Functional Genomics of Human Skeletal Development and the Patterning of Height Heritability

Summary:

Underlying variation in height are regulatory changes to chondrocytes, cartilage cells comprising long-bone growth-plates. Currently, we lack knowledge on epigenetic regulation and gene expression of chondrocytes sampled across the human skeleton, and therefore cannot understand basic regulatory mechanisms controlling height biology. We first rectify this issue, by generating extensive epigenetic and transcriptomic maps from chondrocytes sampled from different growth-plates across developing human skeletons, discovering novel regulatory networks shaping human bone/joint development. Next, using these maps in tandem with heigh GWAS signals, we disentangle the regulatory impacts that skeletal element-specific versus global-acting variants have on skeletal growth, revealing the prime importance of regulatory pleiotropy in controlling height variation. Finally, as height is highly heritable, and thus often the test case for complex trait genetics, we leverage these datasets within a testable Omnigenic Model framework to discover novel chondrocyte developmental modules and peripheral-acting factors shaping height biology and skeletal growth.

Graphical Abstract

In Brief:

A framework for understanding complex-trait genetic uncovers regulatory networks underpinning human height variability

Introduction

In the pursuit of understanding the genetic architecture of complex, highly polygenic traits, genome-wide association studies (GWAS) are being performed on successively larger sample sizes to identify genetic variants associated with trait establishment¹. This approach has been applied to human phenotypes ranging from anthropometric traits (e.g., height, BMI,) to common diseases (e.g., Alzhiemer’s, Diabetes). Indeed, a recent GWAS metanalysis on height², conducted on >5 million individuals, revealed ~12,000 associated loci with putatively causal variants predominantly localizing to non-coding regulatory elements^1,3. The importance of the non-coding genome to complex traits has been corroborated via paired large-scale genetic and transcriptional studies (e.g., TWAS⁴) that revealed how numerous non-coding variants serve to impact the transcriptional regulation of genes via both proximal (i.e., cis) and distal (i.e., trans) mechanisms. Thus, complex traits may be controlled via broad regulatory networks operating at the transcriptional level but impacted by the epigenetic context of variants in relevant cell types.

The aforementioned studies also suggest that for certain complex traits the distribution of genetic associations is diffused across the genome, with an underlying genetic architecture that is massively polygenic, or even ‘Omnigenic’, in which small-effect contributions are dispersed across the genome. Collectively these account for a substantial portion of variability attributed to each trait^5,6. These small effects are proposed to aggregate via trans effects, such that genetic variants impacting genes ‘peripheral’ to a trait integrate onto those ‘core’ genes more proximately involved in trait biology via trait-relevant cell regulatory networks (e.g. through transcriptional regulatory relationships)⁶. In this formulation, the genetic variance of a trait may be influenced by co-regulation between these ‘core’ genes, as covariance terms magnify the aggregate contributions of many small-effect trans-QTLs. Previously, height - given the diffuse, broad nature of its genetic associations⁷ - has been suggested to be extremely polygenic⁵. Yet, the degree to which height truly conforms to expectations derived from this conceptualization of the Omnigenic model remains to be tested using regulatory-network datasets generated on relevant cell types. More broadly, a method that establishes an objective framework for examining whether any given complex trait conforms to omnigenicity has generally been lacking.

Here, we develop a new approach to testing trait omnigenicity on the most studied complex trait, height. We first characterize the epigenetic and transcriptomic landscape of chondrocytes, the key cell-type underlying height biology⁸, across the developing human skeleton. As overall height is often considered to result from the effects of partitioned growth across different body regions (e.g., vertebral column, legs), we elucidate regulatory-networks with regional specificity across the developing skeleton. We next leverage this extensive dataset to explore the relationships between epigenetic modifications, gene expression, regulatory variations, and height in conjunction with a recent large-scale meta-analysis of height GWAS studies generated by the GIANT consortium² and UKBiobank⁹. Leveraging these rich association maps, we describe the genetic underpinnings of height and use them to test a formulation of the Omnigenic model. We then provide a separate independent examination of type 2 diabetes to further demonstrate the broader applicability of our approach. Overall, we reveal that our new approach to understanding complex traits, involving the integration of high-resolution molecular datasets with genetic-association information within a broad regulatory framework, can be extended to the study of other complex traits – further revealing the underlying biological mechanisms linking genotype to phenotype.

Results

Epigenomic assessment of long-bone skeletal development:

As trait variation is predominantly driven by non-coding sequence changes¹⁰, we first characterized the epigenome of human chondrocytes collected in vivo using ATAC-seq on gestational day (E)53/54 samples. This developmental window was chosen because it corresponds to the period when chondrogenic anlage form and long bone cartilage rudiments elongate^11,12. For each sample, we extracted matched growth-plate and epiphyseal zone chondrocytes from each proximal and distal long bone skeletal element (see STAR Methods; Fig.1A), generating chromatin accessibility maps across the fore- and hindlimb skeleton (Table S1 for sequencing details). We used a previously-generated¹² accessibility map of developing brain tissue for this developmental window acting as a tissue-type control, enhancing the specificity of accessibility signals for chondrocytes⁸. Previously, we have shown that filtering for brain signals accounts for the vast majority of signals from non-skeletal developmental tissues¹².

