to mammalian cells, we find that stress granule mRNAs encode ATP binding, cell cycle, and transcription factors. This method opens the possibility to identify stress granule RNAs and other RNA-based assemblies in other single cells and tissues.


INTRODUCTION
Non-membrane-bound compartments represent an important aspect of cell organization. They are formed by phase separation, where solution of seemingly diffuse macromolecules segregates into two distinct phases (Hyman et al., 2014;Gomes and Shorter, 2018). Interestingly, cellular stress induces the formation of many of these compartments (van Leeuwen and Rabouille, 2019), including stress granules (Anderson and Kedersha, 2002;Aulas et al., 2017) that formed around mRNAs after cells have been subjected to many cellular stresses.
Recently, RNAs have also been shown to be structural components of stress granules (Zhang et al., 2015;Van Treeck et al., 2018) both in vitro and in vivo. This is sustained by the demonstration that elements in RNA secondary structure are conducive to phase separation (Langdon et al., 2018). This renders the identification of stress granule RNAs an important biological question.
This has recently been performed in mammalian cultured cells after stress granule core isolation (Namkoong et al., 2018;Khong et al., 2017), cross-linking and immunoprecipitation (Anders et al., 2018), and proximity biotinylation of RNAs (Padró n et al., 2019). These studies revealed that stress granules mainly harbor longer mRNAs that are thought to be more poorly translated by ribosomes. Furthermore, even though a clear enrichment in adenylate-uridylate-rich elements (AREs) has been established in endoplasmic reticulum-stress driven stress granule mRNAs (Namkoong et al., 2018), a true recruitment motif has yet to be identified, although recent efforts show that AREs and Pumilio recognition elements increase mRNA partitioning into stress granules (Matheny et al., 2021).
However valuable, these techniques are somewhat limited by the step of cell fractionation and stress granule purification or by MOTIVATION Sequencing-based methods for identifying stress granule RNAs (Khong et al., 2017;Namkoong et al., 2018;Padró n et al., 2019) have typically required large amounts of input material and can therefore be challenging for single-cell studies on low-volume, heterogeneous tissues. To address this limitation, we have developed an application of hyperTRIBE that allows the identification of stress granule RNAs in single cells and Drosophila neurons.
RNA pulldown. Consequently, they require a substantial amount of input materials Jain et al., 2016), thus preventing the analysis of limited/small tissues and single cells. This represents a critical step, as there are now indications that the RNA content of stress granules might be heterogeneous within a cell population (Khong et al., 2017). Furthermore, it might be important to properly compare the differential response to stress of a given cell in its environment (Aulas et al., 2017).
Here, we adapted a method to identify RNAs that are recruited to stress granules, which does not require fractionation and isolation, and that works in single cells and in tissue in a cell-typespecific manner. This method is based on TRIBE (Target of RNA-binding proteins Identified by Editing) first established by the Rosbash group (McMahon et al., 2016). There, a given RBP is fused to the catalytic domain of the DNA/RNA-editing enzyme ADAR (ADARcd) that deaminates adenosine-to-inosine on RNA molecules, which is read out as A-to-G mutations by sequencing. This method has been used to successfully identify the RNA targets of several RPBs in Drosophila neurons (McMahon et al., 2016). An activating mutation in ADARcd (TRIBE) has since been shown to increase the editing efficiency of ADARcd by 400% when compared with the original (Xu et al., 2018). The purification-free method we adapted from TRIBE enables the identification of stress granule RNAs in single tissue cultured cells, and in specific tissues, the neurons of the Drosophila larval brain.

