CRISPRi screens in human astrocytes elucidate regulators of distinct inflammatory reactive states

In response to central nervous system injury or disease, astrocytes become reactive, adopting context-dependent states and functional outputs. Certain inflammatory insults induce reactive astrocytes that lose homeostatic functions and gain harmful outputs through cellular pathways that are not fully understood. Here, we combined single-cell transcriptomics with CRISPRi screening in human iPSC-derived astrocytes to systematically interrogate inflammatory astrocyte reactivity. We found that autocrine-paracrine IL-6 and interferon signaling downstream of canonical NF-κB activation drove two distinct inflammatory reactive signatures – one promoted by and the other inhibited by STAT3. These signatures overlapped with those observed in other experimental contexts, including mouse models, and their markers were upregulated in the human brain in Alzheimer’s disease and hypoxic ischemic encephalopathy. Furthermore, we validated that these signatures were regulated by Stat3 in vivo. These results and the platform we established have the potential to guide the development of therapeutics to selectively modulate different aspects of inflammatory astrocyte reactivity.


INTRODUCTION
Astrocytes perform critical homeostatic functions in the central nervous system (CNS), such as providing trophic support for neurons, regulating the formation and function of synapses, and maintaining the blood-brain barrier 1 . In CNS injury or disease, astrocytes respond to pathophysiological perturbations by becoming "reactive", which is defined as the adoption of context-specific states with associated transcriptomic signatures and alterations in morphology and function 1, 2 .
Frequently, inflammatory processes play an important, if not central, role in the pathophysiology of CNS injuries, as seen in stroke 3 and trauma 4 , and diseases such as multiple sclerosis 5 and Alzheimer's disease 6 . As an immunocompetent CNS cell type, astrocytes actively participate in inflammatory signaling cascades and interact with both microglia, the CNS resident immune cells, as well as infiltrating immune cells from the periphery 6,7 . Proinflammatory cytokines, such as IL-1α, TNF, and C1q, induce reactive astrocytes that lose homeostatic functions while releasing factors that are potentially harmful in specific contexts [8][9][10] . From here on, we will refer to this form of astrocyte reactivity induced by IL-1α+TNF+C1q as belonging to the category of "inflammatory reactivity", which we will use as an umbrella term that captures a potential multitude of context-specific inflammatory reactive astrocyte signatures.
Given that inflammatory astrocyte reactivity has been implicated in numerous neurodegenerative and neuroinflammatory diseases 8,11 , in addition to being associated with normal aging 12 , it is an attractive target for therapeutic development. However, the cellular pathways that control inflammatory astrocyte reactivity are still not fully understood, partly due to limitations in experimental scalability. On the one hand, animal models have provided key insights into genes required for astrocyte reactivity 9,13,14 , but the throughput to test genetic perturbations in vivo is extremely limited. On the other hand, in vitro astrocyte culture is a powerful alternative, but the scalability of experiments is still limited by the necessity to isolate primary astrocytes or the long duration of the differentiation process (up to 6 months) required to generate mature human iPSC (hiPSC)-derived astrocytes 15 . Lastly, even though molecular profiling approaches such as RNAseq have been used extensively to identify cellular pathways altered in inflammatory reactive astrocytes 9,16 , these assays by themselves can only provide correlative information and cannot uncover causal pathways controlling inflammatory reactivity.
Here, we developed a scalable method to generate hiPSC-derived astrocytes that allowed us to harness the power of pooled CRISPRi screening 17 to systematically identify cellular pathways controlling inflammatory astrocyte reactivity induced by IL-1α+TNF+C1q. Following up on the top hits from CRISPRi screens with single-cell transcriptomics, we found that autocrineparacrine IL-6 and interferon signaling drove two distinct inflammatory reactive signatures -one promoted by and the other inhibited by STAT3. We found that the inflammatory reactive signatures we identified overlapped with those observed in other experimental contexts, both in vitro and in vivo. Furthermore, we validated the induction of these distinct inflammatory reactive signatures by lipopolysaccharide (LPS) in vivo and also their regulation by Stat3. Lastly, we performed immunostaining supporting the existence of these inflammatory reactive signatures in human brains in the context of Alzheimer's disease and hypoxic-ischemic encephalopathy.

iPSC-derived astrocytes (iAstrocytes) perform canonical astrocyte functions and sufficiently model inflammatory reactivity
To generate hiPSC-derived astrocytes in a scalable manner, we modified the protocol from TCW et al. 18 by inducing the expression of the gliogenic transcription factors NFIA and SOX9 during the differentiation process as previously published by Li et al. 19 (see Methods). Our protocol resulted in astrocytes, which we will here on refer to as "iAstrocytes", with increased expression of astrocyte markers ( Fig. 1a-b) as well as astrocyte-specific genes in general (Extended Data Fig. 1a-b) compared to astrocytes generated using the original protocol from TCW et al. 18 ("TCW astrocytes"). Compared to Li et al. 19 , our iAstrocyte protocol differs in the following aspects (see Methods): 1) using a splice isoform of NFIA that is more highly expressed in astrocytes (ENST00000403491 in this study vs. ENST00000371187 in Li et al. 19 ; Extended Data Fig. 1c, 2) ensuring the purity of neural progenitor cells through sorting for CD133+/CD271cells 20 , 3) culturing astrocyte precursors during the differentiation process in monolayer, which is more straightforward in terms of media changes and cell dissociation compared to spheroid culture in Li et al. 19 , and 4) using a commercially available astrocyte media (ScienCell Astrocyte Media) adapted from TCWet al.
We validated our iAstrocyte protocol in a total of three hiPSC lines derived from independent donors of both male and female sex (Extended Data Fig. 2a-b). To confirm that iAstrocytes were capable of performing typical astrocyte functions such as recycling glutamate 21 and phagocytosing synapses 22 , we measured the ability of iAstrocytes to uptake glutamate or phagocytose pHrodo-labeled synaptosomes. We found that iAstrocytes had higher glutamate uptake activity (Fig. 1c) and synaptosome phagocytic activity (Fig. 1d) compared to TCW astrocytes. Furthermore, to see if iAstrocytes were capable of performing other canonical astrocyte functions such as maintaining the blood-brain barrier 23 or promoting neuronal synapse maturation 24 , we co-cultured iAstrocytes with hiPSC-derived brain endothelial-like cells or neurons (see Methods). We found that brain endothelial-like cells co-cultured with iAstrocytes displayed improved barrier formation and integrity compared to mono-culture (Fig. 1e), demonstrating that iAstrocytes can promote the expected functional maturation of these cells. In addition, hiPSC-derived neurons (iNeurons) co-cultured with iAstrocytes developed synchronized calcium oscillations 25,26 compared to iNeurons in mono-culture ( Fig. 1f-h), demonstrating the ability of iAstrocytes to promote neuronal synapse maturation and consistent with the high expression of genes encoding synapse-promoting proteins such as GPC4 27 and CADM1 28 in iAstrocytes (Supplementary Table 1). Furthermore, we found that iAstrocytes cocultured with iNeurons acquired more stellate morphology compared to iAstrocytes in monoculture (Extended Data Fig. 1d).
Having validated that iAstrocytes performed typical astrocyte functions, we next tested whether iAstrocytes could be used to model inflammatory astrocyte reactivity. It has been shown that the cytokines IL-1α+TNF+C1q induce a form of inflammatory reactivity associated with neurotoxicity and loss of homeostatic functions in both primary mouse astrocytes 8 and human cerebral organoid-derived astrocytes 29 . We found that treatment of iAstrocytes generated from multiple hiPSC lines with IL-1α+TNF+C1q recapitulated previously reported in vitro phenotypes such as decreased phagocytosis of CNS substrates, decreased support of neuronal synapse maturation, and neurotoxicity 8,29 . Specifically, iAstrocytes displayed decreased phagocytosis of pHrodo-labeled synaptosomes after treatment with IL-1α+TNF+C1q ( Fig. 1i-j, Extended Data Fig. 2c). In addition, IL-1α+TNF+C1q treatment of iAstrocytes co-cultured with iNeurons abolished the development of synchronized neuronal calcium oscillations ( Fig. 1f-h), demonstrating decreased support of neuronal synapse maturation by iAstrocytes. Finally, we found that the conditioned media from IL-1α+TNF+C1q-treated iAstrocytes was toxic to iNeurons, whereas the conditioned media from vehicle control-treated iAstrocytes was modestly protective compared to unconditioned media (Fig. 1k, Extended Data Fig. 2e).
In addition to recapitulating key functional phenotypes of IL-1α+TNF+C1q-induced inflammatory reactivity, iAstrocytes also responded to IL-1α+TNF+C1q at the transcriptomic level in a similar manner as primary mouse astrocytes and human iPSC-derived astrocytes generated using alternative protocols (Extended Data Fig. 3a-e, Supplementary Table 1). We noted, however, that the differentially expressed genes (DEGs) induced by IL-1α+TNF+C1q in hiPSC-derived astrocytes were not restricted to the previously defined "pan-reactive" and "A1" reactive categories (Extended Data Fig. 3c), which were derived from transcriptomic characterization of astrocytes from LPS-treated mice 8,16 . This finding mirrored similar observations from Barbar et al. 29 . The discrepancy in behavior between hiPSC-derived astrocytes and mouse astrocytes could be due to differences in the experimental conditions or phylogenetic differences between human and mouse astrocytes 30 .
To find a cell-surface marker for inflammatory reactivity that is also functionally involved in neuroinflammation, we examined the DEGs induced by IL-1α+TNF+C1q in iAstrocytes. We found a dramatic increase in the transcript levels of VCAM1 (Fig. 1l), which encodes a cell-adhesion sialoglycoprotein known to facilitate the infiltration of peripheral immune cells into the central nervous system (CNS) during neuroinflammation 31,32 . We subsequently validated by flow cytometry that iAstrocytes generated from multiple hiPSC lines indeed upregulated cellsurface VCAM1 after treatment with IL-1α+TNF+C1q ( Fig. 1m-n, Extended Data Fig. 2d), corroborating previous reports in the literature demonstrating induction of VCAM1 in astrocytes under pro-inflammatory conditions [33][34][35] .

