Morphological diversity of embryonic Drosophila glia

To determine the relationship between glial morphology and transcriptional signature, we began by characterising glial morphology across distinct Drosophila brain regions and developmental stages. We focused on glia in the VNC, akin to the vertebrate spinal cord, in the late embryonic (stage 17) CNS. The developing VNC contains 5 major glial classes: astrocytes, ensheathing glia, cortex glia, and 2 types of surface glia, which are all neuroectodermal in origin and express the marker reversed polarity (repo) [20]. We used enhancer trap Gal4 drivers expressed in each of these glial cell classes to sparsely label cells using the multicolour flip-out (MCFO) cassette. We then visualised single glial cell morphology at 0 hours after larval hatching (0h ALH) to assess morphological diversity both within and between classes (Fig 1). Note that in addition to the 5 major glial classes described above, the VNC contains a distinct class called the midline glia, which are a transient population found only during embryonic and larval stages [28–31]. Although midline glia express wrapper, otherwise known as a cortex glia marker [32–34], they do not resemble cortex glia in form or function but instead ensheath commissural axons and play critical roles in axon guidance and VNC morphogenesis [35]. Moreover, unlike the other major glial classes described above, midline glia are mesectodermal in origin [36,37] and do not express the pan-glial marker repo or the broad glial marker glial cells missing (gcm) [35]. Midline glia have been characterised extensively by several groups [35,38–42]; therefore, given their distinct origin and the ambiguity surrounding their functional classification, we instead focused our analyses on repo+ glia.

Optic lobe glia display morphological diversity within glial classes

In addition to the relatively simple VNC, we also focused on the adult optic lobe, which is more structurally complex than the VNC with 4 distinct neuropils called the lamina, medulla, lobula, and lobula plate (Fig 2A). Several other groups have characterised the morphology of each of the 5 major glial classes (perineurial, subperineurial, cortex, ensheathing, and astrocyte) present in the optic lobe in detail [24,25,27]. Therefore, we used previously characterised Gal4 lines [24,25,54] (S3 Fig) to generate MCFO clones to visualise glial morphologies and validate morphological diversity within each class (summarised in Fig 2; see Materials and methods for more details).

A transcriptomic atlas of embryonic and young adult Drosophila glia

To correlate glial cell morphology and transcriptional identity during development, we performed scRNA-seq on late-stage Drosophila embryos (stage 17). Whole embryos were dissociated into a single cell suspension, filtered, prepared using the 10X Genomics single cell pipeline, and sequenced via Illumina sequencing (see Materials and methods and S5 Fig). At this stage and through larval stage 1, we quantified an average of 528 ± 9 SEM glial cells in the CNS (labelled by the glial-specific marker Reversed polarity or Repo) that were localised in the VNC, and 188 ± 9 SEM glial cells in the brain lobes (N = 8 brains per anatomical region). Thus, our dataset was enriched for VNC glia, which make up 74% of the CNS glia at this time (p < 0.0001, one-way ANOVA). Following sequencing, cells were clustered based on differential gene expression using Seurat, and glial clusters were computationally isolated by expression of the pan-glial markers repo and the astrocyte-specific markers GABA transporter (gat) and astrocytic leucine-rich repeat molecule (alrm), except for the midline glia, which lack repo expression [35] (see Materials and methods and Fig 3 for further details). We bioinformatically isolated midline glia based on wrapper and single minded (sim) expression [34,56] and performed hierarchical cluster analysis on midline glia, all repo+ glial clusters and neuronal clusters (see Materials and methods and S6A–S6F Fig for more details). This analysis revealed that midline glia formed an outgroup to neuronal and repo+ glial clusters, perhaps not surprisingly given their distinct (mesectodermal) origin [36,37] (see S1 Data for genes with enriched expression in midline glia). We therefore continued all further analyses focusing on repo+ glia only but nonetheless validated wrapper expression in midline glia in vivo using wrapper-Gal4 (carried on a BAC insertion; [33]) (S6G and S6H Fig).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Summary of the experimental and computational workflow used in this study.

(Left) We performed scRNA-seq on whole stage 17 embryos. We then subsetted repo+ glial cells and eliminated neuronal contamination (see Materials and methods). This resulted in 1,594 cells and 11 clusters, which we proceeded to annotate by validating marker gene expression in vivo. (Right) We isolated and integrated repo+ glial clusters from 2 published datasets of dissected optic lobes [57,58] (see also S6 Fig). We eliminated neuronal contamination and clusters that contained cells from only the Özel dataset (see Materials and methods) (see also S7 and S8 Figs). This resulted in 2,356 cells and 15 glial clusters, which we went on to annotate by comparisons with cell type-specific bulk RNA sequencing datasets and by validating marker gene expression in vivo. The data underlying this figure can be found at NCBI GEO accessions GSE208324, GSE142787 and GSE156455, and GitHub: https://github.com/AustinSeroka/2022_stage17_glia and https://github.com/VilFernandesLab/2022_DrosophilaGlialAtlas. APF, after puparium formation.

https://doi.org/10.1371/journal.pbio.3002328.g003

For the optic lobes, we used data from 2 recent studies that performed scRNA-seq on whole optic lobe tissue at the young adult stage and throughout pupal development [57,58]. Despite slight differences in experimental approaches between these studies (see Fig 3 for summary), both analyses generated 19 glial clusters for the young adult optic lobes (S7A and S7B Fig); however, since both studies focused on neuronal development, glial clusters were not annotated or analysed [57,58]. Since glial cells constitute only 10% to 15% of neural cells in the Drosophila nervous system [25], we sought to increase the number of cells analysed to maximise our ability to identify rare cell types. Therefore, we isolated glial cells based on their original annotation in each study (i.e., repo expression) and then combined the closest developmental stages using the Seurat integration pipeline to remove batch effects between libraries (Fig 3 and Materials and methods). Here, we focused primarily on the 3-day-old adult dataset from (7,544 cells) [57] integrated with the 96 h after puparium formation (APF) dataset from (766 cells) [58]; hereafter referred to as the young adult dataset (Figs 3, S7C, and S7D).

On uniform manifold approximation and projection (UMAP) visualisations, we noticed that many clusters were connected by streams of cells, which expressed Resistant to dieldrin (Rdl), Frequenin 1 (Frq1), and Nckx30C (S8A Fig, top). These streams also co-expressed embryonic lethal abnormal vision (elav) and found in neurons (fne), whose expression is known to be enriched pan-neuronally (S8A Fig, bottom). Indeed, on UMAP visualisations Rdl, Frq1, and Nckx30C were expressed pan-neuronally (S8B Fig). We also detected elav, fne, Rdl, Frq1, and Nckx30C co-expressed in a subset of the embryonic dataset (S8C and S8D Fig). These data suggested that the Rdl, Frq1, and Nckx30C expressing glial cells could represent neuronal contamination. Therefore, we used in situ hybridization chain reaction (HCR) and MCFO clonal analyses to assess the expression of these genes in adult optic lobes and newly hatched larval VNCs, respectively, but failed to detect any expression in glial cells (S9A–S9D Fig). As glia interact closely with neurons throughout life, we hypothesised that the cells in our datasets that co-express Rdl, Frq1, and Nckx30C represented either glial cells contaminated by neuronal transcripts or neuronal cells contaminated by glial transcripts. Therefore, to circumvent potential clustering artefacts, we removed cells expressing high levels of Rdl, Frq1, and Nckx30C (see Materials and methods; S9E–S9H and S9J–S9N Fig). In addition, we noticed that a few cells expressed high levels of Hemolectin (Hml), a hemocyte-specific marker [59], which likely indicated contamination from a few stray hemocytes; therefore, we eliminated these cells also (see Materials and methods). Lastly, others have reported that glial clustering is sensitive to batch effects [57,58,60], and to further minimise these, we eliminated clusters to which the Kurmangaliyev dataset contributed fewer than 1% of the total number of cells in the cluster (S9I Fig). Following this elimination and reclustering (see Materials and methods), our datasets clustered into 11 embryonic glial cell clusters and 15 adult optic lobe glial clusters (Figs 3, S9N, and S10A).

