Glial cells are essential for proper nervous system development and function. To understand glial development and function, we comprehensively annotated the glial cells from two large single-cell mRNA-sequencing (scRNA-seq) atlases of the developing Drosophila visual system. This allowed us to identify all glial cell types from larval to adult stages and follow their developmental trajectories to understand how the diversity of glial types is generated during development. We show that whereas most glial types, such as chiasm glia, gradually change their transcriptome as they mature during development, neuropil glia that transcriptionally appear as a single-cell class in larvae, splits into ensheathing (EG) and astrocyte-like (ALG) glia types during pupal stages. We have experimentally validated these developmental trajectories and identified the genetic markers expressed through the differentiation between EG and ALG classes. Unexpectedly, our analysis of scRNA-seq datasets allowed us to discover that the transcriptome of glial cell bodies can be distinguished from that of their processes. We have identified that processes are enriched for distinct mRNAs that were validated in vivo. This work provides the most detailed transcriptomic analysis of optic lobe glia during development and helps explain the expansion of glial diversity observed in the adult visual system. We also present an innovative computational approach to identify mRNA species that are differentially localized to cell bodies or cellular processes.

Identification of new glia-specific markers

Marker genes are essential to identify glial clusters in scRNA-seq datasets. In Drosophila, glial cells are most often defined in vivo by the expression of the Repo protein [26]. To identify glial clusters in scRNA-seq, we selected the larval clusters [25] whose transcriptome exhibited a high correlation with the transcriptome of fluorescence-activated (FAC)-sorted cells based on the expression of a repo-GFP transcriptional reporter, and a low correlation with that of neurons FAC-sorted based on Elav expression, a pan-neuronal marker, as previously done for the adult optic lobe dataset [24] (Figs 1A and S1A). Surprisingly, the endogenous repo mRNA was lowly expressed and often absent in many of the glial clusters: repo mRNA was only significantly expressed in 11 out of 23 larval glial clusters and in 5 out of 21 adult glial clusters (Fig 1B and 1D). This low frequency of transcript detection in scRNA-seq experiments is probably due to ‘dropouts’ that are common for lowly expressed genes [27,28]. Consequently, the absence of repo detection in glial clusters in scRNA-seq data may result in many clusters mistakenly not being considered glial cells. For example, in the larva both types of surface glia (subperineurial, spg and perineurial, pg) had low repo expression levels in the dataset (cluster 21 for spg, and 8 and 14 for pg, Fig 1B), which correlates to the low detection of repo mRNA by Hybridization Chain Reaction RNA fluorescence in situ hybridization (HCR RNA FISH) [29,30] even if these cells could easily be identified as glial cells by antibody detection of the Repo protein (S1D Fig).

To overcome this challenge, we looked for other glia-specific genes with higher expression that could allow us to robustly identify glia in scRNA-seq datasets, independently or together with repo expression. By comparing all the larval glial clusters with the other nonglial clusters, we identified several genes that were differentially expressed in glial cells as compared to neurons: Ama, MtnA, AnxB9, CG32032, CG7433, GstE12, CG9686, CG15209, and Npc2a (S1B and S1C Fig). Among these genes, AnxB9, CG32032, and GstE12 were individually expressed in most glial clusters, including those that did not express repo (Figs 1B, 1D and S1C) but were largely absent in neuronal clusters, with few exceptions. AnxB9 belongs to the family of annexins that bind to phospholipids and participate in membrane remodeling, membrane trafficking and exocytosis [31]. Annexins are widely found in the nervous system and show different expression patterns in neurons and glia [32,33]. CG32032 is predicted to belong to the calycin protein superfamily (FlyBase [34], which includes lipocalins and fatty acid-binding proteins often found in glial cells, such as the human FABP7 and PMP2 [6,35]. Finally, GstE12 encodes a cytoplasmic Gluthatione S-Transferase (GST) enzyme [36]. In mice, astrocytes enhance proinflammatory response in the brain through GSTs, and the knockdown of these enzymes in astrocytes compromises microglial activation [7].

We generated antibodies for AnxB9 and CG32032, both of which showed highly glia-specific signal in immunostainings, which confirmed that the protein products of CG32032 and AnxB9 genes are present in glial cells in both L3 larvae (Fig 1F and schematic representation in C) and adults (Fig 1G and schematic representation in E). However, we observed some cell-type specific variability: in L3 brains AnxB9 staining was high in giant chiasma glia (xgo) and low in epithelial and marginal glia (eg/mg), while CG32032 antibody signal was low in xgo and high in eg/mg (S1E and S1F Fig). Thus, CG32032 and AnxB9 are both expressed strongly and mostly specifically in glia and, in combination with repo, are excellent candidates for the identification of glial cell clusters in scRNA-seq datasets. Based on the expression of at least two of these genes, all previously identified glial clusters in both larval (23 out of 23 clusters) and adult (21 out of 21 clusters) datasets were identified, and none of the nonglial clusters expressed these combinations. To test whether these genes could help identify glial types in other datasets, we analyzed their expression in a published optic lobe scRNA-seq dataset from pharate adults (96 hours pupation – P96) [37], and found that repo was lowly expressed in three clusters annotated as glia by the authors (S1G Fig). However, these three clusters expressed at least two of the new glial markers. Moreover, this dataset contained clusters called “C clusters” that were annotated neither as glia nor as neurons. Many of the cells in these “C clusters” expressed the new glial markers. Indeed, out of all 321 cells in the “C clusters”, 103 had transcriptomes that resembled adult subperineurial glial cells. Similarly, cluster N183, previously annotated as an unknown neuronal cluster, also showed expression of these glial markers and was composed of 19 cells, 9 of which also correlated with subperineurial glial cells (S1G Fig). Another published Drosophila scRNA-seq dataset of the adult ventral nerve cord (VNC) was also analyzed [38]. Two of the clusters that the authors had identified as glia/neuron mix due to low expression of repo could be confidently identified as glia when using the combined expression of repo, CG32032, AnxB9, and GstE12 (S1H Fig, arrows) and indeed the authors later managed to identify these clusters as glia, based on specific glial subtypes markers [39]. Interestingly, in this VNC dataset a number of clusters that did not express any of the other glial markers did exhibit AnxB9 expression. Accordingly, we observed strong AnxB9 antibody signal in nonglial cells, presumably neurons, in the larval VNC (S1I Fig). In summary, using four different datasets, we showed that repo, CG32032, AnxB9, and GstE12 are marker genes that can be used in combination to consistently identify glial clusters in Drosophila scRNA-seq datasets.

Annotation of glial clusters in a developmental scRNA-seq atlas of the fly optic lobe

Next, we sought to evaluate glial diversity in the larval and adult stages. High-resolution transcriptomic atlases of the optic lobe throughout development [24,25] had identified more than 200 cell types, including progenitors, neurons and glia, but glial clusters were not annotated as specific glial types. We excluded neurons and re-clustered the bona-fide glial cells (Fig 2B) [40]. Glial cell bodies occupy stereotypical positions in the brain and consistently express Repo protein, which allowed us to match specific gene expression from each cluster to known glial cell types. We annotated and validated larval clusters based on known [7,15,20,23,39,41–48] or new markers identified in this study (summarized in S1 Table and Methods) using transcriptional reporters, protein traps, enhancer traps, antibodies, as well as specific HCR RNA FISH probes (Fig 2A and 2C–2M).