iAstrocytes differentiated from CRISPRi hiPSCs maintain robust knockdown activity
In order to investigate the effect of genetic perturbations on inflammatory reactivity in iAstrocytes, we generated iAstrocytes from hiPSCs stably expressing CRISPRi machinery (Fig.  2a), which enables both specific and non-toxic knockdown of genes targeted by single guide RNAs (sgRNAs) delivered to the cell and large-scale genetic screens 36 . We found that these iAstrocytes maintained robust CRISPRi activity, demonstrating ~100% knockdown of TFRC protein levels 6 days after lentiviral transduction with a single guide RNA (sgRNA) targeting TFRC compared to iAstrocytes transduced with a non-targeting control (NTC) sgRNA, regardless of whether the iAstrocytes were treated with IL-1α+TNF+C1q ( Fig. 2b-c).
For pooled CRISPRi screening, we used the same experimental timeline as described in Fig. 2b. We conducted screens based on two cell-autonomous phenotypes induced by IL-1α+TNF+C1q that can be analyzed by fluorescence activated cell sorting (FACS): decreased phagocytosis of pHrodo-labeled synaptosomes ( Fig. 1i-j) and upregulation of cell-surface of VCAM1 ( Fig. 1mn). For synaptosome phagocytosis, we performed the screens in both vehicle control-treated and IL-1α+TNF+C1q-treated iAstrocytes to find genes whose knockdown could rescue or exacerbate the decrease in synaptosome phagocytosis induced by IL-1α+TNF+C1q (Fig. 3a). For upregulation of cell-surface VCAM1, we performed the screens in only IL-1α+TNF+C1q-treated iAstrocytes (Fig. 3a) given that in the absence of cytokine treatment, very few iAstrocytes expressed cell-surface VCAM1 (Fig. 1m). To ensure that the results from the screens are generalizable, we validated that IL-1α+TNF+C1q also caused decreased synaptosome phagocytosis and upregulation of cell-surface VCAM1 in hiPSC-derived astrocytes generated using alternative protocols (Extended Data Fig. 3f-h), as well as upregulation of cell-surface VCAM1 in primary mouse astrocytes (Extended Data Fig. h-i).
For MRA, the workflow consisted of first reconstructing a gene regulatory network between upstream regulators (e.g. transcription factors, or kinases and phosphatases) and target genes in human astrocytes based on coexpression, which required integrating a large number of human astrocyte expression profiles collected under diverse experimental conditions 41 (see Methods); upstream regulators that control inflammatory reactivity were then predicted by examining the overlap between the regulons of upstream regulators with the DEGs induced by IL-1α+TNF+C1q in iAstrocytes (Fig. 3b).
We performed MRA using the human transcription factors 42 (Fig. 3c) or kinases and phosphatases 43,44 (Fig. 3f) as upstream regulators, and similarly conducted pooled CRISPRi screening using a new custom sgRNA library targeting the human transcription factors ("TF library; Fig. 3d-e) and a library targeting all kinases, phosphatases, and other genes representing the "druggable genome" 45 ("H1 library"; Fig. 3g-h). Together, these two approaches uncovered both expected regulators of inflammatory reactivity as well as potentially novel regulators.
For example, the top hits from both MRA and CRISPRi screens (Supplementary Tables 2-3) included the entire set of genes encoding all potential subunits of the NF-κB transcription factor (REL, RELA, RELB, NFKB1, NFKB2; Fig. 3c-e) as well as genes encoding the upstream kinases of the canonical NF-κB pathway (IKBKB, CHUK, IKBKG; Fig. 3f-h) 46 . For all the above genes (with the exception of NFKB2), the directionality of their screening phenotypes or master regulator analysis activity scores (see Methods) was consistent with the activation of the canonical NF-kB pathway being required for the induction of inflammatory reactivity. As an example, knockdown of RELA, which encodes the p65 subunit of the NF-κB transcription factor, blocked the induction of cell-surface VCAM1 (Fig. 3e) and rescued the decrease in phagocytosis (Fig. 3d) induced by IL-1α+TNF+C1q (also see Extended Data Fig. 4b). In addition, the activity score of RELA from MRA was positive, consistent with the positive average log-fold change of DEGs under the control of RELA (Supplementary Tables 2-3). For the case of NFKB2, knockdown of which actually increased VCAM1 induction by IL-1α+TNF+C1q (Fig. 3e), it is possible that the encoded protein, p100, may act to inhibit RELA-dependent transcription through the cytoplasmic sequestration of p65 47 .
Given that the canonical NF-κB pathway is a well-studied master regulator of cellular responses to inflammatory stimuli 48 , its strong enrichment in the top hits from the screens and MRA serves as a positive control for the technical quality of the screens (see Extended Data Fig. 4a for quality control data) and the validity of the MRA pipeline. As a further control for the technical quality of the screens, we validated the phenotypes of selected top hits in independent experiments (Extended Data Fig. 4c-d).
In addition to the canonical NF-κB pathway, we also recovered genes involved in numerous other cellular pathways known to mediate inflammatory processes. One group of genes (STAT3, CEBPB, CEBPD) consisted of transcription factors classically associated with the acute phase response in the context of systemic inflammation 49,50 . In the central nervous system, these transcription factors control similar responses in reactive astrocytes, such as the production of acute phase proteins C3 and α1-antitrypsin (SERPINA1) by astrocytes challenged with IL-1 and TNF 49 (also see Fig. 1l and Supplementary Table 1). Furthermore, STAT3 is required for the full induction of astrocyte reactivity caused by spinal cord injury 13,51,52 , and CEBPB has been implicated in the pathogenesis of Alzheimer's disease 53,54 . We found that knockdown of STAT3, CEBPB, and CEBPD in the context of IL-1α+TNF+C1q treatment surprisingly promoted the induction VCAM1 while having mixed effects on phagocytosis ( Another group of genes consisted of those involved in the response to interferons (IRF1, STAT1, IKBKE) [55][56][57] . For example, we found that knockdown of IRF1 rescued decreased phagocytosis and reduced VCAM1 upregulation (Fig. 3d,e,h; Extended Data Fig. 4b). Given that interferons have been shown to potentiate the induction of VCAM1 by inflammatory cytokines 33,58 , the reduction of VCAM1 upregulation with IRF1 knockdown suggests that interferons are released by iAstrocytes after IL-1α+TNF+C1q treatment.
The release of interferons by iAstrocytes after IL-1α+TNF+C1q treatment would also explain the effect we observed with knockdown of genes encoding kinases responsible for signal transduction from the IL-1 receptor (IRAK2, IRAK4) 59 . Knockdown of IRAK2 or IRAK4, which should block signaling from the IL-1 receptor, strongly increased VCAM1 induction (Fig. 3h). This can be explained by the fact that IL-1 signaling is known to antagonize the action of interferons 59 .
Lastly, we also found genes involved in cellular pathways that have been relatively less studied in the context of inflammatory astrocyte reactivity. For example, we uncovered genes involved in or related to the mTOR pathway (MTOR, LAMTOR3, LATS2, FOXK1) 60, 61 , the glucocorticoid receptor pathway (NR3C1), the actin cytoskeleton (ARPC3, ACTR2), and also relatively uncharacterized genes such as FOXC2 and ZBTB7A.
Overall, the results from CRISPRi screening and MRA demonstrate that numerous cellular pathways regulate distinct aspects of inflammatory reactivity induced by IL-1α+TNF+C1q. Of particular interest to us was the fact that knockdown of the acute phase response-related genes STAT3, CEBPB, or CEBPD increased the induction of VCAM1 by IL-1α+TNF+C1q, whereas knockdown of the interferon-related gene IRF1 decreased VCAM1 induction. Since all of the above genes are known to be involved in inflammatory responses, their opposing phenotypes suggested to us that VCAM1 upregulation may only be capturing one particular aspect of IL-1α+TNF+C1q-induced inflammatory reactivity, i.e. that IL-1α+TNF+C1q may induce distinct inflammatory reactive signatures, one of which is marked by VCAM1 upregulation.

CROP-seq reveals two distinct inflammatory reactive signatures
To gain deeper insight into how regulators uncovered by CRISPRi screening and MRA may control distinct inflammatory reactive signatures, we turned to CROP-seq 62 . By coupling CRISPRi perturbations to single-cell transcriptomics, CROP-seq enables the recovery of perturbation-associated changes in gene expression from a pooled experiment. We selected 30 regulators (Extended Data Fig. 5d; Supplementary Table 6) that were strong hits from CRISPRi screening and MRA for pooled knockdown in iAstrocytes treated with vehicle control or IL-1α+TNF+C1q, and then performed CROP-seq to characterize the effect of knockdown on IL-1α+TNF+C1q-induced gene expression (see Methods).
Before examining the effect of knocking down different regulators on inflammatory reactivity, we first focused on iAstrocytes transduced with non-targeting control (NTC) sgRNAs. In the absence of IL-1α+TNF+C1q, NTC iAstrocytes partitioned largely by cell cycle in uniform manifold approximation projection (UMAP), with cluster 3 corresponding to dividing cells. A small fraction (<10%) of endothelial-like and stromal-like cells were also present, corresponding to clusters 5 and 6, respectively (Fig. 4a, Extended Data Fig. 5a-c, Supplementary Table 4). Upon treatment with IL-1α+TNF+C1q, NTC iAstrocytes partitioned into two distinct clusters (clusters 1 and 2), with an additional small cluster (cluster 4) corresponding to cycling cells (Fig.  4a, Extended Data Fig. 5a-c, Supplementary Table 4).
Upon further examination of the two major clusters induced by IL-1α+TNF+C1q, we found that iAstrocytes in cluster 1 expressed markers related to the acute phase response such as C3, IL6, and IL1B, whereas iAstrocytes in cluster 2 expressed markers related to interferon (IFN) signaling such as VCAM1, CXCL10, and LTB 63 (Fig. 4b). Enrichment analysis of cluster 2 markers confirmed a strong, specific enrichment (i.e. unique to cluster 2 markers) for interferonrelated pathways (e.g. "Interferon Gamma Response", "Interferon Alpha Response") and transcription factor regulons (e.g. STAT1, IRF1) (Fig. 4c, Supplementary Table 5). On the other hand, cluster 1 markers showed a specific enrichment for pathways and transcription factor regulons related to the mTORC1-HIF1α/MYC axis of metabolic control 64 (e.g. "mTORC1 Signaling", "Hypoxia", "Myc Targets V1"; HIF1A, CEBPD 65 ). For brevity, we will hereon provisionally refer to the inflammatory reactive astrocyte signature (IRAS) corresponding to cluster 1 as IRAS1 and the one corresponding to cluster 2 as IRAS2.
Given that the results from CRISPRi screening and MRA pointed towards the canonical NF-κB pathway being required for the induction of inflammatory reactivity, we examined the effect of knocking down RELA (NF-κB p65 subunit), IKBKG (NEMO), or MAP3K7 (TAK1) on the induction of IRAS1 and IRAS2 by IL-1α+TNF+C1q in our CROP-seq data. Aligned UMAP embedding (see Methods) of RELA, IKBKG, or MAP3K7 knockdown iAstrocytes with NTC iAstrocytes showed that RELA knockdown prevented the full induction of both IRAS1 and IRAS2, with IKBKG and MAP3K7 knockdown showing similar but less complete effects (Fig.  4d). Examining the expression of IRAS1 and IRAS2 markers in IL-1α+TNF+C1q-treated NTC iAstrocytes vs. IL-1α+TNF+C1q-treated RELA, IKBKG, or MAP3K7 knockdown iAstrocytes showed the same pattern (Fig. 4e). To validate the above findings in an independent experiment, we measured the induction of C3, a marker of IRAS1, and VCAM1, a marker of IRAS2, by flow cytometry in NTC iAstrocytes and RELA, IKBKG, or MAP3K7 knockdown iAstrocytes treated with IL-1α+TNF+C1q; we found that RELA, IKBKG, or MAP3K7 knockdown strongly reduced the induction of VCAM1 and C3 (Fig. 4f). Lastly, we performed immunostaining against NF-κB p65 and confirmed that IL-1α+TNF+C1q treatment induced the nuclear localization of p65 in iAstrocytes ( Fig. 4g-h).
To gain a more global view of how knocking down different regulators affected the induction of inflammatory reactivity, we performed differential gene expression analysis for each regulator to find how regulator knockdown altered the differential expression induced by IL-1α+TNF+C1q (Extended Data Fig. 5d, Supplementary Table 6; see Methods). We then performed hierarchical clustering on the log-fold-changes (weighted by statistical significance) of the union of all DEGs (see Methods). We found that regulators segregated into three major groups, with one group (G1) consisting largely of regulators involved in the canonical NF-κB pathway (RELA, IKBKG, NFKB1, REL, CHUK) and upstream signal transduction (IRAK4, RIPK1, MAP3K7), another group (G2) that contained acute phase response-related transcription factors (CEBPB, CEBPD, STAT3), and the last group (G3) consisting of regulators with weak knockdown phenotypes (Extended Data Fig. 5e). In terms of coordinately regulated genes, we recovered 10 gene modules that displayed distinct patterns across knockdown of different regulators (Extended Data Fig. 5e, Extended Data Fig. 6, Supplementary Table 7).
Of particular interest were modules M3, M4, and M9, which all had reduced induction upon knockdown of G1 regulators and thus appeared to be downstream of the canonical NF-κB pathway. Interestingly, M4 genes contained and had similar enrichments as IRAS1 markers (e.g. IL6; see Supplementary Table 7 and Extended Data Fig. 6) and displayed reduced induction upon knockdown of G2 regulators (Extended Data Fig. 5e), whereas M3 and M9 genes contained and had similar enrichments as IRAS2 markers (e.g. VCAM1; see Supplementary Table 7 and Extended Data Fig. 6) and displayed increased induction upon knockdown of G2 regulators (Extended Data Fig. 5e). This led us to hypothesize that the acute phase response-related transcription factors CEBPB, CEBPD, and STAT3 promote IRAS1 while inhibiting IRAS2.