Annotating embryonic glial clusters by validating marker gene expression in vivo

Based on known marker genes, we were able to make predictions about the identity of most of the embryonic clusters (Fig 4A and 4B). For example, cells in cluster 1 strongly expressed the gene wrapper, which other groups have previously reported as a cortex glia-specific marker with wrapper expression validated using a Gal4 driven by an intronic fragment from the wrapper locus (R54H02-Gal4) [61–64]. To validate the identity of each embryonic glial cluster, we acquired Gal4 lines for marker genes that were significantly enriched in one or more clusters (Fig 4A and 4B, see Materials and methods for complete list) and used these lines to generate MCFO clones (Figs 4C–4H and S11–S13). Embryos were heatshocked between 6 and 10 h after egg laying (prior to gliogenesis [65]), and larvae were dissected at 0 h after larval hatching. These MCFO analyses revealed near perfect specificity for the predicted glial cell type (Figs 4B and S11–S13). Indeed, by morphology and marker gene expression, 100% of CG6126-Gal4 MCFO clones and 17.9 ± 1.8 SEM% of pippin-Gal4 MCFO clones were perineurial glia; 100% of CG10702-Gal4 MCFO clones, 80 ± 8.9 SEM% Ntan1-Gal4 MCFO clones, and 100% of moody-Gal4 MCFO clones were exclusive to subperineurial glia. We note that similar to a previous report [66], pippin is also expressed in subperineurial glia at low levels (Fig 4B), though we did not recover any subperineurial clones with the pippin-Gal4 driver (S11I Fig; see below for further resolution of perineurial and subperineurial glial clusters). Furthermore, 98 ± 1 SEM% of Eaat2-Gal4 MCFO clones were ensheathing glia, 96 ± 1.3 SEM% of R54H02(wrapper fragment)-Gal4 MCFO clones were cortex glia, and 98 ± 1.5 SEM% of alrm-Gal4 MCFO clones were astrocytes, consistent with their expression in the scRNA-seq data (Figs 4 and S11–S13). Importantly, drivers inserted in genes that showed expression across multiple clusters always produced MCFO clones with morphologies that matched the predicted cluster identities. In other words, a gene that showed expression in both the predicted astrocyte and ensheathing glial clusters produced both astrocyte and ensheathing glia clones (S11–S13 Figs). In this way, we annotated 6 of the 11 clusters as surface-only perineurial glia, channel-associated perineurial glia, subperineurial glia (all morphological categories), cortex glia, ensheathing glia (including both neuropil ensheathing and tract ensheathing), and astrocytes and uncovered novel marker genes for the major glial classes (and subclasses; summarised in Fig 4B). Lobe (L), a known marker of peripheral nervous system glia [67], was expressed by 3 of the remaining unannotated clusters (clusters #0, #6, and #8). We therefore annotated these clusters as peripheral nervous system glia (PNSg_1, 2, and 3) but did not validate markers for these in vivo. In this way, we extended previous single-cell atlases, which included glial cells of the larval CNS and adult VNC [62,64], by validating marker gene expression in vivo and resolving surface glial classes into perineurial (surface-only and channel-associated) and subperineurial classes, which were indistinguishable in previous datasets.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Annotation of embryonic glial clusters in the VNC.

(A) UMAP of the 11 embryonic glial clusters labelled with both the cluster number (left) and our annotation of the specific glial class (and subtype; right) based on marker gene validation. (B) Expression plot of all marker genes selected for annotation, which were validated in vivo (except for Lobe). The size of the dot represents the percentage of cells with expression in each cluster, while the colour of the dot represents the level of average expression in the cluster. (C–H) MCFO clones (0 h after larval hatching) generated with the Gal4 lines indicated belonged to the main embryonic glial subtypes as accessed by their morphology. (C) Surface-only perineurial (N = 269 clones from N = 11 brains); (D) the newly defined perineurial glial subtype, termed channel-associated perineurial glia (N = 51 clones from N = 13 brains); (E) subperineurial, all morphological categories (N = 95 clones from N = 12 brains); (F) cortex (N = 295 clones from N = 13 brains); (G) ensheathing, both tract (t) and neuropil (n) (N = 190 clones from N = 11 brains); and (H) astrocyte (N = 117 clones from N = 36 brains). All clones labelled in cyan, with Repo in green and Prospero in magenta. Insets in (F–H) show Prospero and Repo in glial nuclei in alrm>MCFO clones in (H) were positive for Prospero. See S11–S13 Figs for additional in vivo marker gene validation. Dashed lines outline the neuropils. Prime panels show MCFO clones alone. Scale bars are 10 μm. The data underlying (A, B) can be found in S2 and S6 Data files. MCFO, multicolour flip-out; VNC, ventral nerve cord; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g004

Perineurial glia morphologies were transcriptionally distinct

Our previous clonal analysis unveiled 2 perineurial glia morphologies: surface-only perineurial glia and channel-associated perineurial glia (Figs 1A–1F and S1). Validating cluster marker expression in vivo revealed that surface-only and channel-associated perineurial glia were transcriptionally distinct, with cluster #4 corresponding to surface-only perineurial glia and cluster #10 corresponding to channel-associated perineurial glia. Clusters #4 and #10 were located adjacent to each other on the UMAP. PRL-1 and pippin were enriched in cluster #4 (Fig 4B) and Gal4 drivers for these markers predominantly labelled clones with surface-only perineurial glial morphology (84.6% of 16 brains contained exclusively surface-only perineurial glia clones labelled by PRL-1-Gal4; 87.5% of 13 brains contained exclusively surface-only perineurial glia clones labelled by pippin-Gal4; Figs 4B, 4C, S11C, S11D, and S11J). By contrast, cluster #10 expressed high levels of CG6126 and CG5080, but low expression of PRL-1 and pippin (Fig 4B). Consistent with these expression patterns, 100% of brains (N = 11 brains) contained both surface-only and channel-associated perineurial glia clones labelled by CG6126-Gal4, and 92.9% of brains (N = 13 brains) contained both surface-only and channel-associated perineurial glia clones labelled by CG5080-Gal4 (Figs 1B–1F, 4D, and S11J). Thus, the 2 perineurial glia morphologies are transcriptionally distinct.

Transcriptional profiles for subperineurial glia, ensheathing glia, and astrocytes did not distinguish morphological categories

Our prior clonal analyses revealed morphological heterogeneity in the subperineurial glia population, along a continuum from surface-only to channel-only (Fig 1H–1L), as well as 2 morphological categories within ensheathing glia: neuropil-only and tract-associated (Fig 1N and 1O), and 2 morphological categories within astrocytes: type 1 and type 2 (Fig 1P–1S). Interestingly, while our annotations revealed separate transcriptional clusters corresponding to surface-only perineurial glia and channel-associated perineurial glia (Figs 4A–4D and S11), we could resolve just 1 transcriptional cluster for subperineurial glia based on strong expression of CG10702, Ntan1, and moody (cluster# 9; Fig 4B and 4E); 1 transcriptional cluster for ensheathing glia based on strong expression of Eaat2 (cluster #7; Figs 1N, 1O, 4B, and 4G); and 1 transcriptional cluster for astrocytes based on strong expression of alrm (cluster #3; Figs 1P–1S and 4B). We questioned whether any of the markers enriched in our astrocyte cluster might distinguish type 1 from type 2 astrocytes. To this end, we generated MCFO clones under the control of enhancers for genes expressed in the astrocyte cluster. All drivers gave rise to clones containing both type 1 and type 2 astrocytes at the expected proportions based on our alrm MCFO study (approximately 38% type 1 and approximately 62% type 2), with the exception of pum, which gave a slightly higher proportion of type 1 astrocytes (52%), but still contained both clones (Figs 4A, 4B, 4H, S12, and S13). Similarly, we failed to identify any markers that distinguished the subperineurial glial morphologies or the ensheathing glial morphologies (Figs 4E, 4G, S11, and S13). Thus, although distinct perineurial glia morphologies corresponded to distinct transcriptional signatures, the same was not true for subperineurial, ensheathing, or astrocyte morphologies, with the caveat that scRNA-seq may fail to detect genes that are expressed at low levels.

Overall, these data suggest that morphological diversity cannot be equated with transcriptional diversity, at least at present levels of detection. As the developing Drosophila VNC contains a single, simple, neuropil, which may not accurately represent the diversity of more complex brain regions, we next turned to the more complex adult Drosophila optic lobe to assess how accurately cellular identity can be gauged by glial morphology.

Annotating young adult optic lobe glial clusters

Validating marker gene expression in vivo.