At the adult stage, we based our annotations on cell-type specific markers established by earlier studies that had identified genes expressed in specific subtypes of glia using enhancer traps, antibody staining, cell-type specific RNA-seq, and scRNA-seq [20,39,46,49–54]. We determined the identity of adult clusters by taking advantage of the spatial distribution of glial cells combined with immunohistochemistry for genes that were expressed in one or a few clusters (Fig 3B and 3C–3L; see S2 Table for the list of markers, annotations, and corresponding references). Ultimately, we could assess the transcriptome of all glial cell types and determine specific markers for each of them (S3 Table). This represents a powerful resource for studying glial cells and their contribution to the development and function of the optic lobe.

Developmental trajectories of glial cells from larva to adult stages

In the course of annotating glial types in larva and adult stages, we noticed that the genetic markers of glial types changed over time, and that there was a significant increase in glial cell diversity in the optic lobe, from 10 cell types in larva to at least 14 distinct cell types in adult (Figs 2B and 3B). To investigate this increase, we integrated all glial cells from larvae [25], P15 (15 hours after pupation), P30, P50, P70, and adults [24] into one dataset (Fig 4A). Interestingly, the same cell types across stages clustered together, such as cortex, perineurial, subperineurial, lamina neuropil, medulla neuropil, inner and outer chiasm glia (S2A and S2B Fig). The transcriptomic similarity of the same cell types throughout development corroborates the robustness of our marker-based cell type annotation (S1 and S2 Tables).

We observed three distinct types of developmental trajectories of glial cell types based on our integrated dataset:

1. (1) Cell types that did not exhibit a trajectory (i.e., minimally changed) from larva to adult, such as cortex and perineurial glia. These are cells that remain transcriptionally stable throughout development (Fig 4A, black circles).
2. (2) Cell types that gradually changed their transcriptome in a linear trajectory throughout developmental stages, like inner and outer chiasm glia, likely reflecting a maturation process (Fig 4A, blue arrows). We also observed certain discrepancies between larval and adult cell types, which can be explained by previous literature. For example, in adults there are two satellite glia in the lamina named proximal and distal [16]. However, only one satellite glia is found in larvae, the larval satellite glia that gradually matures to become proximal satellite glia in adults. In contrast, the adult distal satellite glia has a different origin, corresponding to larval retina wrapping glia [23], which is a larval nerve-wrapping glial type that is repurposed to work as cortex glia in adults. (S2A and S2B Fig).
3. (3) Cell types that formed an initial linear trajectory and then split into at least two different cell fates (Fig 4A, red arrows). Notably, all glial types with a divergent trajectory were neuropil-associated glia that split from a common larval identity to adult ALG and EG cell types.

Specification and diversification of lamina and medulla neuropil glia

In the adult lamina neuropil, there are two types of neuropil glia, known as epithelial glia (eg, a type of ALG) and marginal glia (mg, a type of EG). eg/mg originate in the larva from specialized precursors localized in a region of the OPC called glia precursor center (GPC) [45]. At this stageeg and mg cannot be separated transcriptionally into distinct cell types, but later split into separated trajectories (Figs 4A, S2A and S2B). We further investigated this split by UMAP trajectory analysis (S3A Fig) and found that eg and mg transcriptomes diverge soon after they are born, i.e., between L3 and P15 (S3B and S3C Fig). This apparent split into distinct cell types early during metamorphosis would explain the expanded cell diversity in the adult as compared to the larval scRNA-seq annotations. To confirm this trajectory, we performed memory-tracing experiments using the FLEXAMP memory cassette [55], inducing recombination at late L3 using 10C12-Gal4, a driver expressed in epithelial and marginal glia (ege.g.,/mg) as well as in Lamina wide-field (Lawf) neurons [45] (Fig 4C). In the adult brain, we recovered both epithelial (eg) and marginal glia (mg), confirming that lamina neuropil glia diverge into distinct cell types during pupal development (Fig 4F). This result was independently confirmed using FLEXAMP with the 54H11-Gal4 driver, which is expressed in eg/mg and medulla neuropil glia (see below).

The adult medulla neuropil presents two different neuropil glia subtypes, medulla EG (which have long projections closely associated with photoreceptors that show limited branching) and medulla ALG (that send dense, highly branched fine projections) [16]. Both EG and ALG can present diverse morphologies and positions around the neuropil, defined previously as ‘morphotypes’ [16,17]. Both types derive from larval medulla neuropil glia (mng), which originate from the last division of neuroblasts in the outer proliferation center (OPC) [56,57] and present a single trajectory that remains transcriptionally homogeneous from larva to P30 and then splits into two different trajectories leading to the EG and ALG fates (Figs 4A, 4B, S2A and S2B). To validate this predicted trajectory, we induced FLEXAMP recombination in L3 using the 54H11-Gal4 driver, which is expressed in medulla neuropil glia (mng) and eg/mg glia in larva (Fig 4D). Besides eg/mg lamina glia, we were able to recover adult medulla neuropil glial cells corresponding to both expected subtypes, EG and ALG glia. Furthermore, all adult morphotypes were observed, including distal and lateral medulla EG, as well as distal, lateral, and proximal medulla ALG [16,17] (Fig 4G). To further validate this result, we used another driver, Hand-T2A-Gal4. Although this driver is highly specific for L3 mng, not all larval mng were labeled by GFP (Fig 4E). This sparsity was corroborated by our single-cell analysis, where only 26% of L3 mng expressed Hand (Fig 4B). We used a memory-lacZ cassette [58] to trace L3 mng expressing Hand-T2A-Gal4 in the adult and found that Hand expressing cells were all Repo+ and localized around the medulla neuropil, consistent with the position of adult medulla neuropil glia (S2C Fig). We used FLEXAMP to assess their morphology and were able to identify both medulla EG and ALG cells of all morphotypes [16,17]. These results confirm that L3 mng initially share a common transcriptional identity and diverge into distinct EG and ALG subtypes during pupal development. These findings in lamina and medulla glia support the inferred bifurcation in the single-cell trajectory analysis.

