Identification of astrocyte regulators by nucleic acid cytometry

Abstract

Multiple sclerosis is a chronic inflammatory disease of the central nervous system¹. Astrocytes are heterogeneous glial cells that are resident in the central nervous system and participate in the pathogenesis of multiple sclerosis and its model experimental autoimmune encephalomyelitis^2,3. However, few unique surface markers are available for the isolation of astrocyte subsets, preventing their analysis and the identification of candidate therapeutic targets; these limitations are further amplified by the rarity of pathogenic astrocytes. Here, to address these challenges, we developed focused interrogation of cells by nucleic acid detection and sequencing (FIND-seq), a high-throughput microfluidic cytometry method that combines encapsulation of cells in droplets, PCR-based detection of target nucleic acids and droplet sorting to enable in-depth transcriptomic analyses of cells of interest at single-cell resolution. We applied FIND-seq to study the regulation of astrocytes characterized by the splicing-driven activation of the transcription factor XBP1, which promotes disease pathology in multiple sclerosis and experimental autoimmune encephalomyelitis⁴. Using FIND-seq in combination with conditional-knockout mice, in vivo CRISPR–Cas9-driven genetic perturbation studies and bulk and single-cell RNA sequencing analyses of samples from mouse experimental autoimmune encephalomyelitis and humans with multiple sclerosis, we identified a new role for the nuclear receptor NR3C2 and its corepressor NCOR2 in limiting XBP1-driven pathogenic astrocyte responses. In summary, we used FIND-seq to identify a therapeutically targetable mechanism that limits XBP1-driven pathogenic astrocyte responses. FIND-seq enables the investigation of previously inaccessible cells, including rare cell subsets defined by unique gene expression signatures or other nucleic acid markers.

Rare cells exert substantial influence on tissue physiology and pathology across many domains of biology^5–7. In the context of inflammation in the central nervous system (CNS), small cell subsets have important roles in tissue pathology^8–12. For example, astrocytes are CNS-resident glial cells with multiple roles in development, homeostasis and neurological disorders^2,13,14; small astrocyte subsets identified by single-cell sequencing have been shown to promote or limit CNS pathology in multiple sclerosis and experimental autoimmune encephalomyelitis^4,15–22 (EAE), a pre-clinical model of multiple sclerosis. However, most astrocyte subsets (and many other cells of interest identified using single-cell genomics) lack unique surface markers, limiting their isolation and further investigation. Moreover, single-cell methods must sequence prohibitively large numbers of cells to sufficiently profile rare target cells. In addition to being costly and wasteful, this results in sparse transcriptome coverage is sparse.

Despite an absence of tools to interrogate the properties of many astrocyte subsets, increasingly sophisticated databases of transcriptional markers are being compiled using single-cell approaches^22–24. The sorting of astrocytes defined using these datasets is often not possible because their nucleic acid signatures do not map to unique surface markers that enable their isolation. This is especially true when isolation of nuclei is necessary to achieve adequate RNA sequencing (RNA-seq) data quality, as is the case for many cells in the CNS. Fluorescence in situ hybridization linked to fluorescence-activated cell sorting (FACS) enables the isolation of cells based on the direct detection of RNA markers but is insensitive to low copy targets, and the harsh fixatives used in this method degrade mRNA and interfere with the generation of high-quality sequencing data from the sorted cells²⁵. Moreover, obtaining a detectable signal often requires the use of multiple tiled probes, limiting the detection of sensitive post-transcriptional events such as RNA splicing²⁵. By contrast, droplet methods that use amplification-based nucleic acid detection achieve single-molecule sensitivity and can reliably distinguish transcript splicing isoforms, but mRNA degradation during lysis and thermocycling limits sequencing data quality even when many cells are pooled^26,27. An ideal method for studying rare cells defined by nucleic acid markers would combine the specificity and sensitivity of PCR detection with the ability to obtain high-quality post-sorting sequencing data.

In this study, we describe the development and application of FIND-seq, which enables the high-throughput and in-depth sequencing of rare cell subsets defined by RNA or DNA markers. FIND-seq uses poly-dT oligonucleotide-conjugated agarose to capture cell nucleic acids with high throughput. RNA is reverse transcribed on the agarose bead and subjected to sequence-specific PCR with fluorescent TaqMan probes, followed by droplet sorting of cells containing marker RNA and downstream bulk or single-cell analysis. We combined FIND-seq with cell-specific genetic perturbation studies in vivo and bulk and single-cell transcriptional and genomic analyses of EAE and multiple sclerosis samples and identified NR3C2–NCOR2 signalling as a negative regulator of an astrocyte subset driven by the post-transcriptional splicing-dependent activation of the transcription factor XBP1, which promotes CNS pathology in multiple sclerosis and EAE⁴.

FIND-seq development

A central challenge in the study of astrocyte subsets is the need to isolate populations of interest for in-depth molecular characterization. For example, we recently identified a subset of astrocytes that promotes CNS pathology in EAE and potentially multiple sclerosis, characterized by the activation of the transcription factor XBP1²². Despite the prominent role of this relatively rare subset in promoting CNS pathology²², XBP1⁺ astrocytes lack known surface markers that can be used to isolate and investigate them. Compounding these challenges, XBP1 activation is catalysed by an mRNA splicing mechanism, which removes a premature stop codon from the Xbp1 mRNA transcript²⁸. Thus, XBP1 activation cannot be directly quantified by single-cell RNA-seq (scRNA-seq) because 3′ mRNA capture does not consistently detect Xbp1 mRNA splicing. To analyse XBP1 activity, we assayed Edem1, a marker gene associated with XBP1 activation⁴ in a recent analysis of 18,047 CNS cells during EAE by scRNA-seq; only 14 out of 1,240 astrocytes expressed Edem1 (Fig. 1a and Extended Data Fig. 1a,b). These challenges highlight the need for a method to isolate cells of interest, such as XBP1⁺ astrocytes, on the basis of the expression of mRNA markers.

Fig. 1 |. Development of FIND-seq to study rare astrocyte subsets.

a, Left, clustering of scRNA-seq data from the CNS of EAE mice (re-analysed from ref. ²²). Right, extraction and re-clustering of EAE astrocytes showing Aqp4 and Edem1 expression superimposed on all clusters. b, Schematic of air-triggered droplet generation of molten agarose functionalized with polyT oligonucleotides (step 1). c, Micrographs of cells encapsulated in molten agarose droplets. d, Schematic of reverse transcription of cDNA on agarose beads (left, step 2) followed by agarose bead re-injection for digital PCR detection (right, step 3). e, Bead re-injection. Bottom left, micrographs of agarose hydrogel beads after solidification and removal from oil. Genomes are visualized with SYBR staining. Bottom right, agarose beads re-encapsulated in oil. Droplets maintain single-cell resolution during PCR detection. f, Schematic of dielectrophoretic sorting of droplets into either 100-cell bins or 96-well plates for single-cell analysis (step 4). g, Micrographs of single-cell droplet PCR detection before sorting (bottom left) and after sorting (bottom right). Data are mean ± s.e.m.

Efficient nucleic acid detection, sorting and sequencing of cells requires the simultaneous optimization of target marker detection and mRNA stability in many input cells. FIND-seq addresses these challenges by using a multi-step microfluidic workflow that can take millions of cells as input (Fig. 1b,c and Extended Data Fig. 1c–e) and analyse rare cell populations with <1% abundance in a cost-effective manner compared with scRNA-seq (Extended Data Table 1). The first device encapsulates cells in molten agarose with lysis buffer; the air-triggered droplet generation runs at 10 kHz, encapsulating 10⁶ cells at limiting dilution in about 20 min (Fig. 1b,c and Extended Data Fig. 2a,b). This speed maximizes cell health and reduces transcriptional changes in cells as they await microfluidic encapsulation. To simultaneously capture cellular genomes and transcriptomes, we developed a method for functionalizing primers directly onto agarose (Extended Data Fig. 2c). After lysis, the agarose is cooled, genomes are sterically trapped in the matrix, and cellular mRNA is hybridized with the capture probes (Extended Data Fig. 2d,e). The solid gels are removed from oil and transferred to wash buffer to clear cellular debris and residual lysis buffer, thereby retaining only purified nucleic acids from lysed cells. Captured mRNA is reverse transcribed and genomic DNA remains sterically confined in the matrix (Fig. 1d and Extended Data Fig. 3a). Exploiting bulk processing of hydrogels, we performed technical controls to retain high-quality single-cell messenger RNA and optimize cDNA synthesis yield (Extended Data Fig. 3b–g). Thus, we developed a high-throughput workflow to capture, purify and amplify the transcriptomes of millions of cells in parallel.

We next developed a method to detect nucleic acids from single cells using multiplexed droplet TaqMan PCR assays that also preserved the integrity of the transcriptome (Fig. 1d,e). TaqMan enables the use of a diverse inventory of pre-validated assays while enabling user-driven customization when necessary. Moreover, porous hydrogels simplify bulk processing, since the requisite reagents can be introduced by buffer exchange. Specifically, we re-encapsulated 55-μm hydrogels with PCR reagents into oil using a microfluidic device so that, upon thermal cycling, positive signals are generated within 75 μm drops (Fig. 1d,e and Extended Data Fig. 4a,b). We optimized the detection assay to preserve cDNA during the PCR, as normal thermal cycling conditions reduced downstream whole-transcriptome amplification (WTA) yield and data quality (Extended Data Fig. 4c). The resulting droplet PCR also enabled robust detection of single-copy genomic DNA targets (Extended Data Fig. 4d and the accompanying Article²⁹). To isolate cells that express the target RNA transcript, we sorted fluorescent drops using a concentric microfluidic device previously optimized for high-fidelity sorting³⁰ (Fig. 1f,g and Extended Data Fig. 4e,f). Sorted beads were collected and cDNA was amplified and prepared for RNA-seq analysis (Fig. 1f,g). In summary, we established a high-throughput method that sorts cells on the basis of RNA (or DNA) markers.

Analysis of XBP1⁺ astrocytes with FIND-seq

We recently identified an astrocyte subset that promotes CNS pathology in multiple sclerosis and EAE and is driven by the transcription factor XBP1 induced by the unfolded protein response⁴. A similar astrocyte population contributes to prion-induced CNS pathology³¹. These findings suggest that XBP1-driven astrocytes contribute to the pathology of multiple neurological disorders, but little is known about the mechanisms that regulate them. The Xbp1 gene is first transcribed into an unspliced Xbp1 mRNA (Xbp1u), which harbours a stop codon that results in the synthesis of a truncated, non-functional XBP1³² (Fig. 2a). In response to endoplasmic reticulum stress, activation of the inositol requiring enzyme 1α (IRE1α) promotes Xbp1 mRNA splicing, generating spliced Xbp1 mRNA (Xbp1s) lacking the early stop codon, which thus drives the synthesis of the full-length functional transcription factor XBP1³². However, neither Xbp1 mRNA splicing nor the induction of Edem1³³, one of its primary target genes, is readily detected by scRNA-seq in astrocytes (Fig. 1a and Extended Data Fig. 1b), presumably because of data sparsity and the 3′ bias of single-cell barcoding methods^34,35. We therefore applied FIND-seq to study XBP1-driven astrocytes.