Next, we identified marker genes enriched in each cluster that would enable us to annotate all the clusters exhaustively (see Materials and methods and Fig 5B). We validated the expression of 31 marker genes in vivo using available enhancer trap lines, antibodies, and by in situ HCR (Figs 5C–5H, S10, and S15–S17). Since whole cell morphologies were not always visible when transcripts were visualised by HCR, we validated marker gene expression in specific glial types by visualising transcript expression by HCR with glial type Gal4 lines driving GFP (S3 Fig). In this way, we annotated all 17 clusters with 13 unique glial cell identities—fenestrated glia (lamina-specific perineurial), pseudocartridge glia (lamina-specific subperineurial), chalice glia (lamina-specific perineurial), distal satellite glia (lamina-specific cortex), proximal satellite glia (lamina-specific cortex), epithelial glia (lamina-specific astrocyte), marginal glia (lamina-specific ensheathing), chiasm glia (tract ensheathing), and medulla, lobula and lobula plate perineurial, subperineurial, cortex, ensheathing, and astrocyte glia—with some glial identities mapping to multiple clusters (Figs 5C–5H and S15–S17). Hereon, non-lamina glial classes are referred to as general perineurial, general subperineurial, general cortex, general ensheathing, and general astrocyte glia.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Annotation of young adult optic lobe glial clusters.

(A) UMAP of the 17 clusters obtained for the young adult optic lobe dataset, labelled with both the cluster number (left) and our annotation of the specific glial class based on marker gene expression validated in vivo (right). (B) Expression plot of all marker genes selected for annotation and validated in vivo. The size of the dot represents the percentage of cells with expression in each cluster, while the colour of the dot represents the level of average expression in the cluster. (C–H) Three examples of marker gene validation for the whole optic lobe, focussed on the lamina: (C, D) CG43795 gene trap drove GFP expression (magenta) in epithelial glia specifically. Repo in green and Elav in yellow. (E, F) GstT4 and (G, H) Dtg (both in magenta) expression were detected by in situ HCR in marginal and proximal satellite glia, respectively. GFP was driven by the indicated glial-Gal4 in green. See S10 and S15–S17 Figs for additional in vivo marker gene validation. Dashed lines outline the neuropils. Single focal planes in (C–H). Scale bars are 20 μm in (C, E, G) and 5 μm in (D, F, H). The data underlying (A, B) can be found in S3 and S7 Data files. HCR, hybridization chain reaction; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g005

We confirmed that clusters #9, #16, and #14 corresponded to epithelial, marginal, and proximal satellite glia, respectively, consistent with our previous comparisons to the glial cell type-specific transcriptomes (Figs 5B–5H, 6B, S16J, S17G, S17H, and S17K). Carpet glia did not appear to be represented as a unique cluster, indicating that either they are transcriptionally indistinguishable from another type of surface glia or that they were not sampled in these datasets. The latter is more likely since only 2 carpet glia are present in each optic lobe and they can be uniquely labelled with their own driver lines [70], suggesting that they are a distinct cell type. As well, inner and outer chiasm glia could not be resolved into distinct clusters though unique driver lines do distinguish between them [24]. Instead, both mapped to cluster #10. Given that these are also rare cells, it is possible that cluster #10 contains a heterogeneous population, but that insufficient cells were sampled to resolve them in the present dataset. Thus, our annotations revealed that morphological categories associated with the lamina neuropil across all glial classes were transcriptionally distinct.

General optic lobe glia do not subcluster by neuropil location

While astrocytes could be categorised into 8 distinct morphologies distributed across and within neuropils (including epithelial glia of the lamina) (Fig 2A and 2Q–2W), we only identified 3 transcriptional clusters of putative astrocytes. We annotated cluster #9 as the epithelial glia (astrocytes of the lamina), whereas clusters #0 and #2 co-expressed well-known astrocyte markers including Glutamine synthetase 2 (Gs2), ebony (e), nazgul (naz), alrm, and Eaat1 (and did not express Eaat2, a known ensheathing glia marker; S18 Fig). We wondered whether clusters #0 and #2, which were connected to each other on the UMAP visualisation, corresponded to different morphological subtypes of astrocytes or to astrocytes of different neuropils or neuropil regions. However, most differentially expressed marker genes showed enriched expression in cluster #0 compared to cluster #2; e.g., heartless (htl), which encodes an Fibroblast Growth Factor receptor and myoblast city (mbc) were expressed in cluster #0 only. We could not find unique marker genes for cluster #2 (Fig 6A and 6C). Therefore, we examined the distribution of astrocytes across and within neuropils (excluding the lamina) using a pan-astrocyte-specific Gal4 to drive GFP expression (R86E01>GFP). Since htl and mbc were expressed only in cluster #0 and not cluster #2, we then compared the distribution of GFP-positive astrocytes by region (as indicated in Fig 6G) to the distribution of htl expressing astrocytes. Surprisingly, although htl and mbc were expressed only in one of the 2 astrocyte clusters, we detected their expression in 92.4% and 93.6% of all astrocytes, respectively, and there was no bias in their regional distribution (Fig 6A–6D and 6G). The same was true for wunen 2 (wun 2), a marker expressed across both clusters #0 and #2 (Fig 6E, 6F, and 6G). Therefore, both clusters mapped indistinguishably to general astrocytes, with no apparent correspondence to neuropil or morphology. Similar to clusters #0 and #2, when we validated marker gene expression in vivo, we were also unable to distinguish between clusters #3, #5, and #7, which all appeared to correspond to general ensheathing glia (i.e., ensheathing glia of the medulla, lobula, and lobula plate; S16G–S16J Fig). Likewise, clusters #11 and #13 both mapped to general cortex glia (i.e., cortex glia in the optic lobe excluding the lamina; S15B–S15D Fig). Thus, transcriptional heterogeneity of the general optic lobe glia is not regionally defined.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6. Only 1 cluster within the astrocyte cluster multiplet corresponded to homeostatic astrocytes detected in vivo.

(A) htl expression levels plotted on a UMAP showing only cluster #0 (Astrocyte_1) and #2 (Astrocyte_2). Each dot represents a single cell, and the colour represents the level of expression as indicated. (B) Astrocyte(R86E01)>nls::GFP adult optic lobe in which all astrocyte nuclei were labelled with GFP and where htl expression (magenta) was detected by in situ HCR in most GFP positive nuclei. (C) UMAP of clusters #0 and #2 showing mbc expression. (D) Astrocyte(R86E01)>nls::GFP adult optic lobe where mbc expression (magenta) was detected by in situ HCR in most GFP (green) positive nuclei. (E) UMAP of cluster #0 and #2 showing wun2 expression. (F) Maximum projection of Astrocyte(R86E01)>nls::GFP adult optic lobe where wun2 expression (magenta) was detected by in situ HCR in most GFP (green) positive astrocyte nuclei of the medulla, lobula, and lobula plate. Dashed lines outline the neuropils and scale bars are 20 μm in (B, D, F). (G). Quantification of htl, mbc, and wun2 positive (medulla, lobula, and lobula plate nuclei) astrocyte nuclei (i.e., GFP positive nuclei) from (B, D, F). More than 90% of astrocyte nuclei were positive for the 3 genes. Different regions of the medulla, lobula, and lobula plate were defined, as these captured some differences in astrocyte morphotypes. The percentage of positive nuclei within each region was quantified and no significant difference (Chi-square test; htl p = 0.90, mbc p = 0.89, wun2 p = 0.33) was observed between the distribution of positive nuclei and the distribution of all nuclei in all 5 regions. The data underlying this figure can be found in S5 and S7 Data files and on GitHub at https://github.com/VilFernandesLab/2022_DrosophilaGlialAtlas. HCR, hybridization chain reaction; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g006

Transcriptional diversity in the general optic lobe glia reflects cellular state and not cellular identity

What then is the distinction between these clusters? Across model systems, glia appear to be more susceptible to stress than neurons during tissue dissociation [57,58,60,71]. Indeed, compared with neuronal clusters, glial clusters within the Özel and Kurmangaliyev optic lobe datasets were reported to be enriched for cells with low total gene counts per cell, enriched in mitochondrial transcripts, features of low-quality transcriptomes [57,58,60]. While standard cutoffs (see Materials and methods) for total and mitochondrial gene counts were used to filter out low-quality transcriptomes, we speculated that some of the glial clusters that mapped to the same cell type (hereafter referred to as “cluster multiplets”) likely represented the same cell type split by cell state (e.g., based on transcriptome quality or cellular stress). Therefore, we examined the total number of genes, the total number of reads, and the proportion of mitochondrial genes relative to the total number of genes for each cluster within a multiplet (S19A–S19C Fig). Although ensheathing glial clusters (#3, #5, and #7) and cortex glial clusters (#11 and #13) appeared to separate by transcriptome quality, we found no clear indication that general astrocyte clusters (#0 and #2) differed in this way (S19A–S19C Fig). Therefore, we performed Gene Ontology (GO) enrichment analysis on the genes differentially expressed within the general astrocyte clusters with a 1.2-fold cutoff. Interestingly, this revealed that a wide range of GO terms (from signal transduction and morphogenesis to growth and taxis) were enriched for cluster #0, whereas GO terms associated with mitochondrial regulation, autophagy, and metabolic processes were enriched for cluster #2 (S19D Fig). These data suggested that cells of cluster #2 (astrocyte_2) are unlikely to be a distinct cell type from those of cluster #0 (astrocyte_1), but instead may be cells in distinct metabolic states. Altogether, these data suggest that general astrocyte, ensheathing, and cortex clusters may be segregating based on transcriptome quality and/or cellular state, possibly driven by tissue dissociation. Furthermore, our data indicate that clusters #0 (astrocyte_1), #3 (ensheathing_1), and #11 (cortex_1) likely correspond to a more homeostatic state of general astrocyte, ensheathing, and cortex glia, respectively. We hypothesise that the generation of cluster multiplets occurred in the adult optic lobes due to the requirement to dissect these tissues prior to dissociation, resulting in greater tissue stress compared to our whole embryo dissociations.