Figure 1:

Epigenetic profiling across the developing human skeleton. (A) Experimental schematic. (B) (Left, UpSet plot) Number of called ATAC-seq regions overlapping between different skeletal elements and embryonic brain. (Centre, pie chart) Proportion of ATAC-seq regions overlapping across elements. (Right, heatmap) Jaccard similarity indices for skeletal elements. (C) Proportion of ATAC-seq regions for both common as well as specific regulatory regions or peaks overlapping between E50s and E60s timepoints. (D) GREAT Human Phenotype enrichments for regions sets from E50/E60s timepoints.

After consolidating replicates, we compared accessibility patterns across skeletal sites (Figure 1B, left); interestingly, sites did not cluster by bone type (e.g., proximal and distal femur), but rather by joint identity – results echoed in RNA-seq profiles (below). Across different bone-ends we observed tens-of-thousands of peaks per tissue (Table S1), most peaks sharing activity in two or more bone-ends (Figure 1B, right). Accessible regions shared between chondrogenic and brain datasets likely represent broadly-active (i.e., pleiotropic) regulatory elements. Indeed, locus-associated enrichments¹³ were more specific for skeletal regulatory elements not shared in stage-matched brain tissue, including terms such as ‘collagen binding’, ‘Wnt-protein binding’, and ‘abnormal chondrocyte morphology’ (Table S1). Enrichments for transcription factor (TF) motifs in these elements tended to be similar across element sets, with subsets of TFs demonstrating bone-end-specificity (Table S1, Figure S1).

To address the possibility that these patterns (and later analyses) are specific to this developmental window, we profiled accessibility in later-stage fetal samples, a period when hindlimb elongation occurs (E67, STAR Methods). We observed consistently high correspondence between broad epigenetic activity at earlier and later time points (Figure 1C). However, strictly site-specific regulatory element sets had lower correspondence (Figure 1C); sites tending to cluster more by time-point than site (Figure S1). Comparing time-restricted accessibility at a given site (Figure 1C), we find that early-restricted sites were enriched for ‘accelerated skeletal maturation’, while late-restricted sites were enriched for ‘growth delay’ (Figure 1D, Table S1). It is tempting to speculate that regulatory elements active earlier in skeletal development may be more associated with initiating skeletal growth, whereas later-active elements may instead act to maintain growth programs.

Transcriptomic assessment of long bone skeletal development:

Next, we performed RNA-seq on matched long-bone-end samples from similar developmental stages (STAR Methods), with comparable sample sizes (see Table S2). After accounting for homogenously-expressed genes (STAR Methods, Figure S2), we found that individual biological samples cluster by tissue-type, and further by limb-type (fore/hindlimb) (Figure 2A). When we considered overlaps of differentially expressed genes (DEGs) defined at the level of bone-end specificity (e.g., distal femur), we recapitulated clustering patterns observed via ATAC-seq analysis (Figure 2B), with joint tissues (e.g., distal femur and proximal tibia of the knee, where growth occurs), grouping together rather than by opposing bone-ends (e.g., proximal and distal femur) (Figure 2C). When we further tested for time-biased expression at specific anatomical sites (Figure 2D), most genes were not significantly shifting expression temporally(Table S2).

Figure 2:

Transcriptomic profiling across the developing human skeleton. (A) Expression heatmap of the top 200 genes consistently differentially-expressed across elements; groupings of elements into different anatomical categories (Bone, Limb, End) indicated in top three color bars. (B) Heatmap of overlapping differentially-expressed genes (DEGs) defined in individual bone-end tissues. (C) Heatmap of differentially-expressed genes defined at the anatomical joint level (see STAR Methods). (D) Proportion of DEGs in the indicated bone-end tissues which demonstrate significant bias in expression across early/late gestational timepoints.

To identify biologically-important processes occurring at individual bone-ends or joints, we used our dataset to delve into site-specific expression programs at high-resolution. To showcase this, Figure S2D–E illustrates distal femur-specific gene expression. Genes up-regulated in the distal femur (relative to all other sites) were enriched for ‘skeletal system development’, ‘cartilage development’, and ‘Wnt signaling pathway’ GO terms (Figure S2F, Table S2 for other sites).

Given the clustering of both limb-type and joint tissues by accessibility and expression analyses, we compared expression between aggregated forelimb/hindlimb and all four appendicular joint groups (Table S2, STAR Methods). These analyses yielded complementary sets of DEGs and GO enrichments (between limb-type and joint-site) suggesting that gene expression programs are likely influenced by a combination of membership in different joints and limb types. Genes differentially-expressed in a given joint tended to be exclusively up-regulated for that joint, with the highest degree of DEG overlap between joints of a given limb-type (Figure 3, Table S2). These analyses recovered master regulators of development matching their limbs; e.g.,TBX5 is strongly up-regulated in elbow and shoulder, but not in knee nor hip, whilst TBX4 demonstrates the opposite expression pattern – following their roles in limb-type biology¹⁴. Similarly, the classic hindlimb marker PITX1 was up-regulated in knee (and hip) tissues where it has significant impacts on knee development¹⁵, along with downstream targets HOXC10, HOXC11 and TBX4¹⁶, as was ISL1¹⁷, amongst others. Considering time-sensitive expression patterns at the joint-level (Figure 3A–D, Table S2) we found that, broadly, time-biased genes across joints were associated with different GO enrichments (Figure 3F). We next performed a simplified integration of our datasets to identify possible regulators of joint development (Figure 3E, STAR Methods).

