Spatiotemporal histogenesis of the developing human cerebellum reveals dynamic layering of Bergmann glia

Abstract

Bergmann glia (BG) are a specialized glial population essential for cerebellar development, yet their developmental timeline and molecular identity in the human cerebellum remain poorly understood. Here, we combined detailed histopathological analysis with spatial transcriptomics and single-nucleus RNA sequencing to generate a developmental atlas of human cerebellar BG. Histology revealed that BG emerge around 11 post-conception weeks (PCW), initially serving as a scaffold for Purkinje cells (PCs) migrating into the PC layer of the cerebellar cortex. Following the establishment of a multilayered PC arrangement, BG form a distinct parallel layer separated from the PCs by the lamina dissecans (LD), with both layers merging in the third trimester. This developmental sequence challenges earlier studies that suggested BG appear late in the third trimester. Comparative histology in mice, ferrets, and marmosets indicates that this trilaminar organization, including the LD, is likely unique to humans. Integration of spatial and single-nucleus transcriptomic datasets identified an ASCL1⁺ PTF1A⁺ ventricular zone progenitor cluster giving rise to BG, astrocytes, and oligodendrocytes. Pseudotime analyses delineated three gliogenic lineages and revealed two temporally and transcriptionally distinct BG populations, emerging at 11–12PCW and 17PCW, suggesting multiphasic BG ontogeny. Together, these multimodal data link cellular lineage, spatial organization, and molecular identity of human cerebellar glia, providing a framework for future studies on the role of BG in cerebellar function and their potential contributions to vulnerability in neurodevelopmental disorders.

Introduction

Histogenesis of the cerebellar cortex is largely conserved among mammalian species with a fully developed cerebellar cortex having three distinct layers (Fig. S1A): the outer-most molecular layer which contains dendrites of Purkinje cells (PCs), parallel fibers of granule cells, and inhibitory interneurons like basket and stellate cells; the PC layer in the middle which is composed of single layer of large PCs, and the inner-most internal granule layer (IGL), which is densely packed with granule cells, a class of interneurons called Golgi cells, and unipolar brush cells which are excitatory neurons arising from the rhombic lip. Beneath the IGL is the white matter, which besides containing cerebellar nuclei serving as the primary output centers of the cerebellum, relaying processed information to other brain regions contains afferent and efferent cerebellar fibers. During development, the outermost cerebellar layer is composed of granule cell precursors that make up the external granule layer (EGL), a fourth layer. Granule cell precursors are derived from the rhombic lip and function as a secondary zone of proliferation. Upon differentiation into granule cells they migrate downwards into the IGL (1–3). This migration is aided by a class of specialized glial cells called Bergmann glia (BG) that are essential for cerebellar development and function.

BG serve as structural and signaling hubs that coordinate multiple processes underlying cerebellar morphogenesis. During development, BG provide a critical scaffold that guides the migration of granule cells from the EGL to their final positions in the IGL, ensuring precise laminar organization of the cerebellar cortex (4–6). They also play a central role in cerebellar foliation, particularly at anchoring centers which are key organizing sites that initiate and shape individual folia by integrating proliferative cues, mechanical forces, and cellular interactions (7, 8). BG contribute to this process by linking germinal zones to the cerebellar surface, maintaining structural integrity, and regulating local granule cell proliferation and migration.

Developmentally, BG arise from the same progenitor lineage as basal radial glia and similarly lose their apical contact during differentiation, adopting a specialized radial morphology. Notably, emerging evidence suggests that much like basal radial glia in the cerebral cortex, BG share molecular and functional similarities that may reflect an evolutionary specialization of glial subtypes in the human brain (9, 10). This parallel underscores the broader significance of studying BG not only as key players in cerebellar morphogenesis but also as models for understanding glial diversity and evolution. In their early radial glial form in mice, BG also serve as a scaffold for migrating PCs (11). Beyond development, BG regulate the membrane potential and firing of PCs (12), highlighting their role in modulating cerebellar circuit dynamics, and also influence pain-related behaviors by controlling cerebellar output (13).

While the fundamental scheme of cerebellar histogenesis is largely conserved across species, pioneering work by Rakic and Sidman in the 1970s identified additional transient layers in the developing human cerebellar cortex (14). Their study, which examined human cerebellar development, revealed a five-layered organization between 20 and 30 weeks of gestation consisting of the EGL, Molecular layer, a maturing PC layer, and two additional layers that had not been previously described in detail. One of these layers, termed the Lamina dissecans (LD), was characterized as an acellular band situated immediately beneath the PC layer. Through silver staining and electron microscopy, Rakic and Sidman demonstrated that the LD contained afferent axon terminals as well as the filopodia of PCs and Granule cells. This layer was found to be transient, disappearing by approximately 30 weeks of gestation. Beneath the LD, they described another distinct cellular layer, which they referred to as the granule layer. Based on Golgi staining they suggested that this layer consisted of immature granule cells that had not yet developed their characteristic dendrites. Interestingly, Rakic and Sidman reported that at this stage, no glial cells resembling BG were present within these layers, implying that BG generation likely occurred later in development.

With the advantage of modern molecular techniques such as immunohistochemistry and spatial transcriptomics, we now recognize that BG are indeed present in the developing human cerebellum prior to 30 post-conception weeks (PCW). Notably, rather than being absent at this stage as previously suggested, BG are first seen in the cerebellar cortex around 11PCW near the subventricular PC plate. During the second and third trimesters BGs form a distinct, transiently dense cellular layer situated beneath the PC layer and the LD. This BG-rich layer persists until ~28PCW after which it integrates into the maturing PC layer. Furthermore, beneath this BG layer lies a nascent IGL, composed of granule cells that, at this stage, remain somewhat dispersed rather than tightly packed. Following a cross-species comparison we find that phenotype appears to be human-specific. This refined understanding highlights the dynamic nature of cerebellar histogenesis and underscores the importance of incorporating new and advanced molecular approaches to revisit and expand upon classical neurodevelopmental findings.

Results

Histogenesis of the human cerebellum reveals the presence of multiple transient layers

Traditional models of cerebellar development largely based on rodent studies describe the formation of four distinct layers within the developing cerebellar cortex: a transient EGL, the Molecular layer, the PC layer, and the IGL. However, seminal observations by Rakic and Sidman in the 1970s challenged this framework in the context of human development (14). Their studies revealed the presence of an additional fifth, cell-sparse layer situated just beneath the PC layer during the second trimester, which they termed the lamina dissecans (LD).

