Neuron-specific signatures in the chromosomal connectome associated with schizophrenia risk

Neural progenitor differentiation is associated with dynamic 3DG remodeling

We applied in situ Hi-C to map the 3DG of two male human induced pluripotent stem cell (hiPSC)–derived neural progenitor cells (NPCs) (13), together with isogenic populations of induced excitatory neurons (“neuron”) generated through viral overexpression of the transcription factor NGN2 (14) and differentiations of astrocyte-like glial cells (“glia”) (Fig. 1, A and B, and table S1) (15). Transcriptome RNA sequencing (RNA-seq) comparison with published datasets (16) confirmed that the NPCs, but not glia, from subjects S1 and S2 clustered together with NPCs from independent donors, whereas S1 and S2 NGN2 neurons closely aligned with directed differentiation forebrain neurons (17) and prenatal brain datasets (fig. S1, A and B). As with our transcriptomic datasets, hierarchical clustering of our Hi-C datasets after initial processing (fig. S2A) also showed clear separation by cell type (Fig. 1A and fig. S2B). Genome-scale interaction matrices were enriched for intrachromosomal conformations (fig. S2C), with the exception of the negative control (“No Ligase”) NPC library, in which we omitted the ligase step (Materials and methods) and observed an interaction map with no signal due to the loss of chimeric fragments (fig. S2D). Given the observed correlation between technical replicates of Hi-C assays from the same donor and cell type, and the correlation between cell type–specific Hi-C from the two donors (Pearson correlation of PC1, Rtechnical replicates, range = 0.970 to 0.979; Rsubject1-subject 2 by cell type, range = 0.962 to 0.970), we pooled by cell type for subsequent analyses (fig. S2E).

Fig. 1 Neural differentiation is associated with large-scale remodeling of the 3D genome.

(A) (Top) Derivation scheme of isogenic cell types from two male control cell lines. Pink oval, donor hiPSC; orange, NPC; green, neuron; purple, glia. (Bottom) Hierarchical clustering of intrachromosomal interactions (Materials and methods) from six in situ Hi-C libraries. a and b are technical replicates of the same library; height corresponds to the distance between libraries (Materials and methods) (fig. S2B). (B) Immunofluorescent staining of characteristic cell markers for NPCs (Nestin and SOX2), neurons (TUJ1 and MAP2), and glia (Vimentin and S100β). (C) Venn diagram of loop calls specific to and shared by different subsets of cells, including previously published GM12878 lymphoblastoid Hi-C data. (D) Gene ontology (GO) enrichment (significant terms only) of genes overlapping anchors of loops shared by NPCs, neurons, and glia but absent in GM12878. (E) (Left) Cell-type pooled whole-genome heatmaps at 500-kb resolution (fig. S2C). (Right) “Arc map” showing intrachromosomal interactions at 40-kb resolution of the q-arm of chr17 for isogenic neurons, NPCs, and glia, as indicated, from subject 2. RNA-seq tracks for each cell type shown on top of arc maps. Green, neuron; orange, NPC; purple, glia. (F) FPKM gene expression of CUX2 across three cell types with heatmap zoomed in on CUX2 loop (black arrow) (fig. S3). (G) Number of loops specific to each cell type (not shared with other cell types) with one anchor in an A compartment and another in a B compartment (pink), both in B compartments (red), or both in A compartments (blue). (H) (Left) Box-and-whisker distribution plot of TAD size across four cell types. (Right) Median TAD length for each of the four cell types. (I) Heatmaps at 40-kb resolution for a 3-Mb window at the CDH2 locus on chr18. (Bottom) Nested TAD landscape in glia with multiple subTADs (black arrows) called, which (top) is absent from neuronal Hi-C. RNA-seq tracks: green, neuron; purple, glia (figs. S1 to S5).