Figure 3:

Divergent regulatory activity across developing anatomical joints. (A-D) Description of joint-specific regulatory behaviors, broken down by individual anatomical joints (A = elbow; B = shoulder; C = knee; D = hip). Within each panel: (Top), Sequence motif enrichments for joint-specific enhancers in the indicated anatomical joint. Plots indicate −log10(e-value) of motif enrichment via AME. Red dashed line indicates significance cut-off (e-value < 0.05). (Bottom Left), Expression volcano plot for genes defined as differentially-expressed in the indicated joint. Positive log2FC values indicate up-regulation. Red dots represent significant genes at a log2FC greater than 1. (Bottom Right), (Upper pie chart): Proportion of genes up-regulated in the given joint which are biased towards increased expression earlier in development (E50–54) (pink), later (E55–59) (green), or unchanged over time (blue). (Bottom pie chart): Proportion of joint-upregulated genes which are either exclusively upregulated or also upregulated in at least one other joint. (E) Combining ATAC- and RNA-seq datasets; putative regulatory elements which demonstrated joint-specificity were scanned for differentially-expressed TF motifs, and further subset to joint-specific genes. (F) GO term enrichments for pooled genes biased towards early (left, red) or late (right, green) developing skeletal expression in any given joint.

Finally, at this window (E50-E70) of skeletal development we assayed, the cartilage anlage should be comprised predominantly of actively secreting chondrocytes. Using previous scRNA-seq data from the limb bud¹⁸ (Figure S2G–I, STAR Methods), we confirmed the homogeneity of cell-types within our developing cartilage datasets and note that our transcriptional profiling of bulk cartilage tissues is likely not masking more subtle cell-type-specific behaviors.

Overall, our findings on the epigenomic and transcriptomic landscape of human skeletal development provides insights into the different suites of regulatory programs exhibiting anatomical site-specificity. Given that one would expect adult human height to reflect the contribution of growth programs across the skeleton, we next leveraged our datasets to investigate the links between developmental processes and genetic heritability for height.

Integration of genetic associations for height with skeletal epigenetic and transcriptomic context:

In line with findings in complex trait biology^19,20, we posit that genetic variants falling within chondrocyte regulatory regions should have regulatory effects on gene expression – impacting development and/or growth of individual skeletal elements, as well as the skeleton as a whole and thus height. To explore the interactions between genetic variation, epigenetic activity, and height biology, we utilized a recent meta-analysis of height GWAS studies² spanning millions of individuals and variants. We first considered fine-mapped variants defined through this meta-analysis² (via COJO²¹) and looked for evidence of regulatory activity across the skeleton. Our fine-dissection of skeletal sites allows us to stratify candidate loci into those active across the skeleton, and those whose activity – and possible regulatory effect(s) - are spatially restricted. As expected, given the degree of epigenetic overlap observed with ATAC-seq datasets (Figure 1), we found that most fine-mapped variants within regulatory elements were epigenetically active across the skeleton (Figure 4A), while a subset of variant activity was restricted to specific sites. We observed similar distributions of regulatory pleiotropy when considering variants defined with a stringent (1e-8) and relaxed (1e-6) genome-wide significance threshold (Figure S3A).

Figure 4:

Integration of genetic height associations with functional genomics of skeletogenesis. (A) Proportion of fine-mapped height-associated variants falling within different categories of ATAC-seq regions, from broadly-pleiotropic to highly tissue-specific. (B) Fold-change enrichments/depletions for fine-mapped height-associated variants falling within different ATAC-seq region sets (and aggregates) (Significance stars: n.s.: not-significant, * < 0.05, ** > 0.001, *** < 0.001). (C-E) Fold-change enrichments of significant variants for: (C) sitting height, (D) leg-length, (E) standing height, from UK Biobanks. (F) Fold-change enrichments for variant sets in C-E using axial skeleton epigenetic datasets.

We next tested the significance of overlaps in epigenetic activity with strongly-associated genetic variants. When considering our broadest set of elements (aggregated fore/hindlimb), these sets captured a greater number of significant variants than expected by chance (STAR Methods) (Figure 4B, Table S3). Putative regulatory elements active in brain tissue were also significantly enriched for associated variants - however, this signal was driven largely by pleiotropic elements shared with chondrogenic datasets, with brain-specific regions lacking variant enrichment (Figure 4B, Table S3). Additionally, we found that elements accessible exclusively in the knee (and secondarily elbow) were enriched, while the more proximal hip (and shoulder) was not. We observed similar patterns of enrichments when using variants defined via significance cutoffs (Figure S3C–D). Finally, we observed largely similar patterns of set enrichments using later (E60’s) timepoint accessibility data (Figure S3E–G).

Interestingly, we found that performing enrichment tests using progressively more refined sets (e.g., distal-femur-specific elements), resulted in diminishing enrichments for variant capture (Figure 4B). This does not negate the potential functional significance of individual variants falling within site-specific regulatory elements, as some genetic variants do have large, demonstrable skeletal growth effects²². However, this overall trend does suggest that site-specific regulatory variants of large effect size are the exception for common variants, rather than the rule, for height associations. Instead, variants with broader activity domains across the skeleton tended to be more associated with height (STAR Methods, Figure S3I–J, Table S3), while element sets not filtered for tissue-specificity (i.e., ‘general’ sets) tended to have greater variant enrichments (Table S3). Both results are consistent with a model in which additive regulatory effects on pleiotropic processes (e.g., general cell growth) is the predominant way in which height variation is achieved. We also considered the degree of enhancer modularity at gene loci, again finding that the broader the pleiotropy in regulatory activity, the greater the associations with height (Table S3, STAR Methods). Finally, we compared the per-locus density of fine-mapped height variants (from Yengo et al.²) across loci binned by their prevalence of overlapping ATAC-seq activity, finding a strong significant positive correlation (Figure S3K).