IL-6 and interferons act in an autocrine-paracrine manner to drive IRAS1 and IRAS2
Given that IRAS1 iAstrocytes expressed markers related to the acute phase response, which is driven by IL-6 66 , and that IRAS2 iAstrocytes expressed markers related to the response to interferons, we hypothesized that IL-1α+TNF+C1q induced iAstrocytes to secrete IL-6 and interferons, which then acted on the iAstrocytes in an autocrine-paracrine manner. We found that iAstrocytes derived from multiple hiPSC lines indeed secreted appreciable amounts of IL-6 (~40,000 pg/mL) and IFN-β (~1,000 pg/mL) in response to IL-1α+TNF+C1q (Fig. 5a, Extended Data Fig. 7d). iAstrocytes also secreted appreciable amounts of GM-CSF (~3,500 pg/mL), which is known to act synergistically with IL-6 67 , as well as CXCL10 (~30,000 pg/mL), which is produced in response to interferons 68 .
To test whether the secreted IL-6 and interferons acted on the iAstrocytes in an autocrineparacrine manner, we knocked down their receptors and downstream transcription factors and measured the induction of C3 and VCAM1, markers of IRAS1 and IRAS2 respectively, by flow cytometry. We found that knockdown of IL6R or IL6ST, which respectively encode the IL-6 receptor and its signal transducing partner gp130, decreased the induction of C3+ iAstrocytes; the induction of VCAM1 in iAstrocytes increased slightly for the case of IL6ST knockdown but did not change for IL6R knockdown (Fig. 5b). Knockdown of STAT3, which is activated downstream of IL-6 69 , similarly decreased the induction of C3+ iAstrocytes while increasing the induction of VCAM1+ iAstrocytes (Fig. 5b). We confirmed that IL-1α+TNF+C1q induced STAT3 activation as measured by increased STAT3 phosphorylation at Y705 (Fig. 5c).
On the other hand, concurrent knockdown of IFNAR1 and IFNAR2 (see Methods), which encode subunits of the type I interferon (IFN) receptor, or IFNGR1 and IFNGR2, which encode subunits of the type II interferon receptor, decreased the induction of both C3+ iAstrocytes and VCAM1+ iAstrocytes (Fig. 5b). Concurrent knockdown of STAT1 and STAT2, which are activated downstream of interferons 70 , also resulted in a decrease in the induction of both C3+ iAstrocytes and VCAM1+ iAstrocytes (Fig. 5b). We confirmed that IL-1α+TNF+C1q induced STAT1 activation as measured by increased STAT1 phosphorylation at Y701 (Fig. 5c).
As an alternative approach to assay the induction of IRAS1 and IRAS2, we assayed the production of GM-CSF and CXCL10, which should be produced by IRAS1 and IRAS2 iAstrocytes respectively based on mRNA expression (Fig. 5d). Knockdown of IL6ST or STAT3 decreased the production of GM-CSF but increased the production of CXCL10, whereas knockdown of IFNAR1/2 or STAT1/2 decreased the production of both GM-CSF and CXCL10 (Fig. 5e).
Upon examining the activity of IL-1α, TNF, and C1q alone or in combination, we found that IL-1α by itself induced VCAM1-/C3+ iAstrocytes, whereas TNF by itself induced VCAM1+/C3and VCAM1+/C3+ iAstrocytes (Fig. 6b). Interestingly, IL-1α+TNF decreased the induction of VCAM1+/C3-astrocytes compared to TNF by itself (Fig. 6b). To validate these findings, we performed immunostaining against C3 and IFIT3, an alternative IRAS2 marker (Extended Data Fig. 7a-b), in iAstrocytes derived from three different hiPSC lines and observed similar results (Extended Data Fig. 7c). In terms of cytokine production (Extended Data Fig. 7d), IL-1α by itself, but not TNF by itself induced IL-6 and GM-CSF, whereas TNF by itself induced higher levels of CXCL10 compared to IL-1α by itself. Also, IL-1α+TNF decreased CXCL10 production compared to TNF by itself. These findings were similar between iAstrocytes derived from two different hiPSC lines (Extended Data Fig. 7d).
Taking stock of both VCAM1/C3 induction and cytokine production, the inhibitory effect of IL-1α on the induction of CXCL10 and VCAM1+/C3-iAstrocytes by TNF likely reflects the antagonistic effect of IL-1 on the activity of autocrine-paracrine interferon signaling 59 , which can be induced by both IL-1α and TNF (Extended Data Fig. 7d). As for the activity of C1q, C1q by itself did not induce any cytokine production or any inflammatory reactive astrocytes marked by VCAM1/C3; C1q also had little additional effect in combination with IL-1α, TNF, or IL-1α+TNF ( Fig. 6b; Extended Data Fig. 7c-d). However, C1q may exert effects not detectable by the assays used above, e.g. at the transcriptional level.
With dual staining of VCAM1 and C3, we revisited the effect of knocking down receptors and transcription factors involved in IL-6 or interferon signaling (Fig. 6c) and found that the results were consistent with those obtained with single staining of VCAM1 or C3 (Fig. 5b). In addition, we also tested the effect of knocking down CEBPB, CEBPD, NFKB2, or IRF1, which were strong hits from the VCAM1 CRISPRi screen. Similarly to knockdown of STAT3, knockdown of CEBPB, CEBPD, or NFKB2 reduced the induction of VCAM1-/C3+ iAstrocytes and increased the induction of VCAM1+/C3-or VCAM1+/C3+ iAstrocytes (Fig. 6c). Knockdown of IRF1 reduced the induction of VCAM1+/C3-and VCAM1+/C3+ iAstrocytes and also marginally reduced the induction of VCAM1-/C3+ iAstrocytes (Fig. 6c), similarly to STAT1/2 knockdown. We also tested the effect of STAT3, CEBPB, NFKB2, or IRF1 knockdown in iAstrocytes derived from two additional hiPSC lines and obtained consistent results (Extended Data Fig. 8a-b).
Lastly, to complement our experiments employing CRIPSRi knockdown, we also tested the effect of small molecule known to modulate STAT3 or STAT1/2 activity. We found that napabucasin, which is known to inhibit STAT3-dependent transcription 71 , reduced STAT3 Y705 phosphorylation in a dose-dependent manner (Extended Data Fig. 9b) and at doses above 1.25 μM abrogated the induction of VCAM1-/C3+, VCAM1+/C3-, and VCAM1+/C3+ in IL-1α+TNF+C1q-treated iAstrocytes . The different effect of napabucasin vs. STAT3 knockdown is not surprising given that the mechanism of action of napabucasin is not well understood and likely involves multiple cellular pathways 72 . For modulation of STAT1/2 activity, we found that RGFP966, a HDAC3 inhibitor known to reduce STAT1 activity 73 , boosted the induction of VCAM1-/C3+ iAstrocytes while inhibiting the induction of VCAM1+/C3-and VCAM1+/C3+ iAstrocytes (Extended Data Fig. 9d). The different effect of RGFP966 vs. STAT1/2 knockdown could be attributed to effects of HDAC3 inhibition beyond reducing STAT1 activity.
Based on our findings so far and also existing literature, we created a model (Fig. 6e) of how IL-1α+TNF+C1q treatment induces IRAS1 and IRAS2 through a primary NF-κB-dependent transcriptional response ( Fig. 4d-h) followed by autocrine-paracrine IL-6 and interferon signaling feedback loops involving CEBPB/D, NFKB2, STAT3, STAT1/2, and IRF1 (Fig. 5, Fig.  6). Downstream of NF-κB, CEBPB/D, NFKB2, and STAT3 drive the induction of IRAS1 and inhibit the induction of IRAS2 through IL-6 (also see refs. 74,75 ), whereas STAT1/2 and IRF1 promote the induction of IRAS2 through interferons (also see ref. 76 ), which inhibit the induction of IRAS1 (Fig. 6b-d). IL-1α polarizes astrocytes towards IRAS1 by driving IL-6 production (Extended Data Fig. 7d) and antagonizing interferon signaling (also see ref. 59 ), whereas TNF polarizes astrocytes towards IRAS2 by driving interferon production (Extended Data Fig. 7d), which in turn antagonizes IL-1 signaling (also see ref. 59 ). The discrepancy between the inhibitory effect of IFN-β on IRAS1 induction (Fig. 6b) and the weak requirement of STAT1/2 and IRF1 for IRAS1 induction (Fig. 6c) could be explained by the promiscuity of STAT1/2 in responding to and promoting the production of both IL-6 and IFN-β ( Fig. 6d; also see refs. 76,[77][78][79] ). On the other hand, STAT3 activation has also been reported in response to interferons 80,81 , and our data supports some role of STAT3 in the production of IFN-β (Fig. 6d); however, in our system STAT3 is much more strongly associated with IL-6 signaling given the consistent effect of IL-6 addition and STAT3 knockdown.
We do not intend for the terms IRAS1 and IRAS2 to encompass reactive astrocyte signatures outside those generated within our experimental system. Thus, in light of the cellular pathways driving IRAS1 and IRAS2, we propose that IRAS1 be referred to as "IL-1/IL-6-responsive", and IRAS2 as "TNF/IFN-responsive" reactivity.

Signatures of IL-1/IL-6-responsive and TNF/IFN-responsive reactive astrocytes overlap with reactive astrocyte signatures found across different species and disease contexts
To see if signatures of IL-1/IL-6-responsive and TNF/IFN-responsive reactive astrocytes are found across different experimental paradigms of neuroinflammation, we integrated the NTC sgRNA-transduced iAstrocytes from our single-cell RNA-seq data with previously published astrocyte single-cell RNA-seq datasets using the anchor-based data integration functionality of Seurat 82 (see Methods).
In Barbar et al. 29 , mixed cultures of neurons and glia derived from dissociation of human cerebral organoids were treated with vehicle control or IL-1α+TNF+C1q. After isolating astrocytes from the Barbar et al. dataset and integrating them with NTC iAstrocytes from our study (see Methods), we found that a subset of IL-1α+TNF+C1q-treated astrocytes from Barbar  Table 8). Furthermore, the DEGs between IRAS2 vs. IRAS1 iAstrocytes overlapped strongly with the DEGs between the corresponding co-clustering Barbar et al. astrocytes (Extended Data Fig. 10a).
In Wheeler et al. 83 , astrocytes were isolated by FACS from mice induced with experimental autoimmune encephalomyelitis (EAE), a widely used model of multiple sclerosis (MS), at different stages of EAE progression (naïve, acute, and remitting). After removing low-quality astrocytes from the Wheeler et al. dataset and integrating the remaining astrocytes with NTC iAstrocytes from our study (see Methods), we found that the vast majority of Wheeler et al. astrocytes from the acute stage of EAE co-clustered with IRAS1 and IRAS2 iAstrocytes (  Table 8). Overlap analysis of the DEGs between IRAS2 vs. IRAS1 iAstrocytes and the DEGs between the corresponding co-clustering astrocytes from Wheeler et al. also showed good agreement (Extended Data Fig. 10b), although the degree of overlap (Jaccard index) was lower compared to the integration with Barbar et al likely due to the difference in species.
Lastly, in Hasel et al. 58 , astrocytes were isolated by FACS from mice injected peripherally with saline or LPS, which induces acute neuroinflammation similar to that caused by sepsis. After removing non-astrocyte cells from the Hasel et al. dataset, integration of the Hasel et al. astrocytes with NTC iAstrocytes from our study showed that a subset of Hasel et al. astrocytes from LPS-injected mice co-clustered with IRAS1and IRAS2 iAstrocytes ( Fig. 7i-k) and also correspondingly expressed IL-1/IL-6 responsive markers such as C3, Cxcl5, and Tnfaip2 or TNF/IFN-responsive markers such as Cxcl10, Ifit3, and Isg15 (  Table 8). Overlap analysis of the DEGs between IRAS2 vs. IRAS1 iAstrocytes and the DEGs between the corresponding co-clustering astrocytes from Hasel et al. further supported this correspondence (Extended Data Fig. 10c).
In addition to single-cell RNA-seq datasets, we also reanalyzed bulk RNA-seq data from Anderson et al. 51 , where astrocyte-specific RNA was purified from wild-type or astrocytespecific Stat3 conditional knockout (cKO) mice subject to spinal cord injury (SCI). We found that Stat3-dependent genes (i.e. genes with lower expression in Stat3 cKO SCI vs. WT SCI; e.g. C3) tended to be highly expressed by IRAS1 (IL-1/IL-6-responsive) iAstrocytes (Extended Data Fig. 10h), whereas Stat3-repressed genes (i.e. genes with higher expression in Stat3 cKO SCI vs. WT SCI; e.g. Iigp1) were enriched for genes involved in the response to interferons (Extended Data Fig. 10i), overlapping with IRAS2 (TNF/IFN-responsive) markers (see Supplementary  Table 9 for Stat3 cKO-related DEGs from Anderson et al.). Thus, the results from Anderson et al. further corroborate the regulatory role of STAT3 in promoting IL-1/IL-6-responsive reactivity and inhibiting TNF/IFN-responsive reactivity that we have proposed.

Markers of IL-1/IL-6-responsive and TNF/IFN-responsive reactive astrocytes are upregulated in human disease
Given that we found signatures of IL-1/IL-6 responsive and TNF/IFN-responsive reactive astrocytes in various experimental models of neuroinflammation, we sought to test whether they can be found in human disease. In Alzheimer's disease (AD), neuroinflammation mediated by microglia and astrocytes contributes to disease progression 84 potentially through IL-6 85 and interferons 86 . We performed immunostaining of the IL-1/IL-6 responsive marker C3 and the TNF/IFN-responsive marker VCAM1 in post-mortem brain tissue derived from AD cases (61-92 years old, Braak stage IV-VI) vs. age-matched controls (62-102 years old, Braak stage 0-III; see Supplementary Table 10 for clinical metadata). We found a significant increase in the abundance of C3+ astrocytes in AD ( Fig. 8a-b), consistent with previous studies 8 . Notably, C3+ astrocytes tended to be localized near C3+ plaques (Fig. 8a), which we deduced to be amyloid plaques given the fact that they are known to accumulate complement proteins such as C3 87 . Amyloid plaques are also known to accumulate SERPINA3 88 , which is a IL-1/IL-6 responsive marker and also a marker of disease-associated reactive astrocytes found in mouse models of AD 89 . As for VCAM1, we did not observe a statistically significant difference in the abundance of VCAM1+ astrocytes between AD cases vs. controls (Fig. 8c-d). Lastly, C3+ astrocytes tended to be VCAM1-(416 VCAM1-vs. 31 VCAM1+ astrocytes among C3+ astrocytes aggregated across all individuals; Fig. 8a), and VCAM1+ astrocytes tended to be C3-(135 C3-vs. 31 C3+ astrocytes among VCAM1+ astrocytes aggregated across all individuals; Fig. 8c); there were 31 VCAM1+/C3+ astrocytes among 12,885 total astrocytes.
In addition to AD, we also examined brain tissue derived from age-matched controls (38-41 gestational weeks) vs. cases of neonatal hypoxic-ischemic encephalopathy (HIE; 37-41 gestational weeks), in which neuroinflammation involving IL-1, TNF, IL-6, and interferons [90][91][92] contributes significantly to neuronal injury and negative neurologic sequela 93 . Compared to AD, HIE may be better modeled by our experimental paradigm (hiPSC-derived astrocytes treated with cytokines) given its acute course and the immature status of astrocytes in neonates 94,95 . We observed increased abundance of both C3+ and VCAM1+ astrocytes in HIE (Fig. 8e-h). However, experimental limitations prevented us from staining for C3 and VCAM1 simultaneously.