Fig. 2 |. FIND-seq analysis of XBP1-driven astrocytes.

a, Schematic showing the generation of functional XBP1S and Edem1 expression following Xbp1u mRNA splicing. b, Naive and EAE mice were analysed by FIND-seq using Aqp4 and Edem1 expression as markers of astrocytes and Xbp1 mRNA splicing, respectively. c, Representative droplet cytometry plots. Numbers in each quadrant are the percentage of positive drops. Top quadrants contain Aqp4⁺ cells, and the top-right quadrant contains Aqp4⁺Edem1⁺ cells. d, Percentage and absolute numbers of Aqp4⁺Edem1⁺ cells in naive and EAE mice. n = 3 mice per group. Unpaired two-sided t-test. e, Immunostaining analysis of XBP1S⁺ GFAP⁺ astrocytes in the grey and white matter of the brain and spinal cord in naive and EAE mice. n = 6 images per group from n = 3 mice. Unpaired two-sided t-test. f, Heat map of genes that are differentially expressed between naive and EAE Aqp4⁺ astrocytes (defined by P < 0.05 and |[log₂(fold change)]| > 2). BR, biological replicate. Two technical replicates were analysed for each mouse in each condition. g, Pathways analysed by pre-ranked gene set enrichment analysis (GSEA) of genes from RNA-seq data comparing EAE Aqp4⁺Edem1⁻ with naive Aqp4⁺Edem1⁻ cells. h, Canonical inflammatory pathways in EAE Aqp4⁺Edem1⁻ cells relative to EAE Aqp4⁺Edem1⁻ cells. i, IPA analysis of FIND-seq data comparing Aqp4⁺Edem1⁺ to Aqp4⁺Edem1⁻ cells during EAE showing prediction of XBP1 as an upstream regulator in EAE Aqp4⁺Edem1⁺ cells. This analysis is performed by computing the overlap in the differentially expressed genes of the FIND-seq dataset with the genes regulated by XBP1 by Fisher’s exact test. Data are mean ± s.e.m.

We developed a multiplexed digital droplet TaqMan assay that detected the expression of Edem1 and the astrocyte marker Aqp4, which we validated could accurately detect Aqp4⁺ astrocytes with similar sensitivity and specificity (6.6% of cells) to those of flow cytometry (4.7% of cells) and scRNA-seq (6.8% of cells) (Extended Data Fig. 5a). Moreover, we sorted AQP4⁺ and AQP4⁻ cells using flow cytometry and analysed each population through the FIND-seq workflow (but before transcriptome sequencing), which showed 80% sensitivity to detect Aqp4 in the AQP4⁺ population with a 4.1% detection rate of Aqp4 in the AQP4⁻ population (Extended Data Fig. 5b). We then used FIND-seq to study Aqp4⁺Edem1⁺ astrocytes in EAE (Fig. 2b,c and Supplementary Table 1). In agreement with immunohistochemistry analyses⁴, we detected an increase of approximately tenfold in the number of XBP1-driven astrocytes in the CNS 18 days after EAE induction by immunization with the myelin oligodendrocyte glycoprotein peptide MOG_35–55 (Fig. 2d,e). Bulk FIND-seq (cells sorted into 100-cell bins) produced RNA-seq data (300 cells, 26,171 genes) similar in complexity to that of to Smart-seq2 (ref. ³⁶) (300 cells, 25,076 genes) and showed improved depth and gene diversity compared with techniques such as Probe-seq²⁵ (22,000 cells, 15,118 genes) and scRNA-seq³⁷ (14 cells, 8,190 genes) that enable high-throughput identification of cells based on RNA markers (Extended Data Fig. 5c–e). In addition, RNA-seq analyses of Aqp4⁺ astrocytes from EAE mice detected increased expression of pro-inflammatory pathways linked to the promotion of CNS pathology (Fig. 2f). Aqp4⁺Edem1⁻ astrocytes also displayed increased expression of pro-inflammatory transcriptional programmes during EAE (Fig. 2g), consistent with the existence of Aqp4⁺Edem1⁻ astrocyte subsets that also contribute to CNS pathology^16,18,22,38.

FIND-seq enabled the transcriptional analysis of the Aqp4⁺Edem1⁺ astrocyte subset, detecting increased expression of pro-inflammatory signalling pathways during EAE (Fig. 2h). To identify regulatory mechanisms associated with Aqp4⁺Edem1⁺ astrocytes, we tested for signatures of molecules that putatively control Aqp4⁺Edem1⁺ versus Aqp4⁺Edem1⁻ astrocytes using Ingenuity Pathway Analysis (IPA). Notably, we identified Xbp1 as a regulator of the transcriptional signature we detected in Aqp4⁺Edem1⁺ astrocytes and Edem1 as a downstream gene in this regulon, suggesting that FIND-seq enabled the analysis of XBP1-driven astrocytes (Fig. 2i). Collectively, these findings underscore the ability of FIND-seq to analyse in depth an astrocyte subset of interest with a role in CNS pathology based on the expression of RNA markers.

NR3C2–NCOR2 signalling controls XBP1⁺ astrocytes

To investigate the transcriptional heterogeneity within Aqp4⁺Edem1⁺ astrocytes, we developed single-cell FIND-seq (scFIND-seq). We optimized single-bead sorting into microwells of a 96-well plate (Extended Data Fig. 6a) and performed a species-mixing experiment to validate single-drop sorting fidelity (Extended Data Fig. 6b). To analyse RNA expression in naive or EAE Aqp4⁺Edem1^+/− astrocytes by scFIND-seq, we extended a unique molecular identifier (UMI)-corrected Smart-seq3 approach to individual cells captured in agarose beads (Fig. 3a, Extended Data Fig. 6c–e and Supplementary Table 2). Consistent with our previous analyses validating the quality of bulk FIND-seq data relative to similar technologies, scFIND-seq detected 16,874 unique genes on average compared with Smart-seq2, which detected 25,076 (Extended Data Fig. 5d). scFIND-seq data captured 284 cells representing Aqp4⁺Edem1⁻ and Aqp4⁺Edem1⁺ astrocytes isolated from naive and EAE mice; Aqp4⁺Edem1⁺ astrocytes were predominantly located in a distinct cluster of cells with pro-inflammatory markers that emerged during EAE (Fig. 3b–d). Our bulk FIND-seq datasets identified nine transcriptional regulators predicted to control Aqp4⁺Edem1⁺ astrocytes in EAE when compared with Aqp4⁺Edem1⁻ astrocytes (Extended Data Fig. 7a,b), seven of which were also identified using scFIND-seq (Extended Data Figs. 7c–e). Indeed, we identified two subclusters of Aqp4⁺Edem1⁺ astrocytes (Extended Data Fig. 7c–e) on the basis of single-cell transcriptional analysis, highlighting both the potential heterogeneity within this astrocyte subset and the power of scFIND-seq for the analysis of rare cell populations from complex tissues.

Fig. 3 |. NR3C2–NCOR2 signalling limits disease-promoting astrocyte responses.

a, Schematic of scFIND-seq applied to Aqp4⁺Edem1⁺ astrocytes in EAE. b, Uniform manifold approximation and projection (UMAP) plots of Aqp4⁺Edem1⁺ and Aqp4⁺Edem1⁻ astrocytes from naive and EAE mice analysed by scFIND-seq. c, Analysis of marker genes in Aqp4⁺Edem1⁺ or Aqp4⁺Edem1⁻ astrocytes from naive and EAE mice. d, Pathway analysis of cluster 1 astrocytes analysed by scFIND-seq using IPA. Fisher’s exact test. e, In vivo knockdown of Nr3c2 in astrocytes. Top, schematic of the lentiviral vector containing sgRNA targeting Nr3c2 and Streptococcus pyogenes Cas9 under the control of a Gfap promoter. Bottom, EAE disease progression in mice transduced with sgScrmbl (n = 14) or sgNr3c2 (n = 7) lentiviruses. Two-way repeated measures ANOVA. f, Quantitative PCR with reverse transcription Il1b, Il6 and Ccl2 after 24 h and Xbp1 after 4 h in primary mouse astrocytes treated with or without 1 ng ml⁻¹ IL-1β and TNF in the presence of 1 μM finerenone or vehicle. n = 5 per group. Unpaired two-sided t-test. g, Left, heat map of RNA-seq data from of astrocytes isolated from sgScrmbl and sgNr3c2 mice. n = 3 per group. Right, GSEA pre-ranked analysis of RNA-seq data comparing astrocytes isolated from sgNr3c2 versus sgScrmbl mice. h, FACS analysis of spliced XBP1 (XBP1S⁺) in astrocytes from sgScrmbl and sgNr3c2 mice (n = 3 per group). Unpaired two-sided t-test. i, Left, prediction of decreased NCOR2 activation by IPA in cluster 1 scFIND-seq data compiled from all cells. Fisher’s exact test. Right, differential gene expression of candidate co-regulators of NR3C2 in sgNr3c2 versus sgScrmbl astrocytes showing Ncor2 as the top candidate. j, Analysis of EAE mice transduced with lentivirus co-expressing Gfap::Cas9 and sgNcor2 (n = 9) or sgScrmbl (n = 9). Top, schematic of the lentiviral vector containing sgRNA targeting Ncor2 and S. pyogenes Cas9 under the control of a Gfap promoter. Bottom, disease progression in sgScrmbl versus sgNcor2 mice. Two-way repeated measures ANOVA. k, Heat map of RNA-seq data from astrocytes in sgScrmbl versus sgNcor2 mice. n = 3 per group. l, GSEA pre-ranked analysis of scRNA-seq data in astrocytes from sgScrmbl versus sgNcor2 mice. Data are mean ± s.e.m.

We then performed an in vivo genetic perturbation screen using lentiviral vectors co-expressing CRISP–Cas9 and a single guide RNA (sgRNA) targeting each of the candidate regulators^4,21,22,38 under the control of the Gfap promoter (Extended Data Fig. 7f). Control mice were treated with vectors encoding CRISPR–Cas9 and a non-targeting sgRNA (sgScrmbl). We delivered the lentiviruses via intracerebroventricular injection to mice and monitored EAE development following immunization with MOG_35–55 as described^4,18–22,38. Consistent with our bioinformatic predictions showing downregulation of NR3C2 signalling detected by FIND-seq and scFIND-seq in Aqp4⁺Edem1⁺ astrocytes (Extended Data Fig. 7a–e), astrocyte-specific knockdown of the ligand-activated nuclear receptor Nr3c2 resulted in the worsening of EAE (Fig. 3e), suggesting that NR3C2 limits disease-promoting astrocyte responses.