Molecular analysis of medulla neuropil ALG and EG diversification

To further analyze the basis of this transcriptomic divergence, we examined the trajectory of mng which splits into the medulla ALG and EG subtypes during pupal development. To identify the exact developmental stage of the split, we manually subset the cells at the fork of the trajectory and found that these cells were mostly coming from the P30 dataset (Fig 5C, inset), suggesting that transcriptomic divergence of mng occurs much after their birth, unlike the lamina glia that split shortly after L3 (compare Fig 5A–5D with S3A–S3C Fig). We then embedded a pseudotime trajectory using Monocle3 [59] on the existing UMAP reduction, which allowed us to closely recapitulate the developmental stages of the dataset as a function of pseudotime (Fig 5B, 5D, and see Methods). Since TFs are the main drivers of cell fate, we focused on their expression to explore ALG and EG fate divergence. We found 63 TFs that were differentially and significantly modulated along the pseudotime trajectory (Fig 5E), and we then used K-means to cluster TFs enriched at early, mid, and late pseudotime points (see Methods). We found that TFs known to be required for mng specification, such as gcm and gcm2, as well as Hand, were expressed early in development, as expected [42,57]. Interestingly, we also observed early expression of canonical Notch effector genes such as Hey and five E(spl) complex TFs [60,61], strongly suggesting that mng have active Notch signaling early after specification. This Notch^ON status and initial Hey expression is also shared by lamina neuropil glia [45]. We validated the expression of Hey in eg/mg and mng by immunostaining, which showed that the presence of Hey protein in nascent lamina and medulla neuropil glia is transient and disappears soon after mng migration (Fig 5F and 5F’). K-means clustering of TFs along pseudotime of the entire trajectory revealed that four genes, crc, Eip78C, gem and CG3328 were enriched at the bifurcation point of ALG and EG subtypes, providing a list of candidate TFs that might be involved in mng subtype specification. Furthermore, we expanded this candidate TF list by analyzing the ALG and EG trajectories separately (S3D, S3E and S3D’, S3E’ Fig, respectively), which will serve as the basis for future perturbation-based experiments. Overall, our analysis provides a transcriptomic basis for mng subtype specification during optic lobe development.

Specification of lobula neuropil glia

The glia associated with the lobula complex neuropil (which during pupal development forms the lobula and lobula plate neuropils) seems to develop in a more complex manner than what we found for medulla and lamina neuropil glia.

We annotated and validated one cluster as lobula complex neuropil glia in the larva (Fig 2A, 1B and 1I). However, unique markers to validate lobula neuropil glia in adults were not found (Fig 3A). Since we confidently annotated all remaining glial types, the 6 unannotated glial clusters in the adult were good candidates to be lobula neuropil glia. Moreover, these unknown adult clusters were closely associated with the L3 lobula complex neuropil glia cells in our age-integrated dataset (Fig 4A). To gain further insight into the presumptive lobula neuropil glia, we integrated these clusters across developmental stages (S2E and S2F Fig). We named the presumptive ALG clusters “Astrocytes_1/ 2/ 3/ 4” and presumptive EG clusters “Ensheating_1/ 2/” (S2F Fig). We observed two parallel trajectories: one for lobula neuropil ALG, which originated in the validated L3 lobula complex neuropil glia cluster, and another for lobula neuropil EG, that started at P15 (S2F Fig).

Interestingly, the L3 lobula complex cluster already expressed mature ALG markers and was positioned between adult medulla ALG and the unknown adult astrocyte-like clusters in the integrated data UMAP (see “Lobula C. Neuropil” and “Astrocytes_1/ 2/ 3/ 4”, S2A and S2B Fig). This suggests that the lobula complex neuropil glia found in the larva are already mature, astrocyte-like cells. Intriguingly, the lobula complex is physically closely associated with the central brain during larval development [19]. In the central brain, adult (or secondary) neuropil glia are produced during larval stages from type II neuroblasts to replace larval (or primary) neuropil glia born during embryogenesis, raising the possibility that the L3 lobula complex neuropil glia cluster was composed of mature astrocyte glia migrated from the central brain. Accordingly, we have observed that a population of glial cells surrounding the lobula complex neuropil that expressed 54H11-Gal4 appeared to migrate from the central brain (S2D Fig). It is possible that these cells account for the lobula complex neuropil cluster of larval origin that is at the base of the lobula neuropil ALG trajectory (S2E Fig, yellow cells).

Meanwhile, the presumptive lobula neuropil EG trajectory, which gives rise to “Ensheathing_1/ 2/” adult clusters, started at P15 and did not include any L3 cells (S2E, S2F Fig). We looked for genes that were expressed early in both medulla and lamina neuropil glia developmental trajectories, and found 61 common genes (S4 Table), six of which were unique to these clusters and mostly absent in other glial types: CR30009, gcm, E(spl)mgamma-HLH, CG11670, E(spl)malpha-BFM, and Traf4 (S4A–S4F Fig). Interestingly, these genes were also expressed at the beginning of the lobula neuropil EG trajectory but were almost absent from the lobula neuropil ALG trajectory (S4G–S4I Fig). This observation supports the idea that ALG lobula neuropil glia are more mature than the rest of neuropil glia at L3 and suggests that EG lobula neuropil glia are born at later stages (P15).

Finally, there were three clusters of surface lamina glia for which we could not identify contributions from larva glia: pseudocartridge (a subtype of subperineurial glia) as well as fenestrated and chalice glia, which are both subtypes of perineurial glia (Figs 4A, S2A and S2B, dashed circles) [16]. Pseudocartridge, fenestrated glia and likely chalice glia originate from the eye disc [62] that was not sequenced in the larva dataset [25], thus explaining their absence (summarized on S5 Table).

Together, our data show that the increased number of adult glial clusters compared to the larval dataset is, in part, the result of increased glial cell diversity originated from larval neuropil glia that diverges into different subtypes during pupal development.

scRNA-seq reveals subcellular distribution of RNAs

Next, we sought to investigate three surprising observations in the larval dataset:

- First, we found a cluster corresponding to retina wrapping glia in our larval dataset (Fig 2B and 2K), which was unexpected, since their cell bodies are located in the eye disc and optic stalk [63,64] that were dissected out from the larval optic lobe before sequencing [25] (Fig 6A). In contrast, the other eye disc-derived glial types (chalice, pseudocartridge, and fenestrated, see above) were absent. Interestingly, the retina wrapping glia are the only eye disc glia that extends their processes into the optic lobe. Therefore, we hypothesized that cell processes that were broken off the cell body could form fragments that were sequenced and contain enough mRNA to form clusters of a defined cell identity.

- Second, we found an extra chiasm glia cluster that could also be explained by the sequencing of cellular processes since this glial type also presents long and extensive cellular processes (Fig 6B). At larval stages, unsupervised clustering revealed two chiasm glia clusters: Cluster 12 that could be separated with supervision and annotated as inner and outer chiasm (Figs 2B and 7A), the only two known types of chiasm glia in the optic lobe [65]. On the other hand, cluster 4 shared all the top chiasm glia markers with cluster 12 (S5B Fig), suggesting that cluster 4 was not a different third type of chiasm glia. We thus hypothesized that cluster 12 was composed of cell bodies with nuclei (Chiasm_CB), while cluster 4 was comprised of processes without nuclei (Chiasm_PC) (S5A and S5B Fig).