Our analysis of the developing human cerebellum across 42 samples aged 45 days post-conception (Carnegie stage 19) to 36PCW is consistent with these findings (Fig. 1; Table S1). We observe that the LD first emerges around 15PCW (Fig. 1D, grey line; Fig S1 B–E). Notably, beneath the LD, a distinct cell-dense region running parallel to the PC layer becomes visible using hematoxylin and eosin (H&E) staining, highlighting a clear tripartite organization in this intermediate developmental window (Fig. 1D, yellow line). These three layers: the PC layer (Fig. 1D, blue line), LD, and underlying cell-dense zone remain anatomically distinct until approximately 28PCW. At this stage, the LD gradually dissipates, coinciding with the integration of PCs and the cells situated below the LD.

Fig. 1. Spatiotemporal changes in human cerebellar histogenesis.

(A-C) Hematoxylin & Eosin (H&E) stained midsagittal section of a developing 21 post-conception weeks (PCW) human cerebellum. Black box in (A) outlines the region in (B). (B) Blue box in (B) outlines the region in (C). (D) Timeline of External granule layer (red bar), PC layer (blue bar), LD (grey bar), and a cell-dense layer underneath the LD (yellow bar) from 15 PCW to 32 PCW samples. The lamina dissecans is present from 15PCW to 23PCW and disappears as the cell dense layer merges with the PC layer. The cell dense layer underneath the LD is present at 15 PCW and disappears once it merges with the PC layer at 32PCW. [Scale bar = 100μm (Black), 500 μm (orange)]

We examined anterior, central, and posterior lobes (Fig. S1B–D), as well as horizontal sections (Fig. S1E) of the developing human cerebellum across multiple developmental stages, to assess potential anterior–posterior and mediolateral differences in the persistence of the lamina dissecans. Our analysis indicates that the LD is present across all three lobes and is observed in both the vermis and hemispheres.

These transient layers are developmentally significant because they correspond to discrete stages of progenitor proliferation, neuronal migration routes, and the progressive organization of emerging cerebellar circuits.

Spatial transcriptomics reveal the presence of a transient BG layer beneath the PC layer

In their landmark human study, Rakic and Sidman proposed that the cell-dense layer located beneath the LD in the developing human cerebellum consists predominantly of granule neurons forming the nascent IGL (14). They also noted an absence of identifiable BG at least through 25 weeks of gestation, suggesting that BG likely appear only at later developmental stages. However, our histological analysis indicates that the cellular morphology within this hypercellular zone is inconsistent with that of granule neurons, which typically exhibit a smaller, round, and darkly stained appearance following Hematoxylin and Eosin staining. To resolve the molecular identity of this region, we performed digital spatial profiling using the GeoMx platform on 3 samples aged 17-19PCW (Fig. 2A, B).

Fig. 2: Spatial transcriptomics of developing human cerebellum reveals the presence of a distinct Bergmann glial layer.

(A) Representative section of the developing human cerebellum at 17PCW immunostained with PCNA antibody (Green) and DAPI (Grey) used as input for GeoMx spatial transcriptional profiling. Regions of RNA capture are indicated in dashed colored boxes. [Scale bar = 50 μm.] (B) UMAP of multiple regions of interest corresponding to the External granule layer (EGL), Purkinje cell layer (PCL), Bergmann glia layer (BGL) and Internal granule layer (IGL). (C) Transcriptional heatmap from 17-18PCW GeoMx samples (n=3) reveals a distinct molecular profile for the EGL, PCL, BGL and IGL. (D) Heatmap showing shared molecular signatures between cerebral basal radial glia and Bergmann glia.

Our analysis confirms that the outermost layer expresses canonical markers of the EGL, while the intermediate layer corresponding to the PC layer robustly expresses PC markers. Strikingly, the region immediately beneath the LD shows strong expression of BG markers, despite prior assumptions that BG are absent during mid-gestation (Fig. 2C, D; Data S1–S4). This finding suggests that BG not only emerge earlier than previously reported but also form a distinct molecular and anatomical layer situated directly below the PCs and LD. Furthermore, the deepest of the four layers underlying the BG layer at 17-19PCW exhibits a mixed transcriptional signature, co-expressing markers of both BG and granule neurons. This suggests that this region may represent a transitional zone comprising early-born granule neurons of the nascent IGL alongside BG soma and their radial fibers.

Intriguingly, we also detect the expression of genes typically associated with basal radial glia, a progenitor population described in the developing human and ferret cerebral cortex (15, 16), within the BG (Fig. 2D). This observation is consistent with studies in mice which have shown BG share molecular features with basal radial glia, underscoring a potentially broader role for this radial glial subtype in human brain development (17).

BG undergo dynamic repositioning during human cerebellar histogenesis

To investigate the spatial organization and developmental timeline of BG, we examined the expression of Fatty acid-binding protein 7 (FABP7), also known as brain lipid-binding protein (BLBP) which has long been recognized as a classical marker of BG (18) (Fig. 3A–E). During embryonic stages, FABP7 is expressed in both the ventricular and subventricular zones (Fig. 3A). By 10 PCW, expression is limited to cells lining the ventricular zone (Fig. 3B). However, by 11 PCW, FABP7 expression extends into the cerebellar cortex, where a prominent band of labelled cells appears near the emerging PC layer (Fig. 3C). Between 14 and 17PCW, FABP7 is broadly expressed throughout the cerebellar cortex and white matter, though the dense band of expression adjacent to the PC layer remained consistently visible (Fig. 3D, E).

Fig. 3: Changes in Bergmann glial arrangement across time in the human cerebellum.

(A-E) RNAscope ISH expression of FABP7 in the developing human cerebellar at CS20 (A), 10PCW (B), 11PCW (C), 14PCW (D) and 17PCW (E). Sections were counterstained with methyl green. [Scale bar = 200μm.] (F-M) FABP7 and Calbindin immunostaining in the developing human cerebellum reveals changes in Bergmann glia and Purkinje cell arrangement across time. Mid sagittal sections at (F) 11PCW, (G) 14PCW, (H) 16PCW, (I) 17PCW, (J) 19PCW, (K) 24PCW, (L) 28PCW, and (M) 31PCW. [Scale bar = 100μm.] (N) FABP7 GFAP immunostaining at 17PCW [Scale bar = 20μm.] (O-R) Calbindin FABP7 DAPI staining at 17PCW indicate the presence of PC axons and BG fibers in the LD layer. [Scale bar = 50μm.]. (S-T) SKOR2 and SOX2 immunostaining in the developing human cerebellum at (S) 14PCW, (T) 17PCW. [Scale bar = 100μm.] (U) RNAscope in situ hybridization of MKI67 at 17PCW. [Scale bar = 0.5mm.] (V-X) PCNA and SOX9 immunostaining in the developing human cerebellum at (Q) 17 PCW, (R) 24 PCW, (S) 29 PCW, with the Bergman glial layer indicated by a white box [Scale bar = 100μm.] (Y) Box plot of percentage of PCNA+/SOX9+ double positive nuclei within the BGL across 3 bins of developmental ages. 16-19 PCW (N=4; n=20; mean = 72.4 ± 0.0545), 20-24 PCW (N=3; n=15; mean = 43.3 ± 0.0762), 26-29PCW (N=3; n=15; mean = 19.9 ± 0.0744) Individual cell counts are marked with black dots and asterisks indicate significance (p= 1.224e-10 for 16-19PCW vs 20-24PCW; p= 7.428e-13 for 16-19PCW vs 26-29PCW; p=1.09e-07 26-29PCW vs 20-24PCW; N= number of unique samples. n= number of sections).