Nr3c2 encodes the mineralocorticoid receptor, and belongs to a family of nuclear receptors with known immunomodulatory effects³⁹. Nr3c2 has been associated with early onset hypertension, severe exacerbation of hypertension in pregnancy as well as other peripheral diseases⁴⁰, and neurological disorders such as autism⁴¹. However, the role of NR3C2 in the control of astrocyte subsets is unknown. Consistent with our in vivo findings, treatment with 1 μM of the NR3C2 antagonist finerenone increased gene expression of Il6, Il1b and Ccl2 in primary mouse astrocytes activated in vitro with 1 ng ml⁻¹ IL-1β and TNF (Fig. 3f), cytokines that trigger transcriptional responses similar to those detected in astrocytes during the course of EAE and multiple sclerosis, including XBP1 activation²². Finerenone treatment also boosted Xbp1 splicing (Fig. 3f). Similarly, RNA-seq analyses of astrocytes from EAE mice transduced with Nr3c2-targeting CRISPR–Cas9 lentiviruses detected inhibition of the Nr3c2 regulon compared with mice transduced with non-targeting controls (Extended Data Fig. 8a), concomitant with increased IκB phosphorylation (Fig. 3g and Supplementary Table 3) associated with NF-κB activation, which controls disease-promoting transcriptional modules in XBP1-driven astrocytes⁴. Nr3c2 inactivation in astrocytes was also linked to increased XBP1 activation, as determined by the increased abundance of astrocytes exhibiting spliced XBP1 (Fig. 3h). Nr3c2 inactivation in astrocytes did not affect the T cell response during EAE (Extended Data Fig. 8b,c).

Following its ligand-dependent activation, NR3C2 translocates to the nucleus, where it binds responsive elements in target genes to modulate their expression. However, the transcriptional regulation of target genes by NR3C2 requires additional coactivators and corepressors³⁹. The analysis of the RNA-seq dataset of astrocytes following Nr3c2 inactivation identified the nuclear receptor corepressor 2 (NCOR2) as a potential contributor to the suppression of XBP1-driven astrocyte pro-inflammatory responses by NR3C2 (Fig. 3i), which was also predicted as a regulator of Aqp4⁺Edem1⁺ cells in cluster 1 using scFIND-seq (Fig. 3i). Indeed, NCOR2 inactivation in astrocytes using lentiviral vectors co-expressing CRISPR–Cas9 and a targeting sgRNA under the control of the Gfap promoter resulted in exacerbated EAE (Fig. 3j), concomitant with the increased expression of pro-inflammatory transcriptional modules driven by XBP1, but no changes in the T cell response during EAE (Fig. 3k,l, Extended Data Fig. 8d–f and Supplementary Table 4). Together, these data suggest that NR3C2–NCOR2 signalling limits XBP1-driven pro-inflammatory transcriptional responses in astrocytes.

XBP1 limits NR3C2–NCOR2 signalling in astrocytes

XBP1 is activated in grey and white matter astrocytes in the spinal cord and brain during EAE⁴ (Fig. 2d,e). To investigate the cross-regulation between XBP1 and NR3C2–NCOR2 signalling, we bred mice harbouring a floxed Xbp1 allele⁴² with mice expressing tamoxifen-inducible Cre recombinase in astrocytes under the control of the Aldh1l1 promoter⁴³, generating Xbp1^Astro mice in which Xbp1 is knocked out in astrocytes following tamoxifen administration (Fig. 4a). Xbp1 inactivation in astrocytes resulted in a significant amelioration of EAE (Fig. 4b), but did not affect the T cell response or the number of CNS-resident cells during EAE (Extended Data Fig. 9a–d). However, analysis of the transcriptional programme of Xbp1-deficient astrocytes by RNA-seq showed an increased activation of NR3C2-driven transcriptional modules (Fig. 4c and Supplementary Table 5) concomitant with the decreased activation of disease-promoting transcriptional programmes such as cytokine signalling and the unfolded protein response (Fig. 4c and Extended Data Fig. 9d,e). Consistent with these data, Xbp1 inactivation resulted in increased chromatin accessibility of genes driven by NR3C2 mineralocorticoid signalling, as shown by assay for transposase-accessible chromatin with high-throughput sequencing (ATAC-seq) in bulk sorted astrocytes (Extended Data Fig. 10a), as well as increased chromatin accessibility of Nr3c2 and Ncor2 (Extended Data Fig. 10b). These data suggest that XBP1 activation limits anti-inflammatory NR3C2 signalling in astrocytes, and that XBP1- and NR3C2-driven signalling pathways engage in an antagonistic cross-talk that regulates astrocyte pathogenic activities.

Fig. 4 |. XBP1 limits NR3C2–NCOR2 signalling in EAE and multiple sclerosis.

a, Mice with a floxed Xbp1 allele were crossed to mice expressing creERT2 under control of the Aldh1l1 promoter to generate tamoxifen-inducible Xbp1 astrocyte (Xbp1^Astro) conditional-knockout mice. b, EAE disease progression in wild-type Xbp1 (Xbp1^WT) (n = 8) and Xbp1^Astro knockout (n = 7) mice. c, Analysis of RNA-seq data comparing astrocytes isolated from Xbp1^WT and Xbp1^Astro mice. n = 3 per group. ER UPR, endoplasmic reticulum unfolded protein response; GOBP, Gene Ontology biological process; KO, knockout. d, Re-analysis of scRNA-seq datasets from patients with multiple sclerosis (previously reported in ref. ²²), showing expression of NR3C2 and NCOR2 in human astrocytes. e, GSEA pre-ranked analysis comparing NCOR2⁺ astrocytes to NCOR2⁻ astrocytes and NR3C2⁺ astrocytes to NR3C2⁻ astrocytes in patients with multiple sclerosis. f, Quantification of NR3C2⁺ astrocytes identified by immunostaining of tissue from patients with multiple sclerosis. n = 3 patients and n = 9 images per group. Unpaired two-sided t-test. g, Violin plots depicting the signature score calculated for the indicated gene set in astrocytes isolated from controls or patients with multiple sclerosis, derived from data in d. Data are mean ± s.e.m.

To evaluate the clinical relevance of XBP1–NR3C2 antagonism in human astrocytes, we first analysed CNS scRNA-seq datasets from patients with multiple sclerosis and controls matched for sex, age and brain region²², detecting the expression of NR3C2 and NCOR2 in astrocytes (Fig. 4d). We next analysed the activation of XBP1-driven transcriptional modules in NR3C2⁺ or NCOR2⁺ astrocytes, and found that the expression of NR3C2 or NCOR2 in astrocytes was associated with a decreased activation of the XBP1-driven unfolded protein response and the production of reactive oxygen species (Fig. 4e).

To further evaluate these findings, we used immunostaining to analyse CNS samples from patients with multiple sclerosis and controls. We detected decreased numbers of NR3C2⁺ astrocytes in white matter from patients with multiple sclerosis relative to healthy controls (Fig. 4f). Moreover, the analysis of NR3C2- and NCOR2-driven transcriptional responses in scRNA-seq astrocyte datasets showed a reduction in NR3C2 and NCOR2 signalling in astrocytes from lesions of patients with multiple sclerosis compared with control donors (Fig. 4g). A similar reduction in NR3C2 and NCOR2 signalling was also detected in normal-appearing white matter from patients with multiple sclerosis (Fig. 4g). Together, these findings suggest that the balance between XBP1 and NR3C2–NCOR2 signalling controls disease-promoting astrocyte activities that contribute to CNS pathology in EAE and potentially in multiple sclerosis.

Discussion

Here we describe FIND-seq–a high-throughput method to isolate and sequence rare cells defined by nucleic acid markers–and its use to study an XBP1-driven astrocyte subset associated with CNS pathology in multiple sclerosis and EAE⁴. FIND-seq is a unique tool for the investigation of cell subsets identified in single-cell genomic studies because it can directly target the nucleic acid markers that define them. FIND-seq captures rare cell populations from millions of input cells, produces high-quality transcriptome data, and is compatible with scRNA-seq. FIND-seq eliminates the need for, and the associated high cost of, sequencing large numbers of background cells when focusing on important rare cell populations. Because FIND-seq can sort cells on the basis of RNA markers, including spliced transcripts, non-coding RNA or the presence of viral RNA, it has the potential to accelerate the study of cells for which transgenic animals or antibodies are unavailable. In addition, FIND-seq’s flexible processing of millions of input cells in porous, biocompatible and thermally reversable hydrogels makes it compatible with additional multi-omic readouts in future iterations, including genomic, epigenomic or proteomic measurements. However, further optimization is required to combine FIND-seq detection of transcripts in subcellular compartments such as astrocyte endfeet.

New high-dimensional, genome-wide and single-cell technologies are rapidly uncovering astrocyte subsets associated with a broad array of physiologic functions and neurological disorders^22,23,44,45. However, without techniques to isolate these subsets, it remains challenging to define the molecular mechanisms that control them and discover candidate targets for their therapeutic modulation^2,17. For example, FIND-seq enabled the investigation of rare UPR-driven astrocytes associated with CNS pathology in multiple sclerosis⁴ and prion disease³¹, resulting in the identification of an NR3C2–NCOR2 regulatory pathway that antagonizes XBP1-driven stimulus-induced astrocytes² in the context of CNS inflammatory conditions. To our knowledge, NR3C2 expression and function in astrocytes has not been investigated previously. Our data complement previous reports on the regulatory roles of ligand-activated transcription factors in CNS-resident cells in the context of multiple sclerosis and other neurological diseases. For example, in response to its activation by tryptophan-derived agonists controlled by the commensal flora, the ligand-activated transcription factor aryl hydrocarbon receptor suppresses pro-inflammatory responses in astrocytes and microglia during EAE and potentially, multiple sclerosis^19,20. Similarly, NR4A1 and NR4A2 limit the pro-inflammatory responses of astrocytes, microglia and CNS-recruited macrophages^46,47. On the basis of the potential for selective modulation of NR3C2 activity by small molecules⁴⁸, our findings identify NR3C2 as a candidate target for the therapeutic control of pathogenic, disease-associated astrocytes^2,17,22,23.

NCOR2 facilitates the recruitment of histone deacetylases to target genes recognized by interacting nuclear receptors, introducing epigenetic chromatin modifications that silence gene expression. NR4A2 has been shown to limit NF-κB-driven pro-inflammatory responses in astrocytes and microglia in Parkinson’s disease through epigenetic mechanisms mediated by the histone deacetylase complex CoREST⁴⁶. Moreover, the silencing of antioxidant loci in MAFG⁺ astrocytes contributes to CNS pathology in EAE and multiple sclerosis²². Collectively, these data highlight the potential role of epigenetic modifications in long-term changes in astrocytes that may amplify or perpetuate responses relevant for the development of CNS pathology³, resembling recent reports of trained microglial responses in the context of Alzheimer’s disease^49,50. Thus, the pharmacological targeting of epigenetic programmes associated with NCOR2 may offer novel approaches for the therapeutic modulation of CNS inflammation and neurodegeneration.