There were two striking differences between the putative chiasm cell body and the putative chiasm processes clusters. First, the Chiasm_CB cluster contained cells with significantly more molecules of RNA (UMIs) than Chiasm_PC (Fig 6E); second, only Chiasm_CB contained mRNA for the TF repo (Fig 6F), and the combined content of all TFs was significantly higher in Chiasm_CB compared to Chiasm_PC (Fig 6G). In contrast, Chiasm_PC presented higher levels of Act5C (Fig 6K) and FK506-bp2 (Fig 6L) as well as other genes previously described as having a specific subcellular localization in cell extensions (see full list of differentially expressed genes in S6 Table). For example, the closest human ortholog of Act5C is β-actin, the first gene to be shown to have its mRNA subcellularly localized and transported to the neurites of neurons and lamellipodia of migrating fibroblasts [66–69]. FK506-bp2 is a member of a highly conserved family of immunophilins that are also enriched in neurites in mammalian neurons [70]. To strengthen our hypothesis, reads mapping to introns or exons were compared in both chiasm clusters. Since unspliced mRNAs and introns are essentially only present in the nucleus, Chiasm_CB should contain relatively more reads mapping to introns than Chiasm_PC. Indeed, Chiasm_CB contained, on average, 3% of reads mapping to introns versus 1.3% in Chiasm_PC (Fig 6N).

Next, we decided to confirm that nonnucleated cell fragments were generated during our cell dissociation protocol (Fig 6D). We dissociated the optic lobes of flies expressing membrane-bound GFP in chiasm glia following the same protocol used to generate the larva scRNA-seq dataset [25] and observed similar-sized cell fragments of GFP-expressing chiasm glia that contained (S5C Fig, top panel) or did not contain (S5C Fig, lower panel) nuclei stained with Hoechst.

- Third, we sought to investigate if cell fragmentation could explain the high numbers of cortex glia observed in our scRNA-seq dataset, which was much larger than expected. By nuclear counting using an anti-Repo antibody, we estimated that cortex glia represent less than 7% of all glia in the larval optic lobe (about 100 cortex glia out of about 1,500 glia per optic lobe). However, they comprised the largest cluster in the larval dataset – about 43% of all optic lobe glial cells (1,332 out of 3,117 cells, after excluding central brain glia) (compare Figs 6C, S5A and S5B). Like for chiasm glia, cortex glia clusters could be separated into cell bodies (Cortex_CB) and processes (Cortex_PC). Both clusters were transcriptionally similar (S5B Fig), but Cortex_CB contained many more UMI, a higher amount of repo mRNA and more reads encoding TFs than Cortex_PC (Fig 6H–6J). Cortex glia are very large cells that have processes surrounding the somas of all cells, including neurons, neuroblasts, and the neuroepithelium [16] and might therefore be susceptible to break into multiple fragments by the cell dissociation protocols used to generate scRNA-seq datasets (Fig 6C). This likely explains the high number of Cortex_PC found in the dataset (41% of all glial “cells” in the L3 dataset – 1,270 out of 3,117 cells) compared to Cortex_CB (2% − 62 out of 3,117 cells). Cortex_PC expressed higher levels of fabp and FK506-bp2 than the Cortex_CB cluster (Fig 6O and 6P), among other genes (S6 Table). Fabp is a fatty acid-binding protein whose mouse ortholog Fabp7 mRNA is localized to the tripartite synapses in astrocytes and coimmunoprecipitates with the cytoplasmic polyadenylation element-binding 1 (CPEB1), which is responsible for regulating subcellular mRNA trafficking and translation inside cells [5,71]. In conclusion, the sequencing of cellular processes without nuclei appears to explain the large discrepancy in the number of cortex glia cells in the dataset.

Finally, we sought to determine whether there was indeed a differential subcellular distribution of mRNAs in chiasm glia in vivo. We performed HCR RNA-FISH for repo (Chiasm_CB only) and used as a control CG14598, a gene that is highly and specifically expressed in chiasm glia but not significantly enriched in cell bodies or processes (Figs 2A and 6M). In a fly line expressing membrane-bound GFP in chiasm glia to identify the location of cell processes, repo mRNA was only found around the nucleus in the cell body, while CG14598 mRNA was found throughout the cell (Fig 6Q), which was consistent with the Chiasm_CB cluster corresponding to the cell bodies and Chiasm_PC to the cell processes.

Taken together, these results show that when large cells such as certain glia are dissociated with commonly used scRNA-seq protocols, pieces of cells can form “pseudo cells” that can be mistakenly identified as distinct cells using classical scRNA-seq clustering algorithms. This phenomenon explains the presence of retina wrapping glia, the third cluster of chiasm glia, and the overrepresentation of cortex glia in our larval dataset.

Computational approaches to detect cellular processes in scRNA-seq datasets

We aimed to comprehensively identify and eventually remove cellular processes from our datasets to build an accurate atlas of the developing optic lobe glial cells (S6A Fig). We have previously established that droplets containing cellular process fragments had significantly fewer mRNAs while exhibiting an unambiguous transcriptomic similarity to the clusters annotated as their corresponding cell bodies, and that they were also relatively deprived of TF-encoding transcripts. Therefore, a more robust dataset could be generated by identifying and excluding cellular processes, which are the source of overrepresentation of some glia and of extra clusters in the datasets.

We used two approaches to identify potential cellular process-containing clusters. These clusters contain lower total UMIs and lower TF-encoding reads when compared to either the whole dataset (approach 1) or when compared to the most similar cluster with high UMIs based on Pearson correlation (approach 2) (S6A Fig).

Using both approaches for the larval dataset, we confirmed the previously identified clusters of chiasm and cortex glia that were investigated above as being cellular processes (S6B Fig). Moreover, approach 1 also detected retina wrapping as processes and a newly identified cluster of processes of subperineurial cells. However, because these clusters were the only ones of their identity, they were not removed from the dataset. Moreover, both approaches identified two transcriptionally similar clusters of perineurial glia, one of them likely being the cellular processes of the other. Similarly to retina wrapping, chiasm, and cortex, perineurial glia extend large cellular processes that extend around the whole optic lobe [72]. In summary, for the larval dataset, we excluded clusters comprised of cellular processes for chiasm, cortex, and perineurial glia.

In contrast to the larval dataset, where cell types were subdivided into multiple clusters, the adult dataset had most of its cell types identified as unique clusters; therefore, to better understand the possible presence of cellular process clusters in this dataset, we over-clustered the dataset into smaller subclusters and used both approaches described before (S6C–S6E Fig). Both approaches showed that clusters Astrocytes_1, Astrocytes_2, and Astrocytes_4 were likely entirely comprised of cellular processes. A subcluster of Ensheathing_1 also showed up as processes in both approaches. Adult perineurial glia could also be divided into cell processes and cell bodies, and the same was observed for cortex, distal satellite, and medulla EG. The full list of larval and adult clusters identified as processes can be found in S7 Table. In summary, we excluded 3 clusters of ALG, 1 subcluster of EG, 1 cluster of cortex glia, 2 subclusters of perineurial glia and 1 subcluster of distal satellite glia (S6B, S6C Fig).

Taken together, these different approaches identified cellular processes that were the source of extra clusters in the dataset. This allowed us to build a more accurate atlas of the developing optic lobe glial cells in the larval (Fig 7A and 7B) and adult stages (Fig 7C and 7D) and assess the transcriptome of only the cell bodies (S3 Table).