We next took the expanded genome-wide variant set and performed a partitioned heritability enrichment analysis using LDSC²³ (Figure S3B). Broad sets of elements were enriched for heritability, with our pooled set of skeletal elements capturing approximately 26% of total height heritability (Table S3). We also observed a slightly greater enrichment in hindlimb sets relative to forelimb, enrichment exclusive to the knee, and significant depletion within brain-specific regions (Table S3). We were unable to accurately run this analysis using the most specific element sets, owing to their substantially smaller set size²³. This trend favoring elements of increasing pleiotropic activity was also observed at our later timepoint (Figure S3H).

Previously, high-resolution expression datasets have been used to stratify heritability signals for other complex traits (e.g., schizophrenia²⁴, Parkinson’s disease²⁵, etc.). Therefore, we next asked whether expression patterns across different skeletal elements varied in their association with height heritability (STAR Methods). We found that differential-expression of genes in specific elements (relative to all other elements) was not significantly associated with height heritability via cell-type-specific LDSC (Table S3), similar to our findings on site-specific regulatory elements. We also used a gene-based ranking method to compare the associations between genes broadly expressed in skeletal tissues to those for which expression was extremely low or undetected (see STAR Methods). Skeletally-expressed genes tend to be more strongly associated with height GWAS than those not strongly expressed (Table S3), echoing Wood et al’s.²⁶ previous findings. We also considered the relationship between height-associations of nearby genetic variants and the site-specificity of gene expression, finding that broadly-expressed genes tended to have substantially greater association with height than genes differentially-expressed in only one tissue (Table S3). Likewise, genes up-regulated in hindlimb tissues (relative to forelimb) also had slightly greater association. Finally, given a prior expectation that genes not expressed in the relevant tissue will have lower associations with a given heritable trait^5,26, we considered fetal brain expression²⁷, both collectively as well as genes uniquely expressed in brain (relative to our chondrocyte datasets). Genes expressed in fetal brain were enriched for height heritability, as were genes expressed in both datasets (Table S3). However, these enrichments were driven by pleiotropically-acting genes shared with chondrocyte datasets, as genes exclusively expressed in the brain were significantly-less associated with height (Table S3).

Partitioning genetic associations for height, sitting height, and leg length using human skeletal data:

Thus far, when we discuss ‘height genetics’, we have been referring to the standing height of an individual. However, standing height is a composite of leg-length (reflective of the appendicular skeleton), and trunk-height (reflective of the axial skeleton plus head). Given our regional datasets/analyses across the appendicular skeleton, we sought to map the individual components of height to their respectively-matched skeletal components. To this end, we generated ATAC-seq datasets from thoracic and lumbar vertebra from late E60s/early E70 fetal tissues (Table S1) (STAR Methods). To more explicitly disentangle the epigenetic signals of axial vs. appendicular development, we intersected these axial datasets with later time-point (E67) appendicular ATAC-seq datasets. Briefly, for axial sets we subtracted any region intersecting the set of pooled appendicular ATAC-seq regions, then further intersected thoracic and lumbar regions to define mutually exclusive sets. Similarly, for site-specific appendicular sets we subtracted regions intersecting any axial regions.

A recent re-analysis of UK Biobank data identified fine-mapped variants independently associated with standing height, leg length, and sitting height (i.e., trunk length)²⁸. We tested for variant enrichments across our epigenetic features (Figure 4C–F, Table S3) finding that regulatory regions active in the appendicular skeleton were enriched for both components of height, sitting height (Figure 4C) and leg length (Figure 4D), as well as overall height (Figure 4E). In contrast, axial regions were independently enriched for sitting height and overall height (with no apparent further differentiation between lumbar and thoracic sets) (Figure 4F) but were not enriched for leg-length. Similar patterns were observed when using significance-thresholded variants (Table S3) (Discussion).

Finally, in addition to leveraging our datasets for understanding the overarching architecture of height genetics, we also integrated these data to reveal insights into skeletal biology. To this end we performed several integrative analyses, identifying key pathways and candidate loci possibly involved in skeletal growth and height patterning (Figure S3L–M, Table S3, STAR Methods).

Extending GWAS loci information using regulatory information:

Biological insights from GWAS are often enhanced by integrating regulatory information^29–31. We therefore performed several analyses to extend fine-mapped height associations with our functional genomics datasets (STAR Methods). Integration of such datasets revealed additional variants in fine-mapped loci which had equal-to-better predicted activity to the original lead variant (Figure S4, Table S4). Many of these newly-identified variants overlapped with chondrocyte chromatin-interaction data, further pointing to their functional relevance to regulating skeletal development. Finally, given our substantial RNA-seq datasets, we asked whether we could identify variants which impact gene expression within developing human cartilage. Using the sequenced reads to genotype the transcriptome, as has been done previously^32,33 (n = 70 samples, Table S2, STAR Methods), we identified putative eQTLs (which also show regulatory effects in other tissues via GTEx³⁴) with putative regulatory activity via RegulomeDB³⁰. Several of these variants were predicted to regulate expression of key cartilage development genes; we highlight one example at the ECM1 locus (Figure S4G). ECM1 regulates chondrogenesis, and its genetic manipulation leads to defects in bone formation³⁵. These analyses point to the utility of our trait-relevant datasets to extend the functional interpretation of height GWAS findings.