To further examine the spatial relationship between Bergmann glia and Purkinje cells, we performed double immunolabeling for FABP7 and Calbindin (Fig. 3F–M). Calbindin labels PC somata and neurites, whereas FABP7 marks BG cell bodies, nuclei, and fibers. At 11 PCW, PCs form two distinct plates, one located within the subventricular zone (19). At this stage, PCs are observed in their respective plate (Fig. 3F, Fig. S2A, white arrowhead) and also forming a more superficial PC layer closer to the EGL (Fig. 3F, Fig. S2A, yellow arrowhead). FABP7+ cells are located both within and above the PC plate, extending radial fibers toward the pial surface (Fig. 3F, Fig. S2B, red arrowhead). Notably, Purkinje cell bodies are observed aligned along these fibers in a linear arrangement between 11 and 14PCW, suggesting a close developmental association. These observations imply that BG may serve as a scaffold for PC migration from the deeper PC plate to the more superficial PC layer (Fig. 3F, G). Intriguingly, the uppermost PCs display neurite arborization reminiscent of the dendritic trees typically seen during the third trimester (Fig. 3F). However, our previous studies indicate that this arborization is transient; PC that later integrate into the PC layer only develop their full, elaborate dendritic arbors in the third trimester (19) (Fig. 3M).

By approximately 16PCW, as the PC layer (violet line) consolidates beneath the emerging molecular layer, FABP7-positive BGs (turquoise line) undergo a striking reorganization. They begin to form a distinct layer situated below the PCs, separated by a cell-sparse layer, which we refer to as the LD (grey line). This multilaminar arrangement comprising the EGL and molecular layer above, followed by the PC layer, the intervening LD, and the BG layer beneath is a striking and potentially human-specific feature of cerebellar development (Fig. 3H). Remarkably, this architectural organization is maintained well into the third trimester (Fig. 3H–K). By 28PCW, while FABP7 expression remains robust within the developing IGL, FABP7-positive cells are still observed in proximity to PCs (Fig. 3L). As development progresses, the initially distinct BG and PC layers begin to merge into a more unified radial scaffold (Fig. 3M). This mature arrangement, which remains largely stable throughout the remainder of gestation and into postnatal life, plays a critical role in supporting key developmental processes including granule neuron migration, alignment of PCs, cerebellar foliation and the structural integrity of the cerebellar cortex (7, 8).

SOX2 and SKOR2 expression confirms the presence of SOX2-expressing BG cells within the SKOR2-expressing PCs in the developing PC layer, while by 17PCW, a distinct BG layer emerges, running parallel to the PC layer (Fig. 3S, T).

GFAP and FABP7 staining reveal that BG fibers traverse the ML and EGL with end feet at the pial surface (Figure 3N, red arrowheads). We further examined the composition of the LD. Our findings indicate that the LD is a cell-sparse zone that lacks nuclei from any specific cell type, except possibly migrating granule neurons. High-resolution Calbindin and FABP7 staining reveal that this region is composed primarily of PC neurites, which are likely nascent axons, and basal fibers originating from BG (Fig. 3O–R).

This developmental progression from BGs initially positioned above PCs during early stages to their relocation beneath the PC layer within just a few weeks highlights the dynamic and highly coordinated histogenesis of the human cerebellum. These dynamic temporal and spatial rearrangements reflect the intricate requirements of cerebellar lamination unique to human brain development.

Proliferation within the Bergmann glial layer decreases with age

RNAscope analysis using a MKI67 probe revealed high levels of proliferation across multiple cerebellar layers, including the EGL, and rhombic lip, as expected. Notably, the BG layer also displayed distinct and robust MKI67 expression, suggesting active proliferation (Fig. 3U). To further assess whether this proliferative activity persists over time, we performed SOX9/PCNA double labelling, using SOX9 to identify BG and PCNA as a marker of proliferating cells (Fig. 3V–Y). Between 16 and 20 PCW, approximately 75-85% of SOX9+ BG were proliferative. This proportion declined sharply to 40-55% between 20 and 24PCW (p= 1.224e-10 vs 16-19PCW) and further decreased to 25-30% between 26 and 30PCW (p= 7.428e-13 vs 16-19PCW; 1.09e-07 vs 20-24PCW), indicating that as BG become more spatially restricted and coalesce within the PC layer, their proliferative capacity markedly diminishes while proliferation within the EGL persists. Comparable findings have been reported in the developing mouse cerebellum, where a rapid reduction in KI67+ cells within the SOX9+ BG population is observed between postnatal day 0 (P0) and P10, suggesting a conserved temporal pattern of BG maturation and cell cycle exit across species (20).

Distinct BG organization is a hallmark of human cerebellar development

To assess whether the transient arrangement of BG, initially located above the PCs and later reorganizing into a distinct layer beneath them separated by the LD, is a human-specific feature, we examined cerebellar tissue from developing mice, ferrets, and marmosets. In the mouse cerebellum, between P0 and P6, BGs and PCs remain closely apposed, with no discernible LD or separation between the two cell types (Fig. 4A–D). Similarly, in both ferrets (Fig. 4E, F) and marmosets (Fig. S2F, G), BGs and PCs maintain direct contact without a clear intervening layer. However, in contrast to the relatively thin PC-BG layer observed in mice, both ferrets and marmosets display a noticeably thicker PC-BG region, likely reflecting a higher density of neurons and glia within this compartment. This observation raises an intriguing possibility: the compartmentalization of the BG layer seen in humans may emerge not solely from species-specific architectural programs, but as a consequence of increased PC production and cortical expansion, resulting in the physical separation of these layers during development (19, 21).

Fig. 4: Comparative analysis of Bergmann glial arrangement across multiple species.

(A-H) FABP7, Calbindin, and DAPI immunostained mid-sagittal sections of the cerebellum of a (A) P0 mouse (B) P2 mouse (C) P4 mouse (D) P6 mouse (E) P0 ferret (F) P7 ferret [Scale bar = 50μm.]