Since our data indicated that cluster #0 (astrocyte_1) most likely corresponded to homeostatic astrocytes of the medulla, lobula, and lobula plate, we subclustered it to further probe for any potential heterogeneity (see Materials and methods and Fig 7A). Although we obtained 3 subclusters in this way, they differed only subtly from each other. We could identify only 1 marker gene (Actin 79B; Act79B) that was differentially expressed by at least 4-fold among the 3 subclusters (Figs 7B and S14). Specifically, subclusters #0 and #2 showed enriched expression of Act79B relative to subcluster #1 (Fig 7B). Although 96% of the cells in cluster #0 (astrocyte_1; homeostatic general astrocytes) expressed Act79B according to the scRNA-seq data, in vivo it was expressed very sparsely, in only 10.9% of medulla, lobula, and lobula plate astrocytes (Fig 7C and 7D). These cells were found more frequently in the proximal medulla, lobula, and lobula plate (Fig 7C and 7D). Act79B encodes what is thought to be a muscle-specific isoform of Actin [72,73] and therefore is unlikely to be a cell-fate determinant; we speculate instead that Act79B may be expressed by cells in a particular state of cytoskeletal remodelling (e.g., during process extension or growth). The apparent mismatch between the proportion of Act79B-expressing general astrocytes in scRNA-seq and in vivo may suggest that tissue dissociation pushes cells towards a state where Act79B is up-regulated, but that this state is relatively rare under homeostatic conditions (i.e., Act79B may label a transient cell state under homeostatic conditions) (Fig 7C and 7D). In sum, the unique and stereotyped morphologies of astrocytes of the medulla, lobula, and lobula plate cannot be assigned to unique transcriptional signatures within the present depth of sequencing limits and therefore are unlikely to represent distinct subclasses of astrocytes. Furthermore, any hidden heterogeneity that we have been able to uncover within general astrocytes appears to correspond to a cell state rather than cell type or identity.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7. Analysis of clusters obtained from subclustering cluster #0 (Astrocyte_1; homeostatic astrocytes).

(A) UMAP of the subclusters obtained from the subclustering of Astrocyte_1 (cluster #0). (B) Act79B expression levels plotted on the same UMAP. (C) Maximum projection of astrocyte(R86E01)>nls::GFP adult optic lobe where Act79B expression (magenta) was detected by in situ HCR in a sparse subset of GFP (green) positive astrocyte nuclei of the medulla, lobula, and lobula plate. Dashed lines outline the neuropils and scale bar is 20 μm. (D) Quantification of Act79B positive astrocyte nuclei (i.e., GFP positive nuclei) from (C) approximately 11% of astrocyte nuclei were positive for Act79B in vivo. The percentage of positive nuclei within each indicated region was quantified, showing Act79B was preferentially expressed in astrocytes of the lateral and proximal medulla, lobula, and lobula plate (Chi-squared test, p < 0.0001). The data underlying this figure can be found in S5 and S7 Data files and on GitHub at https://github.com/VilFernandesLab/2022_DrosophilaGlialAtlas. HCR, hybridization chain reaction; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g007

Immune response-related genes are enriched in lamina glia

Lamina glia are the main source of transcriptional diversity in the optic lobes, with more unique cell types associated with the lamina than all other neuropils combined (Fig 8A). Indeed, in addition to its own unique counterparts for the main glial classes, the lamina contains multiple perineurial (fenestrated and chalice), subperineurial (pseudocartridge and carpet), and cortex (distal and proximal satellite) glial subtypes. To investigate lamina glia specialisation further, we performed GO enrichment analysis for biological processes on pooled lamina glia and pooled general optic lobe glia (summarised in Fig 8B and Materials and methods). These comparisons revealed that compared to general glia, lamina glia are enriched for GO terms associated with immune responses and cell junctions (Fig 8B). Indeed, among the markers for lamina glia we validated in vivo in intact optic lobes were Tsf1, GILT1, and JhI-21, which are known to be involved in the antibacterial and/or antifungal immune response (S17D, S17M, and S17N Fig) [74–77]. By contrast, general glia were enriched for terms associated with metabolite transport, intercellular signalling, and migration relative to lamina glia (Fig 8B). Thus, lamina glia appear to be primed to perform immune-related functions relative to their general glial counterparts.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Lamina glia are specialised to perform immune-related functions.

(A) UMAP of the adult optic lobe glia colour-coded for lamina (maroon) and “general” (medulla, lobula, and lobula plate; blue) glial clusters that were included in the GO analysis in (B). Note that the chiasm glia cluster (dark grey) and non-homeostatic clusters belonging to multiplets (light grey) were excluded from the analysis in (B). (B) GO enrichment analysis for biological processes of the lamina glia compared to other (general) optic lobe glia shown as a heatmap. We used a dendrogram to cluster GO terms into groups that were defined by a super-term. The scale indicates the negative log of the FDR. The data underlying this figure can be found in S7 Data and on GitHub at https://github.com/VilFernandesLab/2022_DrosophilaGlialAtlas. FDR, false discovery rate; GO, Gene Ontology; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g008

Motor and sensory-associated glia are transcriptionally similar

Having shown that, apart from the lamina, optic lobe glia within each class are transcriptionally similar irrespective of morphology (e.g., all astrocyte morphotypes are transcriptionally similar, with the exception of lamina astrocytes), we sought to evaluate how different glia are between the young adult optic lobe and embryo, where developmental stage, circuits, and afferent inputs are vastly different. To address this question, we integrated the embryonic dataset with the optic lobe dataset, representing motor and sensory-associated glia, respectively. Note that our embryonic dataset, though generated from whole brains, did not contain any optic lobe glia as they are not born until late larval and pupal development [54]. Strikingly, we found that the embryonic clusters and their corresponding adult general glia clusters (i.e., not the lamina clusters) converged on the UMAP visualisation (Fig 9A–9C). Indeed, several of our validated marker genes were found in both embryonic glia and their counterparts in the adult optic lobe (summarised in Fig 9D). Interestingly, there was no overlap with lamina glia clusters, except for the embryonic channel-associated perineurial glia and the chalice glia, suggesting that they may be functionally analogous (Figs 9B and 2C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 9. Integration of the embryonic stage 17 CNS and young adult optic lobe glial clusters.

(A) UMAP of the integrated dataset, highlighting the embryonic (red) and young adult optic lobe (blue) glial clusters. (B, C) UMAP of the integrated dataset, highlighting embryonic (B, ‘e_’) or the young adult optic lobe (C, ‘a_’) glial clusters in colours by glial class. Grey marks cells from the young adult optic lobe clusters in (B) and from the embryonic clusters in (C). (D) Expression plot of genes with expression in the same glial class of the embryo and adult. The size of the dot represents the percentage of cells with expression in each cluster, while the colour of the dot represents the level of average expression in the cluster. The 5 genes indicated were validated in vivo for both time points by independent strategies: CG6126 for perineurial (embryo only-surface perineurial, S11B, S11I, and S11J Fig; adult perineurial, S16F Fig) and chalice (S16F Fig) and channel-associated perineurial (Figs 4E, S11I, and S11J), ltl for subperineurial (embryo, S11E, S11F, S11I, and S11K Fig; adult, S16C–S16E Fig), ana for ensheathing (embryo, S13A; adult, S16 Fig), wrapper for cortex (embryo, Fig 4F; adult, S15B Fig), and Tet for astrocyte (embryo, S12B, S12D, and S12E Fig; adult, S15A Fig). The data underlying this figure can be found in S6 and S7 Data files and on GitHub at https://github.com/VilFernandesLab/2022_DrosophilaGlialAtlas. CNS, central nervous system; UMAP, uniform manifold approximation and projection.

https://doi.org/10.1371/journal.pbio.3002328.g009

Altogether, our analysis uncovered a surprising transcriptional similarity between glia from different circuit types, brain regions, and time points in development. This suggests that, within each general class (astrocyte, cortex, ensheathing, etc.), glia likely fulfil similar functions from circuit to circuit and across neuropils.