Predictions and experimental setup
To identify the RNA content in stress granules, we adapted the TRIBE method (Xu et al., 2018;McMahon et al., 2016) in such a way that most of the RNA editing takes place in stress granules. The prediction is that the more an RNA is found specifically edited in stress conditions (here arsenite treatment), the more likely it is recruited in stress granules ( Figure 1A). As in TRIBE (Xu et al., 2018), we used the RBP FMR1 tagged to ADARcd-V5 (FMR1-ADARcd-V5). Endogenous (Aguilera-Gomez et al., 2017) and GFP-tagged FMR1 (Zacharogianni et al., 2014) are efficiently recruited to stress granules on arsenite stress. FXR1, the mammalian homolog of Drosophila FMR1, is largely immobile within stress granules (Marmor-Kollet et al., 2020), in a similar manner to FMR1 (Zacharogianni et al., 2014), potentially reducing its dynamics in and out of the stress granules. Last, fusing ADARcd to FMR1 did not appear to disrupt FMR1 function and its binding to other RBPs and RNAs (Xu et al., 2018). In addition, ADARcd remained functional when fused to FMR1 (Xu et al., 2018).
These edited RNAs (predicted to be in stress granules) will likely include FMR1 clients, but we hypothesize that they will also include RNAs that are clients of other stress granule RBPs that would then be edited in trans.
To achieve this, we first verified that FMR1-ADARcd-V5 is indeed efficiently recruited in arsenite-induced stress granules. S2 cell clones stably expressing FMR1-ADARcd-V5 showed its 8.3 ± 2.6-fold concentration in foci also positive for the RBP Caprin, a known marker of stress granules, in 90% of the cells ( Figures 1B and S1A, A'). Arsenite treatment of S2 cells for 4 h is longer than normally used for mammalian cells but we show that it does not result in cell death. S2 cells are still viable after this treatment and grow again when the stress is removed, despite a loss of 20% ( Figure S1B). Interestingly, stress granule dissolution occurs before cells start growing again (Figures S1B and S1C), suggesting that stress granule dissolution precedes growth recovery after stress relief.
Second, to detect a significant differential RNA editing in stress conditions, the RNA editing in basal conditions (non-stressed cells during which FMR1-ADARcd-V5 can edit FMR1 clients independently of their further recruitment in stress granules) should be kept as low as possible. The notion was to find a tradeoff in producing enough of chimeric protein to lead to RNA editing in stress granules without generating a high background in non-stressed conditions. To do so, we expressed FMR1-ADARcd-V5 (which is under control of a copper sulfate inducible metallothionein promoter) for 20 min and 4 h before applying the stress. Four hours of induction of clone 1 resulted in 11.3-fold higher FMR1-ADARcd-V5 expression compared with 20 min ( Figure 1C).
Last, we verified that the induction of FMR1-ADARcd-V5 in clone 1 is able to produce detectable editing on transcripts. To detect these, we used a variant caller that locates specific base changes on the transcripts (McKenna et al., 2010) (see method details section, and below). We identified 1,727 RNAs displaying edited bases. As expected, most of the filtered de novo variants detected were A-to-G events ( Figure 1D), while other possible nucleotide changes were hardly detected ( Figures S2A-S2C). Since the levels of RNA editing after a 4 h induction showed a clear increase when compared with 20 min ( Figures 1D and 1E), we chose a 4 h induction of clone 1 in the rest of the experiments.
RNA editing through ADAR predicts the stress granule transcriptome in S2 cells To identify RNAs present in stress granules, we used the setup conditions described above (4 h induction in basal condition, i.e., Schneider's) followed by cell incubation with arsenite for 4 h ( Figure 2A). Control cells were maintained in Schneider's for the same length of time. An uninduced sample was also included as a control for endogenous editing. Total RNAs from triplicate experiments were generated and sequenced using the new sequencing method VASA-seq (Hollfelder et al., 2022). VASAseq does not rely on primers binding to the 3 0 or 5 0 end of the transcripts. It allows the sequencing of the full transcript and permits the detection of mutations throughout the RNA sequence. This is critical, as editing by ADAR might not be limited to the 3 0 and 5 0 UTR due to FMR1 binding its RNA clients mainly in their coding region (Li et al., 2020).
Mapping the reads (read depth of 1*10 7 reads) to the Drosophila genome ( Figure 2A) revealed that over all conditions 7,357 RNAs were detected with an expression level higher than one transcript per million. Lower expression was deemed too low and those were removed. Pairwise correlation analysis of the expression level of the (total and edited) transcriptome between samples and within triplicates revealed proper clustering ( Figures S3A and S3B), indicating a high reproducibility between libraries.
We noticed FMR1-ADARcd edits at select sites rather than uniformly across a given transcript. When comparing the conditions (uninduced, Schneider's, and arsenite), the same adenosine appear to be edited. For example, along the 500-nt region in the pzg transcript, only five of the 147 adenosines exhibit an increase in mutation frequency between the conditions (vertical lines, Figure S4A). This non-randomness was observed for many of the edited transcripts and was confirmed by a genome wide examination of editing across con- Figure 1. Adaptation of HyperTRIBE using FMR1-ADARcd-V5 to detect stress granule RNAs (A) Schematic overview of our HyperTRIBE adaptation strategy using FMR1-ADARcd-V5 to predict stress granule RNAs. In growing cells, FMR1-ADARcd-V5 binds to its own client RNAs that can be edited in the cytosol. Upon arsenite stress, FMR1-ADARcd-V5 is recruited to stress granules. Allowing ADARcd to edit stress granule RNAs, its clients as well as other stress granule RNAs that are in close proximity. (B) Immunofluorescence visualization of FMR1-ADARcd-V5 (using anti-V5, green) in S2 cells in Schneider's and in arsenite (0.5 mM for 4 h) leading to localization of FMR1-ADARcd-V5 with the known stress granule marker Caprin (red). (C) Western blot of S2 cell extract after 20 min and 4 h induction (+CuSO4) of different clones expressing FMR1-ADARcd-V5 using an anti-V5 antibody. This number indicated is the ratio ''4 h/20 min induction'' of the ratio ''FMR1-ADARcd-V5/tubulin'' for each clone (20 min is therefore 1). (D and E) Graph displaying the frequency (D) and total (E) of A>G editing events in S2 cells in which the expression of FMR1-ADARcd-V5 was induced for 20 min and 4 h. Scale bar: 10 mm (B).
ditions. This reveals that specific adenosines are moderately but specifically edited ( Figure S4B).
Consequently, in order to determine the identity of the specific edited RNAs as well as the level of their editing, we used a variant caller (GATK [McKenna et al., 2010]). For each of these specific edited positions, we extracted the total number of mutated bases in each sample in triplicate, and used these to calculate the editing frequency (e.g., 17 of the 53 bases at position 1 are ''G,'' giving a mutation frequency of 0.32 for arsenite) ( Figure S4C). The editing frequency was calculated for each specific position across all conditions. All identified variant positions and their corresponding editing frequencies for a gene were then averaged (''average editing frequency'') to compare the average editing frequency between conditions ( Figure S4C). Note that the calculation of the editing frequency could be done using alternative strategies (please see more in the limitations of the study and STAR Methods). Using this workflow ( Figure S4C), we first determined the level of endogenous editing by comparing the uninduced samples with those maintained in basal condition (Schneider's). The Euclidean distance between samples was then calculated and visualized. In total, we found 2,844 non-edited RNAs and 4,513  Table S1, Figures S4A and S4B).
When comparing samples in Schneider's with those treated with arsenite, we found that, as predicted, the editing extent was higher in arsenite stress than in Schneider's ( Figures 2B,  S4A, and S4B). In addition, some genes appear only edited in arsenite and not in Schneider's ( Figure 2B). To identify the RNAs only or more edited upon arsenite stress, we set that the average editing frequency of RNAs from the ''arsenite'' triplicates should be significantly higher than the average editing frequency in Schneider's (basal condition) (p < 0.01) using the Empirical Bayes test. Using this definition, 2,657 RNAs (group 1), such as Rack1 and kermit, were found not to be differentially edited (Figures 2C and Table S1).
Conversely, 1,856 RNAs were found to be more edited upon arsenite when compared with Schneider's (Figures 2B and 2C  and Table S1, see also alternative strategy in STAR Methods). These were divided into two groups. Group 2 contains 1,362 RNAs (73%), such as row, Cyt-c-p, red, cbt, and Khc, that were also edited at low level in basal conditions. Conversely, group 3 (496 RNAs, 27%), which includes geminin and Prosb4, were only edited under arsenite stress. According to our predictions, groups 2 and 3 RNAs are predicted to be enriched in stress granules (see below).
We then compared the RNAs we found edited in Schneider's with the list of FMR1 clients established by the Rosbash group (Xu et al., 2018), using a geometric test. The overlap was 86%, showing the reliability of our method in detecting FMR1-ADARcd editing ( Figure 2D).
We then assessed whether the RNAs of groups 2 and 3 are FMR1 clients. As above, we compared these RNAs with the list of FMR1 clients established by the Rosbash group (Xu et al., 2018). Group 2 RNAs (edited in Schneider's but more edited in arsenite) present a 39% overlap (429 RNAs) with these clients ( Figure 2E, Table S2). Group 3 RNAs (only edited only upon arsenite) overlap with the list of the FMR1 clients by 3% (30 RNAs) ( Figure 2F and Table S2). Altogether, only 25% (459 of the 1,856 RNAs) are FMR1 clients.
Taken together, these results show that we have potentially identified 1,856 RNAs that recruited to stress granules upon arsenite stress. Furthermore, the stress granule RNAs we retrieved do not appear to be simply FMR1 clients that are passively bound to FMR1-ADARcd upon its recruitment to the stress granules.