Drosophila strains

Flies were kept at 25 ºC 12-hour light/dark cycles on a standard cornmeal medium. 10xUAS-myr::GFP (BDSC #32197), UAS-RedStinger (#8546), Tret1-1-Gal4 (#67471), Vmat-Gal4 (#66806), rost::LacZ (#12042), CG3036::GFP (#50804), b-Gal4 (#76724), DAT-Gal4 (#84622), CG31663-Gal4 (#76708), ome-Gal4 (#76166), Canton-S (#64349), Hand-T2A-Gal4 (#66795), CG4288-Gal4::GFP (#29208), alphaTub85E::GFP (#60267), Npc2b-Gal4::GFP (#24694), Adgf-D-Gal4 (#77799), R25A01-Gal4 (#49102), UAS-CD8:GFP (from Bloomington), Cyp4g15-Gal4 (#39103), GstT4-Gal4 (#77777), nkd::LacZ (#25111), R94A08-Gal4 (#40673), R10C12-Gal4 (#47841), R54H11-Gal4 (#39094), UAS-AnxB9-RNAi (#38523), UAS-AnxB9-RNAi.GD (VDRC #27493), wg/xg^O-Gal4 (#9488), UAS-CG32032-RNAi.GD (VDRC #5167), Act5C>stop>nuc.lacZ; UAS-Flp (memory-lacZ) (Struhl and Basler 1993), 20XUAS-Flp; Gal80ts; Act>y+>LexA, LexAop-myr:GFP (FLEXAMP) [55], UAS-sfGFPnlsPEST-2A-B2ase (UAS-nGFP, for short across the manuscript) – this is a line developed by Dr. Ben Jiwon Choi (Desplan lab) consisting of a destabilized nuclear-localized GFP that minimizes fluorophore perdurance. The complete genotype for each experiment can be found in S8 Table.

Immunohistochemistry

Fly brains from adult stage or larva were dissected in room temperature 1x PBS. Then fixed in 4% formaldehyde (v/w) in 1x PBS for 20 min at room temperature. Following the fixation, the formaldehyde solution was removed, and the brains were washed 3x in PBSTX (0.5% Triton X-100 in 1x PBS). Then, the brains were blocked in block solution (PBSTX with 5% Horse Serum) for at least 30 min at room temperature. Primary antibody was diluted as specified in block solution and the samples were incubated for two days at 4 ºC with primary antibody. Then the primary HCRantibody solution was removed, and the brains were rinsed 3x with PBSTX and 1x washed with PBSTX for at least 2 hours. Then secondary antibodies were diluted in the block solution and added to the samples, which were incubated for two days at 4 ºC. Then, the samples were washed 2x with PBSTX and 1x with 1x PBS. Slides were mounted using a 0.22 mm spacer and samples immersed in Slowfade. Images were acquired in a Leica SP8 confocal and were processed in Fiji and Illustrator. The antibodies used are listed in S9 Table.

HCR-RNA FISH

To perform FISH for the detection of RNAs, we used custom probes designed for Repo, CG14598, and Cpr51A. All probes were designed by Molecular Instruments [29]. Hybridization buffer, amplification buffer, wash buffer, and fluorophore-labeled amplification hairpins (Alexa 546, 647, 594, and 488) were obtained from Molecular Instruments. We followed the steps and reagents in the published protocol [83]. For combining HCR-RNA FISH and Immunohistochemistry, we performed the Immunohistochemistry first but using Molecular Instrument’s Antibody Buffer instead of PBSTX. Then fixed the tissue again in 4% PFA in PBST for 20 min at room temperature and then performed the HCR-RNA FISH protocol.

Antibody generation

Polyclonal antibodies were generated by Genscript (https://www.genscript.com/). The protein sequences used for each immunization were as follows:

AnxB9 (guinea pig),

MSSAEYYPFKCTPTVYPADPFDPVEDAAILRKAMKGFGTDEKAIIEILARRGIVQRLEIAEAFKTSYGKDLISDLKSELGGKFEDVILALMTPLPQFYAQELHDAISGLGTDEEAIIEILCTLSNYGIKTIAQFYEQSFGKSLESDLKGDTSGHFKRLCVSLVQGNRDENQGVDEAAAIADAQALHDAGEGQWGTDESTFNSILITRSYQQLRQIFLEYENLSGNDIEKAIKREFSGSVEKGFLAIVKCCKSKIDYFSERLHDSMAGMGTKDKTLIRIIVSRSEIDLGDIKEAFQNKYGKSLESWIKMSIKSLTYVAIFGLFWGSIAGTVVDQFGIYGGSPITTTERSNAELRCMNINPQNSVDLEQ;

CG32032 (rabbit), MMGLWYGSEIIVHSQDFPGTYEYDSCVIIHLTDATDQIRLSQANRGYGYGNQDYNRNQNNYGRTTTTQSSYPDSDEYPLRSIQSQQKYLRLIWSERDNNLEYTFNYTTSAPGQWSNIGDQRGSLVTLNTYTQFTGTVQVVKAVNDHLVLTFCGNDVKSSIYTVVLTRNRLGLSLDELRSIRNLLSRRGLYTETIRKVCNGCGRLGGSLFALLALLLVVRLAWGRGQ;

Optix (rabbit),

AFSAAQVEIVCKTLEDSGDIERLARFLWSLPVALPNMHEILNCEAVLRARAVVAYHVGNFRELYAIIENHKFTKASYGKLQAMWLEAHYIEAEKLRGRSLGPVDKYRVRKKFPLPPTIWDGEQKTHCFKERTRSLLREWYLQDPYPNPTKKRELAKATGLNPTQVGNWFKNRRQR;

Repo (rat),

MEHDSFDDPIFGEFGGGPLNPLGAKPLMPTTTAMHPVMLGSVHELCSQQQQQQQQQRLPDCNTILPNGGGGGAGSGGAGGSPNYVTKLDFVNKMGCYSPSQKYEYISAPQKLVEHHHHHH;

Hey (rat),

MDHNMHVNAPSLHHWGYAAGPGVVMPGATATTPQSHWVPPPQSHHSAHSNHSHGHSQGHSHGIGSLKRTLSESDCDDLYSEESSKEQISPSEPGSCQLMSRKKRRGVIEKKRRDRINSSLTELKRLVPSAYEKQGSAKLEKAEILQLTVEHLKSLQSKTLDSLSYDPQRVAMDYHIIGFRECAAEVARYLVTIEGMDIQDPLRLRLMSHLQYFVQQRELSAKSCASPGGWSPAAPSSSGYQPNCAAAPYQSYAAPANPGAYVSSYPTLSASPSQQAQQLGGRTSVSRTSGSAVTESLPSHDLHSDSSSQQQQQQQQQQQQQQQHQQQQHQQQQQRTQTTPQPTQQQHYTHDHSAVHSEQQVPTYIELTNSNRPAAIGSDSLSYSAAPQYPVSGLPGQDYNNSSVLQYATPNGAKPYRPWGAEMAYHHHHHH.