Glial morphological diversity and detectable transcriptional diversity were not correlated

While we must be cautious in interpreting our results given the depth of sequencing constraints associated with scRNA-seq approaches, our data, nevertheless, point to a mismatch between glial morphological and transcriptional diversity. Specifically, glia exhibited more diversity at the morphological level than was detectable at the transcriptional level.

Although we detected striking morphological diversity in embryonic VNC subperineurial glia, this diversity could not be easily divided into discrete categories, as these glia clearly exhibited a continuum of morphologies from one extreme to the other (surface-only to channel-only) (Fig 1H–1L). Similarly, embryonic ensheathing glia also exhibited a range of morphologies from neuropil-only association to both neuropil and axon tract association to varying degrees (Fig 1N and 1O). Nonetheless, we detected only 1 transcriptional cluster corresponding to all subperineurial glia and 1 transcriptional cluster for all ensheathing glia. Even when glial morphologies could be divided into discreet categories such as for the newly hatched larval VNC and the adult optic lobe astrocytes (2 and 8 discreet morphological categories, respectively; Figs 1A, 1P–1S, 2A, 2Q–2W, and S4), detected transcriptional diversity did not match with morphological diversity. While astrocytes of the lamina were transcriptionally distinct, we identified only 1 transcriptional cluster of true, homeostatic astrocytes that corresponded to all remaining astrocyte morphologies at both developmental stages and brain regions. Indeed, when integrated, embryonic and adult glia co-clustered by class (except for lamina glia), despite the difference in developmental stage and origin, highlighting a striking convergence in their transcriptional profiles. This result was unexpected as the embryonic dataset was enriched for glia within a sensorimotor circuit [49–53], and the adult dataset was specific to the visual system. Thus, within the present limits of detection, our data suggest that glial classes are not transcriptionally specialised to the circuit types they associate with.

While at first glance, it may appear that some glial morphologies of the adult optic lobes were associated with unique transcriptional signatures (e.g., lamina astrocytes), the glial morphological categories with unique transcriptional signatures were all associated with the lamina. This suggests that the observed transcriptional specialisation is due to lamina association rather than specialised morphology. Interestingly, the glial transcriptional diversity we defined matches remarkably well with glial diversity described previously by glial-specific drivers from the Janelia GAL4 collection [25,78]. These driver lines also did not distinguish between cortex and ensheathing morphotypes of medulla, lobula, and lobula plate neuropils [25]. It will be interesting to explore if lineage relationships alone account for the transcriptional diversity detected in glia. Glial lineage relationships are characterised with detailed cellular resolution for the embryonic VNC where cell numbers are low, but are still incomplete for the optic lobes [24,45].

We hypothesise that glial morphology may be a plastic cellular feature that is set by the local environment and thus may represent a cellular state rather than an underlying transcriptionally defined identity. In the future, one could leverage genetic manipulations that alter neuropil structure to assess how equivalent glia might alter their morphology and transcriptional profiles in distinct environments. For example, one could assess glial morphologies and transcriptional profiles while (i) altering neuronal diversity and composition by manipulating the temporal series in medulla neuroblasts [79–82], or by blocking lamina neuron differentiation [83,84]; or (ii) by modifying neuropil stratifications with genetic perturbations that disrupt neuronal targeting [85].

Specialisation of the Drosophila lamina glia

Transcriptomic studies in mammals have uncovered regional differences in astrocyte transcriptional profiles (e.g., between and within the cortex and hippocampus) [10,11]. In our optic lobe data, we could only detect within-class transcriptional specialisation for glia associated with the lamina, whereas we could not detect any within-class transcriptional differences between (or within) the other optic lobe regions. This is particularly surprising since others have speculated that circuit complexity (stemming from neuronal diversity) may drive glial specialisation [6,86]; however, the lamina is the simplest of all neuropils, while the medulla is the most complex [55]. Thus, at least for the optic lobes, region-specific neuronal diversity does not appear to be a predictor of glial transcriptional diversity, within the present limits of detection.

Instead, our data revealed an up-regulation of GO terms related to immune functions in lamina glia when compared to other optic lobe glia (Fig 8B). The lamina is the first neuropil to receive input from photoreceptor sensory neurons [55,87] and, as such, its surface glia are densely perforated to allow photoreceptor axons to enter the lamina [22]. Thus, the lamina may be subject to more pathogenic insults compared to other optic lobe neuropils. This finding opens up the intriguing hypothesis that glia that interface with the environment may be specialised to manage environmental insults (e.g., heat, cold, lesions). Recently, a specialised population of cutaneous Schwann cells was identified which sense noxious stimuli and initiate pain sensation [88]. Furthermore, Schwann cells can migrate away from peripheral nerves to aid in cutaneous wound healing following injury, which is accompanied by an up-regulation of genes related to the innate immune response [89]. Defining the extent to which other glial cell populations at environmental boundaries adopt immune-related roles will broaden our understanding of the functional properties of glia not only for nervous system function, but also organismal health.

Limitations of this study

It is possible that all within-class glial morphological categories are transcriptionally distinct, but that the genes that confer this diversity are lowly expressed and therefore were missed due to depth of sequencing limitations. It is also possible that functional specialisation may occur locally in micro-domains of the same cell. For example, astrocytes are intimately associated with neuronal synapses. At the synapse, astrocytes regulate synapse strength and turnover of local neurotransmitters. As a single astrocyte supports many distinct types of synapses simultaneously, each requiring different molecular machinery [86], astrocytes must express a wide variety of channels and receptors (resulting in transcriptional homogeneity). Functional diversity could therefore be conferred by local synaptic activity. In other words, a single astrocyte may have functionally distinct microdomains dictated by the synapses it supports. Indeed, recent studies show that synapse activity regulates positioning of neurotransmitter receptors in astrocyte processes [90–92]. Functional studies, including in vivo Calcium, Glutamate, or GABA imaging will help to resolve glial functional specialisation both within and across cells.

A major anatomical difference between Drosophila and mammals is the size of glial domains relative to the size of the brain. Although only 10% to 15% of the Drosophila brain is made up of glia, astrocytes tile the entire CNS. For example, 6 individual astrocytes are sufficient to tile a single hemisegment of the larval VNC [26]. In contrast, thousands of astrocytes are required to tile an analogous section of the mouse spinal cord [93]. As mammalian astrocytes proportionally tile a much smaller region of the nervous system, it is conceivable that a single mammalian astrocyte can interact with one circuit type. In contrast, Drosophila astrocytes frequently span multiple circuits, which may require more functional flexibility. Thus, it is possible that further exploration of astrocyte diversity in vertebrates will identify unique transcriptional programs that specify morphologically distinct astrocytes with regionalized identities and functions. In sum, our work highlights the necessity for a systematic multimodal approach to characterising glial diversity in other systems [94,95].