We first focused on intrachromosomal loop formations, which are conservatively defined as distinct contacts between two loci in the absence of similar interactions in the surrounding sequences (3). Our comparative analyses included published (3) in situ Hi-C data from the B lymphocyte–derived cell line GM12878 (table S1). When analyzed with the HiCCUPS pipeline (5- and 10-kb loop resolutions combined, subsampled to 372 million valid-intrachromosomal read pairs to reflect the library with the fewest reads after filtration) (3), 17,767 distinct loops were called: n = 3118 (17.5%) were shared among all four cell types, whereas n = 5068 (28.5%) were specific to only one of the four cell types (Fig. 1C). Biologically relevant terms such as “central nervous system development,” “forebrain development,” and “neuron differentiation” were among the top gene ontology (GO) enrichments from genes overlapping loops shared between NPCs, glia, and neurons (brain-specific) but not identified in lymphocytes (Fig. 1D and table S2), indicating strong tissue-specific loop signatures that were also confirmed in individual cell types (fig. S3A and tables S3 to S6).

Unexpectedly, there was a reduction (~40 to 50% decrease) in the total number of chromosomal loops in neurons relative to isogenic glia and NPCs (fig. S3, B and C). Reduced densities of chromosomal conformations were also evident in genome browser visualization of chromosomal arms, including chr17q (Fig. 1E). Although both glia and NPCs harbored ~13,000 loop formations, only 7206 were identified in neurons (Fig. 1C; fig. S3, B and C; and table S1), including 442 neuron-specific loop formations. One such neuron-specific loop was at CUX2, a transcription factor whose expression marks a subset of cortical projection neurons (18) and that is highly expressed in our NGN2-induced neurons (Fig. 1F and fig. S3, D and E). Examples of loops lost in neurons include one spanning the Ca2+ channel and dystonia-risk gene, ANO3 (fig. S3F) (19). Furthermore, NPCs, neurons, and glia had similar proportions of loops anchored in solely active (A) compartments, solely inactive (B) compartments, or in both, indicating no preferential loss of either active or inactive loops in neurons (Fig. 1G). However, among the genes overlapping anchors of loops that underwent pruning during the course of the NPC-to-neuron transition, regulators of cell proliferation, morphogenesis, and neurogenesis ranked prominently in the top 25 GO terms with significant enrichment (Benjamini-Hochberg corrected P 10−6 – 10−12) (fig. S3G and table S4B), which is consistent with a departure from precursor stage toward postmitotic neuronal identity (20). Likewise, loops lost during NPC-to-glia transition were significantly enriched (Benjamini-Hochberg corrected P 10−3 – 10−6) for neuron-specific functions, including “transmission across chemical synapse,” “γ-aminobutyric acid (GABA) receptor activation,” and “postsynapse” (fig. S3G and table S4C), which is consistent with non-neuronal lineage commitment.

We defined “loop genes” as genes that either have gene body or transcription start site (TSS) overlap with a loop anchor (5- or 10-kb bins forming the points of contact in a chromatin loop). Genes with loop-bound gene bodies (one-tailed Z test, Zrange = 42.1 to 59.2, P 10−324 for all) or loop-bound TSS (one-tail Z-test, Zrange = 15.2 to 28.8, Prange 2.32 × 10−52 to 4.40 × 10−182) both showed significantly greater expression [mean log10(FPKM + 1); FPKM, fragments per kilobase of exon per million fragments mapped] than that of background (all genes for all brain cell types) (fig. S4A), suggesting that looping architecture was associated with increased gene expression. Furthermore, 3% of loops shared by NPCs, neurons, and glia (brain-specific loops) interconnected a brain expression quantitative trait locus (eQTL) single-nucleotide polymorphism (SNP) with its destined target gene(s), representing significant enrichment over background as determined with 1000 random distance- and functional annotation–matched loop samplings, (random sampling, one-sided empirical P = 0.012) (Materials and methods) (fig. S4B).