Stat3 drives reactive astrocytes expressing IL-1/IL-6-responsive markers and inhibits those expressing TNF/IFN-responsive markers in vivo
Lastly, to validate in vivo the regulatory role of STAT3 in promoting IL-1/IL-6-responsive reactivity and inhibiting TNF/IFN-responsive reactivity, we performed immunostaining against the IL-1/6-responsive marker C3, the TNF/IFN-responsive marker Isg15, and the astrocyte marker Gfap in mice peripherally injected with vehicle control or LPS to induce acute neuroinflammation (see Methods). We focused on the corpus callosum and the hippocampus, where we observed easily segmentable Gfap immunoreactivity for astrocyte labeling (Fig. 8i,l) so we could verify astrocyte specificity. We found that in wild-type mice, LPS induced Isg15−/C3+ astrocytes (IL-1/IL-6-responsive) as well as Isg15+/C3− and Isg15+/C3+ astrocytes (TNF/IFN-responsive) in the hippocampus (Fig. 8m,n) but not the corpus callosum (Fig. 8j,k). In both brain regions, astrocyte-specific conditional knockout of Stat3 in LPS-treated mice reduced the prevalence of Isg15−/C3+ astrocytes and increased the prevalence of Isg15+/C3− astrocytes (Fig. 8j,k,m,n) compared to wild-type LPS-treated mice, validating our in vitro finding of Stat3 as a key regulator with opposing effects on IL-1/IL-6-responsive and TNF/IFN-responsive reactivity .

DISCUSSION
Inflammatory reactive astrocytes have been implicated in numerous neurodegenerative and neuroinflammatory diseases such as Alzheimer's disease, Parkinson's disease, Huntington's disease, amyotrophic lateral sclerosis, and multiple sclerosis 8,11 . However, we do not yet have a full understanding of the cellular pathways controlling inflammatory reactivity in these contexts. Here, we employed pooled CRISPRi screening to systematically identify genes that control the response of hiPSC-derived astrocytes to the cytokines IL-1α+TNF+C1q, which induce a form of inflammatory reactivity associated with neurotoxicity under certain conditions and loss of homeostatic functions [8][9][10] .
The scalability and homogeneity (see Methods) of our hiPSC-derived astrocytes ("iAstrocytes") was critical to our ability to perform pooled screens. However, it should be noted that iAstrocytes represent one of many available hiPSC-derived astrocyte models, and that there is a tradeoff between the scalability and maturation status of hiPSC-derived astrocytes. For our study, we found that iAstrocytes derived from multiple hiPSC lines sufficiently recapitulated published phenotypes associated with IL-1α+TNF+C1q-induced inflammatory reactivity, and we also ensured that the phenotypes we investigated were conserved across different hiPSC-derived astrocyte models and different hiPSC lines.
In our analysis of the results from CRISPRi screens and computational master regulator analysis (MRA), we gave more weight to the results from the CRISPRi screens, which demonstrate causality, than those from MRA, which are correlational and likely to contain more false positive results. However, given that the CRISPRi screens are more likely to be limited by false negative results (e.g. due to sgRNA dropout), further exploration of the hits from MRA is warranted in future studies. We believe that the gene regulatory network we generated from MRA will have broad utility for identifying potential master regulators of diverse astrocyte phenotypes for which transcriptomic data is available.
Following up on the top hits from our CRISPRi screens with single-cell transcriptomics, we found that IL-1α+TNF+C1q induced two distinct inflammatory reactive astrocyte signaturesprovisionally named "IRAS1" and "IRAS2" -dependent on canonical NF-κB activation and driven by autocrine-paracrine IL-6 and interferon signaling. We observed that CEBPB/D, NFKB2, and STAT3 promoted IRAS1 while inhibiting IRAS2 through IL-6, whereas STAT1/2 and IRF1 promoted both IRAS1 and IRAS2 partially through IFN-β; IL-1α polarized astrocytes towards IRAS1 by driving IL-6 production and antagonizing interferon signaling, while TNF polarized astrocytes towards IRAS2 by driving interferon production and antagonizing IL-1 signaling. Given the findings above, we proposed that IRAS1 be referred to as "IL-1/IL-6responsive" and that IRAS2 be referred to as "TNF/IFN-responsive" reactivity.
Although our data supporting the above conclusions were largely in vitro, which is limited by reductionistic conditions compared to the complex in vivo environment, we validated the regulation of IL-1/IL-6-responsive and TNF/IFN-responsive reactivity by Stat3 in vivo. Furthermore, we found signatures of IL-1/IL-6-responsive or TNF/IFN-responsive astrocyte reactivity in other experimental contexts, both in vitro and in mouse models in vivo, as well as in human Alzheimer's disease (AD) and hypoxic-ischemic encephalopathy (HIE).
It is likely that IL-1/IL-6-responsive and TNF/IFN-responsive reactive astrocytes have distinct functional outputs. Several studies already point towards potential functional outputs of IL-1/IL-6-responsive reactivity, which is dependent upon STAT3. For example, in Anderson et al. 51 , astrocyte-specific deletion of Stat3 was shown to prevent proper axon regeneration after spinal cord injury, suggesting that IL-1/IL-6-responsive reactivity may promote axon regeneration. In addition, in Kim et al. 96 , inhibition of STAT3 in hiPSC-derived astrocytes co-cultured with hiPSC-derived brain endothelial-like cells rescued the decrease in endothelial barrier integrity caused by treatment with TNF, suggesting that IL-1/IL-6-responsive reactivity may perturb blood-brain barrier and cerebrovascular function under inflammatory conditions 97 .
On the other hand, interferon-responsive reactivity may play an important role in EAE and MS. For example, Rothhammer et al. 98 showed that inhibiting type I interferon signaling in astrocytes exacerbated the severity of EAE. Similarly, Hindinger et al. 99 showed that inhibiting type II interferon in in astrocytes exacerbated the severity of clinical symptoms during peak disease. Furthermore, given that interferon-responsive reactive astrocytes upregulate VCAM1 and were shown to be adjacent to vasculature in Hasel et al. 58 , they may be important for controlling the trafficking of peripheral immune cells into the CNS parenchyma.
With regard to AD, literature on mouse models of AD supports a role for both IL-1/IL-6responsive reactivity and TNF/IFN-responsive reactivity. Reichenbach et al. demonstrated that astrocyte-specific deletion of Stat3, which would be expected to reduce IL-1/IL-6-responsive reactivity, decreased the burden of amyloid plaques and dystrophic neurites, lowered inflammatory cytokine levels, and ameliorated cognitive deficits in the APP/PS1 mouse model of AD 100 . Roy et al. detected interferon-responsive astrocytes that were expanded in a amyloid pathology-dependent in the 5XFAD mouse model of AD 86 . Although we did not observe an increase in the percentage of VCAM1+ astrocytes in our examination of human AD brain tissue, this does not rule out the presence of interferon-responsive reactive astrocytes in AD, since VCAM1 may not be an optimal marker of interferon-responsive reactive astrocytes in all circumstances. Interestingly, Sadick et al. detected a strong interferon-responsive signature in astrocytes profiled by single-nucleus RNA-seq from a donor with vascular dementia 101 .
We believe that our iAstrocyte platform will assist in future investigations using patient-derived hiPSCs to dissect the effect of disease-associated mutations on inflammatory astrocyte reactivity. In addition, we believe our work here will serve as a valuable resource for future work to further characterize the functional outputs of IL-1/IL-6-responsive vs. TNF/IFN-responsive astrocyte reactivity in animal models of neuroinflammation and neurodegeneration. More generally, the approach pioneered here could be applied to uncover regulators of different reactive astrocytes signatures induced by other perturbations, which will pave the way for characterizing their functions and targeting them therapeutically.

ACKNOWLEDGEMENTS
We thank Brandon Desousa, Vukasin Jovanovic, Zuzana Krejciova, and Nawei Sun for contributions to preliminary studies and discussions. We thank Professors Anna Molofsky, Aimee Kao, and Michael Oldham for serving on KL's thesis committee. We thank members of the Kampmann lab (Greg Mohl, Sydney Sattler, Olivia Teter) for discussions and feedback on the manuscript. We thank Brian Woo for cloning the transcription factors sgRNA library, and Biswa Ramani for help with obtaining primary mouse astrocytes. We thank the Conklin lab for the gift of the WTC11 hiPSC line.

DATA AVAILABILITY STATEMENT
Bulk RNA-seq data of hiPSC-derived astrocytes generated in this study shown in Extended Data Fig. 3 are available on the Gene Expression Omnibus (GEO) under accession code GSE182307. The raw single-cell RNA-seq data and UMI count matrices from the CROP-seq experiment are available on GEO under accession code GSE182308. Processed data from the CRISPRi screens and CROP-seq experiment can also be interactively explored on CRISPRbrain (https://www.crisprbrain.org/).

CODE AVAILABILITY STATEMENT
The full analysis pipeline (including code and processed data objects) used for master regulator analysis, analysis of CROP-seq data, and integration with previously published single-cell RNA-seq datasets is available at https://kampmannlab.ucsf.edu/inflammatory-reactive-astrocyteanalysis. and then incubated with Versene (ThermoFisher Scientific cat. no. 15040066) for 5-7 min at room temperature; the Versene solution was then aspirated and replaced with E8 + 10 nM Y-27632 dihydrochloride ROCK inhibitor (Tocris cat. no. 125410); hiPSC colonies were then gently detached mechanically using a cell scraper, resuspended gently, and passaged at 1:10-1:30 dilution in E8 + Y-27632, with Y-27632 removed the next day. For near single-cell dissociation, hiPSCs were washed with DPBS, incubated with Accutase for 5-10 min at 37 °C, and then gently triturated with a P1000 pipette tip; the cell suspension was then diluted with PBS, collected into conical tubes, and spun down at 300 g for 3 min; hiPSCs were then resuspended in E8 + Y-27632, counted, and plated onto Matrigel-coated plates at the desired density in E8 + Y-27632; Y-27632 would be maintained until the hiPSC colonies reached the appropriate size (> ~40 cells). Studies with hiPSCs at UCSF were approved by The Human Gamete, Embryo and Stem Cell Research (GESCR) Committee. Informed consent was obtained from the human subjects when the hiPSC lines originally derived.