To determine whether these time points could represent a transitional stage during which the lamina dissecans has diminished we looked at the arrangement of PCs and BG right before birth in both mice and ferrets. We observe that at e17.5 and e18.5 the mouse cerebellum (Fig S2C, D), and e41 in the ferret cerebellum (Fig S2E), PC and BG are interspersed within one layer conspicuously lacking an LD thereby suggesting that presence of the LD is likely human-specific.

Single cell analysis of developing human cerebellum reveals three distinct glial trajectories

To place our histopathological and spatial transcriptomics findings into a broader developmental framework, we first generated a comprehensive reference atlas of human cerebellar progenitors and their derivatives. This was achieved by integrating our datasets containing 52565 nuclei from 9PCW (31470 nuclei), 17PCW (10942 nuclei), 26PCW (8328 nuclei) and importantly a late gestation 36PCW sample (1825 nuclei) with single-nucleus transcriptomic data from 6PCW to 52 years from previously published studies (19, 21–28) (Fig. 5A, C). The resulting integrated atlas comprised 809,502 nuclei, across developmental stages, ranging from early embryonic development to adulthood (Fig. 5B).

Fig. 5: Single-cell analysis of Bergmann glial development in the integrated cerebellum atlas.

(A) UMAP embeddings of 6 published datasets integrated with 9 PCW, 17PCW, 26PCW, 38PCW snRNAseq samples colored by annotation. (B) UMAP of the integrated cerebellum atlas colored by developmental age. (C) UMAP of the integrated atlas grouped by data source. (D) Feature plot showing the Bergmann glia–Astrocyte module score across the integrated atlas. (E) UMAP of cluster positive for Bergmann glia–Astrocyte module scores within the larger progenitor cluster of the integrated atlas. (F) UMAP of Bergmann glia–Astrocyte–positive clusters in Fig 5E colored by developmental age. (G-I) UMAP embeddings of the subset of Bergmann glia-Astrocyte positive clusters from the integrated human cerebellar atlas in Fig. 5E, F colored by cluster identity (G, I) and age (H). (J) Feature plots for Bergmann glia enriched markers (TNC, GDF10, FAM107A, ETV5, SLC1A3, ATP1A2, HOPX, CACNG5, FREM2, COL12A), Astrocytes enriched markers (AQP4, SLC6A11), Oligodendroglial lineage enriched markers (OLIG1, OLIG2, GRIK2), ASCL1+ cell group (ASCL1, PTF1A) and markers including SOX2, SOX9, NES, GFAP, FABP7. (K) Slingshot trajectory analysis of the Bergmann glia–Astrocyte subset revealed three developmental trajectories arising from the ASCL1⁺ PTF1A⁺ root cluster that diverge through an intermediate population into BG, Astrocyte, and Oligodendrocyte lineages.

Within this dataset, we identified the major cerebellar progenitor population as well as both GABAergic and glutamatergic neuronal derivatives (Fig. 5A). To examine glial populations in greater detail, we projected our spatial transcriptomics data onto the integrated atlas by calculating a module score based on the following glial marker genes enriched in BG and astrocytes based on previous studies (17, 29): TNC, FAM107A, ETV5, GDF10, ATP1A2, SLC1A3, CACNG5, AQP4 and SLC6A11 (Fig. 5D). Our analysis revealed two principal glial clusters (Fig. 5E, F). We then performed unsupervised clustering in Seurat v5 resulting in multiple transcriptionally distinct identities (Fig. 5G–I) distributed across eight clusters: (A) An embryonic population from 6–7PCW with a likely ventricular origin (Cluster 5); (B) An ASCL1⁺ PTF1A⁺ cluster (Cluster 4) largely representing ventricular zone progenitors during fetal development, which we identify as the likely source of BG, astrocytes, and oligodendroglial cells; (C) Two transcriptionally distinct BG clusters: an early population present at 11–12 PCW (Cluster 1) and a later population emerging from 17PCW onwards (Cluster 0); (D) Two astrocyte populations marked by AQP4 and SLC6A11 (Clusters 3 and 7); (E) An oligodendrocyte population (Cluster 6) expressing OLIG1, OLIG2, and GRIK2; (F) A transcriptionally intermediate population (Cluster 2) likely representing a transitional stage between glial subtypes (Fig. 5I, J; Figure S4A).

Pseudotime trajectory analysis was then performed using Slingshot (30), with the ASCL1⁺ PTF1A⁺ cluster designated as the root. This analysis revealed three distinct developmental trajectories originating from the ASCL1⁺ PTF1A⁺ cluster, passing through the intermediate population before branching into BG, astrocyte, and oligodendrocyte lineages (Fig. 5K). Each cluster could be segregated by its transcriptomic identity (Fig. S3A–D; Data S5–S12)

Next, RNAscope confirmed that a number of markers including ASCL1, FABP7, SLC1A3, ATP1A2 mark multiple populations including likely a majority of BG (Fig 6, Figure S5, black arrowhead) and Astrocytes (Fig. 6, Figure S5, red arrowhead). However, markers including GDF10, FREM2, COL12A, ETV5, CACNG5, FAM107A expression is largely confined to BG while AQP4 and SLC6A11 can be reliably used to specifically label astrocytes. Although CACNG5 can reliably distinguish BG from astrocytes, its expression was also detected in the PC layer (Fig. 6J, blue arrowhead; Fig S5C). Similarly, SLC6A11 expression is also observed in the PC layer (Fig 6K, blue arrowhead; Fig S5D), in addition to astrocytes. TNC although marking BG within the BG layer, also marks a small subset of cells within the white matter (Fig. 6E). High HOPX expression, although associated with BG is intriguingly stronger in the posterior lobules (Fig. 6I).

Fig. 6: RNAscope in situ hybridization expression of top glial genes in the human cerebellum.

RNAscope ISH expression of following genes at 17PCW: (A) ASCL1, (B) COL12A, (C) SLC1A3, (D) FREM2, (E) TNC, (F) GDF10, (G) ETV5, (H) FAM107, (I) HOPX, (J) CACNG5, (K) SLC6A11, (L) AQP4. Black, red and blue arrowheads point to the Bergmann glial layer, Astrocytes and Purkinje cell layer respectively. Blue, orange and green vertical lines represent the EGL, PCL and BGL respectively. [Scale bar = 500μm (black) and 50μm (red), and 20μm (blue)]

We next sought to contextualize our data with respect to two previous studies. The first reported Nestin-expressing progenitors (NEPs) including two HOPX⁺ gliogenic clusters that give rise to astrocytes and BG, and one neurogenic ASCL1⁺ cluster generating interneurons which are capable of replenishing the EGL after irradiation (29). Using module scores derived from the top marker genes identified in that study, we detected all three populations within our dataset (Fig. S5E). Our results suggest that each of our astrocyte and BG clusters contains a subset of NEPs, although resolving these subsets in detail is beyond the scope of this study. In the more recent second study, transcriptionally distinct BG and astrocyte subpopulations were identified in the mouse cerebellum (31). Applying their marker sets to our human dataset, we find that these mouse-derived markers cannot reliably distinguish the equivalent subtypes in humans (Fig. S5F). Taken together, this integrated analysis defines the temporal and molecular diversity of human cerebellar glia, establishing a reference framework for interpreting spatially resolved datasets within a developmental context.