An approach for assessing omnigenicity and discerning the underlying regulatory architecture of height biology:

The Omnigenic^5,6 model conceives of diffuse genetic heritability across the genome, with the aggregate of many small-effect trans-QTLs integrating on ‘core’ genes (e.g., via protein-protein interactions or regulatory relationships) more proximally involved in controlling a phenotype (Figure 5A). Here, ‘core’ genes may be subject to co-regulation, such that the genetic variance of the trait is dominated by covariance terms between core genes sharing similar directions-of-effect - possibly reflecting regulatory modules⁶. Accordingly, as a key goal of this study was to determine whether height – the ‘classic’ complex polygenic trait – conforms to expectations of omnigenicity^5,6, we developed an objective approach to address this relationship.

Figure 5:

Gene co-expression modules and the genetic architecture of height heritability. (A) Schematic diagram illustrating proposed framework for ‘Omnigenic’ height architecture. Module genes share co-expression patterns (black lines). Gene modules (green circles) are defined by this co-expression and are regulated upstream factors such as transcription factors (TF). These modules may also capture ‘core’ genes most proximately involved in height biology (innermost circle). Locus-based associations with height are expected to increase for genes moving from the periphery to the center of figure. (B) GO Biological Process enrichments for gene modules, showing the top two enriched categories for all modules enriched via MAGMA analysis. (C) Enrichments for height associations of indicated models based on comparisons to randomized size-matched gene sets (see Table S5). Vertical red lines indicate −log10(0.05) threshold for enrichment (right) or depletion (left) of association signal. (D) Distributions of aggregate locus-wise height association statistics for genes in a given module (see STAR Methods). Top left: Heritability (h²) distribution for the CD1 module. Top right: Absolute effect-size (beta) distribution for the CD1 module. Beta-distribution graphs for all tested modules shown in Figure S5.

i). Identifying gene co-expression modules.

We first reasoned that identifying modules of genes sharing co-expression patterns across bone-ends and biological replicates may allow us to objectively capture components of the set of true ‘core’ genes regulating height, similar to how expression modules across psychiatric conditions have been suggested to contain such core genes^6,36. We note that not all such co-expression modules must be categorically involved in height patterning, and indeed many may comprise general cellular processes active in development. By leveraging our large transcriptomic datasets, consisting of many samples and bone-end-types across the skeleton, we defined co-expression modules (Figure S5A, Table S5; CD1-CD30; STAR Methods). Using GO gene-set enrichments, we identified broad biological processes associated with each module, which varied from basic, intrinsic terms such as ‘RNA processing’ and ‘translation’ to more specific terms such as ‘skeletal system morphogenesis’ and ‘ossification’ (Figure 5B, Table S5). There appeared to be little tissue-specific bias in terms of expression across different modules (Figure S5B), though there was heterogeneity in the degree of expression between modules - with some modules more highly expressed across tissues (Figure S5C). We note that the three most highly-expressed modules have enrichments related to translation and protein synthesis, RNA processing, and cartilage/bone development, consistent with our characterization of expression patterns during a period of rapid skeletal growth. Compared to randomly-sampled gene sets, these modules had significantly-elevated expression (Table S5). Comparing the expression patterns of modules relative to one another, there was no apparent higher-order organization/structure - though we noted that CD1 stood apart from other modules (Figure S5, STAR Methods).

We next considered the association of these modules (CD1–30) with height heritability, finding significant enrichment amongst several for height signals (Figure 5C, Table S5). Interestingly, the most strongly enriched modules were not those that had the strongest cross-tissue expression (Figure S5C). This suggests that while, generally, increased skeletal expression is associated with height, it is the composition of gene sets themselves that is also important. Finally, as an orthogonal confirmation of these gene-set associations with height we applied a MAGMA³⁷-based method (Table S5, Figure S5F).

ii). Modules capture trait-relevant biology

Given our findings that the CD1 module had amongst the strongest enrichments for per-gene height associations (Figure 5C) and unambiguous gene-set enrichments (i.e., carbohydrate metabolism, Figure 5B), and the Omnigenic prediction that captured ‘core genes’ underlies module height associations, we explored CD1’s relationship with skeletal biology. We first assessed the contribution of this module to height heritability by using LD blocks from HESS³⁸, looking at the per-block height heritability of LD blocks containing genes from this module (Figure S5G). We found that LD blocks implicated by genes in CD1 captured a disproportionate amount of total height heritability (~26% of genome-wide total, a 1.74-fold enrichment), in line with previous findings of enriched height signal with this module. Across this module, the distribution of heritability was non-normal, with several genes acting as substantial outliers (Figure S5G,I). For example, PPARD fell within a LD block capturing ~1.3% of the genome-wide total. This gene, encoding a nuclear receptor transcription factor protein, has been associated with height in multiple GWAS^39–41. As an additional confirmatory analysis, we considered the proportion of COJO fine-mapped loci around CD1 genes, finding that this module again captured a disproportionate fraction of loci and associated heritability (~30% of genome-wide total, a 1.81-fold enrichment, Figure S5H,J).