Validation of identified stress granule transcriptome by smFISH
To validate our prediction, i.e., to show that group 2 and 3 RNAs are indeed localized and enriched to arsenite-induced stress granules, we used single-molecule fluorescence in situ hybridization (smFISH). Group 1 RNAs (such as Kermit and Rack1) are non-differentially edited upon arsenite, and are therefore predicted not to be enriched in stress granules. Indeed, only 2% of Rack1 RNA and 29% of kermit RNA overlap with stress granules ( Figures 3A and 3D). Interestingly, Rack1 is a clear FMR1 client (Xu et al., 2018). Yet, Rack1 is not recruited in stress granules upon coalescence of its RBP.
Furthermore, we established a significant relationship between the fold change in editing frequency of a given RNA (between arsenite and Schneider's) and the fraction of these RNA molecules localized in stress granules (quantified by smFISH). We observed an increased fold in editing that positively correlates with the localization in stress granules (R 2 = 0.702, Figure 3E). For example, cbt is edited 6.7-fold higher in arsenite when compared with Schneider's, and 62% of cbt RNA molecules localize to stress granules. Conversely, the fold change in editing for Rack1 is 0.7 and the fraction of Rack1 RNAs in stress granules is 2%. This strong correlation indicates that the higher the fold change in editing the most likely RNAs localize in stress granules.
Overall, the successful smFISH validation of these 9 RNAs (all chosen randomly, except Rack1) in stress granules suggests that the 1,856 RNAs that are differentially edited upon arsenite are potentially enriched in stress granules (see discussion).
Arsenite-induced stress granules contain long mRNAs that encode ATP binding, transcription, cell cycle, and splicing factors We then asked whether the 1,856 predicted stress granule RNAs have common and specific features. First, we found that 99.6% are mRNAs and 0.4% are non-coding RNAs (ncRNAs) ( Figure 4A). The stress granule transcripts are 1.3-fold longer than the mRNAs predicted not to be in stress granules (Figure 4B), an increase that is not specific to the coding sequence (CDS), 5 0 UTR, and 3 0 UTR ( Figures 4C-4E). Finally, a gene ontology (GO) analysis on the stress granule mRNAs showed a gene enrichment in ATP binding (kinases, chaperones, and cytoskeleton), transcription factor, RNA splicing, and cell cycle factors ( Figure 4F).
Since VASA-seq is able to detect pre-mRNAs, we use the same variant caller approach as above on intronic RNA sequences. We found a low level of endogenous editing of these sequences. We also did not observe a significant increase in A-to-G editing between the uninduced and Schneider's ( Figure 4G), indicating that pre-mRNAs are not edited by FMR1-ADARcd-V5 in growing conditions. Moreover, A-to-G base changes were not dominantly present when compared with all other nucleotide base changes. Specific A-to-G editing also did not increase upon the arsenite condition. Therefore, we predict that pre-mRNAs are likely not present in stress granules in agreement with (Khong et al., 2017).
Together, these features suggest that stress granules may harbor and protect long mRNAs that encode proteins necessary for the cells to thrive once the stress is relieved.