In summary, we report the development of FIND-seq, a method that can isolate rare cell subsets defined by nucleic acid signatures. Using FIND-seq, we revealed a NR3C2–NCOR2 regulatory mechanism controlling an astrocyte subset that promotes CNS pathogenesis in EAE and, potentially, multiple sclerosis.

Methods

Mice

All experiments were reviewed and approved by the Brigham & Women’s Hospital Institutional Animal Care and Use Committee. Adult mice aged 8–10 weeks and pups aged from postnatal day (P)0–P3 were used on a C57Bl/6J background (000664, The Jackson Laboratory). Aldh1l1-cre/ERT2 mice⁴³ (The Jackson Laboratory, 029655) were bred to Xbp1^fl/fl mice⁴² (a gift from L. Glimcher) to generate Aldh1l1-creERT2;Xbp1^fl/fl mice. Conditional deletion of Xbp1 was induced at 6–10 weeks of age by gavage with 225 mg kg⁻¹ tamoxifen (Sigma-Aldrich, T5648), diluted in corn oil (Sigma-Aldrich, C8267); EAE was induced 4 weeks later.

Analysis of previous data

Previously published data were accessed using the following accession codes: EMBL E-MTAB-9268⁵², GSE135572²⁵, and GSE130119²².

Fabrication of microfluidic devices

Masters were fabricated using soft lithography. Silicon wafers were spin coated with SU-8 2025 photoresist (MicroChem), UV-patterned using a mask aligner (SUSS MJB3), developed for 10 min, and baked at 65 °C overnight. Polydimethylsiloxane (PDMS) prepolymer and curing agent (Momentive; RTV 615) were combined at 10:1, mixed by hand, degassed, poured onto fabricated masters, degassed again, and baked at 65 °C overnight. Cured PDMS was removed from the master mould, punched with a 0.75-mm biopsy punch (Harris Uni-Core 0.75), and plasma bonded (Technics 500-II Plasma Etcher) to a glass slide (75 × 50 × 1.0 mm, Fisher Scientific, 12-550C). Aquapel was flushed through the microfluidic device and allowed to react for 5 min. Channels were purged with air and flushed with Fluorinert FC-40 oil (Sigma, F9755), before being baked at 65 °C for 30 min.

FIND-seq

Generation of oligonucleotide-conjugated agarose.

A 5′-acrydited polyT capture primer was covalently attached to allyl agarose as follows. The primer (/5Acryd/TTTTTTTAAGCAGTGGTATCAACGCAGAGTACT 30VN, Acrydite-T5-SMART-T30VN, IDT) was resuspended in 0.375 M Tris-HCl Buffer pH 8.3 (Teknova) to a concentration of 1,000 μM. SFR allyl agarose (Lucidant Polymers) was resuspended in 0.375 M Tris-HCl buffer (pH 8.3) to a final concentration of 0.5% and heated to 95 °C. Agarose was mixed by vortexing every 15 min for 2 h. Agarose was cooled to 45 °C and allowed to sit under vacuum for 30 min. Acrydite-T5-SMART-T30VN (50 μM), ammonium persulfate (Promega, V3131, 0.1% w/v), and TEMED (Invitrogen, 15524-010, 0.1% v/v) were added to the molten agarose, vortexed, and placed back under vacuum. After four hours, APS and TEMED were added again at their original concentrations, vortexed, and placed under vacuum overnight at 45 °C. Following completion of the reaction, the agarose was heated to 95°C, filtered through a 0.45 μm syringe filter, and ultra-low gelling temperature agarose type IX-A (Sigma-Aldrich, A2576) was added to a final concentration of 2% type IX-A agarose (0.5% SFR and 1.5% agarose). The final 2% conjugated agarose was hardened and allowed to wash overnight in 500 ml of ultra-pure water. Agarose-bound oligonucleotides were quantified by QuBit ssDNA assay (Invitrogen, Q10212) and normalized to 4 μM polyT with 2% ultra-low gelling type IX-A agarose.

Encapsulation and lysis of cells within agarose hydrogels.

Oligo-dT conjugated agarose, cells, lysis buffer, and fluorinated oil were co-flowed onto a bubble triggered droplet microfluidic device to generate 55-μm diameter droplets⁵³. Agarose was heated to 95 °C and vortexed until completely homogenous. Molten agarose was loaded into a 3 ml syringe (BD Biosciences, 309657) and placed into a custom-built syringe heater set to 85 °C. Cells were prepared as previously described²². In brief, mice were perfused with 1× PBS (Gibco, 10010023), the CNS was digested with papain suspension (Worthington, LS003126), mechanically dissociated, filtered through a 70 μm cell strainer (Fisher Scientific, 22363548), centrifuged at 500g for 5 min, resuspended in 10 ml of 30% Percoll solution (GE Healthcare Biosciences, 17-5445-01), and centrifuged at 500g for 25 min with no braking. The cell pellet was washed using HBSS with calcium or magnesium (Gibco, 14170112) and cells were counted manually using a haemocytometer. Cells were resuspended to 6.11 × 10⁶ cells per ml in HBSS with 18% (v/v) OptiPrep (Sigma-Aldrich, D1556-250ML) and 1% Pluronic F-68 non-ionic surfactant (v/v) (Thermo Fisher, 24040032). Lysis buffer containing 20 mM Tris-HCl pH 7.5, 1,000 mM LiCl, 1% LiDS (Sigma, L9781-5G), 10 mM EDTA (VWR, E177), 10 mM DTT (Sigma, 43816), and 0.4 μg μl⁻¹ proteinase K (NEB, P8107S) was prepared and loaded into a 3 ml syringe. A 10 ml syringe was loaded with droplet generation oil for probes (Bio-Rad, 186-3005). All syringes were connected to the microfluidic device using PE/2 tubing (Scientific Commodities, BB31695-PE/2). To make drops, pumps were run at 600 μl h⁻¹ for cells, 1,200 μl h⁻¹ for agarose, at 600 μl h⁻¹ for lysis buffer, and 5,000 μl h⁻¹ for oil. Air was controlled manually at 20 psi and tuned to generate 55-μm droplets. Droplets were collected in a 50 ml Falcon tube, incubated at 55 °C for 2 h, and placed at 4 °C for at least 2 h.

Drop breaking and washes.

After the encapsulation workflow, solid agarose microspheres were removed from oil using a drop-breaking procedure. All steps were carried out at 4 °C to prevent dissociation of mRNA from the polyT oligonucleotides. Oil was removed from the bottom of tubes and 20% ¹H,¹H,²H,²H-perfluoro-1-octanol (PFO) (Sigma-Aldrich, 370533) in HFE-7500 oil was added at a ratio of 3:1 (v/v) PFO:emulsion. Tubes were gently mixed by hand and centrifuged for 1 min at 2,000g. Oil was removed and discarded and the remaining beads were washed in 50 ml of 0.1% SPAN-80 in hexane, shaken gently by hand, and centrifuged at 2,000g for 1 min. Next, beads were subjected to five washes. For each wash, beads were incubated in wash buffer for 5 min on ice, spun down at 4,700 rpm for 10 min, and aspirated before the next wash. Beads were washed 1× with wash buffer 1: 20 mM Tris-HCl pH 7.5, 500 mM LiCl, 0.1% LiDS, and 0.1 mM EDTA. Beads were then washed 2× with wash buffer 2: 20 mM Tris-HCl pH 7.5, and 500 mM NaCl. Beads were then washed 2× with wash buffer 3: 5× reverse transcription buffer (Thermo Fisher); this buffer contains 250 mM Tris-HCl pH 8.3, 375 mM KCl, 15 mM MgCl₂, and 50 mM DTT. Beads in 5× reverse transcription buffer were then filtered using a 100 μm cell strainer.

Synthesis of cDNA using bead-conjugated primers.

Cellular RNA captured on polyT primers conjugated to the agarose hydrogel was reverse transcribed to make cDNA using 1 mM dNTPs (Thermo Scientific, R1121), 2 μM template switch oligo (AAGCAGTGGTATCAACGCAGAGTACATrGrGrG), 6 mM MgCl₂ (Sigma-Aldrich, M1028), 1 M betaine (Sigma, 61962), 7.5% PEG-8000 (Promega, V3011), 2 U μl⁻¹ Maxima H-minus reverse transcriptase, and 0.5 U μl⁻¹ RNAse inhibitor (Lucigen, 30281-2). Reverse transcription was performed on an inverter at room temperature for 30 min, followed by 90 min at 42 °C. Following reverse transcription, agarose beads were washed 5 times in 0.1% Tween-20 (Sigma, P9416) in nuclease-free ultra-pure water (Invitrogen, 10977015). After the final wash, cDNA was amplified from a small aliquot of beads, to ensure reverse transcription was successful. In brief, pelleted agarose beads were diluted fourfold and stained with 10× SYBR green (Sigma, S9430) for 30 min in the dark. Beads were loaded into a haemocytometer using a Leica DMi8 fluorescence microscope and the concentration of genomes was determined based on fluorescence. WTA was performed in a 25 μl reaction containing 750 genomes using 1× KAPA HiFi master mix (Roche, KK2601) and 0.4 μM Smart-seq2 PCR primer (AAGCAGTGGTATCAACGCAGAGT). Reactions were thermocycled under the following conditions: 95 °C for 3 min, 14, 16 or 18 cycles of (98 °C for 15 s, 67 °C for 20 s, 68 °C for 4 min), followed by 72 °C for 5 min, then held at 4 °C. PCR was purified using a 2:1 ratio of AMPure XP beads (Beckman Coulter, A63881), and eluted in 20 μl. The size of the amplified cDNA was validated by a high-sensitivity DNA chip (Agilent, 5067-4626) with an average size of 1.5–2.0 kb and a negligible fraction of fragments below 300 bp.

Validation of TaqMan assays in drops.

To develop a multiplex assay targeting Edem1-expressing astrocytes, we tested combinations of assays targeting astrocyte marker genes (Bio-Rad, dMmuCPE5095061; Thermo Fisher, Mm03048957_m1), and Edem1 (Bio-Rad, dMmuCNS573941944; Thermo Fisher, Mm00551797_m1) with quantitative PCR (qPCR). We performed qPCR on agarose hydrogels containing purified genomes and transcriptomes of primary mouse cells captured from EAE mice, prepared as described above. The reactions consisted of 2 μl of hydrogels, 1× Taqman fast universal PCR master mix (Life Technologies, 4367846), and 1× of each TaqMan assay containing primers and probes. We chose Aqp4 (Bio-Rad, dMmuCPE5095061) and Edem1 (Thermo Fisher, Mm00551797_m1) because single-plex qPCR curves were equivalent to the multiplex curves for both assays.

Single-cell multiplexed digital droplet PCR.