We next characterized the distribution of genes within CD1. Across GO terms represented in CD1 there was moderate heterogeneity in average bone-end expression - with ‘collagen fibril organization’ genes having significantly-elevated expression relative to other terms (Figure S5K, Table S5). This term included several genes involved in skeletal development and chondrogenesis, including COL2A1⁴² (one of the most highly-expressed genes in our expression dataset), LOXL2⁴³, ACAN⁴⁴, and others. The distribution of per-gene heritabilities within this module was homogenous across terms (Figure S5L).

iii). Identifying putative hierarchical regulators acting on gene modules:

Given the proposed concept of core gene co-regulation, we next sought to define possible upstream regulatory factors of our module sets. Using our epigenetic datasets generated across the skeleton, we first looked for enrichments for known TF motifs occurring within regulatory elements around module genes (STAR Methods). Several TFs whose motifs were enriched around skeletally-expressed genes (generally) have been linked to skeletogenesis (Table S5). These include KLF4⁴⁵, EGR2⁴⁶, and FOXP1⁴⁷. We further asked whether different module sets diverged in TF enrichments (Figure S5M). The CD1 module again behaved differently and showed strong enrichments across a broad range of factors despite being comparable in size to other modules. Several TF enrichments were restricted to a subset of modules (Figure S5N). For example, HIF1A is only enriched for three modules, most strongly around CD1 genes; the hypoxia response pathway has been implicated in chondrogenesis and osteogenesis during skeletal development^48–50. The repressive factor ZBTB33 was enriched solely around CD1 module genes. ZBTB33 is a regulator of Wnt/beta-catenin signaling during development⁵¹ and may play a role in vertebrate development⁵². Additionally, the ETS family factor ERG is expressed in developing joints and is important for regulating growth-plate chondrocytes⁵³.

As some TFs should act as regulators of transcriptional networks underlying height biology, we expect that variants falling within motifs for certain factors (e.g., if they modify binding affinity) may be more strongly associated with height. We took the set of all TFs with module-restricted enrichments and asked whether motif-intersected variants were biased in their p-values relative to a randomized background (STAR Methods). After multiple hypothesis correction, we found three TFs biased towards greater association genome-wide, and which have been associated with chondrocyte and bone biology. These included FOXP1⁴⁷, and ZNF384⁵⁴ (Discussion). We next asked whether TF motifs enriched around a given module (e.g., CD1) were also biased in SNPs around genes of that module. Collectively, the set of factors enriched around CD1 genes intersected variants with stronger association p-values than variants in putative regulatory regions generally (Table S5). This was also true using a more stringent test where we constrained our background to variants falling nearby CD1 genes (but not necessarily falling within these TF motifs). Testing these factors individually, we identified nine with significantly-increased association after multiple-hypothesis testing (Table S5, Discussion). Given this, we might expect that the genes encoding these TFs would themselves be strongly-associated loci. However, we found this to not be the case (Table S5). Finally, we considered the possibility that variants hitting TF binding motifs may be biased in their effect – i.e. disrupting or improving predicted binding affinity (STAR Methods). While we did not observe global biases in predicted variant effects for any of the TFs we identified, we found that height variants intersecting putative FOXP1 sites around CD1 genes were biased in their effects – with variants improving FOXP1 motifs tending to be height-increasing, and CD1 genes enriched for FOXP1-regulated targets via CHEA⁵⁵ (Figure S5O, Table S5). Broader trends in signed-variant effects were observed using GO terms (Figure S5P).

iv). Regulatory networks around modules integrate genetic heritability signals

Through our analyses we defined and characterized gene modules, which we propose contain ‘core’ genes proximately involved in height biology, and whose expression is co-regulated via upstream trans factors. As a key tenet of the proposed Omnigenic framework, we considered the propagation of height heritability via network effects, choosing to focus on regulatory relationships between gene loci. Combining multi-tissue eQTL datasets³⁴ with TF-target databases^56,57, we aggregated regulatory variants which may regulate a given module, conditioned on epigenetic activity in developing skeletal tissue (STAR Methods). We considered per-variant associations with height (Figure 5D), observing a unimodal distribution of heritabilities. Next, we compared heritability distributions of our modules to those generated using randomized module sets. Importantly, we established our background distributions by sampling from the pooled set of module genes to define randomly-scrambled modules – that is, we consider sets of genes expressed in skeleton, but with no underlying genetic architecture as predicted by the Omnigenic model. This ensures that any given TF is equally likely to regulate a randomized ‘module’ expressed in cartilage, and accounts for the observed positive relationship between skeletal gene activity and height heritability. Furthermore, we only considered putative trans-regulatory factors (TFs) that were expressed in cartilage, to avoid a scenario in which cartilage TFs, regulating cartilage-expressed genes, exhibit larger height heritability than TFs not expressed in cartilage. Across several modules, the aggregate cis-QTL heritability of upstream regulators was significantly higher than expected relative to modules contained randomized genes (Figure 5D) – despite these genes also being expressed in the skeleton (along with their predicted upstream regulators).