Identification of stress granule RNAs in single cells
There are clear indications that the RNA content of stress granules might be heterogeneous within a cell population (Khong et al., 2017). Our own smFISH data revealed that, overall, each cell does not have the same stress granule RNA enrichment as its neighbor ( Figure S5A). The existing published techniques investigating the RNA content of stress granules require a substantial amount of material (see introduction), which makes single-cell analysis difficult. Given that our adapted hy-perTRIBE technique does not require purification and pulldown, we tested the ability of this method (when combined with VASA-seq) (Hollfelder et al., 2022) to identify stress granule RNAs in single cells.
The editing efficiency in bulk S2 cells was high enough to attempt detecting RNA editing in single cells. The same procedures as above resulted an average read depth of 70,000 reads per cell. After applying the variant caller, we observed a significant enrichment in A-to-G editing compared with all other possible nucleotide base changes ( Figure 5A), indicating that hy-perTRIBE can be used to detect RNA editing in single cells (pseudobulk of combined 360 analyzed single cells). As in bulk (see above), we determined the mean editing frequency per gene per cell. Interestingly, as in bulk ( Figures 2B, S4A, and S4B), the editing activity upon arsenite treatment was found significantly higher when compared with Schneider's ( Figure 5B). Yet, not all single cells display high levels of editing. Some did not show editing, correlating with a very low expression of FMR1-ADARcd ( Figure 5C).
In the pseudobulk, we identified 2,303 RNAs that were more edited upon arsenite, which show a significant overlap (64%) with the 1,856 RNAs found by bulk sequencing ( Figure 5D). Notably, we found that most of the RNAs validated with smFISH to be enriched in stress granules were also significantly edited in pseudobulk, showing the reliability of the single-cell analysis ( Figure 5E).
By comparing the editing frequency per gene in all cells, we observed the largest variation between conditions (Schneider's and arsenite), not within a given condition. There, only a small degree of variation in the editing level between the cells is observed, suggesting a low level of heterogeneity ( Figure 5F). Taken together, using our adaptation of the hyperTRIBE method, stress granule RNAs can be reliably identified in single cells, as they are similar to the stress granule RNAs found in bulk S2 cells.

