From the early 20^th century up to the 1980s, the morphological appearance of astrocytes was the only readout of their role in neuropathology. Hypertrophy and increased GFAP content were generally regarded as reflections of a detrimental astrocyte phenotype. The advent of genetic engineering in the early 1990s opened a new phase of research based on astrocyte-targeted manipulation of gene expression. For example, depletion or overexpression of receptors, membrane proteins^33,34, cytoskeleton proteins³⁵, acute-phase proteins³⁶, heat-shock proteins³⁷, and transcription factors^38–40 in astrocytes or ablation of proliferative scar-border forming astrocytes⁴¹ was reported to modify (protect or exacerbate) the course of neurological diseases in mouse models. An important conclusion drawn from these studies is that the morphological appearance of astrocytes does not correlate with functional phenotypes or with their impact on other cell types. Moreover, the overall impact of reactive astrocytes on each disease is complex. For example, the manipulation of reactive astrocytes has resulted in improved outcomes^38,42,43, worse³⁵ outcomes, and no change⁴⁴ in mouse models of AD and MS^40,45,46. Plausibly, such differences arise from several scenarios: (i) pathways that ultimately exacerbate, attenuate, or have no impact on ongoing pathology occur in the same astrocyte, such that the selective manipulation of one pathway may mask, or secondarily impact, the manifestation of others; (ii) coexisting astrocyte subpopulations may have opposing effects on pathology⁴⁵; (iii) in neurodegenerative diseases, a spectrum of reactive-astrocyte phenotypes conceivably coexist in the same brain at a given time point because of the asynchronous progression of neuropathology in different brain regions; or (iv) the pathological impact of astrocytes is stage-dependent, as shown in mouse models of MS^40,45,46. Finally, pathways inducing astrocyte reactivity may be beneficial in one disease and detrimental in another. For example, activation of STAT3-dependent transcription is beneficial in neonatal white matter injury⁴⁷, traumatic brain injury³⁰, spinal cord injury^48,49, and motor neuron injury⁵⁰, but detrimental in AD models^42,43. That is, STAT3-mediated transcriptional programs may contribute to malfunctional astrocyte states in AD models and to resilient states in other conditions. We broadly define ‘astrocyte resilience’ as the set of successful astroprotective responses that maintain cell-intrinsic homeostatic functions in neural circuits (Table 2) while promoting both neuronal and astrocyte survival. Lastly, responses of reactive astrocytes may be maladaptive and result in malfunctional astrocytes, which, in addition to losing homeostatic functions, may also gain detrimental functions, thus exacerbating ongoing pathology⁶. Numerous mixed scenarios of malfunctional and resilient astrocytes plausibly exist, with multidirectional transitions among them.

Research in the last decade has begun to unravel specific functional alterations in reactive astrocytes underlying complex phenotypic changes. In normal conditions, astrocyte Ca²⁺-based responses, and downstream signaling via neuroactive mediators, exert multifarious effects on synaptic function and plasticity, neural-network oscillations, and, ultimately, on behavior^51,52. In pathology, various functional changes emerge. Astrocyte Ca²⁺ dynamics and network responses become aberrant in mouse models of HD⁵³, AD⁵⁴, and ALS⁵⁵, possibly contributing to cognitive impairment and neuropathology^43,53,56. Reactive microglia may shift astrocyte signaling from physiological to pathological by increasing production of tumor necrosis factor α (TNFα), thus altering synaptic functions and behavior⁵⁷. Functions lost or altered in reactive astrocytes include neurotransmitter and ion buffering in mouse HD models⁵⁸, communication via gap junctions in the sclerotic hippocampus of patients with epilepsy⁵⁹, phagocytic clearance of dystrophic neurites⁶⁰, and metabolic coupling by glycolysis-derived d-serine⁶¹ and lactate⁶² in mouse AD models. The excessive release of GABA by reactive astrocytes in AD⁶³ and Parkinson’s disease⁶⁴ may be a case of gain of detrimental function. Another example may be what is sometimes called ‘astrocyte neurotoxicity’, but we recommend using this term only when increased neuronal death is due to the verified release of an identified toxic factor by reactive astrocytes, and not merely due to loss of trophic or antioxidant support from astrocytes. An example is neuronal damage due to nitrosative stress caused by astrocyte-derived nitric oxide in MS³³. Finally, a classical gain of beneficial function is the restriction of immune cell infiltration in open injuries by scar-border forming reactive astrocytes⁷.