Many of these modules also had significantly increased distributions in terms of signed and absolute effect-size (beta-coefficients) (Figure S5Q–R). As a point of comparison, we also considered skeletal genes with human phenotypes from OMIM as an artificial ‘module’. Remarkably, several of our co-expression modules had significantly-greater aggregated heritability compared to OMIM genes (Figure 5D). In fact, OMIM genes tended to not be significantly-shifted relative to randomized gene sets (Figure 5D, Table S5) – this, despite strong skews in absolute effect-size (Figure S5R). This behaviour may reflect the nature of OMIM genes and how they are historically discovered/defined (Discussion). Through this network analysis, we suggest that the observed biases in network-associated QTL heritabilities of some module sets reflect their aggregate role in height biology (Discussion).

v). A general framework for understanding complex trait genetics

Our work focused on studying height to develop a conceptual framework for understanding the genetic architecture of complex traits; generating and leveraging functional genomics datasets to ground this framework in the proximate cellular biology acting to shape/influence the emergent phenotype (i.e. standing height). This framework, which was built on the Omnigenic model^5,6, need not be limited to skeletal development and the patterning of human height; here, we propose its potential for broader utility to the understanding of other complex traits. To illustrate this, we applied our framework to type II diabetes (T2D). T2D is highly polygenic⁵⁸, has well-studied trait biology proximately associated with the pancreas⁵⁹ (though other organs also contribute⁶⁰), and whose genetic architecture supports an Omnigenic model⁶¹. We similarly defined co-expression modules for genes expressed in the adult pancreas which were enriched for terms related to protein synthesis (Figure 6A, Table S6, STAR Methods). Several of these modules were significantly enriched for T2D genetic association signals (Figure 6B, Figure S6A). We similarly defined sets of upstream transcriptional regulators (TFs) for each module and, conditioning on pancreatic chromatin accessibility, aggregated QTL variant data across modules. As seen in our above height analysis, we observed that several of our pancreas modules yielded aggregate per-variant association distributions significantly increased relative to randomized size-matched module sets (Figure 6C, Figure S6B). All module definitions and statistical results are in Table S6.

Figure 6.

Application of framework to genetic associations of Type 2 Diabetes. (A) Top two enriched GO terms for modules yielding significant association tests (see below). Full GO enrichments provided in Table S6. (B) Enrichments of modules for gene-set associations with Type 2 diabetes GWAS variants. Vertical red lines indicate −log10(0.05) for either enrichment (right) or depletion (less) of association signal. See Figure S6A for MAGMA-based enrichment tests. (C) Distributions of aggregate locus-wise T2D association statistics for genes in a given module. Top: Distribution of per-locus averaged −log10(p-value) for a given module, compared to the distribution obtained from a randomly generated, size-matched ‘module’. P-value shown is the rate at which a randomized module has a significantly greater distribution than the indicated module via t-test (see STAR Methods, Table S6). Middle: For 100K random size-matched modules, the distribution of set-averaged locus −log10(p-value) is shown. Vertical red and turquoise bars indicate average of background and indicated module distributions, respectively. P-value shown is the frequency of random modules with an average value greater than the target set and acts as an additional stringent test. Bottom: For the same set of 100K random size-matched modules, the distribution of set-averaged locus absolute effect size (beta coefficient) is shown. Vertical bars and p-values are as in middle. For modules falling to left of distribution, p-value is calculated as 1 – (frequency).

Discussion

Here, we generated a large, high-resolution dataset of epigenetic and transcriptomic activity throughout the appendicular and axial skeleton for a developmental period when skeletal elements markedly and differentially grow. This approach allowed us to tease apart individual genes with anatomical specificity in their expression domains, along with the putative regulatory regions underlying the development of chondrocyte-mediated skeletal regionalization. Leveraging this knowledge of site-specific skeletal regulatory networks, we help disentangle the developmental genetic regulation of skeletal growth as it pertains to standing height, sitting height, and leg length. More broadly, we explore the genome-wide relationship between regulatory activity and height heritability during skeletogenesis.

This was inspired by other works in complex trait genetics, wherein genomic heritability implicated by relevant tissues, anatomical regions, and cell-types for related trait associations have led to deeper understandings of how traits manifest^19,62,63. Specifically, we sought to characterize the contributions of individual bone-ends to height heritability by integrating our data with a recent GWAS meta-analysis for height². Counter to our expectations, we did not observe substantial differences in association between specific anatomical sites. Instead, we observed stronger genomic enrichments for broader trends in accessibility, namely biasing towards hindlimb and knee tissues, and greater gene-wise enrichments for genes with less-restricted expression patterns. In general, we found that the number of regulatory regions in a locus (aggregated across tissues), along with average skeletal gene expression, had the strongest influence on gene-set associations with height, and that, on a per-variant basis, associations with height trended with increased pleiotropy in accessibility across the skeleton.

Our results suggest that general growth programs active across the skeleton may have the greatest impact on patterning adult height; the broader the domains of expression/accessibility of genes and regulatory elements (respectively), the greater the influence on skeletal growth (and thus, height). Of course, we cannot rule out that site-specific activity may play a greater role in patterning height at some later developmental time-point, i.e., post-natally. Indeed, accelerated growth at the hindlimb epiphyseal growth-plates contributes gains in height observed in pre-pubescent individuals⁶⁴. However, such effects are complicated by the role of environmental factors, e.g., nutrient status, sex hormone signaling, immune function, etc., which are known to impact post-natal skeletal growth⁶⁵.