Identification of stress granule RNAs in Drosophila neurons
We then tested whether this hyperTRIBE-based method could be used to establish the RNA content of stress granules in specific cell types in primary tissues, and we focused on neurons of the Drosophila third instar larval brain. We sought to specifically express the FMR1-ADARcd-V5 fusion in Drosophila neurons, while allowing for precise temporal control to limit basal editing in unstressed conditions. To achieve this, we used the geneswitch/Gal4 system (Osterwalder et al., 2001) in which a modified Gal4/UAS system (Phelps and Brand, 1998) can lead to increased protein expression upon addition of the drug RU486 (here for 14 h). To limit the expression to neurons, we selected the pan-neuronal elav-geneswitch-Gal4. Thus, FMR1-ADARcd-V5 was specifically expressed in Drosophila neurons by bathing third instar larvae in RU486 ( Figure 6A).
Dissection and dissociation of brains into isolated cells (neurons and glia) showed that 80% of them expressed FMR1-ADARcd-V5 ( Figures 6B and 6B'), whereas there was no detectable expression in the absence of the drug. We confirmed that FMR1-ADARcd-V5 was solely expressed in neurons and not in glial cells. Indeed, FMR1-ADARcd-V5 was only expressed 9% in Repo positive (glial) cells ( Figures 6C  and 6C'), while 93% of the elav-positive cells did express FMR1-ADARcd-V5 ( Figures 6D and 6D'). These data indicate that the elav-GeneSwitch-Gal4 driver allows the expression of FMR1-ADARcd-V5 specifically in neurons. Last, we tested whether dissociated neurons form FMR1-ADARcd-V5 positive stress granules upon arsenite treatment. Ninety-one percent of the dissociated neurons form stress granules (Caprin positive) very efficiently (i.e., they form at least one stress granule per cell) upon 4 h incubation with arsenite ( Figure 6E). As expected, these stress granules contain polyadenylated RNAs as shown by RNA FISH using an oligo(dT) probe ( Figure 6F). These results show that when stressed by arsenite, larva brain neurons form bona fide stress granules where FMR1-ADARcd-V5 is recruited, as in S2 cells.
RNA editing through ADAR predicts the stress granule transcriptome of dissociated Drosophila neurons We sequenced the total RNA isolated from biological triplicates of dissociated neurons from the uninduced, Schneider's, and arsenite-stressed conditions. Across all conditions, we detected the editing of 4,137 RNAs including 870 endogenous edited RNAs at very low levels ( Figure 7A and Table S3). Of the 4,137 edited RNAs, 3,739 RNAs, such as Rack1, were not differentially edited (group 1) and are predicted not to be in stress granules ( Figure 7B and Table S3). Conversely, 398 RNAs were significantly more edited upon arsenite stress (p < 0.01, Empirical Bayes test) and are predicted to be in stress granules (Figures 7B and Table S3).
As for S2 cells, RNAs from both group 2 and 3 are predicted to be in stress granules. To validate this, we used smFISH as above, and show that row, wdn, and wal RNA are localized in arsenite-induced stress granules in the neurons of wild-type flies ( Figures 7C and 7C'). Conversely, and similar to S2 cells, Rack1 is not enriched in stress granules, as predicted. These data indicate that, as for S2 cells, transcripts that were predicted to be in stress granules through RNA editing are indeed localized in stress granules. Arsenite-induced stress granules in S2 cells and neurons share the same RNAs The majority (99.7%) of RNAs predicted to be recruited to neuronal stress granules are coding mRNAs with only 0.3% of ncRNAs ( Figure 7D). They are on average longer than mRNA not recruited in stress granules ( Figures 7F-7H). The gene analysis revealed that, as in S2 cells, the predicted stress granule RNAs show an enrichment for genes encoding ATP binding, transcription, cell cycle factors ( Figure 7I).
Performing a parallel analysis in Drosophila S2 cells and neurons, we established that 61% of the 398 RNAs predicted in neuronal stress granules are also predicted to be in stress gran- Figure 6. Stress granule formation in Drosophila neurons (A) Schematic overview of the workflow. elav-Gen-eSwitch-Gal4/FMR1-ADARcd-V5 expressing third instar larvae were exposed to RU486 by larval bathing to allow the expression of FMR1-ADARcd-V5 specifically in neurons. The dissociated brain cells were exposed to 0.5 mM arsenite for 4 h. Finally, the total RNA was isolated, sequenced, and the RNA editing analyzed. (B) Immunofluorescence visualization of FMR1-ADARcd-V5 (anti-V5, red) in dissociated Drosophila brain cells from mock (80% ethanol) and RU486 (3 mg/mL) bathed third instar larvae. Quantification in (B'). (C and D) Immunofluorescence visualization of FMR1-ADARcd-V5 (anti-V5, red), the glial cell marker Repo (cyan, C) and the neuronal marker Elav (cyan, D) in dissociated Drosophila brain cells. Quantification in (C' and D'). (E) Immunofluorescence visualization of FMR1-ADARcd-V5 (anti-V5, red) in dissociated Drosophila brain cells upon incubation in Schneider's and upon arsenite (0.5 mM for 4 h) leading to FMR1-ADARcd-V5 localization in stress granules together with Caprin (green). (F) Visualization of polyadenylated mRNAs (red) in arsenite-stressed (0.5 mM for 4 h) neurons by RNA FISH using an oligo(T) probe, in FMR1-ADARcd-V5 positive stress granules. Scale bar: 10 mm (B, C, D, E); 5 mm (F). ules in S2 cells ( Figure 7J). GO-term analysis confirmed an enrichment for neuronrelated factors/proteins, such as mira and Nrt in the remaining 39%.

DISCUSSION
Here, using our adaptation of RNA editing via hyperTRIBE (using FMR1-ADARcd) combined to VASA-seq (a sequencing method allowing the detection of editing events along the whole RNA sequence) (Hollfelder et al., 2022), and smFISH, we show that the RNA content of stress granules can be reliably identified not only in bulk cells but also in single Drosophila S2 cells, and in tissues (Drosophila neurons).