Discussion

The human cerebellum develops on a much longer timeline than that of model organisms, and detailed descriptions of its lamination provide critical biological insights by revealing when, where, and how specific neuronal populations and layers emerge (2). Such anatomical resolution clarifies both normal developmental mechanisms and windows of disease vulnerability. As in our prior work on the human rhombic lip, where precise mapping of transient zones uncovered a previously unrecognized germinal expansion, characterizing lamination can similarly identify human- or primate-specific transient layers that are not captured in rodent studies (32). These transient layers illuminate species-specific programs of growth and circuit assembly. Because lamination is tightly linked to axonal targeting, dendritic patterning, and the formation of the glial scaffold, precise descriptions of how layers interface provide a foundation for understanding how circuits such as Purkinje-to-deep-nuclei or mossy-fiber pathways are organized during development. Finally, many congenital cerebellar disorders, including Dandy–Walker malformation, Joubert syndrome, and congenital ataxias, reflect disruptions in the timing or fidelity of lamination; thus, defining the “normal” sequence of lamination at each developmental stage makes deviations easier to detect, classify, and mechanistically interpret.

Our study delineates the temporal and spatial progression of Bergmann glial development in the human cerebellum, integrating histopathology, RNAscope, spatial and single-cell transcriptomics to map their emergence and molecular specialization. We show that FABP7-expressing progenitors, initially confined to the ventricular zone during early embryonic stages and early fetal development, undergo a gradual lineage shift toward BG, concomitant with their alignment within and above the PC plate of the cerebellar cortex around 11PCW. This transformation is followed by a striking reorganization of BG first forming their own layer underneath the PC layer and LD around 16PCW, followed by their eventual merging into the PC layer by 28PCW establishing the scaffold that supports granule cell migration which ultimately leads to laminar architecture characteristic of the mature cerebellum (Fig. 7A, B).

Fig. 7: Illustration of Bergmann glial development in the human cerebellum.

(A) Illustration depicting changes in histogenesis between 11PCW, 17-24PCW, and 28-32PCW developing human cerebellum based our findings. Outer external granule layer (EGL) composed of proliferating granule cell precursors represented in orange; Purkinje cell layer (PCL) represented in green, Proliferating Bergmann glia (BG) and Differentiated BG represented in violet and red respectively and differentiated granule cells of the inner EGL and internal granule layer (IGL) represented in yellow. (B) Timeline of Bergmann glial development in humans.

The need for a separate BG layer

While certain aspects of cerebellar histogenesis are evolutionarily conserved, notable differences between humans and mice reveal intriguing developmental distinctions (2, 19, 32–34). A key factor underlying these differences is the spatiotemporal expansion of progenitor zones, which are much more prominent and elaborate in humans (2, 19, 32, 33). However, in the specific context of BG development, the maturation and organization of PCs play an even more pivotal role in shaping cerebellar architecture.

In humans, PCs are generated in substantially larger numbers than in mice, most likely due to subventricular zone expansion (19, 21). Remarkably, this production occurs early, before 8PCW, when the cerebellum is still relatively small (19). Throughout the first and much of the second trimester, these PCs remain in multilayered configurations. The eventual transition to the characteristic monolayer is a protracted process, driven predominantly by the dramatic expansion of cerebellar volume, surface area, and circumference that occurs during the late second and third trimesters (19, 35).

Studies in mice show that BGs are derived from ventricular radial glia following loss of their apical connections (17). In humans, BGs are also born early in development first seen around 11PCW. This early wave of BGs appears to serve as a scaffold for migrating PCs. In mice, PCs originate from a single PC plate adjacent to the VZ and migrate directly into the PC layer along radial glial fibers (11). In humans, however, the expanded SVZ produces a second PC plate visible between 8-11PCW. PCs from this second ‘upper’ plate must migrate towards the pial surface to form the nascent PC layer (19). We hypothesize that the first set of BGs which occupies the space within and above the PC plate, provides a scaffold for these migrating PCs. Interestingly, the earliest PCs to form the PC layer exhibit dendritic arbors that disappear once the multilayered PC arrangement is established with their characteristic dendritic arbors reappearing in the third trimester (19).

Following the early wave of BG production around 11PCW, a second wave of BG production occurs by 16PCW, followed by local proliferation within the BG layer. These later-born BGs form their own layer beneath the PCs and LD, establishing the PC-LD-BG laminar arrangement that is a hallmark of the developing human cerebellum. We hypothesize that this arrangement occurs due to the extremely high density of PCs observed during the second trimester, with very little space between PCs. As a result, BGs settle in a distinct zone beneath the PC layer and the transient LD. However, as the cerebellum grows driven by EGL expansion, the density of PCs in the PC layer decreases, creating more intercellular space. This structural reorganization, occurring mainly in the last twelve weeks of gestation, enables BG to migrate into the PC layer, where they ultimately form their mature radial scaffold that aids in the migration of newly-born granule neurons into the developing IGL.

Bergmann glia share molecular signatures with cerebral basal radial glia

BG in the cerebellum and basal radial glia in the cerebral cortex exhibit both shared signatures and mechanisms of generation, though initially they also display distinct differences in their spatial arrangement (17). Both cell types lose their apical contact with the ventricular surface and migrate basally during development (15–17). This basal migration is a key feature of their role in neurogenesis, as they transition from an apical progenitor to a basally located cell. In our previous study, basal radial glia-like basally mitotic cells in the developing cerebellar subventricular zone were shown to be arranged in a more haphazard (32), disorganized manner compared to their counterparts in the cerebral cortex, where basal radial glia cells are more consistently aligned and form a characteristic, radial structure (10). However, when examining BG between 17–28PCW, a clear pattern of alignment emerges, closely resembling the more layered arrangement of basal radial glia in the cerebral cortex. This suggests that BG may adopt a similar organizational structure as basal radial glia during a critical window of cerebellar development. Furthermore, a subset of BG in this period were found to be proliferative, as indicated by their expression of PCNA, suggesting that some BG retain a basal proliferative capacity, similar to basal radial glia in the cerebral cortex. However, while cerebral basal radial glia are associated with the generation of upper layer neurons (33), BG mostly function as a scaffold for inwardly migrating granule cells (17). In mice BG Nestin-expressing progenitors have also been shown to produce astrocytes, BG, as well as replenish the EGL by getting converted into granule cell precursors in the event of irradiation inducted injury of the EGL (29, 36). Whereas expansion in outer-subventricular zone cell-types including basal radial glia has been associated with cerebral expansion and gyrification (37–40), BG play a crucial role in establishing anchoring centers that drive cerebellar foliation (7, 8). This comparison highlights how both BG and cerebral basal radial glia, while functionally distinct in their respective regions, share developmental mechanisms involving basal migration and proliferation, which ultimately shapes their roles in cortical and cerebellar organization. Understanding these relationships could shed light on conserved glial mechanisms across brain regions and species.