Cloning of NFIA and SOX9 cDNA into dox-inducible cassette
To obtain NFIA cDNA, we designed PCR primers to amplify cDNA corresponding to transcript ENST00000403491(NFIA isoform 1) from astrocyte cDNA (Forward primer complementary sequence: ATGTATTCTCCGCTCTGTCTCAC; reverse primer complementary sequence: TCCCAGGTACCAGGACTGTG). We chose to amplify cDNA corresponding to NFIA isoform 1 because the cDNA clone (BC022264) used in Li et al. 19 corresponds to transcript ENST00000371187 (NFIA isoform 2), which we found was not expressed highly in human astrocytes. To obtain SOX9 cDNA, we ordered cDNA clone OHu19789 (which corresponds to MGC clone BC056420) from GeneScript and then amplified SOX9 cDNA from the plasmid (Forward primer complementary sequence: ATGAATCTCCTGGACCCCTTCA; reverse primer complementary sequence: TCAAGGTCGAGTGAGCTGTGT). We then inserted NFIA and SOX9 cDNA joined by a T2A sequence into an AAVS1 safe-harbor plasmid containing a doxinducible cassette (Addgene plasmid no. 105840, gift from Michael Ward; digested with AflIIand ClaI) using Gibson assembly (New England Biolabs; cat. no. E2611L), resulting in pKL100.
Generation of hiPSC lines with stable integration of NFIA-SOX9 cassette and CRISPRi cassette WTC11, TCW-1E44, and 162D hiPSCs were transfected with pC13N-dCas9-BFP-KRAB 105 (Addgene plasmid no. 127968) to stably integrate constitutive CRISPRi machinery into the CLYBL locus using TALEN-based editing as previously described 105 . CRISPRi WTC11, TCW-1E44, and 162D hiPSCs were then transfected with pKL100 to stably integrate the dox-inducible NFIA-SOX9 cassette into the AAVS1 locus using the same TALEN-based editing approach. Briefly, CRISPRi WTC11, TCW-1E44, and 162D hiPSCs were dissociated with Accutase to a single-cell suspension and plated at 500,000 cells per well in a Matrigel-coated 6-well plate in E8 + 10 µM Y-27632. The next day, a media change with E8 + Y-27632 was performed and then the hiPSCs were transfected with 2.5 ug of pKL100, 1.25 ug of left and right AAVS1 TALEN plasmids (Addgene plasmid no. 59025 and 59026, gift from Danwei Huangfu), and 0.5 ug of Bcl-XL plasmid (pEF1-BCL-XL-wpre-polyA P1102, gift from Xiaobing Zhang, described in ref 106 ) using Lipofectamine Stem Transfection Reagent (ThermoFisher Scientific cat. no. STEM00015) following the manufacture's protocol. After hiPSCs reached confluence, they were dissociated with Accutase and passaged to a 10 cm dish in E8 + 10 uM Y-27632 + 0.1 ug/mL puromycin (ThermoFisher Scientific cat. no. A1113803) to select for clones with stable integration. Y-27632 was maintained until stable colonies formed, and puromycin was maintained for 5-7 days. The hiPSCs were then sorted for mCherry+ cells, which were then plated at 500,000 cells per well in a Matrigel-coated 6-well plate for transfection with 1.2 ug of Cre recombinase mRNA (TriLink Biotechnologies cat. no. L-7211) using Lipofectamine Stem to remove the puromycin resistance gene and mCherry. The Cre-transfected hiPSCs were then expanded and sorted for mCherry-cells, which were then plated for colony picking to generate monoclonal hiPSC lines. hiPSC clones were tested for integration of the NFIA-SOX9 cassette and removal of the puromycin resistance and mCherry by genomic PCR with the following pairs of primers:

Neural induction of hiPSCs
Embryoid body (EB)-based neural induction of WTC11, TCW-1E44, and 162D hiPSCs stably integrated with CRISPRi machinery and dox-inducible NFIA-SOX9 was performed as previously described 107 with some modifications. Briefly, on day 0, hiPSCs were dissociated to a near single-cell suspension with Accutase, resuspended in neural induction media (NIM; see recipe below) + 10 nM Y-27632, and then transferred to an Aggrewell 800 plate (StemCell Technologies cat. no. 34815) pre-coated with Anti-adherence rinsing solution (StemCell Technologies cat. no. 07010) at 3 million cells per well for EB formation. The next day (day 1), a half media change with NIM was performed and LDN193189 (LDN; Tocris cat. no. 6053) and SB431542 (SB; Tocris cat. no. 1614) were added to final concentrations of 0.1 μM and 10 μM, respectively. A half media change of NIM + LDN + SB was performed every day until day 7, when EBs were transferred to a Matrigel-coated 6-well plate after performing a half media change (one well from the Aggrewell plate would be transferred to one well in the 6-well plate). The next day (day 8), a full media change with NIM + LDN + SB was performed, and then every other day afterwards until day 14, during which time neural rosettes would appear in the attached EBs. On day 14, neural rosettes were detached non-enzymatically with the following method: the attached EBs were incubated Neural Rosette Selection Reagent (NRSR; StemCell Technologies cat. no. 05832) for 1 hour at 37 °C to weaken the attachment of the neural rosettes, the NRSR was aspirated and replaced with DMEM/F12, and the rosettes were detached with targeted jetting of DMEM/F12 using a wide-orifice P1000 pipette tip and then collected into a conical tube; targeted jetting with DMEM/F12 and collection of released rosettes was repeated until the majority of rosettes had detached. Rosettes were then spun down at 100 g for 3 min, resuspended in NPC media (see recipe below), and transferred to a Matrigel-coated 6-well plate (rosettes collected from 1-2 wells would be replated into 1 well depending on the yield). From day 15 to 21, a full media change with NPC media was performed every other day, during which time neural progenitor cells (NPCs) would spread out from the attached rosettes and cover the well completely. Once NPCs reached confluency, they were dissociated with Accutase and replated at high density (at least 1 million cells per well of a 6-well plate) in Matrigel-coated plates for expansion.

Purification of NPCs
To remove contaminating neural crest cells from the NPC cultures, we used fluorescence activated cell sorting (FACS) to select for CD133+/CD271-cells following the protocol described in Cheng et al. 107 with modifications. Briefly, NPCs were dissociated to a single-cell suspension with Accutase, resuspended in FACS buffer (see recipe below), and then incubated with PE-conjugated CD133 antibody (1:50 dilution; Miltenyi Biotec cat. no. 130-113-108) and PerCP-Cy5.5-conjugated CD271 antibody (1:50 dilution; BD Biosciences cat. no. 560834), including single antibody-stained and unstained controls. After antibody incubation, the cell suspension was diluted 10x with FACS buffer, spun down, and resuspended in FACS buffer for sorting. CD133+/CD271-cells were sorted using a BD FACSAria Fusion cell sorter at 5,000-10,000 events per second, and then plated at 100,000 cells per cm 2 onto Matrigel-coated plates. Media was then changed every other day until the NPCs reached confluency, at which point the NPCs were passaged for expansion. After 2-3 additional expansion passages after sorting, NPCs were characterized by immunostaining and qPCR for NPC markers such as PAX6 and Nestin, and then cryopreserved in NPC media + 10% DMSO.

Generation of iAstrocytes from NPCs
For iAstrocyte differentiation, CD133+/CD271-sorted NPCs generated from WTC11, TCW-1E44, and 162D iPCS with stably integrated CRISPRi machinery and dox-inducible NFIA-SOX9 were dissociated to a single-cell suspension with Accutase and then replated at 7,500-15,000 cells per cm 2 in NPC media onto a 10-cm or 15-cm dish coated with Matrigel diluted at 1:200 in DMEM/F12. The next day, media was changed to ScienCell Astrocyte Media (ScienCell Research Laboratories cat. no. 1801) + 2 μg/mL doxycycline (Millipore Sigma cat. no. D9891) to initiate iAstrocyte differentiation. A full media was then changed every other day, with doxycycline maintained at 2 μg/mL throughout the differentiation process. When the differentiating NPCs reached confluency within 3-4 days, the culture was dissociated with Accutase and split 1:4-1:10 onto new Matrigel coated dishes for expansion, with some cells saved for cryopreservation. After confluency was reached within 5-6 days, the cultures were dissociated and split 1:4-1:8, saving some cells for cryopreservation. Expansion of the cultures at 1:4-1:8 split with cryopreservation of cells after each split was continued until day 20 of differentiation, yielding iAstrocytes. To maintain the post-split cell density between 10,000-20,000 cells per cm 2 , the split ratio was adjusted based on the proliferative tendency of the differentiating astrocyte precursors, which varied those generated from WTC11, TCW-1E44, and 162D hiPSCs. In terms of yield, starting from ~400,000 NPCs plated into a 10 cm dish, in theory ~100 million iAstrocytes could be obtained by around day 20 assuming that the cells are split 1:4 for a total of four times. However, the yield may vary depending on the proliferative tendency of the astrocyte precursors.
For cryopreservation, iAstrocytes were resuspended in ScienCell Astrocyte media + 10% DMSO, transferred into Corning CoolCell alcohol-free freezing containers (Corning cat. no. 432001) in cryogenic vials, and then transferred to liquid nitrogen. For experiments, cryopreserved iAstrocytes were thawed by warming cryovials in a 37 °C water bath until no ice was left, transferring the contents to a 15 mL conical tube, adding 4 volumes of DPBS, equilibrating at RT for 3 min, spinning down cells at 300 g for 3 min, and then resuspending the pellet in ScienCell Astrocyte Media.
Experiments were performed using iAstrocytes derived from WTC11 hiPSCs unless otherwise indicated.

Generation of hiPSC-derived astrocytes using alternative protocols
In parallel to the generation of iAstrocytes, WTC11 hiPSC-derived astrocytes were also generated from CD133+/CD271-sorted NPCs according to the protocol published in TCW et al. 18 , which mirrored the process described above for iAstrocyte generation with the exception of adding doxycycline. In addition to the protocol published in TCW et al., hiPSC-derived astrocytes were also generated from WTC11 iPSCs with stably integrated CRISPRi machinery and dox-inducible NFIA-SOX9 according to Li et al. 19 , or from WTC11 iPSCs with stably integrated CRISPRi machinery according to Krencik et al. 15 .
Primary mouse astrocyte culture P0 C57BL/6J pups were decapitated and the brains were removed and placed into ice cold HBSS. Cortices were removed from the rest of the brain in HBSS and dissociated using the Pierce Primary Neuron Isolation Kit (ThermoFisher Scientific cat. no. 88280). Instead of using the Neuronal Culture Medium from the kit, cells from the dissociated cortices were plated onto PDL-coated 96-well plates in primary astrocyte medium (see recipe below) at the recommended density. Media was changed every other day for 6 days, at which point the culture consisted mainly of astrocytes.
Primary astrocyte medium: DMEM (ThermoFisher Scientific cat. no. 11965084) 10% FBS (VWR cat. no. 89510-186; lot no. 184B19) 1% Penicillin-streptomycin (ThermoFisher Scientific cat no. 15140122) Induction of inflammatory reactivity in hiPSC-derived astrocytes or primary mouse astrocytes iAstrocytes or hiPSC-derived astrocytes generated according to TCW et al. 18 were plated onto Matrigel-coated (1:200 diluted Matrigel) 96-well plates, 24-well plates or 6-well plates at 20,000 cells per cm 2 in ScienCell Astrocyte Media without addition of doxycycline; hiPSC-derived astrocytes generated according to Li et al. 19 or Krencik et al. 15 were plated onto Matrigel-coated (1:200 diluted Matrigel) 96-well plates at ~62,500 cells per cm 2 in Astrocyte Maturation Media (AMM; see recipe below) without addition of doxycycline. A full media change was performed the next day, and then every other day afterwards. Five days after plating, hiPSC-derived astrocytes were treated with vehicle control or IL-1α (3 ng/mL; Peprotech cat. no. AF-200-01A), TNF (30 ng/mL; Peprotech cat. no. AF-300-01A), and C1q (400 ng/mL; Complement Technology cat. no. A099) with a full media change in the appropriate media to induce inflammatory reactivity according to Liddelow et al. 8 . For experiments involving all possible combinations of IL-1α, TNF, and C1q, the same concentrations as above were used. For experiments involving treatment with additional IL-6 or IFN-β, a IL-6/IL-6R chimeric protein (25 ng/mL; R&D systems cat. no. 8954-SR-025) or IFN-β (5 ng/mL; Peprotech cat. no. AF-300-02B) was added concurrently with all possible combinations of IL-1α, TNF, and C1q, which were added at the same concentration as above. We found that using a IL-6/IL-6R chimeric protein was necessary given our observation that addition of IL-6 itself had no additional effect on iAstrocytes, which is likely due to the fact that IL-6R is not expressed at high levels constitutively in most cell types 108 .
For primary mouse astrocytes, IL-1α+TNF+C1q treatment (same concentration as above) occurred after 6 days of culture. All assays were performed at 24 hours after cytokine treatment (see Fig. 1g

CRISPRi-mediated gene knockdown in hiPSC-derived astrocytes using lentiviral sgRNA delivery
For experiments involving CRISPRi-mediated gene knockdown, iAstrocytes were transduced with lentivirus containing single-guide RNAs (sgRNAs) at the time of plating (see Fig. 2b). CRISPRi sgRNAs were cloned into pMK1334 105 (Addgene cat. no. 127965) digested with BstXI and BlpI as previously described in Gilbert et al. 36 . Lentivirus containing sgRNAs was produced by transfecting HEK293T cells with pMK1334 and 3 rd generation lentiviral packaging plasmids with TransIT-Lenti Transfection Reagent (Mirus cat. no. MIR6606) according to the manufacturer's instructions. The lentivirus was then precipitated using Lentivirus Precipitation Solution (ALSTEM cat. no. VC150) according to the manufacturer's instructions, resuspended in DPBS at 1/10 of the original volume, and then aliquoted and stored at -80 °C. The functional titer of the lentivirus was then tested on iAstrocytes by serial dilution followed by measurement of BFP+ cells 48 hours after transduction. For all experiments involving CRISPRi-mediated gene knockdown in iAstrocytes, sufficient sgRNA lentivirus was added to transduced >70% of iAstrocytes.
Samples were imaged with an IN Cell Analyzer 6000, using a 20X 0.45 NA objective, 2x2 binning, 100-400 ms exposure, an aperture width of ~1 Airy unit, and 16 fields per well.