Identification of stress granule transcriptome in bulk S2 cells
In bulk S2 cells, we identified 1,856 potential stress granule RNAs, seven of them validated by smFISH. Importantly, only 25% of these RNAs are expected to be clients of FMR1. As in mammalian (Khong et al., 2017;Namkoong et al., 2018;Padró n et al., 2019) and yeast cells (Khong et al., 2017), the majority of the stress granule RNAs in S2 cells are mRNAs that are longer (here 30%) than their non-recruited counterparts. Longer Article ll OPEN ACCESS mRNAs may potentially have more binding sites for RBPs that are proposed to be major driving factors in stress granule partitioning (Maharana et al., 2018;Matheny et al., 2021;Banani et al., 2017;Jain and Vale, 2017;Molliex et al., 2015;Murray et al., 2017;Patel et al., 2015). They may also extend beyond the stress granule physical boundaries where they might interact with, and pull, cytoplasmic RNAs within stress granules (Moon and Parker, 2018). Indeed, RNA-RNA interactions are proposed to contribute to stress granule assembly (Van Treeck and Parker, 2018), and this is reinforced by the notion that RNAs themselves can phase separate (Langdon et al., 2018).
Conversely, we found fewer ncRNAs (0.4%) in the Drosophila S2 stress granule transcriptome when compared with what is reported for mammalian stress granules (22%) (Khong et al., 2017). This may partially be explained by the fact that Drosophila cells contain a number of ncRNA loci 8-fold lower than mammalian cells (1,119 and 9,277, respectively) (Derrien et al., 2012;Young et al., 2012). Last, In agreement in mammalian cells (Khong et al., 2017), pre-mRNAs are not predicted to be in stress granules.
GO analysis of the S2 cell stress granule mRNAs shows enrichment in ATP binding, transcription, cell cycle, and RNA splicing factors that also compare well the stress granule mRNA complement in NIH3T3 cells (Namkoong et al., 2018). These similarities indicate that a common core of stress granule RNAs might be conserved across evolution.

Stress granule transcriptome in single S2 cells
When compared with other published techniques, the RNA-editing-based method we used here has the major advantage to require neither stress granule isolation, nor the RNA immunoprecipitation. We found that the edited RNAs overlap by 64% between the pseudobulk (360 cells) and the bulk (2 million cells), showing the reliability and the potential of the approach. The analysis of these editing events did not reveal a heterogeneity between cells, either those kept in basal conditions, or those exposed to arsenite. Perhaps sequencing deeper would reveal a heterogeneity within a condition. However, in cultured S2 cells, this is not necessarily expected. Our method could therefore be used for the detection of stress granule RNAs in rare cells in vivo, either forming a small tissue or disseminated into larger tissues.

Drosophila neurons
In this regard, using genetic tools to control spatial and temporal expression of ADARcd, we show that hyperTRIBE can be adapted to specifically identify the arsenite-triggered stress granule transcriptome in Drosophila neurons. Neuronal stress granules contain fewer RNAs (5% of the expressed transcriptome) than S2 cells (25%). Ruling out technical differences in ribosome removal and sequencing, these differences might be due to a lower editing efficiency per transcript in neurons. Indeed, there were fewer editing events per transcript in neurons when compared with S2 cells, although the number of edited transcripts was roughly the same in both, as was the expression of FMR1-ADARcd (estimated by immunofluorescence). The reasons for this lower editing efficiency are not clearly understood.
Despite these differences, 398 RNAs predicted to be in neuronal stress granules have been identified and three were validated (row, wdn, and wal). As in S2 cell stress granules, they are mainly longer mRNAs displaying a similar enrichment for ATP binding, transcription, and cell cycle factors.
The identification of the stress granule transcriptome could be extended to mammalian cells and tissues by using TET-on inducible systems or by specific conditional CRE promotors to ensure tissue and time specificity. Recently, hyperTRIBE was shown to be able to detect protein-RNA interactions in mammalian cells (Biswas et al., 2020). This could also be easily followed by single-cell analysis, to not only investigate the stress granule RNA content in specific cell type, but also to address their heterogeneity of their response.