Transcriptomics has contributed to a fundamental discovery: astrocytes in the healthy brain are diverse and specialized to perform specific roles in distinct CNS circuits^14,65. Astrocyte diversity in healthy tissue arises from embryonic patterning programs or local neuronal cues¹⁴. Likewise, reactive astrocytes are also diverse, as unequivocally demonstrated by microarray-based^66–68 and RNAseq-based^48,69–71 transcriptomic profiling of mouse bulk astrocytes^48,66–70 or of astrocyte populations preselected according to cell-surface markers⁷¹. Such transcriptomic profiling specifically shows that reactive astrocytes adopt distinct molecular states in different disease models^48,66–70, in different CNS regions⁷⁰, and in brain tumors⁷¹. These studies also suggested complex functional changes in reactive astrocytes, including novel regenerative functions⁷⁰, proliferation, and neural stem cell potential⁶⁸, as well as loss of homeostatic functions⁶⁶. They have also identified drug candidates to establish the impact of altered astrocytic pathways in mouse models^68,70. Whether baseline astrocyte heterogeneity influences astrocyte reactivity is an outstanding question.

In one early transcriptome study⁶⁶ and its follow-up⁷², it was proposed that mouse astrocytes adopted an ‘A1’ neurotoxic phenotype after exposure to specific cytokines secreted by microglia exposed to lipopolysaccharide (LPS), whereas they acquire an ‘A2’ neuroprotective phenotype after ischemic stroke——two acute pathological conditions. Two correlative signatures of 12 genes with 14 pan-reactive genes were proposed as fingerprints identifying these phenotypes and, for A1 astrocytes, combined with thorough functional analyses in vitro⁷². Although the A1 and A2 phenotypes were not proposed to be universal or all-encompassing, they became widely misinterpreted as evidence for a binary polarization of reactive astrocytes in either neurotoxic or neuroprotective states, which could be readily identified in any CNS disease, acute or chronic, by their correlative marker genes in a manner similar to the once-popular, but now discarded, Th1–Th2 lymphocyte and M1–M2 microglia polarization theories⁷³. For multiple reasons, we now collectively recommend moving beyond the A1–A2 labels and the misuse of their marker genes. Importantly, only a subset, often a mix of A1 and A2 or pan-reactive transcripts, are upregulated in astrocytes from brains of study participants with HD⁷⁴ or AD^75,76, or from several mouse models of acute injuries and chronic diseases of the CNS^42,69,76,77. Moreover, the functions of these genes are not known, because, to date, no experimental evidence has causally linked any of the proposed marker genes of A1 or A2 astrocytes to either toxic or protective functions. Thus, the mere expression of some or even all these marker genes does not prove the presence of functions that these genes have not been demonstrated to exert. Specifically, complement factor 3 (C3) should not be regarded as a single and definitive marker that unequivocally labels astrocytes with a net detrimental effect. In addition, steadily increasing evidence indicates that any binary polarization of reactive astrocytes falls short of capturing their phenotypic diversity across disorders. For example, single-cell and single-nucleus RNAseq (scRNAseq and snRNAseq, respectively) studies in mouse models and human brains of chronic neurodegenerative diseases have unraveled numerous stage-dependent transcriptomic states in HD⁷⁴, AD^75,78, and MS⁴⁰ that do not clearly comply with A1–A2 profiles. In addition, advanced statistics using multidimensional data and co-clustering approaches reveals that the A1 and A2 transcriptomes represent only two of many potential astrocyte transcriptomes segregating along several latent variables⁷⁹. The analyses also indicate that multidimensional data are necessary to establish the distinctiveness of astrocyte phenotypes (Fig. 2). Characterization of the potentially extensive and subtle functional diversity of reactive astrocytes suggested by transcriptomic data is an important future goal.