Generation of hiPSC-derived brain endothelial-like cells and measurement of barrier integrity
hiPSCs were differentiated to brain endothelial-like cells as previously described 109 . Briefly, hiPSCs were dissociated with Accutase and seeded on Matrigel-coated plates in E8 medium containing 10 μ M Y27632 at a density of 15,000 cells/cm 2 . Differentiation was initiated 24 hours after seeding by changing to E6 medium, with daily medium changes for 4 days. Next, cells were expanded with serum-free basal endothelial cell medium (EC medium) supplemented with 50x diluted B27 (Thermo Fisher Scientific), 1x GlutaMAX (Thermo Fisher Scientific), 10 μ M retinoic acid (Sigma Aldrich), and 20 μ g/ml FGF2 for 2 days without a media change. Following this treatment, cells were collected by a 20-minute incubation in Accutase and seeded onto Transwell filters (1.1 cm 2 polyethylene terephthalate membranes with 0.4 μ m pores; Fisher Scientific) coated with a mixture of 400 μ g/ml collagen IV (Sigma Aldrich) and 100 μ g/ml fibronectin (Sigma Aldrich). The following day, cells were switched to EC medium lacking FGF2 and RA. For co-culture with iAstrocytes, filters were transferred to 12-well plates containing iAstrocytes and the same medium was utilized. Starting at this time, transendothelial electrical resistance (TEER) was measured using STX2 chopstick electrodes and an EVOM2 voltameter (World precision Instruments) approximately every 24 hours. TEER readings on empty Transwell filters were subtracted from all measurements to reflect the resistance of only the cultured cells.

Generation of hiPSC-derived neurons (iNeurons) and GCaMP iNeurons
hiPSC-derived neurons (iNeurons) were generated from WTC11 hiPSCs with stably integrated dox-inducible NGN2 (NGN2 iPSCs) according to Fernandopulle et al. 110 . To generate GCaMP iNeurons, NGN2 iPSCs were transduced with a lentivirus delivering GCaMP6m (gift from Dr. Michael Ward). To facilitate segmentation of neuron soma, an additional lentivirus transduction was performed to deliver pMK1334 containing a non-targeting sgRNA, which confers BFP expression localized to the nucleus. Clonal lines were then isolated by colony picking, differentiated to neurons according to Fernandopulle et al. 110 , and evaluated for homogeneity of GCaMP6m expression and the presence of spontaneous calcium oscillations. A clonal line satisfying the above criteria was selected for GCaMP imaging experiments.

Measurement of calcium activity in GCaMP iNeurons
Briefly, GCaMP iNeurons at day 3 of differentiation (see Fernandopulle et al. 110 ) were replated onto poly-D-lysine-coated 96-well plates (Corning cat. no. 354640) at 62,5000 cells per cm 2 in neuron media + 2 μg/mL doxycycline. For co-culture experiments, iAstrocytes were added on day 3 at 10,000 cells per cm 2 in an equivalent volume of ScienCell Astrocyte Media + 2 μg/mL doxycycline. For mono-culture experiments, an equivalent volume of neuron media + 2 μg/mL doxycycline was added. On day 6, half of the media was replaced with fresh neuron media + 2 μg/mL doxycycline; on day 10, half of the media was replaced with fresh neuron media without doxycycline. On day 17, the cultures were treated with vehicle control or IL-1α+TNF+C1q (using the same final concentration as described above for astrocyte experiments) by performing a half media change with fresh neuron media. Calcium activity in GCaMP iNeurons was recorded on day 18 and day 20 with an IN Cell Analyzer 6000, using a 20X 0.45 NA objective, 2x2 binning, environmental control set to 37 °C and 5% CO 2 , an aperture width of ~1 Airy unit, and 1 frame (800 ms exposure) collected per second for 40 seconds per field (1 field per well).

Isolation of synaptosomes and labeling with pHrodo
Synaptosomes were isolated from fresh Innovative Grade US Origin Rat Sprague Dawley Brains (Innovative Research, Inc.; Cat. No. IGRTSDBR) with the Syn-PER™ Synaptic Protein Extraction Reagent (ThermoFisher Scientific cat. no. 87793) according to the manufacture's protocol with minor changes. Briefly, 10 mL of Syn-PER Reagent supplemented with 1x protease inhibitor cOmplete Mini, EDTA free (Roche cat. no. 11836170001) and 1x phosphatase inhibitor PhosSTOP (Roche cat. no. 4906845001) were added per gram of brain tissue. Dounce homogenization was performed on ice and homogenate was transferred to a conical tube and centrifuged at 1200 × g for 10 minutes at 4°C. The pellet was discarded, the supernatant was transferred to a new tube, and the centrifugation step was repeated. The supernatant was then centrifuged at 15,000 × g for 20 minutes at 4°C. The supernatant was removed and the wet pellet was weighed. The synaptosome fractions were resuspended at a concentration of 50 mg/ml. 3 μ M of pHrodo Red, succinimidyl ester (ThermoFisher Scientific cat. no. P36600) was added to the synaptosome fraction and incubated for 45 min at room temperature in the dark. After diluting the solution 1:10 in DPBS, the synaptosomes were spun down at 2500 × g for 5 min. The supernatant was removed and then the synaptosomes were washed two times with DPBS. The pHrodo-labelled synaptosomes were resuspended in DMEM/F12 + 5% DMSO at a stock concentration of 50 mg/ml, aliquoted, and then frozen in liquid nitrogen for later use.

Measurement of synaptosome phagocytosis
For synaptosome phagocytosis experiments, pHrodo-labeled rat synaptosomes were used for iAstrocytes, and pHrodo-labeled iNeuron synaptosomes were used for Li et al. 19 and Krencik et al. 15 hiPSC-derived astrocytes. Briefly, astrocytes were incubated with pHrodo-labeled synaptosomes resuspended in the appropriate astrocyte media (ScienCell Astrocyte media for iAstrocytes, AMM for Li et al. and Krencik et al. astrocytes) at 1 mg/mL for 3 hours at 37 °C; for negative controls, some samples were pre-treated with 10 uM cytochalasin D (Millipore Sigma cat. no. C8273) for 15 min and also incubated with synaptosomes in the presence of 10 uM cytochalasin D to inhibit phagocytosis. After incubation with pHrodo-labeled synaptosomes, astrocytes were washed with DPBS, dissociated with Accutase, and pHrodo fluorescence was measured by flow cytometry. The gating strategy to determine the percent of phagocytic cells was based on the separation between the fluorescence histograms of samples treated or not treated with cytochalasin D.

Measurement of iNeuron viability in the presence of astrocyte conditioned media
Conditioned media was collected from iAstrocytes treated with vehicle control or IL-1α+TNF+C1q for 24 hours, spun down at 300 g for 10 min to remove dead cells, and transferred to day 17 iNeurons after removing the original iNeuron media. Unconditioned ScienCell Astrocyte Media was used as a negative control. After 72 hours, iNeuron viability was assessed by adding 10 ug/mL Hoechst and 1 μΜ TO-PRO-3 (ThermoFisher Scientific cat. no. T3605) in DPBS, incubating for 10 min at 37 °C, and then imaging on an IN Cell Analyzer 6000, using a 10X 0.45 NA objective, 2x2 binning, environmental control set to 37 °C and 5% CO 2 , an aperture width of ~1 Airy unit, 200 ms exposure, and 4-9 fields per well. The percent of dead neurons (stained by TO-PRO-3) was calculated after image processing and segmentation with CellProfiler (see Data Analysis section).

Antibody staining for flow cytometry
For antibody staining of cell-surface proteins (VCAM1, TFRC), hiPSC-derived or primary mouse astrocytes were dissociated with Accutase, washed with DPBS, incubated with conjugated primary antibodies for 20 min on ice in DPBS + 1% BSA, washed with DPBS, and then resuspended in DPBS + 1% BSA for flow cytometry. For antibody staining of intracellular proteins (C3), hiPSC-derived astrocytes were dissociated with Accutase, washed with DPBS, fixed with 2% paraformaldehyde for 10 min at RT, washed twice with DPBS + 0.5% Tween 20 (Millipore Sigma cat. no. P9461), incubated with unconjugated primary antibody for 20 min at RT, washed with DPBS + 0.5% Tween 20, incubated with conjugated secondary antibody for 20 min at RT, washed with DPBS + 0.5% Tween 20, and then resuspended in DPBS + 0.5% Tween 20 for flow cytometry. For dual staining of VCAM1 and C3, the protocol for intracellular protein staining was used.

Pooled CRISPRi screening
To identify transcriptional regulators of inflammatory reactivity, we created a custom sgRNA library targeting the human transcription factors 42 , using 5 sgRNAs with the highest predicted activity scores from Horlbeck et al. 45 per gene. The library was created by cloning a pool of sgRNA-containing oligonucleotides custom-synthesized by Agilent Technologies into our optimized sgRNA expression vector as previously described 36 . To screen against the druggable genome, we used the H1 sub-library from Horlbeck et al. 45 . The transcription factor and druggable genome libraries were packaged into lentivirus as previously described 105 . For each experimental replicate, ~10 million iAstrocytes were plated onto 4 Matrigel-coated 15-cm dishes, transduced with the lentiviral transcription factor or H1 sgRNA library so that 70-80% of cells were transduced, treated with vehicle control or IL-1α+TNF+C1q for 24 hours, and then incubated with pHrodo-labeled rat synaptosomes or stained for cell-surface VCAM1 (using the PE-Cy7 mouse anti-VCAM1 antibody) as shown in Fig. 3a. iAstrocytes were sorted into pHrodo high vs. low or VCAM1 high vs. low (top and bottom 30% of cells) using a BD FACSAria Fusion cell sorter at 5,000-10,000 events per second, and then pelleted for genomic DNA extraction. sgRNA abundances were then measured using next-generation sequencing as previously described 105 . The screens were performed with two experimental replicates per condition.
We chose to transduce iAstrocytes with sgRNA libraries at 70-80% transduction because we found that we could not enrich for sgRNA-transduced iAstrocytes with puromycin selection, which changed the morphology and behavior of iAstrocytes. Although a transduction rate of 70-80% is higher than the usual target of 30%, it is not problematic for the following reasons: (i) in this regime, most cells will only be transduced with 1 or 2 sgRNAs, (ii) the majority of sgRNAs do not have a phenotype because most genes don't have a phenotype in the screen; thus, the rate of cells with 2 sgRNAs that both cause phenotypes is very low, (iii) the screens are carried out at very high representation, i.e. each sgRNA is represented in hundreds of independent cells. Since pairings of sgRNAs during double infections can be assumed to be random, each sgRNA will mostly occur by itself or in combination with sgRNAs that don't cause a phenotype, and the few instances where it is paired with other sgRNAs that do have a phenotype constitute a negligibly small fraction that will not substantially affect the average phenotype measured in a pooled screen.
CROP-seq sgRNAs targeting the top hits from the CRISPRi screens and also candidate regulators selected based on literature were cloned into pMK1334. The concentration of each sgRNA plasmid was measured using the Qubit dsDNA HS Assay Kit on a Qubit 2.0 Fluorometer, and then the plasmids were pooled. iAstrocytes were transduced with lentivirus generated from the sgRNA pool at low multiplicity of infection so that <30% of cells were transduced, treated with vehicle control or IL-1α+TNF+C1q for 24 hours, and then sorted for sgRNA-transduced cells via FACS. Sorted iAstrocytes were then used as input for single-cell RNA-seq using Chromium Next GEM Single Cell 3' v3.1 reagents (10X Genomics cat. no. PN-1000121, PN-1000127, and PN-1000213), loading ~45,000 cells per reaction into four reactions, with two reactions for vehicle control-treated iAstrocytes and two reactions for IL-1α+TNF+C1q-treated iAstrocytes. To facilitate association of single-cell transcriptomes with sgRNAs, sgRNA-containing transcripts were amplified as described in Tian et al. 105 . The sgRNA-enrichment libraries were separately indexed and sequenced as spike-ins alongside the whole-transcriptome single-cell RNA-seq libraries on a NovaSeq 6000, recovering on average ~29,000 transcriptome reads and ~5,000 sgRNA-containing transcript reads per cell.

Measurement of cytokine concentrations
Conditioned media was collected from iAstrocytes, spun down at 300 g for 10 min to remove debris, and then frozen at -80 °C until analysis. The concentration of selected cytokines (IFN-α2a, IFN-β, IFN-γ, IL-6, CXCL10, GM-CSF) was measured by multi-array electrochemiluminescence using custom U-PLEX plates from Meso Scale Discovery following the manufacturer's instructions.