Agarose hydrogels containing purified genomes and transcriptomes from single cells were mixed with PCR reagents to achieve a final concentration of 1× TaqPath master mix, 2.5% PEG, 2.5% Tween-20, and 1× of each TaqMan assay. The bead solution was mixed in the dark for an hour on a shaker, transferred to a 3 ml syringe, and spun at 4,700 rpm for 10 min in a centrifuge (Thermo Sorvall XTR) in custom 3D-printed syringe holders. The supernatant was removed, transferred to a second 3 ml syringe, and the remaining beads in the first syringe were pushed into PE/5 tubing (Scientific Commodities, BB31695-PE/5). Both syringes, and fluorinated oil (Bio-Rad, 1863006) were re-injected into a microfluidic device at 400 μl h⁻¹ for beads, 400 μl h⁻¹ for 1× PCR mix, and 2,000 μl h⁻¹ for oil, and collected in 30 μl aliquots in PCR tubes. Drops in PCR tubes were then thermocycled under the following conditions: 88 °C for 10 min, then 50 cycles of (88 °C for 15 s and 60 °C for 1 min). After PCR amplification of the TaqMan targeted genes, the emulsions were pooled into a 3 ml syringe and sorted by droplet cytometry.

Microfluidic droplet cytometry.

Droplet microfluidic sorting was performed using a custom droplet cytometer controlled by LabView software. Three lasers (OptoEngine, MLL-FN-473, MLL-III-532, MRL-III-640) were aligned through dichroic mirrors (Semrock, FF552, R488), diverted into an inverted microscope (Motic, AE31) using a quad-pass dichroic mirror (Semrock, Di01-R405/488/532/635) and focused onto the microfluidic chip using a 20× objective. Photons from fluorescent droplets were collected by the same objective, filtered using a series of dichroic mirrors and bandpass filters (GFP channel FF552 and FF01-517/20; RFP channel FF593 and FF01-572/28; Cy5 channel FF01-673/11), and detected with photomultiplier tubes (PMTs) (Thorlabs, PMM01, PMM02). An embedded controller (National Instruments, cRIO-9056) programmed in LabView FPGA detected droplet peaks, integrated time series fluorescence data, and triggered on-chip sorting with a high-voltage amplifier (Trek, 609E-6). Gates, sorting parameters, and data acquisition occurred on a host computer graphical user interface built using LabView. Thermocycled droplets were re-injected into the concentric sorter³⁰ with the following flow rates: 100–150 μl h⁻¹ for drops, 400 μl h⁻¹ for drop spacing oil (Bio-Rad, 1863006), 2,000 μl h⁻¹ for HFE used as additional spacer between drops, 3,000 μl h⁻¹ for bias oil. The sorting parameters were 20 cycles of 1,000 V at 4,000 Hz. Drop sorting was verified by inspecting slow-motion videos taken of each sort using a high-speed camera (Vision Research, Miro110). For bulk FIND-seq, drops were collected in aliquots of 100 in 1.5 ml microcentrifuge tubes.

Validation of transcriptome amplification by FIND-seq.

Before amplifying sorted samples, the correct number was determined to ensure that drops were not over- or under-cycled during WTA. We sorted 100 genomes based on the percent of agarose hydrogels that contained cells, determined using fluorescence microscopy as described above. A 50 μl aqueous overlay of distilled nuclease-free water was added before spinning the tubes at 20,000g for 5 min. Tubes were then frozen at −80 °C for at least 2 h and 100 μl PCR reactions were set up with a final concentration of 1× KAPA HiFi master mix (Roche, KK2601) and 0.4 μM Smart-seq2 primer (AAGCAGTGGTATCAACGCAGAGT). For 100-aliquot sorts, we amplified triplicate aliquots of 100-cell sorts using the following PCR conditions: 95 °C for 3 min, N cycles of (98 °C for 15 s, 67 °C for 20 s, and 68 °C for 4 min), followed by 72 °C for 5-minutes, then held at 4 °C, where N varied as 18, 20, or 22 for primary mouse CNS cells to identify the optimal cycle number. Libraries were purified using AMPure XP magnetic beads (Beckman Coulter, A63881) at 1.2× ratio according to the manufacturer’s protocol, then eluted in 20 μl nuclease-free water, and run on a Bioanalyzer 2100 high sensitivity DNA chip (Agilent Technologies, 5067-4626) to assess size and yield. The minimum cycle number that resulted in a concentration of greater than 2 ng μl⁻¹ cDNA was then performed on the sorted samples.

Whole-transcriptome processing by FIND-seq.

After these validation steps, sorted samples were processed as above: 100-drop sorts were overlayed with 50 μl of distilled nuclease-free water, centrifuged at 20,000g for 5 min, frozen at −80 °C for 2 h, heated to 60 °C for 10 min, mixed gently by pipette and PCR amplified using the appropriate cycle number that was identified on cells from the same batch. After amplification, cDNA yield was quantified using a Qubit 4 fluorometer (Thermo Fisher) and DNA size distribution was assayed using a Bioanalyzer 2100 high sensitivity DNA chip (Agilent Technologies, 5067-4626). Libraries were prepared using the Nextera XT library preparation kit with v2 indices (Illumina, FC-131-1096) according to the manufacturer’s instructions. Individual sample libraries were pooled in equimolar amounts to produce a single library. The library was quantified using KAPA fast universal qPCR kit (Kapa Biosystems, KK4824). Library concentration was validated by Bioanalyzer. The library was diluted and denatured in accordance with standard Illumina protocols. The library was sequenced by paired-end 75-bp reads on one lane of NovaSeq S2.

Bulk FIND-seq analysis.

The reads from the raw fastq files were trimmed using Trim Galore⁵⁴ in order to remove adapter contents and low-quality bases. The trimmed reads were aligned to GRCm38 mouse genome using STAR (v2.7.3)⁵⁵ with the default parameters for paired-end RNA-seq samples; the alignment process generated transcriptome-level alignment bam files. The alignment results were used by the quantification software RSEM⁵⁶ to yield transcript-level quantification, and tximport⁵⁷ was used to aggregate the gene-level quantification from the transcript-level quantification into a count matrix. FIND-seq detected an average of 17,566 genes per 100 cells. DESeq2⁵⁸ was used to complete the differential expression analysis, and ApeGLM⁵⁹ was used to shrink the log2 fold change results to remove noise from the differential expression analysis. The results of differential expression analysis were analysed using GSEA⁶⁰ and IPA⁶¹. To search for DNA binding motifs in the upstream elements of genes, the upstream 2000 base pair sequences were extracted from the transcriptional start site of each gene using UCSC Genome Browser. Then the sequences were analysed using the FIMO algorithm⁶² from MEME suite⁶³. The motifs in the JASPAR database⁶⁴ were used in the analysis.

scFIND-seq.

A Smart-seq3 approach⁶⁵ was used for scFIND-seq. The protocol remained largely identical, with the following modifications. A Smart-seq3 bead-side mRNA capture and reverse transcription primer (/5Acryd/TTTTTTACGAGCATCAGCAGCATACGATTTTTTTTTTTTTTTTTTTTTTTTTTTTTTV) was conjugated onto allyl agarose as described above. Cell encapsulation and washing were identical, but a Smart-seq3 template switch oligonucleotide (/5Biosg/AGAGACAGATTGCGCAATGNNNNNNNNrGrGrG) containing an 8-bp UMI was used during reverse transcription. Bead re-injection and ddPCR followed the 100-cell sort procedures described above. Thereafter, single-drop sorts were performed using an automatic xyz stage to accurately deposit single sorted drops into a 96-well plate. Drops were overlayed with 25 μl of nuclease-free water, and frozen at −80 °C overnight. The following day, cells were defrosted for 10 min at 60 °C (Eppendorf, Thermomixer C), and centrifuged for 5 min at 1,000g. Whole-transcriptome cDNA amplification was performed in a 50 μl reaction for single cells using KAPA HiFi master mix (Roche, KK2601) and 1 μM Smart-seq3 forward primer (TCGTCGGCAGCGTCAGATGTGTATAAGAGACAGATTGCGCAA*T*G) and 0.5 μM Smart-seq3 reverse primer (ACGAGCATCAGCAGCATAC*G*A). Reactions were thermocycled at 95 °C for 3 min, then 26 cycles of (98 °C for 15 s, 65 °C for 20 s, and 68 °C for 4 min), followed by 72 °C for 5 min, and held at 4 °C. PCR reactions were purified using a 2× ratio of AMPure XP beads, followed by a second cleanup using 0.7× AMPure XP beads followed by elution in 10 μl of nuclease-free water. The size of the amplified cDNA was randomly sampled from individual wells of the 96-well plate by Bioanalyzer (Agilent Technologies, 5067-4626). Libraries were prepared using the Nextera XT DNA library preparation kit for 96 samples (Illumina, FC-131-1096) with 10-bp dual indices. Reaction volume was scaled down tenfold and automated with robotics (SPTLabtech Dragonfly). Individual sample libraries were pooled, purified by AMPure XP beads at a 0.9× ration, and sequenced on two separate NextSeq runs with 150-bp paired-end reads.

scFIND-seq analysis.

Kallisto⁶⁶ was used to align the reads from scFIND-seq by specifying the technology as Smartseq3 using ‘–technology SmartSeq3’. The 8-bp UMI in read 1 was used to assign unique transcript molecules to individual cells. The alignment results were quantified using bustools⁶⁷ by default parameters, and the quantification results were analysed in Seurat⁶⁸ as follows. The top 2,000 variable genes were used to conduct the principal component analysis reduction of the dataset, and the top 25 PCs were used to perform the nearest neighbourhood construction and UMAP reduction. The Leiden community clustering algorithm was used to cluster the dataset using resolution of 1.4. scFIND-seq had an average of 520 genes per cell, and 5,749 transcripts per cell.

Cell line validation studies

The JLat full-length human T cell line (6.3, ARP-9846) or Jurkat human T cells (Jurkat Clone E6-1, ATCC, TIB-152) were cultured in Gibco RPMI Media 1640 (Thermo Fisher, 11875093) with penicillin and streptomycin (Thermo Fisher, 15140122) and 10% fetal bovine serum (FBS). Mouse fibroblast cells (NIH/3T3, ATCC, CRL-1658) were cultured in DMEM with penicillin and streptomycin (Thermo Fisher, 15140122) and 10% FBS. 3T3 cells were dissociated using 0.25% trypsin-EDTA (Thermo Fisher, 25200-072) and neutralized in DMEM with 10% FBS. Cells were washed twice with HBSS, no calcium, no magnesium (Thermo Fisher, 14170112) and counted.

Mouse–human mixing studies.