Single-molecule FISH
Of the potential stress granule RNAs that we have identified, nine have been validated by smFISH in S2 cells and neurons, plus two, which are predicted not to be. An extended validation could be performed using multiplex smFISH (Maynard et al., 2020), potentially revealing whether stress granules are populated with all RNAs identified, or a subset of them, and in which ratio. Our smFISH in S2 cells has begun to reveal such a stress granule heterogeneity ( Figures S5A and S5B). If this was confirmed in a systematic manner, understanding what governs this heterogeneity could reveal important principles governing stress granule formation.
Why are RNAs recruited to stress granules? One of the large questions in the stress granule field is why mRNAs are recruited in stress granules. Besides their structural role that has been proposed recently (Van Treeck and Parker, 2018;Langdon et al., 2018), one general hypothesis is that stress granules form upon stress to protect RNAs from degradation after translation stalls and that RNAs are no longer covered in ribosomes. When stress is relieved, these protected and capped mRNAs recruited to stress granules could be immediately translated, enabling the cells to thrive again and gain a fitness advantage. In this regard, stress granule formation has been shown to allow cell survival upon stress and thriving upon stress relief (Jevtov et al., 2015;Riback et al., 2017;Kroschwald and Alberti, 2017). The enrichment in RNAs encoding ATP binding, cell cycle, and transcription factors may be consistent with this notion.
However, translation initiation of the RNAs stored in stress granules has been shown not to be required for growth, as translation can resume before stress granule clearance is complete (Loschi et al., 2009;Walters et al., 2015). Furthermore, a recent study suggests that the mRNAs that are upregulated upon oxidative stress are vastly excluded from stress granules (Singh et al., 2021), strengthening the notion that stress granules might not be needed to protect RNA from degradation, but simply store them. Last, contrary to the finding that stress granules do not contain fully assembled ribosomes (only 40S) (Kedersha et al., 1999;Anderson and Kedersha, 2002), single-molecule imaging has revealed that the translation machinery can be activated in stress granules (Mateju et al., 2020). This suggests that although general translation is inhibited, stress granules could be a hub for translating specific RNAs, perhaps related to coping with stress.
As a proof of principle, using HyperTRIBE opens avenues to refine the specific core of common stress granule RNAs upon different stress and understand their role and their features. It is the first method to identify stress granule RNAs in single cells and in tissues, and could be particularly important to reveal heterogeneity between cells, within primary cells or tissues in their response to different stress. Importantly, this method could also allow the identification the RNA content of other RNA-based phase-separated assemblies, whether formed upon stress or existing in basal conditions, such as neuronal granules (Formicola et al., 2019;Kiebler and Bassell, 2006), posterior granules in Drosophila oocytes (Trcek and Lehmann, 2019;Bose et al., 2021) and P granules in Caenorhabditis elegans (Aoki et al., 2021;Lee et al., 2020), whose RNA content has so far been difficult to isolate and study.
Limitations of the study However successful, this proof of principle presents several limitations. First, cells and tissues need to express ADARcd through transfection or generation of transgenic animals. Second, one of the requirements is to limit the basal editing (in the absence of stress even if RNA do not have a long half-life) by controlling the timing and temporal (in vivo) expression of ADARcd. In Drosophila cells and tissues, we used the metallothionein-inducible promoter and the geneswitch system, respectively. Photoactivatable (light oxygen voltage) FRM1-ADARcd would also be envisioned for specific activation in stress granules, thus, to restrict even more the background (Wu et al., 2009). Third, the calculation editing frequencies could be done differently. Here, we calculated the average editing frequencies for a gene by extracting the percentage of mutated bases at each identified variant position, for each sample, and averaged the numbers (see method details section, Figure S4D). However, the editing frequency could be calculated by extracting the number of all edits over all variant positions for one gene and dividing this by the sum of the total positions (see STAR Methods for further discussion of this, with examples). This strategy could in particular be important for genes with considerably varying read depths at edited positions across a gene. In this study, using this alternative strategy gave slightly different results but did not affect the number and the types of RNAs predicted to be localized in stress granules (99.99% overlap). Fourth, certain RNAs may be covered with RBPs, therefore blocking the editing sites. Fifth, even though our validation by smFISH indicates that the RNAs predicted to be in stress granules are localized to these coalescences, it is possible that due to the translational arrest during arsenite treatment, these RNAs are edited in the cytosol by FMR1-ADARcd-V5 before they coalesce into stress granules. To assess how many and which RNAs are edited by ADARcd in the cytoplasm in arsenite irrespective of their coalescence into stress granules, NES-ADAR (instead of FMR1-ADARcd) could be expressed. Sixth, once edited, a given RNA molecule might leave the stress granule and return to the cytoplasm or to P-bodies (Zhang et al., 2011;Mollet et al., 2008), leading to the identification of RNAs that are only transiently residing in stress granules. However, the predicted stress granule RNAs we have tested by smFISH are enriched in stress granules, not only present, suggesting that they are true resident. Last, ADARcd was only tagged to one RBP (FMR1), thus potentially biasing our analysis toward FMR1 clients. Even if it is not the case, using another stress granule protein that is not an RBP would further validate our findings.