Immunostaining of human neuropathological samples
Alzheimer's disease Brain tissue collection procedures were approved by the Institutional Review Board at Vanderbilt University Medical Center, and written consent for brain donation was obtained from patients or their surrogate decision makers. A diagnostic post-mortem evaluation was performed to confirm the presence of Alzheimer's disease following the National Alzheimer's Coordinating Center (NACC) Neuropathology Data Form. Tissue for this study was flash-frozen in liquid nitrogen at the time of donation. Cryosectioned tissues were fixed with 4% paraformaldehyde, photobleached with a broad-spectrum LED array for two days, and then processed for immunofluorescence staining. Samples were permeabilized with 0.3% Triton X-100 in PBS, blocked with 10% goat serum in PBS, and incubated at 4 overnight with the following primary antibodies: mouse anti-C3 (1:200 dilution; Biolegend cat. no. 846302), chicken anti-GFAP (1:300 dilution; Aves Labs cat. no. GFAP), rabbit anti-VCAM1 (1:200 dilution; Abcam cat. no. ab134047). The following day, after washing with PBS, immunostaining was completed by a 1 hour room temperature incubation with secondary antibodies (goat anti-rabbit Alexa Fluor 488, goat anti-mouse Alexa Fluor 546, and goat anti-chicken Alexa Fluor 647; 1:1,000 dilution; ThermoFisher Scientific). After additional washes, tissue sections were mounted with the antifade Fluoromount-G medium containing DAPI (Southern Biotechnology). Images were acquired with a Leica DMi8 epifluorescence microscope. 6-8 fields were collected per sample.

Hypoxic ischemic encephalopathy
All human HIE tissue was collected with informed consent and in accordance with guidelines established by UCSF Committee on Human Research (H11170-19113-07) as previously described 111 . Immediately after procurement, all brains were immersed in PBS with 4% paraformaldehyde for 3 d. On day 3, the brain was cut in the coronal plane at the level of the mammillary body and immersed in fresh 4% paraformaldehyde and PBS for an additional 3 d. After fixation, all tissue samples were equilibrated in PBS with 30% sucrose for at least 2 d. After sucrose equilibration, tissue was placed into molds and embedded with optimal cutting temperature medium for 30 min at room temperature followed by freezing in dry ice-chilled ethanol. UCSF neuropathology staff performed brain dissection and its evaluation. The diagnosis of HIE requires clinical and pathological correlation; no widely accepted diagnostic criteria are present for the pathological diagnosis of HIE. HIE cases showed consistent evidence of diffuse white matter gliosis, as evaluated by the qualitative increase in the number of GFAP-positive cells in addition to the increased intensity of GFAP staining. For immunostaining of reactive astrocyte markers, tissue slides were bleached with UV overnight, rinsed with PBS for 10 minutes, incubated with blocking solution (10% normal goat serum + 0.2% Triton X-100 in PBS) for 1 hour at room temperature, and then incubated with the following primary antibodies overnight at 4 : mouse anti-C3 (1:200 dilution; Biolegend cat. no. 846302) + rabbit anti-GFAP (1:500 dilution; Agilent Dako cat. no. Z0334), or mouse anti-GFAP (1:500 dilution; Milipore Sigma cat. no. G3893) + rabbit anti-VCAM1 (1:200 dilution; Abcam cat. no. ab134047). Afterwards, the samples were rinsed three times with PBS + 0.2% Triton X-100 for 10 minutes each time, and then incubated with secondary antibodies for 1 hour at room temperature in the dark.

Immunostaining of mouse tissue Animals
All in vivo experiments were conducted according to protocols approved by the Chancellor's Animal Research Committee of the Office for Protection of Research Subjects at the University of California, Los Angeles. Mice were housed in a 12 h light/dark cycle with controlled temperature and humidity and allowed ad libitum access to food and water. Wild-type C57BL/6J mice (RRID: IMSR_JAX:000664) were compared with transgenic mice with the Stat3 gene conditionally deleted from astrocytes using a well-characterized 73.12 mGfap-Cre-Stat3-loxP mice on a C57BL/6J background 13 .
LPS-mediated neuroinflammation, euthanasia, and histology preparation E. coli-derived lipopolysaccharide (LPS) (Sigma L5024, O127:B8; Lot 0000084086 with >500,000 endotoxin units per mg) was dissolved and diluted to 1 mg/ml in sterile PBS and stored in aliquots at -80°C. LPS (5 mg/kg) was administered to mice by intraperitoneal injection and mice were euthanized 24 h later. Young adult male and female mice were used between two and four months of age. After terminal anesthesia by barbiturate overdose, mice were perfused transcardially with 4% paraformaldehyde (Electron Microscopy Sciences 15714). Brains were removed, post-fixed 4 hours, and cryoprotected in buffered 30% sucrose for 48 h.

Data analysis Pooled CRISPRi screens
We analyzed the data from the pooled CRISPRi screens using the MAGeCK-iNC bioinformatic pipeline previously described in Tian et al. 105 . Briefly, reads from next-generation sequencing were cropped and aligned to the reference using Bowtie to determine sgRNA counts in each sample. Next, counts files of samples subject to comparison were input into MAGeCK 112 and log 2 fold changes (LFCs) and P values were calculated for each sgRNA using the 'mageck testk' command, which takes into account variance among replicates. Following that, gene level knockdown phenotype scores were determined by averaging LFCs of the top 3 sgRNAs targeting a given gene with the most significant P values. The P value associated with each gene was determined by comparing the set of P values for sgRNAs targeting that gene with the set of p values for non-targeting control sgRNAs using the Mann-Whitney U test. To call hit genes, we first generated a null distribution by performing random sampling of 5 (the number of sgRNAs against each gene in the library) with replacement from non-targeting control sgRNAs to generate 'negative-control-quasi-genes' and calculated knockdown phenotype scores and P values for each of them. Then, we calculated the hit strength (gene score), defined as the product of knockdown phenotype score and -log (P value), for all genes in the library and for 'negativecontrol-quasi-genes' generated above. Based on the distribution of the gene scores, a cutoff value was chosen so that less than 10% of hit genes were negative control quasi-genes.

Master regulator analysis (MRA)
To collect human astrocyte RNA-seq data for co-expression network reconstruction, we downloaded gene-level counts from samples annotated as "Astrocyte" in ARCHS4 113 . For published human astrocyte RNA-seq datasets not found in ARCHS4, we manually downloaded FASTQ files from GEO, and then processed the FASTQ data using Elysium 114 , which implements the alignment pipeline used in ARCHS4. In addition to published human astrocyte RNA-seq datasets, we also used Elysium to process additional human astrocyte RNA-seq datasets from our labs (Ye Zhang lab, UCLA 41 , and Kampmann lab; all datasets used for MRA available at https://kampmannlab.ucsf.edu/inflammatory-reactive-astrocyte-analysis). We merged the gene-level count matrices from the above sources into a single matrix, which was then used for MRA (see Supplementary Table 11 for the metadata corresponding to all samples used for co-expression network reconstruction). Batch correction was first performed across datasets using ComBat-seq 115 , which is a part of the R (version 4.0.3) package "sva" (version 3.38.0). Gene-level counts were then transformed to log-scale using the variance-stabilizing transformation in DESeq2 (version 1.30.1). A list of human transcription factors (TFs) was obtained from Lambert et al. 42 , and a list of human kinases and phosphatases was obtained from Manning et al. 43 and Liberti et al. 44 . Genes and regulators (TFs, kinases and phosphatases) with low expression in hiPSC-derived astrocytes used in our study (mean transcripts per million < 1) were removed from the matrix. MRA was then performed using the R package "RTN" (version 2.16.0) [38][39][40] , following the RTN vignette available on Bioconductor.

CROP-seq
Alignment and quantification were performed on 10X single-cell RNA-seq libraries and sgRNAenriched libraries using Cell Ranger (version 5.0.1) with default parameters and reference genome GRCh38-3.0.0. Cellranger aggr was used to aggregate counts belonging to the same sample across different GEM wells. sgRNA unique molecular identifier (UMI) counts for each cell barcode were obtained using the pipeline described in Hill et al. 116 . sgRNAs were assigned to cells using demuxEM 117 . The gene vs. cell barcode matrix outputted by Cell Ranger was converted into a SingleCellExperiment (SCE) object using the read10xCounts function from the DropletUtils R package (version 1.10.388). sgRNA assignments were appended to the SCE metadata and filtered to only include cells with a single sgRNA. The SCE object was then converted into a Seurat object for subsequent analysis. Data normalization, log-transformation, and identification of highly variable genes were performed using Seurat::SCTransform 118 .
For exploratory analysis of the effect of knocking down each regulator targeted in the CROP-seq experiment, a separate Seurat object was created for each regulator consisting of iAstrocytes transduced with a non-targeting (NTC) sgRNA and iAstrocytes transduced with the sgRNA targeting that regulator. For each of these Seurat objects, the data was renormalized with SCTransform, dimensionality reduction was performed with Seurat::RunPCA (30 PC's retained) and Seurat::RunUMAP, and clustering was performed with Seurat::FindNeighbors and Seurat::FindClusters (resolution = 0.5). Regulators whose knockdown resulted in clear spatial separation between knockdown iAstrocytes and NTC iAstrocytes in UMAP were selected for further analysis. The PCA embeddings from the Seurat objects corresponding to the selected regulators were used for aligned UMAP 119 , which allows the effect of knocking down different regulators to be visualized in the same UMAP embedding. This aligned UMAP embedding was used for Fig. 4a,b,d and Fig. 5c, which displayed only NTC iAstrocytes. Markers of IRAS1 and IRAS2 in NTC iAstrocytes were identified by performing Student's t-tests on the Pearson residuals from SCTransform using Seurat::FindMarkers.
To find genes whose differential expression induced by IL-1α+TNF+C1q treatment was changed by regulator knockdown, we used the R package limma (version 3.46.0) to perform linear regression on the Pearson residuals from running SCTransform on the Seurat object containing all iAstrocytes assigned with a single sgRNA. We used the design formula y ~ regulatorKnockdown + cytokineTreatment + regulatorKnockdown:cytokineTreatment, where the interaction term regulatorKnockdown:cytokineTreatment reflects the effect of knocking down a given regulator on the differential expression induced by IL-1α+TNF+C1q treatment. For each regulator knockdown, genes with a statistically significant interaction term were extracted for the analysis presented in Extended Data Fig. 4d-e.
To construct the heatmap in Extended Data Fig. 4e, the log 2 -fold-change (LFC) values of the above DEGs were weighted by their associated multiple testing-corrected P-values (P adj ) to generate a gene score, where gene score = LFC * (1 -P adj ). Hierarchical clustering (Ward's method) was then performed on both the rows (genes) and columns (regulators) of the gene score matrix, using Pearson correlation as the distance metric. The heatmap was drawn using R package ComplexHeatmap (version 2.6.2) 120 .
Integration of iAstrocyte single-cell RNA-seq data with external single-cell RNA-seq datasets UMI count matrices of single-cell RNA-seq data from Barbar et al. 29 were downloaded from Synapse (syn21861229). Processed single-cell RNA-seq data objects corresponding to data from Wheeler et al. 83 were requested from Zhaorong Li and Michael Wheeler. UMI count matrices of single-cell RNA-seq data from Hasel et al. 58 were downloaded from GEO (GSE148611). Integration of NTC iAstrocytes from the CROP-seq experiment with the above previously published single-cell RNA-seq datasets was performed by first renormalizing each dataset with Seurat::SCTransform and re-running Seurat::RunPCA, then selecting features for integration using Seurat::SelectIntegrationFeatures (nfeatures = 3000) and Seurat::PrepSCTIntegration, and then performing data integration using reciprocal PCA with Seurat::FindIntegrationAnchors (normalization.method = 'SCT', reduction = 'rpca') and Seurat::IntegrateData (normalization.method = 'SCT'), using NTC iAstrocytes as the reference dataset. The integrated PCA embeddings were then used for clustering (resolution = 0.4) and UMAP. For integration with Barbar et al. 29 , the k.anchor parameter of Seurat::FindAnchors was set to 20; for Wheeler et al. 83 , k.anchor was set to 50; for Hasel et al. 58 , k.anchor was set to 100. Seurat::AddModuleScore was used to generate module expression scores for IRAS1 vs. IRAS2 markers in astrocytes from external single-cell datasets.
Bulk RNA-seq of hiPSC-derived astrocytes generated in this study Alignment and quantification were performed using Salmon (version 1.4.082), with the --noLengthCorrection flag and an index generated from the human transcriptome (GRCh38, Gencode release 37). The R package Tximport (version 1.18.083) was used to obtain gene-level count estimates. Genes with zero counts across all samples were removed from the analysis. Differential gene expression analysis was performed with DESeq2 (version 1.30.1).