We aimed to confirm that the observed net loss of loop formations during the NPC-to-neuron transition could be replicated across a variety of independent cell culture and in vivo approaches and was not specific to our methodological choice of NGN2-induction. We conducted an additional Hi-C experiment on cells differentiated from hiPSC-NPCs by means of a non-NGN2 protocol that used only differentiation medium and yielded a heterogeneous population of hiPSC-forebrain-neurons in addition to a small subset of glia (17). In addition, we reanalyzed Hi-C datasets generated from a mouse model of neural differentiation, consisting of mouse embryonic stem cell (mESCs), mESC-derived NPCs (mNPC), and cortical neurons (mCN) differentiated from the mNPCs via inhibition of the Sonic Hedgehog (SHH) pathway (21). To examine whether such genome-wide chromosomal loop remodeling also occurred in the developing brain in vivo, we reanalyzed Hi-C data from human fetal cortical plate (CP), mostly composed of young neurons, and forebrain germinal zone (GZ), primarily harboring dividing neural precursor cells in addition to a smaller subset of newly generated neurons (7). Across both the hiPSC-NPC-to-forebrain neuron and mESC-mNPC-mCN differentiation, in vitro neurons showed a 20% decrease in loops compared with their neural progenitors (fig. S4, C and D). Consistent with this, in vivo CP (neuron) compared with GZ (progenitor) showed a 13% decrease in loops genome-wide (fig. S4E). The highly replicative cell types included here, mouse ESCs and human lymphoblastoid GM12878 cells, exhibited loop numbers very similar to their neuronal counterparts (fig. S4, D and E), suggesting that the changes in 3DG architecture from NPC to neurons do not simply reflect a generalized effect explained by mitotic potential.

Along with having fewer total loops, neurons exhibited a greater proportion of longer-range (100 kb) loops than did NPCs or glia (two-sample two-tailed Kolmogorov-Smirnov test, KSrange = 0.1269 to 0.2317, P 2.2 × 10−16 for three comparisons: Neu versus NPC/Glia/GM) (fig. S5A). Likewise, in each of the alternative in vitro and in vivo analyses considered above, neurons exhibited a greater proportion of longer-range (100 kb) loops than did NPCs or glia [two-sample two-tailed Kolmogorov-Smirnov test, KS = 0.0427, P = 1.5 × 10−5 for hiPSC-NPC versus forebrain neuron; KS = 0.0936, P = 1.1 × 10−16 for mESC-NPC versus mCN; KS = 0.0663, P = 2.04 × 10−8 for fetal CP (neuron) compared with GZ (progenitor)] (fig. S5, B, C, D, and E). Therefore, multiple in vitro and in vivo approaches comparing, in human and mouse, neural precursors to young neurons consistently show a reduced number of loops in neuron-enriched cultures and tissues, primarily affecting shorter-range loops.

Consistent with studies in peripheral tissues reporting conservation of the overall loop-independent TAD landscape across developmental stages, tissues, and species (when considering syntenic loci) (10, 22), overall TAD landscapes (3) remained similar between neurons, glia, and NPCs. Nonetheless, TADs also showed a subtle (~10%) increase in average size in neurons compared with isogenic NPCs, independent of the differentiation protocol applied (Wilcoxon-Mann-Whitney test, P 5.3 × 10−6) (Fig. 1H and fig. S5, F and G), as highlighted here at a 3.4-Mb TAD at the CDH2 cell adhesion gene locus (Fig. 1I). TAD remodeling may therefore reflect restructuring of nested subdomains within larger neuronal TADs (tables S7 and S8). To examine whether such developmental reorganization of the brain’s spatial genomes was associated with a generalized shift in chromatin structure, we applied the assay for transposase accessible chromatin with high-throughput sequencing (ATAC-seq) to map open chromatin sequences before and after NGN2-neuronal induction (table S1). Genome-wide distribution profiles for transposase-accessible chromatin were only minimally different between NPCs and neurons (fig. S5H) and further revealed that both NPCs and neurons showed low to moderate chromatin accessibility [–2.5 log2(ATAC signal) 1] for ≥89% of the anchor sequences comprising cell type–specific and shared “brain” loops in our cell culture system (fig. S5I). These findings, taken together, point to widespread 3DG changes during the NPC-to-neuron transition and NPC-to-glia transition in human and mouse brain that are unlikely attributable to global chromatin accessibility differences. This includes highly cell type–specific signatures in gene ontologies of differentiation-induced loop prunings, reflecting neuronal and glial (non-neuronal) lineage commitment (fig. S3, A and G, and table S4, B and C), and a subtle widening of average loop and TAD length in young neurons (Fig. 1H and fig. S5, A to G).