A comparative framework of BG development in human and mouse

Our single-nucleus RNAseq analyses recapitulate a core developmental architecture that has previously been described in mouse Ascl1 genetic inducible fate-mapping (GIFM) studies (41), while also revealing clear species-specific differences in timing and molecular identity. In mouse, Ascl1 is transiently expressed by ventricular zone progenitors that give rise to virtually all VZ-derived lineages (PCs, early BG, GABAergic interneurons, oligodendrocytes and astrocytes), with GIFM marking early BG that leave the ventricular zone around ~E14 but not later-dividing BG in the PC layer (41). Mutant analyses further show that Ascl1 is necessary for the full complement of GABAergic interneurons and biases gliogenic progenitors toward oligodendrocyte fates while repressing astrocytic expansion (41). These observations define Ascl1 as a temporal marker of ventricular zone neurogenesis that continues to function in gliogenic progenitors, and they provide a useful comparative framework for our human data.

Several aspects of our human dataset align directly with the mouse paradigm. The ASCL1⁺ PTF1A⁺ cluster we detect in our single-cell data occupies the same conceptual position as the mouse Ascl1-expressing ventricular zone pool: it sits transcriptionally upstream of BG, astrocyte and oligodendrocyte branches, and Slingshot pseudotime rooted in this cluster yields three principal gliogenic trajectories that pass through an intermediate state before resolving into BG, astrocyte, and oligodendrocyte lineages (Fig. 5G–K; Figure S4A). Likewise, the coexistence of NEP-like signatures (29) within our astrocyte and BG clusters mirrors reports of Nestin-expressing progenitors and HOPX⁺ gliogenic NEPs in mouse that contribute to BG and astrocyte lineages after injury or during development.

Crucially, however, our data define important species-specific differences. First, the timing and multiphasic nature of BG production in humans are distinct. We resolve an early BG population present at 11–12PCW, a second population emerging ≥17 PCW, followed by continued local proliferation of BG within their own layer, whereas mouse genetic inducible fate-mapping largely describes a single early cohort of BG arising with embryonic PCs and later local divisions within the PC layer (41). Second, although ASCL1 marks a conserved progenitor module, the downstream molecular programs and marker repertoires diverge. For example, marker sets that distinguish mouse BG/astrocyte subtypes fail to cleanly map onto human subpopulations in our dataset. This molecular divergence is exemplified by the regionalised expression of HOPX in human posterior lobules, patterns not readily paralleled by the published mouse markers. Third, while mouse gliogenesis, especially of oligodendrocytes and a significant proportion of astrocytes, is heavily postnatal, our human data indicate that substantial gliogenic specification and the emergence of oligodendroglial and astrocytic transcriptional identities begins prenatally, consistent with the protracted and expanded time course of human cerebellar development and maturation. It is important to consider that the differences between mice and humans may also stem from comparing developmental stages that are not truly analogous, as neuronal and likely glial maturation occurs at a different pace in humans than in mice (42).

Functionally, these similarities and differences lead to complementary interpretations. The conservation of an ASCL1-rooted progenitor that branches into BG, astrocyte and oligodendrocyte lineages supports a common developmental pattern across mammals: a ventricular zone-derived pool transiently expresses Ascl1 to coordinate early neurogenesis and later gliogenesis, and Ascl1 activity can bias glial fate decisions. In human cerebellum, however, this program is adapted to an extended developmental window and altered histogenic arrangements (expanded subventricular zone, multilayered PC stage), producing two temporally and molecularly distinct BG cohorts and earlier prenatal specification of glial subtypes. These adaptations likely reflect the mechanical and signaling constraints imposed by high PC density, extended migratory routes (two PC plates), and the need to establish mature circuitry over a longer gestational interval.

Finally, our integrated findings suggest several testable predictions and experimental priorities. First, the ASCL1⁺ PTF1A⁺ cluster in human tissue is a prime candidate for functional perturbation in cellular model systems like organoids, ex vivo slices or in vivo models like chimeric human-mouse platforms to determine whether ASCL1 similarly biases oligodendrocyte versus astrocyte fates in human progenitors. Second, the early versus late BG transcriptional programs identified here should be compared against temporal genetic inducible fate-mapping in mouse and against lineage tracing in human-relevant models to clarify whether the later BG cohort arises by local proliferation within the PC layer or by delayed migration from ASCL1-positive ventricular zone progenitors. Third, the failure of mouse marker sets to map onto human BG/astrocyte subtypes argues for human-centric marker panels when interpreting developmental samples. In summary, the mouse Ascl1 genetic inducible fate-mapping literature provides a valuable developmental template: a transient ASCL1-expressing ventricular zone pool that seeds neuronal and glial lineages and modulates glial fate balance. Our human atlas preserves this overarching architecture but reveals species-specific timing, regionally patterned marker expression, and multiphasic BG ontogeny that together reflect the unique structural and temporal demands of human cerebellar morphogenesis.

In conclusion, this study provides a detailed molecular and developmental framework for human BG ontogeny, revealing distinct temporal populations and lineage relationships within the broader context of cerebellar gliogenesis. By integrating spatial transcriptomics with single-nucleus sequencing, we have defined key progenitor populations and highlighted species-specific differences relative to mouse models, underscoring the critical importance of human-focused research. Despite these advances, significant questions remain regarding BG development, particularly under conditions of prenatal adversity such as preterm birth where a decrease in BG has been reported (43, 44). Such insults may disrupt the tightly regulated processes of BG maturation and migration, potentially leading to long-term alterations in cerebellar structure and function. Moreover, considering the intimate association between BG and PCs, investigating BG function in the postnatal and adult human cerebellum is equally imperative.

Finally, we acknowledge several limitations in our study, including reliance on postmortem human tissue which are snapshots of development, and challenges in fully resolving the cellular heterogeneity of complex glial populations. Additionally, functional validation of the candidate progenitor states and lineage trajectories using enriched populations, live imaging and in vitro cellular models remains an important future goal. Nevertheless, this work establishes a crucial foundation and highlights the urgent need for further investigation into human BG development, particularly in pathological contexts. Such studies are essential to deepen our understanding of cerebellar vulnerability to disease and to inform therapeutic strategies targeting neurodevelopmental disorders associated with glial dysfunction.

