Life-long Restructuring of 3D Genome Architecture in Cerebellar Granule Cells

The cerebellum contains most of the neurons in the human brain, and exhibits unique modes of development and aging. Here, by developing our single-cell 3D genome assay Dip-C into population-scale (Pop-C) and virus-enriched (vDip-C) modes, we resolved the first 3D genome structures of single cerebellar cells, created life-spanning 3D genome atlases for both human and mouse, and jointly measured transcriptome and chromatin accessibility during development. We found that while the transcriptome and chromatin accessibility of cerebellar granule neurons mature in early postnatal life, 3D genome architecture gradually remodels throughout life, establishing ultra-long-range intra-chromosomal contacts and specific inter-chromosomal contacts that are rarely seen in neurons. These results reveal unexpected evolutionarily-conserved molecular processes underlying distinctive features of neural development and aging across the mammalian lifespan.

Competing interests: LT is an inventor on the Dip-C patent US 11,530,436 ("Multiplex end-tagging amplification of nucleic acids").KD is a co-founder and Scientific Advisory Board member of Stellaromics and Maplight Therapeutics, and Scientific Advisory Board member of BrightMinds Biosciences.

Main Text:
Different cell types within the same organism can mature along highly-distinctive developmental and aging trajectories.At the molecular level, cell type-specific gene transcription can be orchestrated by diversity in genome architecture (folding of chromosomes in 3D) (1).Yet the genome architecture over the lifespan has not been elucidated, limiting our understanding of the life-spanning cellular dynamics of brain function and dysfunction.
Previously, using our single-cell 3D genome assay (diploid chromosome conformation capture, or Dip-C) (2-4), we found that cells in the mouse forebrain (cerebral cortex and hippocampus) undergo cell type-specific transformation in transcriptome and genome architecture during the first month of life (5).However, technological limitation hindered generalization of this finding across brain regions, species, and lifespan.Here we broaden our perspective along all three of these dimensions, by turning our focus across the neuraxis, from forebrain to hindbrain-specifically to the cerebellum, which contains ~80% of all neurons in the human brain.The cerebellum, a powerful yet compact processing unit that has expanded over evolution (6), exhibits unique characteristics including prolonged development after birth, malformation in autism (7), and degeneration during aging (8).Understanding genome dynamics of the cerebellum may provide insights into these unique features, as well as into motor control and cognition (9).
Prior work has revealed nuclear morphological features of cerebellar cells in vitro (10), but the 3D genome structures remains to be fully solved.Recently, chromosome conformation capture (3C/Hi-C) was performed on the adult cerebellum (11)(12)(13); however, a comprehensive, cross-species, single-cell 3D genome atlas of the developing and aging cerebellum is lacking.In addition, simultaneous analysis of transcriptome (14) and chromatin accessibility (15) in the cerebellum would provide additional valuable information.Here we show that the cerebellum undergoes an extraordinary, lifelong 3D genome transformation that is conserved between human and mouse, and is far greater in magnitude than forebrain (5), revealing genome rewiring as a potential molecular hallmark of aging.

A 3D genome atlas of the developing and aging cerebellum
Granule cells (the vast majority of cerebellar neurons) are generated between postnatal days P0-P21 in mouse, and between the third trimester of pregnancy and ~1 year after birth in human-much later than the same process in the cerebral cortex (5).During this period, granule cell progenitors divide and migrate from the external granular layer (EGL) to the internal granular layer (IGL), expanding >100-fold in number.Toward the other end of the lifespan, the cerebellum is also known to exhibit a slow epigenetic aging clock of DNA methylation (16).To explore the genomic underpinnings of this entire timeline, we created a 3D genome atlas extending across the human and mouse lifespan, alongside a multi-ome atlas focused on human development (Fig. 1A).
We first simultaneously profiled transcriptome and chromatin accessibility during postnatal development of the human cerebellum, sequencing 63,768 cells from 7 donors (17) (6 between the ages of 0.1-2.3years, and 1 adult) and detecting a median of 645-1,617 genes (944-3,845 unique molecular identifiers (UMIs)) and 12-34 k assay for transposaseaccessible chromatin (ATAC) fragments per cell from each donor (Table S1, Table S2).We additionally profiled a critical age in mouse-P14, when cells are present in both the EGL and the IGL-sequencing 7,182 cells and detecting a median of 618 genes (944 UMIs) and 22 k ATAC fragments per cell.
We then comprehensively profiled 3D genome architecture across the human and mouse lifespan, sequencing 11,207 cells (Fig. 1A).In human, we sequenced 5,202 cells from 24 donors (0.1-86 years) and obtained a median of 608,000 chromatin contacts per cell.Among these cells, 3,580 came from the cerebellum (chiefly lateral; vermis if lateral not available), and 1,622 from the cerebral cortex (Brodmann area (BA) 46 of the dorsolateral prefrontal cortex (DLPFC)) of the same donors (Table S2, Table S3, Table S4).In mouse, we sequenced 6,005 cells from cerebellum (birth to 21 months)-obtaining a median of 496,000 contacts per cell, and incorporated our prior dataset of 1,075 and 879 cells from mouse cerebral cortex and hippocampus, respectively (5).