Cells were mixed at a 1:100 JLat:3T3 ratio and resuspended in HBSS containing 18% OptiPrep density gradient medium (Sigma-Aldrich) for microfluidic encapsulation. Mouse–human mixing studies used agarose conjugated with acrydite-T5-SMART-T30VN and the Smart-seq2 TSO (AAGCAGTGGTATCAACGCAGAGTACATrGrGrG) for reverse transcription. The digital droplet genomic PCR was performed using 900 nM HIV gag forward primer (CACTGTGTTTAGCATGGTGTTT), 900 nM HIV gag reverse primer (TCAGCCCAGAAGTAATACCCATGT) and 250 nM HIV gag TaqMan probe (CY5-ATTATCAGAAGGAGCCACCCCACAAGA-3′ Iowa Black RQ). Drops were thermocycled under the following conditions: 88 °C for 10 min, then 55 cycles of (88 °C for 15 s and 60 °C for 1 min) and sorted into single wells. PCR strip tubes were frozen at −80 °C for at least 2 h and 50 μl WTA reactions were set up with a final concentration of 1× KAPA HiFi master mix (Roche, KK2601) and 0.4 μM Smart-seq2 primer (AAGCAGTGGTATCAACGCAGAGT). Library preparation was performed as described in the scFIND-seq section.

Validation of detection of a single genomic target in single cells.

JLat (contains a near full-length HIV provirus sequence) and Jurkat (HIV-negative) human T cells were independently cultured, washed, and encapsulated. FIND-seq was completed as described, with the following changes: No reverse transcription was performed to ensure that only a single copy of the HIV genomic target was present in the JLat cell line; detection PCR was performed as described above, using HIV gag-specific droplet TaqMan PCR. Quantification of positive droplets used a custom microfluidic droplet cytometer.

Generation of lentivirus

Constructs were generated using the pLenti-U6-sgScrmbl-Gfap-Cas9-2A-EGFP-WPRE lentiviral backbone, described previously⁴. Cas9 was driven by expression of the ABC₁D gfa2 GFAP promoter⁶⁹. Two PCR products were generated using Phusion flash HF 2× master mix (Thermo Fisher, F548L) and the pLenti-U6-sgScrmbl-Gfap-Cas9-2A-EGFP-WPRE plasmid template. The U6-PCR-F and U6-PCR-R amplicon was purified (Qiagen, 28104) and digested with BsaI-HF (NEB, R3535/R3733), AscI (NEB, R0558) and DpnI (NEB, R0176S). The cr-RNA-F and cr-RNA-R amplicon, which included the sgRNA sequence, was purified and digested with BsaI-HF (NEB, R3535/R3733), SbfI-HF (NEB, R3642) and DpnI (NEB, R0176S). The backbone was digested with SbfI-HF (NEB, R3642) and AscI (NEB, R0558). All digestions were purified (Qiagen, 28104) and a three-way ligation performed overnight at 16 °C using T4 DNA Ligase (NEB, M0202L). Ligations were transformed into NEB Stable Cells (NEB, C3040), recovered at 30 °C for 1 h, and plated at 30 °C overnight. Six single colonies per sgRNA construct were picked, grown in SOB media, and purified (QIAprep Spin Miniprep, 27104). The genes targeted for knockdown using the lentivirus delivery system were: E2f1(GCAGCACGTCAGAATCGCGA), Rgs19(TCATGGCTGGACATCGAGGG), Etv5 (ACAGGACGACAACTCGGAGG), Bcl3(TGAGTAGGCAGGTTCAGCAG), Thrb(GCAGCGATAGTGGTACCCTG), Ss18(GCTCACCTTCTGGATGGCGG), Nr3c2(CAAGGAACTTTCAGCCACGG), Hira(CATGTGTGAACTGTGTGCGG), and Angpt1 (CGCCGAAATCCAGAAAACGG). Lentiviral plasmids were co-transfected into HEK293FT cells with helper plasmids (pLP1, pLP2, and pLP/VSVG) according to the manufacturer’s protocol (Thermo Fisher Scientific, K497500), concentrated overnight at 4 °C (Clontech, 631231) and resuspended in 150 μl of nuclease-free water.

In vivo CRISPR transduction

Intracerebroventricular injection was used to deliver concentrated lentivirus to mice as previously described⁴. Mice were anaesthetized with 1–3% isoflurane in oxygen. Heads were shaved and cleaned with ethanol and betadine (Thermo Fisher, 19-027132) and a small incision was made in the skin. Two injections of 10 μl were made bilaterally with a 25 μl Hamilton syringe (Sigma-Aldrich, 20787) at ±1.0 (lateral), −0.44 (posterior), −2.2 (ventral) relative to Bregma using a stereotaxic alignment system (Kopf, 1900). Mice were sutured and allowed to recover in a clean cage on a heating pad for 1 h. Mice were further recovered for 6 days before the induction of EAE.

Experimental autoimmune encephalomyelitis

Induction of EAE was performed as previously described⁴. In brief, 150 μg MOG_35–55 (Genemed Synthesis, 110582) per mouse was prepared from a 10 mg ml⁻¹ stock diluted in 1× PBS. Complete Freund’s adjuvant (CFA) was prepared by mixing 100 mg Mycobacterium tuberculosis H-37Ra (BD Biosciences, 231141) with 20 ml Incomplete Freund’s adjuvant (BD Biosciences, BD263910). An emulsion was prepared by mixing MOG and CFA at 1:1 v/v using 20 ml Hamilton syringes until the emulsion was a thick, milky white solution. The MOG/CFA was delivered to mice with two subcutaneous injections of 100 μl each. Pertussis toxin (List Biological Laboratories, 180) was delivered with an intraperitoneal injection at a concentration of 2 ng μl⁻¹ in 200 μl of PBS. EAE clinical scores were defined as follows: 0, no signs; 1, fully limp tail; 2, hindlimb weakness; 3, hindlimb paralysis; 4, forelimb paralysis; 5, moribund–as described previously^4,19–22,38.

Bulk RNA-seq of sorted astrocytes

Bulk RNA isolated from flow cytometry-sorted cells was processed using the PicoPure kit (Thermo Fisher, KIT0204) and used as input with the kit (NEB, E6420) according to the manufacturer’s protocol. Reverse transcription was performed according to the Smart protocol using a template switching oligo. Then, cDNA was amplified and cleaned using Ampure XP beads (Beckman Coulter, A63881) and quantified using a Bioanalyzer DNA HS assay (Agilent, 50674626). Libraries were then fragmented, end-repaired, and ligated to Illumina compatible adapters followed by sample barcoding using NEBNext Multiplex Oligos for Illumina (E7335S, E7500S). Samples were selected again using Ampure XP beads and quantified using a Bioanalyzer. Libraries were quantified using a Kapa library quantification kit (Kapa Biosystems, KK4824) and run on an Illumina NextSeq550 as 1 × 75 bp reads with 6-bp index read and de-multiplexed into FASTQ files. Alternatively, 5 ng of RNA was resuspended 5 μl of nuclease-free H2O and processed using the Smart-seq2 workflow at Broad Technology Labs and the Broad Genomics Platform. Processed RNA-seq data was filtered, removing genes with low read counts. The fastq files of each RNA-seq data sample were aligned to the Mus musculus GRCm38 transcriptome using Kallisto (v0.46.1), and the same software was used to quantify the alignment results. Differential expression analysis was conducted using DESeq2, and the log2 fold change was adjusted using apeGLM for downstream analysis.

Pathway analysis

Genes ranked by fold change were analysed using GSEA pre-ranked analyses⁶⁰. GSEA analysis used the KEGG/Reactome/Biocarta (c2.cp.all), Gene ontology (c5.cp.all), and Hallmark (h.all) gene sets from Molecular Signatures Database v7.1. ENRICHR was used on lists of differentially expressed genes^70,71. Upstream regulator analysis was completed using IPA (Qiagen)⁶¹. Transcriptional signatures were developed based on several publicly available gene sets: ‘GOBP response to mineralocorticoid’, ‘GOBP cellular response to mineralocorticoid stimulus’ and ‘Jepsen smrt targets’.

Primary astrocyte cultures and gene expression analysis

Procedures were performed largely as described previously^4,21,22,38. Brains of mice aged P0–P3 were dissected into PBS on ice. Brains were pooled, centrifuged at 500g for 10 min at 4 °C and resuspended in 0.25% Trypsin-EDTA (Thermo Fisher Scientific, 25200-072) at 37 °C for 10 min. DNase I (Thermo Fisher Scientific, 90083) was then added to the solution, and the brains were digested for 10 more minutes at 37 °C. Trypsin was neutralized by adding DMEM/F12 + GlutaMAX (Thermo Fisher Scientific, 10565018) supplemented with 10% FBS (Thermo Fisher Scientific, 10438026) and 1% penicillin/streptomycin (Thermo Fisher Scientific, 15140148), and cells were passed through a 70-μm cell strainer. Cells were centrifuged at 500g for 10 min at 4 °C, resuspended in DMEM/F12+GlutaMAX with 10% FBS/1% penicillin/streptomycin and cultured in T-75 flasks (Falcon, 353136) at 37 °C in a humidified incubator with 5% CO₂, for 7–10 days until confluency was reached. Astrocytes were shaken for 30 min at 180 rpm, the supernatant was aspirated and the media was changed, then astrocytes were shaken for at least 2 h at 220 rpm and the supernatant was aspirated and the media was changed again. Medium was replaced every 2–3 days. Primary astrocytes were lysed in Buffer RLT (Qiagen) and RNA was isolated from cultured astrocytes using the RNeasy Mini kit (Qiagen, 74106). cDNA was transcribed using the High-Capacity cDNA Reverse Transcription Kit (Life Technologies, 4368813). Gene expression was then measured by qPCR using TaqMan fast universal PCR master mix (Life Technologies, 4367846). TaqMan probes used in this study are: Actb (Mm02619580_g1), Il1b (Mm00434228_m1), Ccl2 (Mm00441242_m1), Il6 (Mm00446190_ m1), and Xbp1 (Mm03464496_m1, Mm00457357_m1). qPCR data were analysed by the ddCt method by normalizing the expression of each gene for each replicate to the Actb housekeeping gene and then to the control group. Compound treatment was performed for 24 or 4 h with compounds diluted in DMEM/F12+GlutaMAX (Life Technologies, 10565042) that was supplemented with 10% FBS (Life Technologies, 10438026) and 1% penicillin/streptomycin (Life Technologies, 15140122). Compounds used in these studies are: 1 ng ml⁻¹ IL-1β (R&D Systems, 401-ML-005, 100 μg ml⁻¹ stock in PBS), 1 ng ml⁻¹ TNF (R&D Systems, 410-MT-010, 100 μg ml⁻¹ stock in PBS), and 1 μM finerenone (MedChem Express, HY-111372, 1mM stock in DMSO).