Materials and Methods

Human tissue collection:

This study analyzed a total of 42 histologically normal human cerebellar specimens, spanning developmental stages from 48 post-conceptional days to full term (see Table S1). All human tissues were obtained with appropriate ethical oversight, following full approval by the Institutional Review Board (IRB) at Seattle Children’s Research Institute. Early-stage embryonic tissues were provided by the Human Developmental Biology Resource (HDBR) at University College London and Newcastle University (United Kingdom). Mid-gestation and late-gestation samples were sourced from Hôpital Necker–Enfants Malades in Paris, France. In all cases, tissue collection was carried out in strict accordance with institutional, national, and international ethical guidelines, including appropriate informed consent protocols. Following collection, tissues were fixed in neutral buffered formalin (pH 7.6) and processed for paraffin embedding. Sagittal midline sections measuring 4–5 μm in thickness were cut using a Leica RM2135 microtome and mounted onto Superfrost Plus white glass slides (VWR International, USA). Prepared slides were stored at room temperature until immunohistochemical staining procedures were performed. For developmental classification, specimens younger than 8 post-conceptional weeks (PCW) were categorized as embryonic, whereas samples from 8 PCW and older were considered fetal.

Animal tissue collection:

All mouse cerebella used in the study were collected in accordance with the guidelines laid down by the Institutional Animal Care and Use Committee (IACUC), of Seattle Children’s Research Institute, Seattle, WA, USA. All ferret cerebella used were collected in alignment with Spanish and European Union regulations, and experimental protocols were approved by the Universidad Miguel Hernández Institutional Animal Care and Use Committee and Consejo Superior de Investigaciones Científicas (CSIC) Comité de Ética. Marmoset cerebellar tissue was obtained from the Southwest National Primate Research Center (SNPRC), San Antonio, Texas. Brains were obtained from animals that had been humanely euthanized as part of a separate, IACUC-approved research protocol at SNPRC. Tissue samples from all three species were fixed in 4% paraformaldeyde (PFA) for 24-72 hours, washed in PBS and sunk in 30% sucrose. Cerebella were subsequently embedded in optimum cutting temperature (OCT) compound. Mid-sagittal cryo-sections of 11 microns were taken and processed for immunohistochemistry.

Histology, Immunohistochemistry and in situ hybridization:

Human cerebellar tissue samples were fixed in neutral buffered formalin, then processed through a graded series of ethanol solutions followed by xylene to facilitate paraffin infiltration. Tissues were subsequently embedded in paraffin wax, and mid-sagittal sections of 4 μm thickness were obtained using a rotary microtome.

Immunohistochemistry was performed following standard protocols (32). Briefly, tissue sections were first deparaffinized in xylene, rehydrated through descending alcohol concentrations, and subjected to heat-mediated antigen retrieval. Blocking and permeabilization were carried out using 5% normal goat serum (Vector Laboratories, S-1000) containing 0.35% Triton X-100. Sections were incubated overnight at 4°C with primary antibodies (Table S2). The following day, fluorescently conjugated secondary antibodies (Table S2) were applied. Sections were counterstained with DAPI and mounted in Vectashield antifade medium (Vector Laboratories, H-1200).

RNAscope in situ hybridization was performed using commercially available probes (Advanced Cell Diagnostics/Bio-Techne), following the manufacturer’s instructions without modification. Probes used in this study have been listed in Table S3. Following hybridization, sections were counterstained with either hematoxylin or methyl green prior to imaging.

Microscopy:

Slides from fluorescent immunohistochemical assays were imaged using a Zeiss LSM-Meta confocal microscope equipped with ZEN 2009 software (Zeiss). For brightfield imaging of in situ hybridization (ISH) slides, a NanoZoomer Digital Pathology slide scanner (Hamamatsu, Bridgewater, New Jersey) was used at 20× magnification. Minor, uniform adjustments to contrast and brightness were applied across the entire image; no other image modifications were made. All figures were assembled and finalized using Adobe Illustrator.

Cell counting:

Cell counts were performed using the StarDist plugin for ImageJ/Fiji, applying the Versatile (fluorescent nuclei) model to identify individual nuclei. Cell counts by the StarDist model were manually reviewed and corrected for accuracy. Samples were grouped into 3 age bins: 16pcw-19pcw (n=4), 20pcw-24pcw (n=3), and 25-29pcw (n=3). For each age bin, a minimum of 5 sections from each case across analogous regions of the anterior vermis were used. The proportion of SOX9+/PCNA+ double positive cells and SOX9+/PCNA− single positive cells relative to the total SOX9 cells in the BGL were determined.

GeoMx spatial transcriptional profiling:

5-micron FFPE medial embryonic cerebellar sections from 3 samples aged 17-19PCW were collected on glass microscopic slides within the recommended scan area. The slides were baked in a 60°C drying oven overnight before deparaffinization. The sections were deparaffinized, and an immunofluorescent marker for proliferation, KI67 (Thermofisher Catalog# MA5-14520, 1:100), and Cyto DNA 532 for nuclear visualization were used for region of interest (ROI) selection. The DSP whole transcriptome assay (WTA) was then used to assess genes collected in each ROI. For DSP processing, slides were prepared following the automated Leica Bond RNA Slide Preparation Protocol for FFPE samples, which included digesting samples with 0.1 μg/ml proteinase K for 15 minutes and performing antigen retrieval for 20 minutes at 100°C. In situ hybridization with the GeoMx Whole Transcriptome Atlas Panel (WTA, 18,676 genes) at a final concentration of 4 nM were performed in Buffer R (NanoString) overnight at 37°C. The slides were then incubated with the antibody solution at room temperature for 1-3 hours, then washed in SSC and loaded onto the NanoString DSP instrument. A 20x scan was used to select freeform ROIs to guide the selection of the EGL, PC layer, BG layer and IGL. GeoMx WTA sequencing reads from NextSeq 2000 were compiled into FASTQ files corresponding to each ROI. These FASTQ files were then converted to digital count conversion (DCC) files using the NanoString GeoMx NGS Pipeline. Analysis of raw digital count conversion files was performed according to the GeomxTools vignette from NanoString. In brief, quality control was performed to filter probes with probe ratios < 0.1 or if an outlier according to Grubb’s test in more than 20% of regions. Next, samples were normalized for differential expression with regression for inter-slide batch effects.