3D genome profiling of diverse populations and rare cells with Pop-C and vDip-C
We next focused on genome architecture.Cerebellar cells exhibit unique genome morphology, beginning with nuclear dimensions; granule cell nuclei are among the smallest in the brain (5-6 μm diameter), whereas Purkinje cells have large nuclei (~12 μm diameter) (21).During differentiation, cultured mouse granule cell progenitors reduce nuclear volume and spatially redistribute histone H3.3 (10).However, 3D genome structures of cerebellar cells have remains unclear and little is known about lifetime-spanning dynamics in vivo.
In the simplest case, many of our mouse samples were a pool of males and females, which we demultiplexed based on the ratio of reads between X chromosome and autosomes (Fig. S9).In a more complex case, we pooled one mouse each from the 8 founder strains of the JAX Diversity Outbred (DO) collection (23), and demultiplexed based on known singlenucleotide polymorphisms (SNPs) (Table S4).In the most challenging case, we pooled 3-13 unrelated human individuals and demultiplexed them (22) based on common SNPs among populations, without prior knowledge about donor genotypes (Fig. 3A).Pop-C was thereby shown to provide a robust method for profiling single-cell 3D genome at scale (Fig. S11).
We used vDip-C to solve the 3D genome structures of Purkinje cells (Fig. 3B).Although Purkinje cells are abundant at birth (P0), they quickly become outnumbered by granule cells (Fig. 3D).To isolate this rare (<0.5%) cell type from adults, we constructed a vDip-C vector with a Purkinje cell-specific promoter (Pcp2) (29), administered the viral vector to wild-type mice, and isolated nuclei by fluorescence-activated cell sorting (FACS) (Fig. S12).

Lifelong 3D genome transformation of granule cells
We created a high-resolution, cross-species single-cell 3D genome atlas and resolved 3D genome structures for a subset of cells (from F1 hybrid mice) (Fig. 1A, Fig. 3C, Fig. 3D).Similar to our previous studies (2, 3, 5), single-cell chromatin A/B compartment (scA/B) analysis revealed 3D genome structure types corresponding to diverse cerebellar cell typesincluding granule cells, astrocytes, oligodendrocytes, and microglia in both species, as well as MLIs and Purkinje cells in mouse.Replicates yielded reproducible scA/B and contact patterns (Fig. 3A, Fig. S11, Fig. S21).Note that 3 of our 24 donors were diagnosed with autism, Alzheimer's disease, and/or Lewy body disease; excluding them did not affect our conclusions (Fig. S28).
Granule cells exhibited by far the most dramatic structural transformation.Granule cells of both species were born with an immature structure type, termed structural (S) stage S1, that resembled forebrain neurons (Fig. 3C, Fig. 3D, Fig. S14, Fig. S19).As the cerebellum developed and aged, granule cells continuously and progressively evolved into new structure types S2-S5, which increasingly differed from forebrain neurons (Fig. S14, Fig. S19).This transformation was the primary source of scA/B variations (the first principal component (PC)) and could be visualized regardless of the analysis method (Fig. S15, Fig. S16).

Life-spanning scA/B changes associated with granule cell-specific marker genes
We previously showed that scA/B generally correlates with cell type-specific gene expression, although discordance can be observed at the single-gene level and regarding temporal dynamics (3,5), and it has remained unclear how scA/B interacts with gene expression during aging.In granule cells, we found the predominant mode of scA/B changes to be progressive up-or down-regulation.We calculated the mean scA/B of each 1-Mb genomic region at S1-S5, and identified the top 20% dynamic regions.In both species, these ~500 dynamic regions either gradually increased or decreased in scA/B across S1-S5 (Fig. 5A).

Robust 3D genome maturation despite functional perturbations
To test robustness of this genome restructuring, we explored functional perturbation of chromatin remodeling (Fig. S26).Using bulk Dip-C, we observed little effect on 3D genome maturation in mice with clinically-relevant heterozygous deletion of Arid1b (35) or Chd8, although we cannot rule out more subtle differences.Granule cell-specific, homozygous deletion of Chd4 caused moderate 3D changes (12); however, these changes had little overlap with (and were much smaller than) our observed architectural maturation.