Isolation of cells from adult mouse CNS

Astrocytes were isolated by flow cytometry as described^4,21,22,38. In brief, mice were perfused with 1× PBS and the CNS was isolated into 10 ml of enzyme digestion solution consisting of 75 μl papain suspension (Worthington, LS003126) diluted in enzyme stock solution (ESS) and equilibrated to 37 °C. ESS consisted of 10 ml 10× EBSS (Sigma-Aldrich, E7510), 2.4 ml 30% D(+)-Glucose (Sigma-Aldrich, G8769), 5.2 ml 1 M NaHCO₃ (VWR, AAJ62495-AP), 200 μl 500 mM EDTA (Thermo Fisher Scientific, 15575020), and 168.2 ml ddH2O, filter-sterilized through a 0.22 μm filter. Samples were shaken at 80 rpm for 30–40 min at 37 °C. Enzymatic digestion was stopped with 1 ml of 10× hi ovomucoid inhibitor solution and 20 μl 0.4% DNase (Worthington, LS002007) diluted in 10 ml inhibitor stock solution (ISS). 10× hi ovomucoid inhibitor stock solution contained 300 mg BSA (Sigma-Aldrich, A8806), 300 mg ovomucoid trypsin inhibitor (Worthington, LS003086) diluted in 10 ml 1× PBS and filter-sterilized using at 0.22 μm filter. ISS contained 50 ml 10× EBSS (Sigma-Aldrich, E7510), 6 ml 30% D(+)-Glucose (Sigma-Aldrich, G8769), 13 ml 1 M NaHCO₃ (VWR, AAJ62495-AP) diluted in 170.4 ml ddH2O and filter-sterilized through a 0.22 μm filter. Tissue was mechanically dissociated using a 5 ml serological pipette and filtered through a 70 μm cell strainer (Fisher Scientific, 22363548) into a fresh 50 ml conical. Tissue was centrifuged at 500g for 5 min and resuspended in 10 ml of 30% Percoll solution (9 ml Percoll (GE Healthcare Biosciences, 17-5445-01), 3 ml 10× PBS, 18 ml ddH2O). Percoll suspension was centrifuged at 500g for 25 min with no brakes. Supernatant was discarded and the cell pellet was washed 1× with 1× PBS, centrifuged at 500g for 5 min and prepared for downstream applications.

Flow cytometry

Cells were stained in the dark on ice for 15 min with flow cytometry antibodies. Cells were then washed once with 1× PBS and resuspended in 1× PBS for sorting as described previously^4,21,22,38. Antibodies used in this study were: PE anti-mouse CD45R/B220 (BD Biosciences, 553089, 1:100), PE anti-mouse TER-119 (Biolegend, 116207, 1:100), PE anti-O4 (R&D Systems, FAB1326P, 1:100), PE anti-CD105 (eBioscience, 12-1051-82, 1:100), PE anti-CD140a (eBioscience, 12-1401-81, 1:100), PE anti-Ly-6G (Biolegend, 127608, 1:100), PerCP anti-Ly-6C (Biolegend, 128028, 1:100), APC anti-CD45 (eBioscience, 17-0451-83, 1:100), APC-Cy7 anti-CD11c (BD Biosciences, 561241, 1:100), and FITC anti-CD11b (eBioscience, 11-0112-85, 1:100). All cells were gated on the following parameters: CD105^negCD140a^negO4^negTer119^negLy-6G^negCD45R^neg. Astrocytes were subsequently gated on: CD11b^negCD45^negLy-6C^negCD11c^neg. Microglia were subsequently gated on: CD11b^highCD45^lowLy-6C^low. Pro-inflammatory monocytes were subsequently gated on: CD11b^highCD45^highLy-6C^high. Compensation was performed on single-stained samples of cells and an unstained control. Cells were sorted on a FACS Aria IIu (BD Biosciences). Sorted astrocytes were subsequently stained for intracellular XBP1S using the antibody PE Mouse anti-XBP-1S (BD, 562642). For AQP4 sorting experiments, isolated CNS cells were stained with AmCyan dye (Thermo Fisher, L34957, 1:2,000) for 30 min on ice, washed 1× with 1× PBS, then stained with anti-AQP4-PE (Bioss, BS-0634R-PE, 1:100) for 20 min on ice. Thereafter, cells were washed with 1× PBS and subjected to flow cytometry. For FIND-seq analysis ofAqp4⁺ cells, bead loading was estimated at 70% to determine number of positive cells within identified droplets.

T cell FACS analysis

To analyse T cell populations, CNS and splenic cell suspensions were stimulated with 50 ng ml⁻¹ phorbol 12-myristate 13-acetate (PMA, Sigma-Aldrich, P8139), 1 μM ionomycin (Sigma-Aldrich, I3909-1ML), GolgiStop (BD Biosciences, 554724, 1:1500) and GolgiPlug (BD Biosciences, 555029, 1:1500) diluted in T cell culture medium (RPMI (Life Technologies, 11875119) containing 10% FBS, 1% penicillin/streptomycin, 50 μM 2-metcaptoethanol (Sigma-Aldrich, M6250), and 1% non-essential amino acids (Life Technologies, 11140050)). After 4 h, cell suspensions were washed with 0.5% BSA, 2 mM EDTA in 1× PBS and incubated with surface antibodies and a live/dead cell marker on ice. After 30 min, cells were washed with 0.5% BSA, 2 mM EDTA in 1× PBS and fixed according to the manufacturer’s protocol of an intracellular labeling kit (eBiosciences, 00-5523-00). Surface antibodies used in this study were: BUV661 anti-mouse CD45 (BD Biosciences, 565079, 1:100), BV750 anti-mouse CD3 (Biolegend, 100249, 1:50), PE-Cy7 anti-mouse CD4 (eBioscience, 25-0041-82, 1:100), BUV737 anti-mouse CD11b (BD Biosciences, 564443, 1:100), BV570 anti-mouse Ly6C (Biolegend, 128030, 1:100), BUV805 anti-mouse CD8a (BD Bioscience, 564920, 1:100), BUV563 anti-mouse Ly6G (BD Biosciences, 565707, 1:100), BUV737 anti-mouse CD11c (BD Biosciences, 612797, 1:100), Pe/Cy5 anti-mouse CD44 (BioLegend, 103010, 1:100). Intracellular antibodies were: APC anti-mouse IFN-γ (BD Biosciences, 554413, 1:100), PE anti-mouse IL-17A (eBiosciences, 12-7177-81, 1:100), BV421 anti-mouse GM-CSF (BD Biosciences, 564747, 1:100), PE/Dazzle 594 anti-mouse IL-10 (BioLegend 505033), and FITC anti-mouse FoxP3 (eBiosciences, 11-5773-82, 1:100). Gating of CNS and splenic cells was performed on >50,000 live CD4⁺ cells. FACs was performed on a Symphony A5 (BD Biosciences).

Analysis of human brain tissue by immunofluorescence

Use of tissue from patients with multiple sclerosis was approved by the Neuroimmunology Research Laboratory, Centre de Recherche du Centre Hospitalier de l’Université de Montréal (CRCHUM) under ethical approval number BH07.001. Human brain tissue was obtained from patients with clinical and neuropathological multiple sclerosis diagnosis according to the revised 2010 McDonald’s criteria⁷². Tissue samples were collected from healthy donors and patients with multiple sclerosis with full ethical approval (BH07.001) and informed consent as approved by the local ethics committee. Flash frozen 7 μm tissue sections were fixed in 4% PFA for 10 min at room temperature. Next, sections were washed 3× with 1× PBS, followed by three washes with 0.3% PBS-T. Sections were blocked with 5% donkey serum (D9663, Sigma-Aldrich) in 0.3% PBS-T at room temperature for 30 min. Sections were then incubated with primary antibody (mouse anti-NR3C2, Abcam, ab2774, Clone: H10E4C9F, 1:100) diluted in blocking buffer overnight at 4 °C. Following primary antibody incubation, sections were washed 3× with 0.3% PBS-T and incubated with secondary antibody (Alexa Fluor 488 AffiniPure Fab Fragment Goat Anti-Mouse IgG (H+L), Jackson ImmunoResearch, 115-547-003, 1:500) diluted in blocking buffer for 2 h at room temperature. Sections were then washed 6× with 0.3% PBS-T and stained with pre-conjugated GFAP-CY3 (Sigma-Aldrich, C9205, 1:100). Sections were then dried and mounted using Fluoromount-G with DAPI (SouthernBiotech, 0100-20) and imaged using a Zeiss LSM710 confocal microscope. Staining was quantified by first selecting DAPI⁺GFAP⁺ cells then quantifying the number of those cells expressing NR3C2.

Analysis of mouse CNS tissue by immunostaining

Mice were intracardially perfused with ice cold 1× PBS and CNS tissue was excised from the mouse then post-fixed in 4% PFA overnight at 4 °C, then the tissue was dehydrated with 30% sucrose for one week at 4 °C. If whole spinal columns were collected containing bone, they were de-calcified in 20% EDTA (pH = 7.4) for one week at 4 °C with inversion prior to sucrose dehydration. Tissue was frozen in OCT (Sakura, 4583) and 20 μm sections were prepared by cryostat on SuperFrost Plus slides (Fisher Scientific, 15-188-48). Sections were permeabilized with 1× permeabilization buffer (BD Biosciences, 554723) for 10 min at room temperature, then blocked using serum-free protein block (Agilent, X0909) for 10 min at room temperature. Sections were then incubated with primary antibodies diluted in 1× permeabilization buffer overnight at 4 °C. Following primary antibody incubation, sections were washed 3× with 1× permeabilization buffer and incubated with secondary antibodies diluted in 1× permeabilization buffer for 2 h at room temperature. After secondary and conjugated antibody incubation, sections were stained with 1 μg ml⁻¹ DAPI (Sigma-Aldrich, D9542) diluted in 1× permeabilization buffer for 5 min at room temperature, then washed 3× with 1× permeabilization buffer and mounted with ProLong Gold Antifade Mountant (Fisher Scientific, P36930) or alternatively, Fluoromount-G with DAPI (SouthernBiotech, 0100-20). Primary antibodies used in this study were: rabbit anti-XBP1S (Biolegend, 619502, 1:50), mouse anti-NR3C2, (Abcam, ab2774, clone: H10E4C9F, 1:100), rabbit anti-NCOR2 (Abcam, ab5802, 1:100), chicken anti-GFAP (Abcam, ab4674, 1:100), or mouse anti-GFAP (Millipore, MAB360, 1:500). Secondary antibodies used in this study were: Rhodamine Red-X-AffiniPure Fab Fragment donkey anti-mouse IgG (H+L) (715-297-003, Jackson ImmunoResearch, 2 mg ml⁻¹ stock), Goat anti-Chicken IgY (H+L) Secondary Antibody Alexa Fluor 647 (Jackson Immunoresearch, 703-475-155), Goat anti-Chicken IgY (H+L) Alexa Fluor 488 (Life Technologies, A11039), and Goat Anti-Rabbit IgG H&L Alexa Fluor 488 (Abcam, ab150077) all at 1:500 working dilution. Sections were imaged on an LSM-880-AiryScan confocal microscope (Zeiss) or LSM710 confocal microscope (Zeiss) and image analysis was performed with FIJI⁷³ first by selecting GFAP⁺DAPI⁺ cells and then quantifying the proportion positive for a given marker protein.