Reanalysis of external bulk RNA-seq datasets
For Guttenplan et al. 9 , we downloaded the table of DEGs induced by IL-1α+TNF+C1q in immunopanned astrocytes from wild-type mice provided in GSE143598. We used BioJupies 121 to reanalyze bulk RNA-seq data and obtain differentially expressed genes from Perriot et al. 122 (GSE120411; hiPSC-derived astrocytes treated with IL1β+TNF), Barbar et al. 29 (syn21861229; CD49f+ astrocytes sorted from cerebral organoids treated with vehicle control or IL-1α+TNF+C1q), and Anderson et al. 51 (GSE76097; astrocyte-specific RNA from Stat3 cKO vs. wild-type mice subject to spinal cord injury). For Anderson et al., sample GSM1974209 was removed because it was an outlier in PCA visualization of the samples.

Pathway enrichment analysis
We used Enrichr [123][124][125] to perform enrichment analysis of gene lists. For display of enrichment results against the TRRUST gene set library, terms corresponding to mouse gene sets were removed for genes lists derived from human astrocytes; for genes lists derived from mouse astrocytes, terms corresponding to human gene sets were removed.

Image segmentation and analysis
We used CellProfiler (v3.15) 126 to segment and quantify neuron viability, neuron GCaMP recordings, and iAstrocyte immunostaining images. For neuron viability, monochrome images of Hoechst and TO-PRO3 were segmented to nuclei and TO-PRO3+ objects, respectively; TO-PRO3+ neurons (i.e. dead neurons) were determined by overlap of nuclei with TO-PRO3+ objects using the RelateObjects module in CellProfiler. For neuron GCaMP recordings, neuronal soma were segmented on the maximum intensity projection of all GCaMP images in the recording, nuclei were segmented on a single BFP image collected at the end of recording, and only nuclei that were contained within a GCaMP+ soma were retained using the RelatedObjects and FilterObjects modules in CellProfiler. The median GCaMP intensity per soma at each time point was then extracted for analysis. For each neuronal soma, ΔF was calculated as F t − F t-1 and F 0 was calculated as the average of a 10-second rolling window around F t .
For immunostaining of astrocyte markers, the total image intensity of GFAP, S100β, GLAST, NFIA, Cx43, glutamine synthetase, or vimentin was calculated and divided by the total image intensity of Hoechst to normalize for cell number. For immunostaning of C3 and IFIT3, nuclei were segmented from monochrome images of Hoechst followed by segmentation of cell boundaries from monochrome images of Na + /K + ATPase. The integrated intensity of C3 and IFIT3 was then calculated on a per-cell basis. Thresholds to define C3+ and IFIT3+ cells based on integrated C3 and IFIT3 intensity were determined from samples treated with vehicle control.
For analysis of images from immunostaining of human neuropathological tissue, monochrome images of DAPI, VCAM1, C3, and GFAP staining were converted to probability maps of nuclei, VCAM1+ objects, C3+ objects, and GFAP+ objects respectively using ilastik 127 ; annotations for representative nuclei, VCAM1+ objects, C3+ objects, and GFAP+ objects were inputted manually to train the classifier. The probability maps were then inputted into CellProfiler to segment nuclei, VCAM1+ objects, C3+ objects, and astrocytes (defined as nuclei-containing GFAP+ objects), and the VCAM1 or C3 staining status of astrocytes was determined by overlap of astrocytes with VCAM1+ or C3+ objects using the RelateObjects module in CellProfiler.
For analysis of images from immunostaining of mouse tissue, images from the Axio Scan.Z1 were imported into QuPath (v0.3.2) 128 . The corpus callosum or hippocampus were outlined, and then nuclei were segmented on the DAPI channel using the "Cell detection" module, with virtual cell boundaries created by expansion of the nuclei segmentation. Classifiers were then created to distinguish Gfap+, C3+, and Isg15+ cells by thresholding on the maximum intensity of the respective markers contained within the virtual cell boundary. The classifiers were then applied sequentially to obtain classify astrocytes (GFAP+ cells) as Isg15−/C3+, Isg15+/C3−, or Isg15+/C3+.

Beta regression, linear regression, and multiple testing adjustment of P values
When multiple comparisons needed to be made in a single experiment, we used beta regression for data consisting of percentages and linear regression for fluorescence values, cytokine concentrations, etc. For beta regression, which is appropriate for data ranged from 0 to 1, we first converted percentages to proportions. In cases where data occupied the extremal values of 0 or 1, we applied the transformation (y · (n − 1) + 0.5)/n, where n is the sample size, as previously proposed by Smithson & Verkuilen 129 . We then calculated P values using the betareg package in R (v3.1-4) 130 , using a logit link function (link = 'logit') and the bias-corrected maximum likelihood estimator (type = 'BR'). For linear regression, we used the function lm in R with parameters set to default values. Where appropriate, P values obtained via beta or linear regression were adjusted for multiple hypothesis testing with Holm's method using the function p.adjust in R (method = 'Holm').

Statistics and reproducibility
Sample sizes were determined by referencing existing studies in the field. Major findings were validated using independent samples and orthogonal approaches. Numbers of replicates are listed  . c, f, Clustering of transcription factors (c) or kinases and phosphatases (f) predicted to regulate inflammatory reactivity based on regulon overlap (see Methods); the activity score and regulon mean log-fold-change (LFC) associated with each predicted regulator (see Methods) are shown below the dendrogram. d, g, Scatterplot of gene scores (see Methods) of transcription factors (d) or the druggable genome (g) from synaptosome phagocytosis screens on iAstrocytes treated with vehicle control vs. IL-1α+TNF+C1q. e, h, Volcano plot of phenotype scores and associated log-scaled P values (see Methods) of transcription factors (e) or the druggable genome (h) from cell-surface VCAM1 screens on iAstrocytes treated with IL-1α+TNF+C1q. Fig. 4 (legend overleaf)   Fig. 4 | CROP-seq of iAstrocytes reveals two distinct inflammatory reactive signatures dependent on the canonical NF-kB pathway. a, Uniform manifold approximation projection (UMAP) of single-cell transcriptomes of iAstrocytes transduced with a non-targeting control (NTC) sgRNA treated with vehicle control or IL-1α+TNF+C1q, colored by cytokine treatment or cluster assignment. b, Visualization of transcript levels of selected Cluster 1 and Cluster 2 markers in the iAstrocytes shown in panel a, overlaid onto the same UMAP embedding. c, Cellular pathway (MSigDB 131 ) and upstream transcription factor (TRRUST 132 ) enrichment analysis of Cluster 1 and Cluster 2 markers; TF -transcription factor. d, Aligned UMAP embedding (see Methods) of NTC sgRNA transduced iAstrocytes with iAstrocytes transduced with sgRNAs knocking down RELA, IKBKG, or MAP3K7. e, Heatmap of Cluster 1 and Cluster 2 marker transcript levels in IL-1α+TNF+C1q-treated NTC sgRNA iAstrocytes compared to IL-1α+TNF+C1q-treated RELA, IKBKG, or MAP3K7 sgRNA iAstrocytes. f, VCAM1/C3 levels (MFI: median fluorescence intensity) or percent positive cells measured by flow cytometry in NTC, RELA, IKBKG, or MAP3K7 sgRNA iAstrocytes treated with vehicle control or IL-1α+TNF+C1q (n = 3 wells). g, Representative images of NF-κB p65 immunostaining or Hoechst counterstain in iAstrocytes treated with vehicle control or IL-1α+TNF+C1q. Scale bar: 60 μm. h, Quantification of NF-κB p65 nuclear localization (integrated fluorescence intensity masked to Hoechst; n = 3 wells). In panel c, the adjusted P values were derived from Enrichr 123 . In panel f, P values were calculated using linear regression for MFI values or beta regression for percentages (see Methods) and adjusted for multiple testing (P adj ; Holm's method) per family of tests (all comparisons made within a plot). In panel h, the P value was calculated using the twosided student's t-test. In panels b and d, NTC sgRNA astrocytes in Cluster 1 or Cluster 2 are circled by colored dotted lines.

Fig. 5 | IL-6 and interferons act in an autocrine-paracrine manner downstream of IL-
1α+TNF+C1q. a, Cytokine concentrations in conditioned media (measured by multi-spot electrochemiluminescence) from iAstrocytes treated with vehicle control or IL-1α+TNF+C1q (n = 3 wells). b, VCAM1/C3 levels (MFI: median fluorescence intensity) or percent positive cells measured by flow cytometry from NTC sgRNA iAstrocytes compared to iAstrocytes transduced with sgRNAs knocking down genes or gene pairs involved in IL-6 or interferon (IFN) signaling as indicated (n = 3 wells). c, Levels of phosphorylated STAT3 (Y705; p-STAT3) or phosphorylated STAT1 (Y701; p-STAT1) measured by flow cytometry in iAstrocytes treated with vehicle control or IL-1α+TNF+C1q (n = 6 wells for p-STAT3 and n = 9 wells for p-STAT1). d, Transcript levels of CSF2 and GM-CSF overlaid onto the UMAP embedding from Fig. 3a. e, Concentration of GM-CSF or CXCL10 in conditioned media from iAstrocytes transduced with sgRNAs knocking down genes or gene pairs involved in IL-6 or IFN signaling as indicated (n = 4 wells). In panels a and c, P values were calculated using the two-sided Student's t-test. In panels b and e, P values were calculated using linear regression for MFI values or beta regression for percentages (see Methods), and only comparisons with statistically significant differences are marked. Where appropriate, P values were adjusted for multiple testing (P adj ; Holm's method) per family of tests (all comparisons made within a plot). Fig. 6 | Differential regulation of distinct inflammatory reactive signatures by cytokines and cellular factors. a, Representative contour plot of VCAM1 and C3 levels measured by flow cytometry in iAstrocytes treated with vehicle control or IL-1α+TNF+C1q. b, Percent VCAM1−/C3+, VCAM1+/C3−, or VCAM1+/C3+ cells measured by flow cytometry in iAstrocytes treated with vehicle control vs. all possible combinations of IL-1α, TNF, and C1q, in the absence or presence of additional IL-6/IL6R chimera (25 ng/mL) or IFN-β (5 ng/mL) added concurrently (n = 6 wells). c, Percent VCAM1−/C3+, VCAM1+/C3−, or VCAM1+/C3+ cells measured by flow cytometry in NTC sgRNA iAstrocytes compared to iAstrocytes transduced with sgRNAs knocking down genes involved in IL-6 or IFN signaling (n = 5 wells). d, IL-6 or IFN-β concentration in conditioned media (measured by multi-spot electrochemiluminescence) from NTC iAstrocytes vs. iAstrocytes with knockdown of CEBPB, STAT3, IRF1, or STAT1/2 treated with vehicle control or IL-1α+TNF+C1q (n = 4 wells). e, Model of transcription factors and signaling pathways controlling inflammatory reactivity induced by IL-1α+TNF+C1q. In panels b, c, and d, P values were calculated using linear regression for MFI values or beta regression for percentages (see Methods), and only comparisons with statistically significant differences are marked. Where appropriate, P values were adjusted for multiple testing (P adj ; Holm's method) per family of tests (all comparisons made within a plot).  and hypoxic-ischemic encephalopathy (HIE) and are regulated by Stat3 in a mouse model of neuroinflammation. a-h, Post-mortem human brain sections. a,c, Representative immunofluorescence images of C3+ (a) or VCAM1+ (c) astrocytes (marked by arrowheads) in an AD case compared to an age-matched control. Staining for VCAM1 and C3 was performed simultaneously (the same image is shown twice with different marker combinations). Scale bars correspond to 20 μm. b,d, Quantification of the percentage of C3+ (b) or VCAM1+ (d) astrocytes in AD cases (n = 4 individuals) vs. age-matched controls (n = 4 individuals). e,g, Representative immunofluorescence images of C3+ (e) or VCAM1+ (g) astrocytes (marked by arrowheads) in a HIE case compared to an age-matched control. Staining for VCAM1 or C3 was performed separately. Scale bars correspond to 10 μm. f,h, Quantification of the percentage of C3+ (f) or VCAM1+ (h) astrocytes in HIE cases (n = 4 individuals) vs. age-matched controls (n = 4 individuals). i-n, Brain sections from mice injected with vehicle or LPS. i,l, Representative immunofluorescence images of GFAP in the corpus callosum (i) or hippocampus (l). Scale bars correspond to 100 μm. j,m, Representative images of C3, Isg15, and Gfap staining in the corpus callosum (j) or hippocampus (m) in wild-type (WT) or Stat3 astrocyte-specific conditional knockout (Stat3-cKO) mice (n = 4-6 mice per group). Staining for C3 and Isg15 was performed simultaneously (the same image is shown twice with different marker combinations). Representative Isg15−/C3+ (solid arrowheads), Isg15+/C3− (empty arrowheads), and Isg15+/C3+ astrocytes (striped arrowheads) are labeled. k,n, Quantification of the percentage of Isg15−/C3+, Isg15+/C3−, and Isg15+/C3+ astrocytes in the corpus callosum (k) or hippocampus (n). P values were calculated using the two-sided Mann-Whitney U-test in b, d, f, and h and beta regression in k and n. For beta regression in k and n, sex was included as a covariate.