Interference RNAi expression was performed to validate AnxB9 and CG32032 antibodies. The appropriate UAS-RNAi lines were crossed to wg/xg^O- or 10C12-Gal4, respectively. Crosses were performed at 25 °C for 2–3 days and vials were then transferred at 29 °C to maximize RNAi efficiency until late L3 larvae were dissected.

Glia isolation and annotation from scRNA-seq in the adult stage

Glial clusters were isolated from the previously published dataset [24]. We subset the clusters that were annotated as glial clusters by the authors based on high transcriptome correlation of these clusters with the bulk transcriptome of repo-positive sorted cells and low correlation with that of elav-positive sorted cells (Ozel, Simon and colleagues 2021). We subset the following clusters from the final annotations found in the whole optic lobe adult dataset: “G04”, “G08”, “G03”, “G09”, “G06”, “G/LQ3”, “G02”, “G05”, “G14ab”, “G/LQ1”, “G11”, “G/LQ4”, “G10”, “G07”, “G01”, “G/LQ2”, “G12”, “G13”, “G16/LQ”. Then, we analyzed subset data using Seurat package in R. Specifically, genes that exhibit variable expression in the subset were identified with FindVariableFeatures, and the data were scaled with ScaleData before principal component analysis (RunPCA). The first 20 principal components were used to embed cells in a UMAP space for visualization (RunUMAP(dims = 1:20)) and to construct a nearest neighbor graph (FindNeighbors). Re-clustering of the subset was performed with FindClusters(resolution = 0.6). This split the adult glial cells into 21 clusters, which were annotated accordingly to S2 Table. For S6 Fig, we ran FindClusters(resolution = 2). (Script 2_Glia_Adult).

Glia isolation and annotation from scRNA-seq in the L3

Glial clusters were isolated from a previously published dataset (https://github.com/NikosKonst/larva_scSeq2022/blob/main/1_preparation_of_larval_pupal_dataset.R - larvaOL.integrated150.rds) [25]. We subset the clusters that presented a high transcriptome correlation with the bulk transcriptome of repo-positive FACS sorted cells and low correlation with that of elav-positive FACS-sorted cells [24] (S1A Fig) (Script 1b_Glia_Larva_Correlations). Then, genes that show variable expression in the subset were identified with FindVariableFeatures. Principal component analysis was performed on scaled data (ScaleData followed by RunPCA). The first 30 principal components were used to embed cells in a UMAP space (RunUMAP(dims = 1:30)) and to construct a nearest neighbor graph (FindNeighbors). The subset data was then re-clustered with a resolution of 0.8 (FindClusters(resolution = 0.8)). This split the L3 glial cells into 23 clusters, which were annotated accordingly to S1 Table. (Script 1_Glia_Larva)

Glia annotation in adult and larva

Larval glia are less studied than neurons, and we, therefore, sought specific markers to annotate and validate the clusters based on immunohistochemistry and/or RNA-FISH (Fig 2A). Subperineurial glia, exclusively expressed rost (Fig 2C), and perineurial glia expressed svp but not rost (Fig 2C). Satellite glia expressed Optix (Fig 2D) but did not express Dll, while lamina neuropil glia expressed both Optix and Dll (Fig 2D and 2E).

- Chiasm glia (aka giant glia) expressed CG14598 and DAT, a dopamine transporter responsible for the uptake of dopamine from synaptic sites (Figs 2F and 6P). However, chiasm glia can be divided anatomically into two glia subtypes: outer chiasm (xgo) and inner chiasm (xgi). We manually subset and re-analyzed the chiasm cluster in order to separate it into two unsupervised clusters. We obtained two clusters that differentially expressed hth and were able to identify the two clusters as being xgo (hth-low) and xgi (hth-high) (Fig 2G).

- Medulla neuropil glia were originally identified in three different closely associated clusters that changed gene expression rapidly through time. This may be the consequence of these glia originating from neuroblasts after their neurogenesis window. They then migrate past all neuron cell bodies to contact the neuropil. We merged the three clusters of medulla neuropil glia into one that exhibited expression of Hand and Dll (Fig 2E and 2H).

- Lobula complex neuropil glia were validated by the expression of Cpr51A (Fig 2I) CG4288 (Fig 2J) and Dll (Fig 2E).

- Cortex glia expressed CG4288 (aka MFS9) (Fig 2J), and Cyp4g15 (Fig 6C). Optic lobe cortex glia could be distinguished from central-brain cortex glia that expressed Npc2b (Fig 2L).

- Additionally, we validated another two clusters as glial cells originating from the central brain based on their expression of Adgf-D (Fig 2M) and their high correlation with each other (Fig 2B). These central-brain glial clusters, together with the cortex central brain cluster, were excluded from the present analysis that focuses on the optic lobe.

- Wrapping glia (wg) was identified by the expression of alphaTub85E (Fig 2K). The observed cluster corresponds to cell projections from wg located in the eye imaginal disc, as discussed above.

In the adult, we could rely more on known markers from the literature (Fig 3A). For example, we could identify perineurial glia based on the expression of Tret1-1 [46] (Fig 3C) and Vmat [18], which is also expressed in fenestrated glia [49] (Fig 3D); subperineurial glia, as well as in the larva, exclusively expresses rost (Fig 3E), while pseudocartridge glia share the same subperineurial type markers but do not express rost, being the subperineurial glia from the lamina. Chalice glia expressed CG3036 and Optix (Fig 3F). Distal satellite glia were validated by the expression of vvl and b but neither Optix nor CG31663, while proximal satellite glia expressed vvl, b, Optix, and CG31663 (Fig 3F, 3G and 3J). Cortex clusters were annotated based on the expression of wrapper [46] and validated with CG31663 (Fig 3J). Both outer and inner chiasm expressed CG31663, while outer chiasm expressed DAT and inner chiasm expressed GstT4 and ome (Fig 3H and 3K). Both Epithelial and Marginal glia expresses Optix, these clusters were manually separated and compared to the transcriptome of FAC-sorted epithelial and marginal glia (Script 2b_Glia_Adult_Corr_Epi_Marg), then it was validated by epithelial expressing b and marginal expressing GstT4 and ome (Fig 3G, 3H and 3K). Neuropil glia (astrocytes-like and ensheathing) are harder to separate and annotate because it is difficult to find specific markers that are exclusively expressed in one cluster compared to the other. We found two clusters of ensheathing that we annotated as medulla ensheathing based on the expression of known EG markers, such as CG9657 [20] (more markers in S2 Table), these clusters expressed CG31663 (also around lobula/lobula plate), nkd, and ome (both only around medulla, not lobula/lobula plate), which we found expressed in some of the medulla neuropil glia but not all. The medulla astrocytes-like cluster we annotated based on known astrocytes-like genes expression, such as wun2 [50,51,53], we annotated this cluster as medulla based on nkd expression (Fig 3L) and because our integration showed that this cluster has a common trajectory with that of medulla EG. Other neuropil glia were annotated as astrocyte-like or ensheathing based on the same known markers mentioned above, but we could not validate their identity. Because we did not find lobula/lobula plate neuropil glia in our adult dataset and because these other glial types formed a different trajectory that of medulla neuropil, we speculate these other neuropil glia are lobula/lobula plate neuropil.