Statistical analysis

Statistical analyses were performed with Prism software (GraphPad, v9.4.0). Data plotted in Extended Data Table 1 were generated in Excel (Microsoft, v16.64). All data represented shown are represented as mean ± s.e.m., unless indicated otherwise. The specific statistical test used for each comparison is indicated in the respective panel in the figure legend and the exact P values are indicated in each panel. In brief, for the comparison of two datasets, an unpaired two-sided t-test was used. To assess changes in a dependent variable across a range of conditions a one-way ANOVA was used. For analyses involving two variables, such as EAE time courses based on genotype, a two-way repeated measures ANOVA was used. Data were tested for normality by the D’Agostino and Pearson test, but if the number of replicates per experiment was insufficient to test for normality, a Gaussian distribution was assumed. Testing for activation of pathways was determined by z-score measured using IPA, which is a metric that represents predicted relative activity of the indicated pathway in an input gene list. The z-score for each pathway is specifically calculated using a list of curated genes assigned to each pathway; a simplified z-score is calculated as follows: $z=\frac{N^{+}-N^{-}}{\sqrt{N}}$, where N is the total number of genes and N⁺ and N⁻ are the number of genes where the direction of regulation is concordant or discordant, respectively, with predictions from the literature within the curated pathway. A positive z-score suggests pathway activation while a negative z-score suggests pathway inhibition. Statistical significance of the enrichment of a pathway or control by an upstream regulator was determined using a Fisher’s exact test based on overlap between the curated gene list and the input gene list as previously described⁶¹.

Extended Data

Extended Data Fig. 1 |. Analysis of rare cells by FIND-seq.

(a) Featureplots of Aqp4 and Edem1 expression in cells isolated from the mouse CNS during EAE reanalyzed from ref.²². (b) Violin plots of Aqp4 and Edem1 by cluster from astrocytes isolated from the EAE CNS in a dataset re-analyzed from ref. ²². (c) Micrograph images of agarose beads containing captured nucleic acids from encapsulated cells. (d) Droplet cytometry plots showing an estimate of the number of cells captured (>1 million) in the bead volume shown in (c). (e) Negative control of droplet cytometry SYBR fluorescence.

Extended Data Fig. 2 |. Droplet formation and capture of mRNA by FIND-seq.

(a) Schematic of bubble trigger device that drives air-triggered droplet formation during cell encapsulation. (b) Air-triggered droplet formation enables kHz generation of beads from viscous molten agarose. Images from time-lapse videos of agarose jet breakup. (c) Functionalization of agarose with allyl groups is used to directly link acrydited primers. (d) Microscope images demonstrating successful conjugation of polyT primers to agarose with polyA-FAM probes. (e) Quantification of agarose-bound polyT.

Extended Data Fig. 3 |. Optimization of first-strand synthesis of agarose captured mRNA.

(a) Steric capture of cellular genomic DNA inside the agarose matrix. Left: brightfield image of agarose gels. Right: SYBR green fluorescence of stained genomes inside agarose gels. Cells are loaded at limiting dilution to ensure single-cell encapsulation; approximately one in every ten agarose hydrogels contains a cell. (b) Estimated number of cells per drop based on Poisson statistics for microfluidic loading during FIND-seq. (c) Quantification of live cells by flow cytometry using AmCyan live/dead cell dye. n = 4 mice. (d) Whole transcriptome amplification (WTA) of cDNA covalently attached to agarose beads shows full length material is captured and reverse transcribed. (e) WTA yield as a function of PCR cycle number. (f) Optimization of cDNA capture with buffer composition, enzyme, template switch oligonucleotide concentration and additives (6 mM Mg²⁺, 1M betaine, and 7.5% PEG-8000). (g) Quantification of the percent mitochondrial reads in bulk FIND-seq data for each replicate. n = 22 samples. (h) Calculation of score per cell from astrocytes derived from scRNA-seq from ref. ²² or each bulk FIND-seq replicate for the pathway GOBP: Execution phase of apoptosis (GO: 0097194).

Extended Data Fig. 4 |. Agarose bead re-injection and sorting for single-cell detection of nucleic acid markers.

(a) Schematic of microfluidic for re-injection of agarose-captured genomes. (b) Microscope image of single agarose hydrogel beads inside droplets with measured size for each droplet and agarose sphere. (c) Optimization of droplet detection PCR to preserve cDNA quality during thermocycling. (d) Detection of a single-copy HIV genomic DNA target by FIND-seq in infected human JLat cells, but not in uninfected Jurkat control cells, as a proof-of-concept experiment testing FIND-seq sensitivity and specificity. The DNA target was amplified using TaqMan PCR in the FIND-seq workflow (Step 3 in Fig. 1d) followed by detection by droplet cytometry (Step 4 in Fig. 1f). (e) Schematic of microfluidic for droplet sorting using a concentric dielectrophoretic design. (f) Micrographs of droplet sorting. Top: time-lapse images from a droplet sorting video showing droplet deflection into the collection channel. Bottom: In the absence of FPGA sort-triggering, droplets are maintained, via bias oil flow, in the outer waste channel.

Extended Data Fig. 5 |. Benchmarking FIND-seq as a technology.

(a) Quantification of percent aquaporin-4-expressing cells in naïve mice by antibody-based flow cytometry (n = 4); FIND-seq (n = 3), or scRNA-seq reanalyzed from ref. ²². (b) Validation of FIND-seq specificity and sensitivity through the analysis of AQP4+ and AQP4− cells isolated by flow cytometry and subjected to FIND-seq. Percentages show number of cells expressing Aqp4 in each population based on 70% bead loading. (c) PCA plot of bulk FIND-seq analysis of Aqp4+Edem1− cells. (d) Comparison of FIND-seq detection sensitivity with comparable technologies. (e) Correlation of raw expression counts per gene between bulk FIND-seq-sorted Aqp4+Edem1+ cells and Edem1+ astrocytes extracted from droplet-based scRNA-seq data that we previously reported in²². (f) Quantification of fraction of duplicate reads across bulk RNA-seq platforms. (g) Quantification of Ern1 in bulk FIND-seq data as a function of EAE and Edem1 expression. (h) Calculation of a signature score for transcripts enriched in astrocyte endfeet as reported by ref. ⁵¹, analyzed in astrocytes from ref. ²² and bulk FIND-seq.

Extended Data Fig. 6 |. Single-cell analysis by FIND-seq.

(a) Visualization of a single droplet in flat-bottom microwell plates. Refraction of light at well edges obscures imaging. This is solved by sorting directly into hexadecane. HFE oil sinks, forming a convex shape that forces droplets to the center so that they can be imaged. (b) The percentage of reads mapping to mouse or human cells after single-cell sorting cell mixtures. Mouse (3T3) and human (JLat) cells were mixed 1:100, sorted based on a TaqMan PCR targeting JLat cells, and the transcriptome was sequenced. (c) Quality control analyses for scFIND-seq. (d) Marker genes of Aqp4+Edem1+/− cells from naïve or EAE mice analyzed by scFIND-seq. (e) Elbow plot of principal components detected by scFIND-seq.

Extended Data Fig. 7 |. In vivo screening of FIND-seq-identified candidate regulators of XBP1+ astrocytes.

(a) Predicted upstream regulator analysis showing Nr3c2 and Xbp1 from bulk FIND-seq data using Qiagen IPA. Differentially expressed genes were used as input and the overlap with the regulon controlled by each molecule was computed. Fisher’s exact test. (b) Left: Prediction of Nr3c2 as an upstream regulator in Aqp4+Edem1+ cells analyzed by FIND-seq during EAE using Qiagen IPA as in (a). Fisher’s exact test. Right: Identification of an NR3C2 motif by SeqPos in genes downregulated in Aqp4+Edem1+ versus Aqp4+Edem1− cells in EAE. (c) UMAP plot of Aqp4+Edem1+ cells from EAE mice analyzed by scFIND-seq. (d–e) Prediction of upstream regulators (d) and pathway analysis (e) based on Qiagen IPA in Cluster 1 astrocytes analyzed from Aqp4+Edem1+ cells in EAE mice shown in (c). Fisher’s exact test. (f) Schematic of lentiviral vector containing a sgRNA targeting candidate genes and spCas9 under the control of a Gfap promoter. EAE disease progression in mice transduced with sgScrmbl and candidate sgRNA lentiviruses. n = 4–5 mice per condition.

Extended Data Fig. 8 |. Control analyses of Nr3c2 and Ncor2 knockdown.

(a) Left: Upstream regulator analysis of RNA-seq data by Qiagen IPA from sgNr3c2-targeted versus sgScrmble-targeted mice shows NR3C2 downregulation. Fisher’s exact test. Right: Validation of NR3C2 knockdown by immunostaining quantification. n = 6 images per group from n = 3 mice. Unpaired two-sided t-test. (b–c) FACS analysis of total CD4+ cells or FoxP3+, IFN𝛾+, IL17+, and IL10+ CD4 cell subsets in the (b) the spleen and (c) CNS. (d) Validation of NCOR2 knockdown by immunostaining. n = 6 images per group from n = 3 mice. Unpaired two-sided t-test. (e–f) FACS analysis of total CD4+ cells or FoxP3+, IFNγ +, IL17+, and GM-CSF+ CD4 cell subsets in the (e) the spleen and (f) CNS.

Extended Data Fig. 9 |. Control analysis of cell subsets from Xbp1^WT and Xbp1^Astro mice.

(a) FACS analysis of astrocytes and microglia in the CNS. n = 3 per group. (b) Validation of XBP1 KO by immunostaining. n = 6 images from n = 3 mice per group. Unpaired two-sided t-test. (c–d) FACS analysis of total CD4+ cells or FoxP3+, IFN𝛾+, and GM-CSF+ CD4+ T cell subsets from the (c) the spleen and (d) CNS of Xbp1^WT and Xbp1^Astro mice. (e) Pathways analyzed by pre-ranked gene set enrichment analysis (GSEA) of genes from RNA-seq data comparing Xbp1^WT and Xbp1^Astro mice.

Extended Data Fig. 10 |. Analysis of chromatin accessibility in genes responsive to mineralocorticoid signaling as a function of XBP1 expression in astrocytes.

(a–b) Re-analysis of ATAC-seq data on bulk flow cytometry-sorted astrocytes that we reported in ref. ⁴, showing increased chromatin accessibility in NR3C2 responsive genes (a) and in Nr3c2 and Ncor2 (b) as a function of Gfap-specific shRNA-driven lentiviral knockdown of Xbp1.

Extended Data Table 1 |. FIND-seq practical considerations.

(a) Table of the molecular reagents required to perform a FIND-seq experiment on 1M cells itemized by cost, part number, and stage of the workflow. (a) Cost-benefit analysis of performing FIND-seq relative to 10X Genomics 3′ scRNA-seq. The comparison shown represents the cost associated with sequencing all target cells in a sample at a sequencing depth of 20k reads per cell. 10X Genomics requires sequencing every cell to find rare target cells whereas FIND-seq sorts and sequences only target cells.

Data availability

Genomic data have been deposited in the GEO database under accession number GSE198971. All other data are available from the corresponding authors on reasonable request. Correspondence and request for materials should be addressed to F.J.Q. or A.R.A. Source data are provided with this paper.