Drosophila stocks and maintenance

Drosophila melanogaster strains and crosses were reared on standard cornmeal medium and raised at 25°C unless stated otherwise. We used the following strains in this study (see S3 File for more details):

hs-FLPG5;; 10×UAS(FRT.stop)myr::smGdP-HA, 10×UAS(FRT.stop)myr::smGdP-V5, 10×UAS(FRT.stop)myr::smGdP-Flag (hsMCFO; BDSC 64085), pBPhsFlp2::PEST in attP3;; HA_V5_FLAG_OLLAS (BDSC: 64086), CG6126-Gal4 (BDSC 67505), CG5080-Gal4 (BDSC 83275), CG10702/CG17343-Gal4 (BDSC 76234), Ntan1^Mz97-Gal4 (BDSC 9488), PRL-1-Gal4 (BDSC 65566), R54H02-Gal4 (BDSC 45784), Eaat2-Gal4 (BDSC 78932), alrm-Gal4 (BDSC 67031 and 67032), Rdl-Gal4 (BDSC 65421 and 66509), moody-Gal4 (BDSC 90883), R25H07-Gal4 (BDSC 49145), Tet-Gal4 (BDSC 19427), pum-Gal4 (BDSC 63368), w[*] TI{RFP[3xP3.cUa] = TI}Tre1[attP] (BDSC 84582), ana-Gal4 (BDSC 86394), CG9657-Gal4 (BDSC 78971), CG9449-Gal4 (BDSC 91274), pippin-Gal4 (BDSC 86401), ltl-Gal4 (BDSC 76144), PRL-Gal4 (BDSC 81151), hs-flp;lexAop-myr::tdTomato; repo-lexA, R85G01-Gal4 (BDSC: 40436), R54C07-Gal4 (BDSC: 50472), R10C12-Gal4 (BDSC: 47841), R47G01-Gal4 (BDSC: 45768), R50A12-Gal4 (BDSC: 47618), R46H12-Gal4 (BDSC: 50285), R53H12-Gal4 (BDSC: 50456), R56F03-Gal4 (BDSC: 39157), R35E04-Gal4 (BDSC: 48127), R86E01-Gal4 (BDSC: 45914), 55B03-Gal4 (BDSC: 39101), Repo-Gal4 (BDSC: 7415), Mi{MIC}CG43795-GFP[MI03737] (BDSC: 41395), 10XUAS-mCD8::GFP (BDSC 32184 and 32186), 10XUAS-IVS-myr::GFP (BDSC 32197) and UAS-nls::GFP (BDSC 4775 and 4776).

Embryo collections

For timed collections of embryonic and larval stages, crosses were reared in collection bottles fitted with 3% agarose apple caps coated with yeast paste. Embryos were then collected for 1.5-hour (h) windows and reared at 25°C until the desired developmental stage.

Multicolour flip-out clonal analyses (see S3 File for list of genotypes)

Embryonic/larval ventral nerve cord: To generate MCFO clones, we crossed virgin females of the hsMCFO line to males of Gal4 lines that were known to be expressed [26,47,96,97] or found to be enriched in a given transcriptional cluster (see S3 File for full genotype list). Embryos were collected for 1.5 h, aged for 6 h (embryonic stage 11/12, prior to gliogenesis [65]), and then heatshocked at 37°C for 15 min to induce FLP expression. Embryos were then transferred to 4°C for 5 min to halt FLP expression and transferred to 25°C until hatching. At 25°C, hatching occurs at 21 h after egg laying [98].

Adult optic lobe: For MCFO experiments of adult brains, we utilised Janelia Gal4 driver lines with specific glial expression, as previously described [25] (S3 Fig). We crossed males of the glial-Gal4 lines with virgin females of a 3-tag (HA, V5, and FLAG) MCFO line, except the cortex and marginal Gal4s which we crossed with a 4-tag (HA, V5, FLAG, and OLLAS) MCFO (S3 File). We raised progeny at 18°C and induced FLP expression in adult flies (0 to 5 days old) by heat-shocking at 37°C. Length of heat shocks varied depending on the Gal4 driver (S3 File). After heat-shocking, flies were placed back at 18°C for at least 1 night, before being dissected and stained as detailed in our immunohistochemistry protocol.

Immunohistochemistry

Larval VNC: For VNC studies, larvae were dissected at 0 h after larval hatching. We dissected larval brains in sterile-filtered, ice-cold 1× PBS. Brains were then mounted on 12 mm #1 thickness poly-d-lysine coated round coverslips (Neuvitro Corporation, GG-12-pdl) and fixed in fresh 4% paraformaldehyde (Electron Microscopy Sciences, 15710) in 1× PBS with 0.3% Triton detergent (0.3% PBST). We then washed the coverslips in 0.3% PBST to remove fixative and blocked overnight at 4°C in 0.3% PBST supplemented with 1% BSA (Fisher, BP1600-100), 1% normal donkey serum, and 1% normal goat serum (Jackson ImmunoResearch Laboratories 017-000-121 and 005-000-121). Brains were then incubated in primary antibody overnight at 4°C, washed overnight at 4°C with 0.3% PBST, and then incubated in secondary antibodies overnight at 4°C. We then removed the secondary antibodies, transferred the coverslips to 0.3% PBST overnight, and mounted in DPX. To mount in DPX, brains were dehydrated with an ethanol series: 30%, 50%, 70%, and 90%, each for 5 min, then twice in 100% ethanol for 10 min each (Decon Labs, 2716GEA). Samples were then transferred to glass-bottomed depression slides with xylenes (Fisher Chemical, X5-1) for 2 × 10 min. Finally, samples were mounted onto slides containing DPX mountant (Millipore Sigma, 06552) and cured for 1 to 2 days before imaging.

Adult optic lobe: For immunocytochemistry experiments of the adult optic lobe, we dissected whole brains in 1× phosphate buffered saline (PBS), fixed in 4% paraformaldehyde for 30 min, then washed with 0.5% PBTx (1× PBS with 0.5% TritonX). We next incubated the samples with primary antibodies diluted in block (5% normal horse serum), for 2 nights. Samples were washed with 0.5% PBTx, incubated for a further 2 nights with secondary antibodies diluted in block, washed again, and mounted in SlowFade (Life Technologies).

We used the following primary antibodies: Rabbit anti-Pros (1:1,000) [99], Rat anti-HA (1:100; Millipore Sigma, 11867423001), Chicken anti-V5 (1:1,000; Bethyl Laboratories, A190-118A), Mouse anti-Repo (1:50, DSHB 8D12), Mouse anti-Cherry (1:500; Clontech, 632543), Rabbit anti-Gat (1:4,000; a gift from M. Freeman) [23], Rat anti-FLAG (1:400, Novus NBP1-06712), Rabbit anti-HA-tag (1:400, Cell Signalling Technology, C29F4), Rat anti-Elav (1:20, DSHB 7E8A10), Mouse anti-V5-Tag:DyLight-550 (1:400, Bio-Rad), Chicken anti-GFP (1:400, EMD Millipore), Rabbit anti-GFP (1:400, Thermo Fisher A6455), DyLight 405 conjugated Goat anti-Horseradish peroxidase (HRP) (1:50, Jackson ImmunoResearch, 123-475-021), Rabbit anti-dsRed (1:500, Takara, 632496), and Mouse anti-FasIII (1:20, DSHB 7G10).

We used the following secondary antibodies at 1:400: Alexa Fluor Rhodamine Red-X Donkey anti-Mouse (Jackson ImmunoResearch, 715-295-151), Alexa Fluor 488 Donkey anti-Rabbit (Jackson ImmunoResearch, 711-545-152), DyLight 405 Donkey-anti Rabbit 405 (Jackson ImmunoResearch, 711-475-152), Alexa Fluor 488 Donkey anti-Chicken (Jackson ImmunoResearch, 703-545-155), Alexa Fluor Rhodamine Red-X Donkey anti-Rat (Jackson ImmunoResearch, 712-295-153), Alexa Fluor 647 Donkey anti-Mouse (Jackson ImmunoResearch, 715-605-151), and Alexa Fluor 647 donkey anti-Rat (Jackson Immunolabs, 712-605-153).

Embryo dissociation for single-cell RNA sequencing (scRNA-seq)

We prepared cell dissociates from embryos at 17 to 18.5 h after egg laying (stage 17). We washed embryos in deionized water before surface sterilising them in 30% bleach for 2 min. We then homogenised them in Chan-Gehring (C + G) medium by 6 to 8 strokes of a loose-fitting dounce. We filtered the cell suspension through a 40 μm Nitex mesh and pelleted cells in a clinical centrifuge at 4°C (setting 5, IEC). We washed the cell pellet twice by pouring off the supernatant and gently resuspending the pellet in fresh C + G, pelleting between each rinse as above. We determined the cell-survival proportion for each dissociate using the BioRad TC-20 trypan-blue assay. Samples that met a threshold of 80% viability were submitted for sequencing at a concentration of 1,000 cells per microliter.

Single-cell RNA sequencing of stage 17 embryos

The University of Oregon Genomics and Cell Characterization core facility (https://gc3f.uoregon.edu/) prepared embryonic cell samples for scRNA-seq. We ran dissociated cells on a 10X Chromium platform using 10X NextGem v3.1 chemistry targeting 10,000 cells per sample. Following cDNA library preparation, we amplified the library with 15 cycles of PCR before sequencing on 2 separate Illumina Hi-seq lanes, providing 2 technical replicates. Following examination for batch effects between technical replicates (S5 Fig), we merged the datasets using the CellRanger Aggregate function prior to quality control and downstream analysis. Reads were aligned to the Drosophila genome (BDGP6.22) and protein coding reads were counted.