Glia integration larva, pupa and adult

We integrated our annotated larval glia and adult datasets with glial cells from P15, P30, P50, and P70 stages. We used all clusters that were considered to be glia by the authors (Özel and colleagues, 2021): “198”, “212”, “210”, “201”, “197”, “207”, “203”, “196”, “213”, “219”, “204”, “200”, “202”, “185”, “186”, “187”, “188”, “189”, “190”, “199”, “205”, “206”, “208”, “209”, “211”, “216”, “217”. However, in the P15 stage, there was a cluster label as being of identity “206”, which was clustering by itself after integration, with no cell from other stages clustering together. Based on our analysis using the genes found in this paper as glial markers (repo, CG32032, GstE12, and AnxB9), the cluster “206” in P15 is indeed not a glial cluster, we excluded this cluster from integration. After integrating all stages, we had a cluster in the center of the UMAP which was composed of cell from many different clusters and also expressed some neuronal markers (e.g., brp, syt1), we excluded this cluster (Script 3_All_Glia_Integration).

We subset medulla neuropil (Figs 4B and 5) and the unknown neuropil (S2E and S2F Fig) from the dataset with all stages. We used these clusters representing the early trajectory to look for early neuropil differentiation genes (S4G–S4I Fig).

To look for common markers for early medulla and lamina neuropil glia, we used the integrated data and selected the clusters representing the beginning of the trajectory and used the Seurat function FindMarkers (logfc.threshold = 1, min.diff.pct = 0.4).

Validation of glial trajectories

To validate the divergence of unique glial cell types in L3 into different adult glial types, we used known Gal4 markers to sparsely and permanently label lamina neuropil and medulla neuropil glia in L3 and observe their adult morphologies. We crossed the FLEXAMP memory cassette (Bertet, Li and colleagues 2014) line which carries a tub-Gal80^TS with R10C12-Gal4 and R54H11-Gal4. To prevent unspecific labeling, we let flies mate for 2–3 days at 18 °C and kept the crosses at 18 °C, and then we shifted vials to the permissive temperature (29 °C) for 12–24 h. After the temperature shift, white pupae were selected and returned to 18 °C until adults hatched and were dissected. In the case of Hand-T2A-Gal4, the aforementioned protocol did not yield any labeled adult glia, most likely due to a weak expression of Gal4 that was insufficient to flip out the FLEXAMP cassette. In this case, we let flies mate for 2–3 days at room temperature and vials where then incubated at 29 °C until adults hatched. Even in these conditions, no more than 2–3 mng glial cells were recovered per optic lobe. For Hand-T2A-Gal4>>memory-lacZ glial tracing, we followed the same protocol as described for FLEXAMP.

Pseudotime analysis for medulla neuropil glia

The integrated Seurat object for medulla neuropil glia consists of 3,168 single-cells from 6 developmental stages: Larva, P15, P30, P50, P70 and Adult. We used this object to make a Monocle3-based [59] cell_data_set object and projected the UMAP reduction on the same coordinates as the original mng Seurat object, allowing us to perform trajectory analysis along the developmental stages used for integration. We established a pseudotime by choosing the root of the UMAP, enriched for Larva stage cells, as the starting node. We used the Monocle3 graph_test function to identify genes that are modulated along the pseudotime and intersected them with annotated TFs in the Drosophila genome [82] to find TFs that vary in expression along mng maturation. Finally, we used k-means clustering (k = 3) to group TFs expressed in early, mid, and late pseudotime that correspond to different developmental stages. To find genes modulated specifically along the EG and ALG branch of mng (Fig 3D and 3E), we choose the principal graph nodes along the respective trajectories and then repeated the K-means clustering as described above.

Identification of cell processes

We used two approaches based on a comparison of UMI and TFs. Approach 1 compared each cluster (or subcluster after over-clustering adult glia) with the whole dataset. To be considered processes, a cluster/ subcluster must have significantly fewer UMIs than the dataset (t test p.adjusted < 0.005) and significantly fewer reads mapping to TFs than the dataset (t test p.adjusted < 0.005). For approach 2, we compared each cluster/subcluster with another cluster/subcluster. To be considered processes, a cluster/subcluster must have significantly fewer UMIs (t test p.adjusted < 0.005) than another close related cluster/subcluster (Pearson Correlation based on 2,000 variable genes in the dataset > 0.9). Additionally, this cluster/ subcluster must also have fewer reads mapping to TFs (t test p.adjusted < 0.005) than the cluster/subcluster that it is being compared to (Script Functions).

Besides the results described in the result section, we had four clusters that showed up as being processes of other clusters by our approach 2: Inner Chiasm (processes) - > Medulla Ensheathing (cell body), Chalice - > Perineurial_CB, Epithelial and Marginal - > Cortex, and Ensheathing_2 - > Cortex glia. We considered these events false positives since, with the exception of Ensheathing_2, all these clusters were validated as having separated cell identities (Fig 3), and cluster Ensheathing_2 had its own markers (S3 Table). Interestingly, out of these five clusters, only inner chiasm glia was shown as processes also using approach 1.

S1 Fig. Newly identified glia-specific genes.

(A) Pearson correlation coefficient of larval clusters from Konstantinides and colleagues, 2022 with elav-positive bulk transcriptome (x-axis) and repo-positive transcriptome (y-axis) from Özel and colleagues, 2021. (A’) UMAP representation of all cells in the larval dataset from Konstantinides and colleagues, 2022. Glial cells are in red and nonglial cells in dark gray. (B) Table showing the top differentially expressed genes between Glial and NonGlial cells. (C) Dot plot representing the expression of all genes from (B) by glial clusters in the larva dataset. (D) Combined HCR-RNA FISH and antibody staining for detection of both repo mRNA (cyan and separate channel below) and Repo protein (magenta and separate channel below). Note that surface glia presents a strong protein staining but weak mRNA signal. (E) Double staining of AnxB9 (cyan and separate channel to the right) and Repo (magenta) antibodies. (F) Double staining of CG32032 (cyan and separate channel to the right) and Repo (magenta) antibodies. (G) Dot plot showing all nonannotated clusters from Kurmangaliyev and colleagues, 2020—a late pupa (P96) Optic Lobe dataset. Arrows point to clusters with glial cells that were not previously identified as glia due to lack of repo expression. (H) Dot plot showing the three newly identified glial genes by cluster of a ventral nerve cord (VNC) dataset (Allen and colleagues, 2020). Arrows point to clusters that were not identified as glia by regular pan-glial markers that were later annotated as glia by specific cell type markers by the authors. Open arrowheads show clusters that express high levels of AnxB9 that are most likely not glia, see (I). (I) Cross section of a larval VNC stained for AnxB9 (cyan and separate channel to the right) and Repo (magenta). Note that some AnxB9⁺ in the cortex do not present Repo antibody (yellow arrowheads) and may be neurons. pg = perineurial glia, spg = subperineurial glia, eg = epithelial glia, mg = marginal glia, mng = medulla neuropil glia, xgi = inner chiasm glia and np = neuropil. The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s001

(TIFF)

S2 Fig. Cluster identity in the integrated dataset and the origin of lobula neuropil glia.