Discussion
Once born, most neurons must last for a lifetime; however, we know little about how underlying genomic information may be structurally organized.Here we discovered unique genome architecture in cerebellar granule cells: ultra-long-range contacts that are uncommon in neurons, specific inter-chromosomal contacts reminiscent of those in nasal tissue (3), and remodeling over decades that may be stabilized by cell type-specific gene transcription.We showed that mouse is an excellent animal model of this process, despite substantial differences from human beings in lifespan.
We provided mechanistic insights into the principles of this reorganization.For example, both granule cells and Purkinje cells lack neuron-specific non-CpG DNA methylation (13), revealing that non-CpG methylation was neither required for our previously discovered, neuron-specific radial genome movement (5)-which we observed in both cell types (Fig. S27) (36), nor required for suppressing ultra-long-range contacts.
A potential function of this architecture might be to manage space and energy expenditure.Our brains are 80% cerebellar granule cells by neuron number.If each granule cell consumed the same volume and energy as a typical neuron in the cerebral cortex, metabolic costs could become prohibitive.Consistent with this idea, granule cells are quiet by firing rate (~0.1 Hz) (37), in contrast to Purkinje cells (~50 Hz) (38).Granule cells might therefore have adopted an energy-saving state: physiologically, transcriptionally, and architecturally.It is worth noting that cerebellar and hippocampal granule cells adopt different structural strategies, despite similar nomenclature.Hippocampal granule cells were more similar to other forebrain neurons than to cerebellar granule cells (Fig. S14) and have larger nuclei (9-10 μm) (39,40), although both are similarly inactive (firing rate 0.1-0.2Hz) (41).It remains to be determined how granule cells in the olfactory bulb organize their genome.
More broadly, this approach showcases how life-spanning 3D genome profiling of a complex, living tissue can provide unprecedented dimensions of information.This lifelong structural transformation may point the way to new therapeutic targets for developmental and aging-related disorders.Wide application of the 3D genome technologies developed here to many brain regions and tissues of the body may contribute to solving longstanding challenges such as dissecting the genetic basis of inter-individual variability, characterizing ultra-rare cell types, and revealing the full diversity and dynamics of 3D genome organization across the life of mammals.
This study has certain important limitations.For example, we used frozen human samples, which might differ from fresh samples.We also note that vDip-C does not apply to human; however, human Purkinje cells could alternatively be isolated by flow cytometry based on size.Finally, future work will be required to test functional relationships between structural and transcriptional changes.

Fig. 1 .
Fig. 1. 3D genome atlas across lifespan for human and mouse cerebellum with multi-ome atlas of postnatal development.(A) Study design.(B) Integrative transcriptome analysis of human multi-ome samples.(C) Representative expression profiles of marker genes.

Fig. 2 .
Fig. 2. Simultaneous transcriptome and chromatin accessibility profiling revealed continuous maturation of cerebellar granule cells over the first postnatal year.(A) Stages of granule cell maturation, their marker genes (ranked by specificity), and enriched pathways (summarized for the top 100 genes).(B) Maturation pseudotime analysis of each sample.

Fig. 3 .
Fig. 3. High-throughput, high-precision 3D genome profiling uncovered lifelong genome remodeling in the human and mouse cerebellum.(A) Pop-C method.(B) vDip-C method.(C-D) Cross-species 3D genome atlas for the developing and aging cerebellum (with cerebral cortex as counterpoint).Pearson's r (and p-value) was calculated from logarithm of age.

Fig. 4 .
Fig. 4. Cerebellar granule cells formed ultra-long-range intra-chromosomal contacts and specific inter-chromosomal contacts during development and aging.(A) Distribution of genomic distances of chromatin contacts.(B) Aggregated contact maps for an example chromosome and a zoomed-in region (upper right triangles).Zoomed-in regions are homologous.Dashed boxes highlight prominent changes.Bin size: 250 kb.(C) Aggregated inter-chromosomal contact maps.Bin sizes: 6 Mb (human); 5 Mb (mouse); 500 kb (zoom-in).

Fig. 5 .
Fig. 5. Lifelong maturation of chromatin A/B compartments was associated with cerebellar granule cell-specific genes.(A) Mean scA/B of each dynamic 1-Mb region at each stage (left).Rows are ordered by hierarchical clustering; the two clusters were visualized on t-SNE plots (right).As in (5), scA/B calculation excludes contacts within each region and thus primarily reports on longrange interactions.(B) Mean scA/B of each 1-Mb region harboring granule cell-specific marker genes (14) at each stage (left), with aggregated scA/B shown on t-SNE plots (right).(C).Aggregated contact maps for an example gene.Bin size: 100 kb.(D) Schematic of transcriptional etching.