Initial quality control and cell clustering of glia from embryonic scRNA-seq

We analysed the resulting sequencing data with the 10X CellRanger pipeline, version 3.1.0 [100], R version 3.6.3 and Seurat [101] version 3.1.2 using standard quality control, normalisation, and analysis steps. We filtered cells by the percentage of mitochondrial genes expressed (relative to the total number of genes expressed), indicating a high stress state. Only cells expressing <10% mitochondrial reads were retained for analysis. Additionally, cells containing reads for <50 and >3,000 unique genes were filtered out of downstream analysis. For each gene, expression levels were normalised by total expression, multiplied by a scale factor (10,000) and log-transformed, and the top 3,000 variable genes were identified for downstream principal component analysis (PCA). Clustering was performed using 50 PCs [FindClusters resolution 1.0], resulting in a Seurat object containing 52,881 cells. To isolate neuronal and repo+ glial cell clusters from all somatic tissues, we visualised the expression of elav, repo, Gat, and alrm in UMAP space. We then selected embryonic clusters 18, 2, 0, 4, 24, 26, 22, 16, 13, 14, 11, and 27 to be subset for further analysis. The resulting subset of cells were re-clustered using the previous parameters, resulting in the stage 17 embryonic neurons and repo+ glia dataset (19,600 cells). To isolate embryonic repo+ glia from embryonic neurons, we again visualised the expression of elav, repo, Gat, and alrm in UMAP space. From this object, clusters 23, 26, 34, 38, 17, 41, 27, 32, and 19 were identified as glial/non-neuronal. All the scripts described above are available at https://github.com/AustinSeroka/2022_stage17_glia. The resulting dataset (3,221 cells) was exported (.Rds) for subsequent analyses.

Midline glia are known to express sim and wrapper, but do not express elav or repo [35]. Therefore, to identify midline glia from the full dataset, we visualised sim, wrapper, elav, and repo expression in UMAP space. We identified a small group of cells meeting these expression criteria, which were clustered together and belonged to a larger cluster (cluster 19). We selected these cells as putative midline glia and defined them as a new cluster within the whole embryonic dataset. We pooled midline glia together with the stage 17 embryonic neurons and repo+ glia dataset described above and performed hierarchical clustering analysis [hclust] on this combined dataset, using average cluster gene expression to produce a Euclidean dissimilarity matrix [dist]. We used [FindAllMarkers] (default parameters, except only.pos = TRUE) to obtain the most highly expressed genes for putative midline glia relative to all other embryonic cells (see S1 Data).

Preliminary integration of published optic lobe glial single-cell RNA sequencing datasets

Two independent studies published scRNA-seq datasets of the optic lobes throughout pupal development with a focus on neuronal development and therefore did not analyse glial clusters [57,58]. We took advantage of the glial cells within these datasets here. Özel and colleagues performed scRNA-seq on optic lobes at specific stages in triplicate, profiling a total of 275,000 cells [57], whereas Kurmangaliyev and colleagues performed multiplexed scRNA-seq of many developing brain samples in parallel and profiled 51,000 cells in total [58]. The 2 datasets (hereafter referred to as the Özel and Kurmangaliyev datasets, respectively) thus varied dramatically in the number of cells they each profiled. Here, we chose to focus specifically on optic lobe glia from developmental stages that most closely match the young adult, i.e., young adult dataset from [57] and 96h APF dataset from [58] We used R (version 4.1.0) and Seurat (version 4.0.3) [101] to analyse the optic lobe scRNA-seq datasets from NCBI GEO accession GSE142787NCBI GEO accession GSE156455 [57,58], and [] enclose the specific Seurat functions used.

We converted the Kurmangaliyev 10X Genomics data to an RDS file [CreateSeuratObject] and subsetted [subset] the cells belonging to the 96h APF time point (within metadata, under “time”). Glial clusters were subsetted from both the 96h APF-Kurmangaliyev dataset (within metadata, under “class”) and adult-Özel dataset based on the original annotation; i.e., repo expression. Since the 2 datasets represent glia from the same structure and from very close developmental stages, we used Seurat’s Integration pipeline to find corresponding cells and batch correct the 2 datasets (Fig 3). We normalised each glial dataset [NormalizeData] and selected the 2,000 most variable features [FindVariableFeatures]. We defined Anchors [FindIntegrationAnchors, dims = 1:30] and integrated the datasets [IntegrateData, dims = 1:30]. We then scaled the integrated dataset [ScaleData] and ran PCA [RunPCA, npcs = 30]. We checked elbow plots [ElbowPlot] to determine the number of dimensions to use in [FindNeighbors, dims = 1:16], followed by [FindClusters, resolution = 0.5]. We obtained 20 clusters and used a UMAP dimensionality reduction [RunUMAP, reduction = pca, dims = 1:20] to plot the cell clusters and visualise their distribution as a 2D representation [DimPlot] (Figs 3, S7C, and S7D).

Further clean-up of young adult optic lobe and embryonic glial datasets

Re-clustering the embryonic glial dataset

Following the clean-up described above for the embryonic glial dataset, we normalized the data [NormalizeData], selected the 2,000 most variable features [FindVariableFeatures], scaled the data [ScaleData], and ran PCA [RunPCA, npcs = 30]. We examined elbow plots [ElbowPlot] to determine the number of dimensions to use in [FindNeighbors, dims = 1:16], followed by [FindClusters, resolution = 0.5]. We obtained 11 clusters and plotted them as a 2D representation ([RunUMAP, reduction = pca, dims = 1:20]; [DimPlot]; Fig 3; see also S9L–S9N Fig). We used [FindAllMarkers] (default parameters, except only.pos = TRUE) to obtain the most highly expressed genes for the embryonic glia (see S2 Data).

Re-integrating the young adult optic lobe dataset and eliminating clusters originating from a single dataset

Following the clean-up described above for the young adult optic lobe glial dataset, we obtained a list of the cells remaining using [WhichCells] (all identities, by default ident argument) and isolated these cells from the original 96h APF-Kurmangaliyev and young adult-Özel datasets (using [subset] and lists of cells used as cells argument). We then integrated these as described previously, [FindNeighbors, dims = 1:19] and [FindClusters, resolution = 0.5] (S9I Fig). We obtained 19 clusters. Since others have reported that glial clustering is sensitive to batch effects [57,58,60], we eliminated clusters to which the Kurmangaliyev dataset contributed fewer than 1% of the total number of cells in the cluster (S9I Fig; a total of 850 cells were eliminated) and reintegrated the remaining cells as described above (Fig 3). Following this, the young adult optic lobe glial dataset consisted of 15 clusters (Fig 3).

Comparing cell type-specific bulk RNA sequencing to scRNA-seq

Cell type-specific transcriptomes, obtained by Tandem-affinity purification of intact nuclei sequencing, were published for the marginal glia, epithelial glia, and the proximal satellite glia [68] (NCBI GEO accession #GSE116969). We simulated the FACS-sorted glial type-specific transcriptomes as single-cell transcriptomes to enable us to compare these deep bulk RNA sequencing datasets with the shallower scRNA-seq dataset, as described in [63]. Briefly, a random number of reads were assigned to each of 900 simulated cells, from a normal distribution with the same mean and standard deviation as the cells in the final Adult-96hAPF dataset. The probability of expression of each gene was obtained by dividing the number of transcripts per million (TPM) of each gene by the total, in each cell-type, and this probability was used to allocate the number of reads of each simulated cell. The simulated matrix of expression was transformed into a RDS object, normalised [NormalizeData] and average expression of each gene was calculated [AverageExpression]. We then examined the Pearson correlations between the average expression of the simulated single-cell and the average expression of the scRNA-seq clusters to determine the best match.

Subclustering individual glial clusters

To analyse individual young adult optic lobe glial clusters, we used [WhichCells] to obtain the list of cells belonging to a specific cluster and subsetted those cells from the original Kurmangaliyev and Özel datasets [subset]. We then integrated these as described above. The arguments dims, k.score and k.filter in [FindIntegrationAnchors], and dims and k.weight in [IntegrateData], were assigned a value of 1 unit lower than the smallest number of cells in either the Kurmangaliyev or Özel subsetted cells. Next, we used [FindNeighbors] and [FindClusters] (see Fig S14 for specific dimensions and resolutions used).