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

Data analysis -Workflow variant identification and calculation mutation frequency
The non-randomness of the edited sites leads us to the following data processing workflow, resulting in a gene-wise editing metric ( Figure S4C). Sequencing reads were first trimmed, aligned, and de-duplicated. The resulting de-duplicated alignments were merged and enter the GATK RNA-seq short variant discovery best-practices workflow using GATK (version 4.1.9.0). Reads with N in the alignment cigar were split into multiple alignments (SplitNCigarReads). Next, there were two rounds of base quality score recalibration (BaseRecalibrator, ApplyBQSR, and AnalyzeCovariates) and variant calling (HaplotypeCaller). The first round of recalibration masked all known reference variants, and the second additionally masked all novel variants. The exonic variants from the second round of variant calling were used for further downstream analyses. The variant positions were further selected based on hard cut-offs for quality (QUAL >250) and mutation frequency (AF < 0.95). Manual inspection on several genes revealed that the resulting positions were indeed capturing the rare, moderately edited positions that we previously observed ( Figure S4A; vertical light grey bars highlight identified variants). For each of these variant positions, we then extracted the total number of mutated bases for each sample and used these to calculate their editing frequency. The editing frequencies of all identified variant positions for a gene were then averaged; this is reported as the ''average editing frequency''. For instance, with reference to Figure S4D Figure S4C, the editing frequency would be calculated as (17 + 23)/(53 + 58) = 0.36. Note that the frequencies are very slightly different, but it did not affect the number and the types of RNAs predicted to be localized in stress granules (1856 in the used strategy versus 1855 using this alternative strategy, with 99.99% overlap). As mentioned in the discussion, using this alternative analysis strategy might be important when many genes are detected with low or varying read depths at edited positions across a gene.

Immunofluorescence (IF)
After treatments, S2 cells and dissociated brain cells were fixed with 4% paraformaldehyde (Sigma Aldrich) in PBS (pH 7.4) for 20 min. Cells were stained following the same protocol as described in (Zhang et al., 2021).

Single-molecule FISH (smFISH)
Plated Wildtype S2 cells were first treated with arsenite (see above), fixed and labeled for endogenous FMR1 as described above in the IF section. After incubation with the secondary antibody, cells were washed 3 times with PBS and cells were post-fixed in 4% paraformaldehyde in PBS (pH 7.4) for 10 min. Following a washing 3 times in PBS, cells were further incubated for 5 min in 10% formamide (Thermofisher) in DEPC-treated water. They were then incubated overnight on a droplet containing one fluorescent smFISH probe (125 nM in 1% dextransulfate (D8906, Sigma Aldrich), 10% formamide (Thermofisher) in DEPC-treated water at 37 C) in a moistened chamber to avoid drying. Cells were washed 2 times for 30 min with 10% formamide (Thermofisher) in DEPC-treated water and mounted with Prolong antifade media (+DAPI) (ThermoFisher) on a microscope slide. All smFISH probes (Table S4) were labeled with Atto565 following (Gaspar et al., 2017). The DNA oligos were purchased from IDT. The TMR-oligo(dT)30x was also purchased from IDT.

Microscopy and image analysis
Immunofluorescence images were acquired on a TSC SP8 confocal microscope (Leica GmbH) with a 633 lens. For smFISH, a widefield Leica MM-AF microscope was used with a 1003 lens. smFISH images were deconvoluted using the Regularized Inverse Filter algorithm with the DeconvolutionLab2 plugin (Sage et al., 2017) in ImageJ. The smFISH spots (each corresponding to a single RNA molecule, (Raj et al., 2008)) were counted using the spot count tool in the Imaris Image Analysis Software (Bitplane), and the percentage of spots in stress granules was estimated by their overlap in fluorescent signal with FMR1. At least, 20 randomly selected cells (N) were quantified per smFISH probe. The minimum and maximum display values were adjusted for each channel for visualization purposes.
For calculating the fraction of FMR1-ADARcd-V5 in stress granules, we used FIJI's ''plot profile'' plugin. The ratio of fluorescence intensity of FMR1 in stress granules and cytoplasm was calculated for 10 cells.
S2 cell death in arsenite stress and upon relief A total of 2 3 10 6 S2 cells per well in a 6-wells were plated and allowed to adhere for 1 h at 26 C in Schneider's. Cells were treated with either Schneider's or with arsenite for 4 h at 26 C in duplicate. Directly after the treatment, dead cells were stained with trypan blue Cell Reports Methods 2, 100235, June 20, 2022 e5 Article ll OPEN ACCESS and living cells were counted using a cell counter. After 4 h of arsenite, the growing medium was changed for Schneider's, and living cells were counted at several time points after stress relief every 24 h up to 168 h.
In parallel, stress granule formation and dissolution were monitored by immunofluorescence (using FMR1 as marker). The number of cells displaying stress granules was scored again the total number of living cells in each time point.

QUANTIFICATION AND STATISTICAL ANALYSIS
To identify the RNAs more edited upon stress conditions, we set that the editing frequency of RNAs from the ''arsenite'' triplicates should be significantly higher (p < 0.01 using the Empirical Bayes test (ebbr package in R)) than the editing frequency of same RNAs from the ''basal triplicate''. The correlation (R2) between samples was calculated using the function pandas.dataframe.corr() in Python. A Mann-Whitney test (scipy.stats.mannwhitneyu) was used in Python to determine the significance between transcript length. To calculate the overlap between two datasets, we applied a hypergeometric test using the phyper function in R.