(A, B) UMAP representing the integration of all glial cells from L3, P15, P30, P50, P70, and adult stages (see Fig 4A). Only larval cells are colored in (A), and adult cells in (B). Note that pseudocartridge, fenestrated, and chalice glia are not present in the L3 dataset (dashed circles) and the expansion of neuropil glia cell types. (C) Hand-T2A-Gal4 drives the expression of a memory-lacZ cassette. β-Gal (cyan and separate channel below) positive cells all express Repo (magenta), and are located in the proximal, lateral, and distal surfaces of the medulla neuropil, labeled by anti-Brp antibody (yellow and yellow dashed line below). (D) Section of a third instar larva optic lobe showing the lobula complex neuropil (Ncad, yellow and dashed contour). The border between the central brain (CB) and the optic lobe (OL) is shown by a dashed line. The expression of UAS-LifeActGFP (cyan, and separate channel to the right) under the control of the 54H11-Gal4 driver labels numerous cells that cross from the CB towards the OL and associate with the lobula complex neuropil. Repo antibody (magenta, and separate channel to the right) is used to label glial cells. (E, F) UMAP showing the integration of the clusters we identified as lobula neuropil glia. Cells are colored by their developmental stage of origin in (E) and their adult identities in (F). Note that astrocyte-like clusters (Astrocytes_1/2/3/4) derive from the validated L3 lobula complex neuropil glia cluster (red arrow), whereas EG (Ensheathing_1/2) are only represented by cells from P15 onwards (dashed-line arrow). The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s002

(TIFF)

S3 Fig. Comparison of lamina and medulla glia lineages across development.

(A) UMAP embedding of isolated lamina neuropil glia integrated from all developmental stages (L3, P15, P30, P50, P70, and adult). Cells are colored by their stage of origin, showing the progressive divergence between epithelial and marginal glia fates (circled) through development. (B) UMAP-based subsetting of maturing lamina neuropil glial cells at the bifurcation (colored red) of epithelial and marginal glia adult fates. (C) Bar graph showing the proportion of red cells in (B) that belong to each developmental stage, showing that most cells at the fork belong to L3 and P15. (D, D’) UMAP-based trajectory selection (purple cells) illustrating maturation paths for medulla neuropil ALG and EG fates. (E, E’) Heatmap showing transcription factors with differential expression at early (right) mid (center) and late (left) pseudotime points for medulla neuropil EG (E) and medulla ALG glia (E’) developmental trajectories. The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s003

(TIFF)

S4 Fig. Common genes expressed in early neuropil glia.

(A–F) UMAPs representing the integration of all glial cells from L3, P15, P30, P50, P70, and adult stages. Panels show the expression levels of genes that are common to the initial trajectory of neuropil glia lineages. Genes shown were the first six genes selected from S4 Table by highest avg_log2FC and pct.2 < 0.1 (cells from other clusters that expresses the same marker). Cells expressing the genes are shown on top of the other cells (order = T). (G–I) UMAPs showing only lobula neuropil glia integrated from all stages, displaying the expression levels of CR30009, gcm and E(spl)malpha-BFM. Cells expressing the genes are shown on top of the other cells (order = T). The proposed start for the lobula neuropil ALG (red arrow) and lobula neuropil EG (green arrow) glia are shown in each panel. Note the relative absence of early neuropil glia markers from the lobula neuropil ALG trajectory. The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s004

(TIFF)

S5 Fig. scRNA-seq reveals subcellular distribution of mRNAs.

(A) UMAP representing the L3 glial clusters annotated accordingly to cluster numbers prior to annotation (Fig 2B and S1 Table). Chiasm glia clusters are highlighted in blue are cortex glia clusters in pink. PC = Cellular Processes, CB = Cell Body. (B) Heatmap representing the top markers for each cluster (high expression in yellow, and low expression in purple). The chiasm glia cluster 12 and 4 are highly similar, sharing the same markers, as is the case for the cortex glia clusters 20, 7, 1, and 3. (C, D) Dissociated cells from L3 optic lobe expressing membrane-bound GFP in chiasm glia. Hoechst labels cellular nuclei. Chiasm glia with nucleus (C) and Chiasm glia fragments without nucleus (D). Scale bar = 20 µm. The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s005

(TIFF)

S6 Fig. Computational approach to identify clusters of cellular processes.

(A) Flowchart describing the two different approaches used to identify clusters comprised of cell processes. Both approaches were used in parallel in our work. (B) Violin Plot showing each cluster from the larval dataset. Red bars below indicate whether each cluster was identified as processes by Approach 1, Approach 2, and if the cluster was excluded from the final atlas. (C) Violin Plot showing each subcluster from the adult dataset after over-clustering. Red bars below indicate whether each subcluster was identified as processes by Approach 1, Approach 2 and if the subcluster was excluded from the final atlas. It also shows which original ID cluster the subclusters come from. (D, E) UMAP plots showing the original ID (D) and the over-clustering (E) of the adult dataset. The code used to generate the graphs can be found at https://zenodo.org/records/21890456.

https://doi.org/10.1371/journal.pbio.3004015.s006

(TIFF)

S7 Fig. Validation of AnxB9 and CG32032 antibodies.

(A) wg/xg^O-Gal4 driver is used to express UAS-nlsGFP-PEST (cyan) in wrapping and outer chiasma glia (cyan dashed contour in lower panels), which show higher levels of AnxB9 antibody (magenta and separate channel below). Repo (gray) is used to identify glial cells. Control UAS-mCherry-RNAi (A), and two different interference RNA constructs against AnxB9 (A’, A’’) are expressed, showing a cell type-specific reduction of AnxB9 signal in outer chiasma glia. (B) 10C12-Gal4 driver is used to express UAS-nlsGFP-PEST (cyan) in lamina neuropil glia, which display higher levels of CG32032 antibody. We did not observe any difference in CG32032 signal in optic lobes expressing UAS-mCherry-RNAi controls (B) or a UAS-CG32032-RNAi construct (B’) in lamina neuropil glia. Scale bar = 20 µm.

https://doi.org/10.1371/journal.pbio.3004015.s007

(TIFF)

S8 Fig. Developmental expression of ALG-related genes.

UMAP of medulla neuropil glia (mng) cells showing the expression of transmembrane protein encoding genes previously described as relevant for ALG mng morphogenesis (A, B) or as ALG-specific markers (C). While lapsyn and htl expression do not show ALG-specific expression, wun2 transcription strongly correlates with ALG fate after the bifurcation (upper branch).

https://doi.org/10.1371/journal.pbio.3004015.s008

(TIFF)