Adding subclusters to main young adult optic lobe glial dataset

S14 Fig summarises the number of subclusters obtained following subclustering of individual glial clusters. To account for over-clustering artefacts, we examined the number of genes that were differentially expressed between subclusters. Only the subclusters of clusters 8 and 9 expressed greater than 20 genes differentially (4-fold change or higher), whereas all other subclusters expressed fewer than 15 genes differentially. Therefore, we proceeded to manually divide clusters 8 and 9 accordingly (S10F–S10Q Fig). We obtained lists of cells belonging to each of these subclusters using [WhichCells] and used these lists to define new cell clusters in the full dataset, thus generating 17 clusters in total (Fig 5A).

Finding markers for clusters in the embryonic and young adult optic lobe glial datasets

We identified marker genes expressed by each glial cluster using [FindMarkers]. To find genes with highly specific expression, we focused on the top 30 differentially expressed transcripts. We then examined their expression within and across clusters [FeaturePlot] to choose transcripts that were expressed by most cells in a given cluster (see Figs 4B and 5B for a summary of the marker genes validated in vivo). We used [FindAllMarkers] (default parameters, except only.pos = TRUE) to obtain the most highly expressed genes for the young adult optic lobe glia (see S3 Data).

In situ hybridisation chain reaction (HCR) probe design

To assess the expression of marker genes in vivo, we designed HCR probes against chosen marker genes (Fig 5B). We designed 6–21 antisense probe pairs against each target gene, tiled along the annotated transcripts but excluding regions of strong sequence similarity to other transcripts (see S4 Data), with the corresponding initiator sequences for amplifiers B3 and B5 [102]. We purchased HCR probes as DNA oligos from Thermo Fisher (at 100 μm in water and frozen).

In situ hybridisation chain reaction (HCR)

We dissected optic lobe–central brain complexes from adult flies (female and male) in 1× PBS. We then fixed them in 4% formaldehyde for 35 min at room temperature. We then rinsed (3×) and washed (3× 30 min) them in PBSTx (1× PBS with 0.5% Triton X-100, Fisher BioReagents). Next, we transferred the optic lobe–central brain complexes to 1.5 mL Eppendorf tubes and followed the Multiplexed HCR RNA-FISH protocol (Detection and Amplification stages) for whole-mount fruit fly embryos, from Molecular Instruments (molecularinstruments.com), also described in [102] with the following modifications: we used 2 pmol of stock probe set and 12 pmol of each hairpin. Before proceeding with the 30 min washes with 5× sodium chloride sodium citrate (SSC) with 0.1% Tween 20 (SSCT), we incubated samples with in 5× SSCT (with DAPI) for 2 h at room temperature. All HCR buffers and hairpins were purchased from Molecular Instruments. Samples were stored at 4°C and mounted in SlowFade Gold Antifade Mountant (Thermo Fisher) within 3 days of completing the protocol.

Microscopy and image processing

Larval VNC: We used a Zeiss LSM700 point-scanning or an Intelligent Imaging Innovations (3i) spinning disc confocal microscope with a 63× objective to image all samples with a step size of 0.2 to 0.38 μm and the following laser lines: 405, 488, 555, and 647. Images were acquired using Zen Black or Slidebook, respectively. Samples were imaged to encompass the entire dorsal-ventral extent of the VNC.

Young adult optic lobe: We used a Zeiss LSM800 or LSM880 confocal microscope with 20× or 40× objectives to image whole optic lobes with a maximum step size of 1 μm. Individual astrocytes for morphological quantifications were imaged on a Zeiss LSM980 confocal microscope with AiryScan 2. Briefly, we acquired z-stacks (step size 3 to 4 μm) of whole optic lobes containing isolated MCFO-labelled astrocytes to determine astrocyte cell body position (using anti-Repo) along with information on the neuropil layers occupied by individual astrocytes (using anti-Brp). We then acquired z-stacks (optimal step size) of individual astrocytes at high resolution.

Quantifications and statistical analyses

Embryonic astrocyte morphological quantifications: To quantify the volume of type 1 versus type 2 astrocytes, we used Bitplane Imaris (version 10.0.0) to render 3D constructions of individual astrocyte MCFO clones (genotype: hsMCFO + alrm-Gal4). We used the [Surfaces] function with default parameters to measure total cell volume. Primary branches were quantified manually by segmenting astrocytes, 5 microns at a time, using the Imaris [Ortho Slicer] function.

Adult optic lobe astrocyte volumetric quantifications: We used Bitplane Imaris (version 9.9.0) to render 3D reconstructions of individual astrocytes. We used the [Surfaces] function with default parameters to measure total cell volume. We then aligned cells using the [Frame] function and cut the object (the cell) by neuropil layers (binning 2 layers together), and measured the volume for each was section of the cell. We quantified 8 cells of each astrocyte morphological category. Primary branches were quantified manually by segmenting astrocytes as detailed above. We used Graphpad Prism 9 to analyse the data for statistical significance as indicated in the main text.

Larval CNS: In Bitplane Imaris (version 9.6.1), a 3D projection was created for each individual VNC. For quantification of Repo+ cells in the VNC versus brain lobes, we used the Imaris “Spots” function to automatically reconstruct and quantify the number of Repo+ nuclei (xy diameter set to 3 μm, z spread set to 5 μm, manual thresholding). We quantified Repo+ cells in 7 independent VNC samples and 7 independent brain lobe samples at hatching (Genotype: hsMCFO + fne-Gal4). VNC enrichment was then determined via a one-way ANOVA in GraphPad Prism 9. For morphometric analyses, subtypes and morphotypes were quantified manually, blinded to genotype, and assessed statistically in Graphpad Prism 9 by Chi-squared statistical tests.

Adult optic lobe astrocyte cluster validation: We used the following approach to quantify the proportion of total astrocytes expressing a particular marker gene and their positions within the optic lobe. In Fiji (ImageJ2 version 3.2.0), we chose 35 to 50 optical slices [Make Substack] from each Z-stack. We performed a standard background subtraction [Subtract Background] (default parameters). We then created an average intensity projection and measured the mean fluorescence intensity of the HCR probe channel to obtain a Normalisation value (see below). We then used Icy (version 2; icy.bioimageanalysis.org) for further analysis. In Icy, we defined neuropil regions (as in Figs 6 and 7 for astrocytes), we then used the Spot Detector plugin (spot scale of 13 or 25 pixels in size and 40 to 100 sensitivity) to identify glial nuclei marked by nuclear GFP (R86E01-Gal4>nls::GFP) within these neuropil regions. We manually removed nuclei from the central brain that were detected by this method to restrict our analysis to the medulla, lobula, and lobula plate only. We then quantified the mean fluorescence intensity of mRNAs detected by HCR in the defined ROIs. We scaled these measures to the normalisation value (defined above) and used a threshold of 2.25, above which cells were considered positive for expression of the specific transcript examined. We settled on this threshold empirically by analysing several samples for markers that these cells were known to express or not express. We then calculated the proportion of nuclei positive for a given transcript in each optic lobe (excluding the lamina) as well as the proportion of positive nuclei within different neuropil regions. We used GraphPad Prism 9 to analyse these data with Chi-squared statistical tests.

Gene Ontology enrichment analyses

All GO enrichment analyses were carried out in R studio (version 1.4.1717), using R (version 4.1.1). We identified differentially expressed genes in our cluster/s of interest using [FindMarkers] from Seurat (version 4.1.1) [101]. For comparisons of one single cluster to another, markers were selected with a log2FC ≥ 0.25 (fold change of 1.2). When comparing larger groups of clusters (i.e., the pooled lamina glia clusters to the pooled general glia clusters), a more stringent threshold of 1 log2FC (fold change of 2) was used to select differentially expressed genes. GO enrichment analysis for biological processes (BP) were carried out using [enrichGO] from clusterProfiler (version 4.2.2) [101]. Within this function, the adjusted p-values associated with the GO terms were calculated, using the Benjamini–Hochberg (BH) adjustment method for multiple comparisons. Only terms with adjusted p-value < 0.05 were considered. We next used [simplify] from clusterProfiler [103], with a p-adjust threshold of 0.7, which acts to reduce the redundancy in the enriched GO terms. The top 20 GO terms for each group were selected based on their significance (adjusted p-value) and plotted as a heatmap using ggplot2 (version 3.3.5) [104]. A dendrogram of the enriched GO terms was built using the Ward D agglomerative method with [hclust; dist], based on their pairwise similarity calculated with the Wang method using [pairwise_termsim] from the enrichplot package (version 1.14.2). This dendrogram was used to arrange and group the GO terms, enabling us to manually define super terms for the enrichment. For several poorly annotated GO terms, we examined the annotated genes and manually renamed the GO terms (see R script available on our GitHub page for full details).