GeoMx DSP preprocessing, QC, and normalization:

The R package [NanoStringGeoMXSet] (version) was used for quality control and analysis of GeoMx data. First, transcript counts were shifted by a pseudocount of 1 before probe-level quality control using the following filters: minProbeRatio = 0.1, percentFailGrubbs = 20, removeLocalOutliers = TRUE). Remaining probe-level counts were then aggregated to gene targets using the aggregateCounts() function. Normalization was performing using quantile normalization to the 75th percentile. Raw and normalized matrices were then exported for downstream analysis and visualization. We used the Seurat toolkit for subsequent processing and visualization of GeoMx data. Normalized counts were scaled, and dimensionality reduction was performed following the default Seurat dimensionality reduction pipline with Harmony integration to account for inter-slide batch effects. Differential expression of GeoMx ROIs annotated to anatomical compartments of interest was then performed using the FindMarkers() function from Seurat with the argument test.use = “DESeq2.” Visualization of these comparisons were performed by generating volcano plots and heatmaps in ggplot2 and Seurat (45).

Single-nucleus RNAseq sample preparation, library preparation, and sequencing:

Four samples at 9 PCW, 17 PCW, 26 PCW, and 36 PCW were processed using the 10x FLEX protocol. The 17–36 PCW samples were FFPE, from which multiple 50-micron scrolls were collected and deparaffinized prior to nuclei isolation. The frozen 9PCW sample was crushed on dry ice using a ceramic mortar and pestle. The resulting pulverized samples were then placed into chilled microcentrifuge tubes and stored at −80 °C until needed. Nuclei were extracted from either 150 mg of pulverized tissue or the entire dissected cerebellum using a previously published protocol (22, 46). Single nucleus RNA-seq libraries were generated using the protocol for 10x Genomics Chromium 3’ single-cell gene expression assay and libraries were sequenced on an Illumina Novaseq System.

10X Flex data Single-cell RNAseq alignment, and Quality control:

Gene-level counts were generated from FASTQs with CellRanger (v.3.1.0) using first the ‘mkfastq’ then ‘count’ functions (47). Alignment was performed to the human reference genome hg38 v.3.0.0 from 10X Genomics. All downstream pre-processing and analyses were performed using the R package [Seurat] v5.1.0 (45, 48). Quality control was performed to filter out low-quality cells. Cells were retained that had greater than 500 genes, and fewer than the sum of the median number of genes plus 5X the median absolute deviation number of genes. This latter metric was also used to determine sample-specific cutoffs for the fraction of mitochondrial gene counts. Low-quality cells were excluded from further analysis. Cells were also omitted that included counts for HOX genes HOXA1, HOXA2, HOXA3, HOXB1, HOXB2, HOXB3, HOXC4, HOXD1, HOXD3. This was done to remove any cells included in the original dissection that may have arisen from the brainstem/rhombomeres 2-7 that express HOX family genes, compared to the cerebellum/rhombomere 1, which does not. Genes expressed in fewer than ten cells were also removed. Dimensionality reduction was performed by calculating 20 principal components used to construct a uniform manifold approximation and projection (UMAP) embedding. Cell clusters were identified using Seurat’s shared-nearest neighbour algorithm following modularity optimization using the Louvain algorithm with multilevel refinement.

Human cerebellar integrated atlas:

We assembled cerebellar datasets from multiple studies including our Carnegie-stage 10x ,10x FLEX and 10x multiome RNA (21–28, 49). For balance, cells from Cao et al (24) was randomly down-sampled to 40,000 cells; other datasets were proportionally down-sampled. Counts from each dataset underwent independent pre-processing before integration using the Seurat RPCA pipeline. Following integration, preliminary annotation of clusters was performed based on the expression of canonical marker genes. To confidently identify glial populations of interest from this integrated dataset, we quantified per-cell enrichment for known Bergmann-glia/astroglial markers using the AddModuleScore() function using the following gene set: a curated list of significant BG genes from GeoMx and previous studies (17)(TNC, FAM107A, ETV5, GDF10, ATP1A2, SLC1A3, SLC1A3, CACNG5) and astrocyte-specific genes described previously (29) (SLA6A11, AQP4). The resulting score was plotted on the UMAP to identify these groups which were then subsetted and reclustered.

Pseudotime inference:

To estimate the pseudotime of each cell along developmental lineages of interest, we used the R package [slingshot] (v2.16.0). Having identified glial lineages from among the larger integrated atlas of public cerebellar scRNAseq data, we manually identified a root cluster based on the co-expression of PTF1A and ASCL1 (41). Downstream lineages were calculated using the slingshot() function (30) and visualized using the R package [ggplot2].

Supplementary Material

Data S1

Data S1, Tab1: Top differentially expressed genes (DEGs) in the EGL (GeoMx)

Data S1, Tab2: Top DEGs in the PCL (GeoMx)

Data S1, Tab 3: Top DEGs in the BGL (GeoMx)

Data S1, Tab 4: Top DEGs in the IGL (GeoMx)

Data S1, Tab 5: List of DEGs in Cluster-0 (Late BG)

Data S1, Tab 6: List of DEGs in Cluster-1 (Early BG)

Data S1, Tab 7: List of DEGs in Cluster-2 (Intermediate cell state)

Data S1, Tab 8: List of DEGs in Cluster-3 (Astrocytes)

Data S1, Tab 9: List of DEGs in Cluster-4 (VZ-fetal development)

Data S1, Tab 10: List of DEGs in Cluster-5 (Embryonic VZ)

Data S1, Tab 11: List of DEGs in Cluster-6 (Oligodendroglial lineage)

Data S1, Tab 12: List of DEGs in Cluster-7 (Astrocytes)

Significance Statement.

Bergmann glia are critical for shaping the cerebellum, but their development in humans has been unclear. Using histology, spatial transcriptomics, and single-nucleus RNA sequencing, we built the first developmental atlas of human BG. We show that BG appear much earlier than previously thought (around 11PCW) and undergo distinct phases of development, including the formation of a unique trilaminar structure not seen in other mammals. We identify progenitor cells that generate BG alongside other glial types and reveal two waves of BG development. This work links cell lineage, spatial organization, and molecular identity, offering key insights into how human cerebellar circuits form and why they may be vulnerable in disease.

Data, Materials and Software availability:

Normal human material was provided by the Joint Medical Research Council/Wellcome (MR/R006237/1) HDBR (www.hdbr.org). Human tissue used in this study was covered by a material transfer agreement between SCRI and HDBR. Marmoset tissue was obtained from the Southwest National Primate Research Center who are supported by the Office of Research Infrastructure Programs, National Institutes of Health (P51OD011133). RNAseq data and custom code used in the analyses have been deposited in Github (50). All study data are included in the article and/or supporting information. Previously published cerebellar scRNAseq data was integrated with data generated in the lab. All papers have been cited appropriately in the text and methods section. Human tissue used in this study was covered by a material transfer agreement between SCRI and HDBR. All GeoMx and sn-RNAseq data are publicly available on Zenodo (https://zenodo.org/records/18167058) (50).