Advances in human induced pluripotent stem cell (hiPSC) technology are being adapted to astrocyte research. Interestingly, astrocytes generated from hiPSC derived from fibroblasts obtained from patients with CNS diseases (usually with a genetic mutation causative of disease or a risk polymorphism) show pathological phenotypes, including dysregulation of lipid metabolism¹¹, alteration in the contents of the extracellular vesicles released by astrocytes⁸⁰, reduced autophagy⁸¹, or altered STAT3 signaling⁸². hiPSC-derived astrocytes are also amenable to study responses to viral infection⁸³ and to specific stimuli⁸⁴. Nevertheless, caution is in order, for more research is needed to establish hiPSC-derived astrocytes as bona fide models of human astrocytes and to determine whether they recapitulate the maturity as well as the temporal, regional, and individual heterogeneity of in vivo astrocytes. Importantly, not only are these cells removed from their original milieu, but the serum pervasively used in culture media may render them reactive⁸⁴. In addition, generation of astrocytes from neural stem cells is inherently difficult, and derivation and culture conditions have not yet been standardized, leading to diversity of clone phenotypes. Finally, aging-related neurodegenerative diseases should be modeled with astrocytes derived from cells from aged individuals, but, in this case, the epigenetic rejuvenation intrinsic to the reprogramming of adult cells arises as a confounding factor to be controlled for.

To define astrocyte phenotypes, all sources of heterogeneity should be considered and integrated with multidimensional statistical analyses (Fig. 2). ScRNAseq and snRNAseq are becoming established as valuable tools to gain insight into basal⁹⁴ and reactive-astrocyte heterogeneity (Fig. 1e)^40,78,95. Notably, isolation protocols may not always be optimal for astrocytes, resulting in low numbers of cells or nuclei being sequenced, and some highly relevant but weakly expressed transcripts, such as transcription factors and plasma-membrane receptors, may be overlooked, particularly in snRNAseq. Translation from scRNAseq or snRNAseq data to in situ immunohistochemical detection and functional validations is far from trivial, because the molecular profiles of astrocyte clusters and subpopulations partly overlap. Thus, instead of individual markers, signatures composed of a combination of markers with specified levels of expression or relative fold-changes are required to identify astrocyte phenotypes⁷⁴. Such signatures must be statistically validated to the point of predicting phenotypes. Alternatively, the diversity within astrocyte populations from mouse models may be dissected out by combining fluorescence-activated cell sorting (FACS) and cell-surface markers identified in screens⁷¹. Further, emerging spatial transcriptomics that allow the simultaneous in situ detection of numerous genes will be of value to study the heterogeneity of reactive astrocytes at local and topographical levels (Fig. 1f)⁹⁶. Importantly, molecular signatures based on the expression of genes or proteins need to be validated by assessing specific astrocyte functions (Table 2), since post-transcriptional and post-translational events critically shape functional outcomes. Functional validations should preferably be performed in vivo or using in vitro models closely mimicking human diseases. Classical knockout-, knockdown-, or CRISPR-based approaches to inactivate gene expression are available to gain insight into the impact on disease of a given pathway within previously identified astrocyte subsets⁴⁰.

An important implication of the disease-specific induction of distinct reactive astrocyte states is that the damage- and pathogen-associated stimuli from one disorder cannot be assumed to be active in another. For example, the now widely-used cocktail of factors released by LPS-treated neonatal microglia⁷² cannot be simply assumed to model reactive astrocytes in diseases other than neonatal septic shock due to infection by gram-negative bacteria. Likewise, exposure to Tau, amyloid-β, or α-synuclein needs to be carefully designed in vivo and in vitro to replicate the concentration, protein species, and combinations thereof found in patient brains. Acute metabolic damage with the mitochondrial toxin MPTP does not replicate chronic Parkinson’s disease (PD), to cite another example of in vivo inappropriate modelling. To complicate things further, the outcome of activating a signaling pathway may depend on the upstream stimuli⁸² or priming caused by previous exposure to other stimuli⁹⁷, perhaps through epigenetic control⁴⁰. Thus, careful selection of upstream stimuli is essential for appropriate in vivo and in vitro modelling of disease-specific reactive astrocytes. Finally, interventional strategies such as classical pharmacology^56,98, genetic manipulation^42,56, and biomaterials⁹⁹ are available tools to modify pathological signaling in reactive astrocytes for therapeutic purposes. Optogenetics²⁵ and designer receptors exclusively activated by designer drugs (DREADDs)²⁵ are potential tools to manipulate reactive astrocytes or to restore their aberrant Ca²⁺ signaling observed in mouse models of neurodegenerative diseases^53–55. However, it is unknown whether and how the changes in Na⁺, K⁺, Cl⁻, and Ca²⁺ fluxes and second messengers triggered by these approaches²⁵ modulate signaling cascades driving phenotypical changes of reactive astrocytes (for example, JAK–STAT and NF-κB pathways)⁶.

Although some basic functional properties of astrocytes have been shown to be evolutionarily conserved between humans and rodents¹⁰⁰, it is still critical to study patient samples and develop models of human reactive astrocytes, because morphological and transcriptomic comparisons have revealed prominent differences between mice and humans^101–103. In addition to astrocytes from postmortem samples and biopsies⁵⁹ (Fig. 1b), hiPSC-derived astrocytes, which can be generated with a fast protocol in 2D layers¹⁰⁴ or integrated in 3D systems such as spheroids and organoids^105–108, are rapidly becoming commonplace in basic research^11,82 and therapy development¹⁰⁹. Researchers need to be aware of the pros and cons of the various protocols available, as discussed in previous sections and elsewhere^110–112. Also, hiPSC glial mouse chimeric brains, in which hiPSC differentiate into human astrocytes, oligodendrocytes, and their progenitors, offer the possibility to study human astrocytes from patients in contexts amenable to in vivo experimentation^113,114. In addition, proteins released by injured astrocytes are currently being considered as fluid biomarkers of neurological trauma³¹. Biomarkers of reactive astrocytes in human disease will indeed be needed to demonstrate target engagement of future astrocyte-directed therapies in clinical trials. Emerging reactive-astrocyte biomarkers are either measured in blood or cerebrospinal fluid (for example, YKL-40)¹¹⁵ or used for brain imaging, such as MAO-B-based positron emission tomography (PET)¹¹⁶, which provides important topographical information (Table 1)¹¹⁷. Plausibly, disease-specific biomarker signatures rather than single ubiquitous biomarkers will be needed.

Computerized tools, including systems biology and artificial intelligence, are essential to organizing and interpreting the increasing wealth of high-throughput, multidimensional molecular and functional data from reactive astrocytes. Currently, molecular data (for example, -omics) can be transformed into mathematical maps by artificial intelligence¹¹⁸, thereby providing quantitative representations of the otherwise-vague notion of phenotypes. An example of functional data is 2D and 3D Ca²⁺ imaging that generates kinetic profiles and maps for single astrocytes and 2D or 3D networks (Fig. 1c)^119,120. Artificial intelligence can identify patterns of Ca²⁺ signaling in astrocytes^55,120. Multidimensional molecular and functional data have then two applications. First, multivariate analysis may unravel molecules, pathways, and variables shaping astrocyte phenotypes in acute versus chronic degenerative conditions, different disease stages, sexes, and CNS regions (Fig. 2). Second, these data can be used to predict the net functional outcome of a complex mix of potentially protective or deleterious pathways and identification of hubs, such as master transcription factors or epigenetic regulators, that, when activated, promote globally beneficial transformations. Importantly, the inhibition of detrimental pathways must not secondarily impair protective ones or damage basic astrocyte functions. Finally, no astrocyte-targeting therapy can be successful if it does not consider the complex interactions of reactive astrocytes with other CNS cells.