Based on the above, we more deeply investigated the polygenic architecture of growth processes underlying height biology and variation^5,7,26,66 using the Omnigenic model^5,6. In this context, we developed a conceptual framework that we suggest can be applied to any complex trait for which suitable regulatory, transcriptomic, and genetic-association datasets are available to address trait omnigenicity; all-the-while revealing important biology for that trait.

First, reasoning that sets of genes impacting trait biology, if exhibiting coordinated expression as modules, would have outsized impacts on trait heritability collectively (Figure 5A). We identified one module, CD1, which had amongst the strongest enrichments for per-gene associations with height (Table S5). CD1 was strongly enriched for genes involved in carbohydrate metabolism and small-molecule metabolic processes (e.g., ATP synthesis) (Table S5, Figure 5B). Carbohydrates are a key metabolite for chondrocytes⁶⁷ during the endochondral ossification process that governs long-bone growth. Glycolytic enzymes have been associated with skeletal defects in mice⁶⁸, including those of the aldolase family, of which CD1 contains two members. Such sugars also serve as precursors to more complex biomolecules such as proteoglycans, key structural components of developing cartilage models and growth-plates⁶⁹. CD1 also contained key cartilage proteoglycans and glycoproteins, including ACAN⁴⁴, COL2A1⁴², and COMP⁷⁰, as well as genes involved in glycosylation and proteoglycan modification, such as B3GALT6 and CHST14, which are associated with developmental disorders and connective tissues⁷¹. Furthermore, this module captured a disproportionate number of genes known to be involved in abnormal skeletal development (via OMIM)⁷². We suggest that modules such as CD1 reflect a part of the ‘core’ set of genes involved in height biology.

Second, we considered possible upstream regulators of these modules. We identified three TFs whose motifs tended to harbor variants with greater height association globally, and nine TFs with biases around CD1 genes (Table S5). That the loci encoding these TFs were collectively not strongly associated with height may indicate heightened selective constraint, as previously suggested⁶. Furthermore, one of these factors, FOXP1, had significant skews in both the height-association and direction-of-effect for motif-hitting variants – with variants improving FOXP1 motifs tending to be height-increasing. This was specific to certain modules (i.e. CD1) and GO terms (i.e., those involved in developmental and general cellular processes). FOXP1 inhibits ossification⁴⁷, with mutations leading to substantial and growth-related skeletal phenotypes in mice⁴⁷ and possibly humans⁷³. A recent scRNA-seq study of human long-bones at a similar developmental time-point¹⁸ found that cells in the osteochondral lineage (which includes chondrocytes) were highly enriched for FOXP1/2 network activity, while FOXP1/2 expression was detected in the perichondrium of long-bone ends. We posit that FOXP1 may be one example of a key ‘master regulator’ of height biology, acting to regulate gene sets involved in height biology, such as cellular metabolism and proliferation.

Third, we tested a key expectation of the Omnigenic model: that such modules, if impacting height biology as a set via their co-regulated expression, would have an aggregate effect on height heritability via regulatory network effects. By considering the putative upstream regulators controlling module expression, we found that several co-expression modules exhibited increased distributions in aggregate height-heritability, as well as effect size, relative to randomized module sets (Figure 5C, Figure S5). We compared these modules to a set of genes with known impacts on human skeletal development (from OMIM⁷²). Unexpectedly, we found that these genes collectively had a weaker distribution of associations than our modules, performing little better than randomized sets of genes. Phenotypes of OMIM genes are, by definition, of large effect size, and many of these loci are individually strongly associated with height. However, given that these genes are often discovered through clinical observation of skeletal abnormalities, they may not necessarily share co-expression patterns with one another, nor share upstream regulators. In contrast, given our interest in uncovering new height biology, our gene modules are explicitly defined by co-expression, not biasing based on known height-associated genes. Thus, while the individual height effects of module members may be more moderate than OMIM genes, their network effects on height genetics are more likely to be directional, and additive, in aggregate.

Finally, we asked whether the framework we developed and tested in the context of height is applicable to the study of complex trait genetics more broadly. We considered another complex trait, Type 2 Diabetes^58,59, for which relevant epigenetic, transcriptomic, and genetic association datasets are available. Applying our framework, we found that the patterns/distributions of genetic associations we observed in studying T2D genetics behaved similarly to those identified in our analyses on height genetics and aligned with expectations based on our conceptualization of the Omnigenic model. We note that T2D is largely associated with one organ/tissue, similar to height (though a potential simplification of additional complexity in T2D disease biology⁶⁰). Testing and expanding our framework to the study of complex traits with more integrative aetiology (e.g. hypertension⁷⁴) is a compelling direction of future research.

Resource Availability

Lead contact

• Further information and requests for resources, reagents and analysis code should be directed to and will be fulfilled by the Lead Contact, Terence Capellini (tcapellini@fas.harvard.edu).

Materials availability

• This study did not generate new unique reagents.

Data and Code availability

• All sequencing datasets (ATAC-seq and RNA-seq) raw fastq files and processed data have been deposited on NCBI GEO (GSE252288 and GSE252289) and are publicly available as of the date of publication. Additional external datasets used in this paper are outlined in the Key Resources Table and STAR Methods.

• This paper does not report original code.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Bacterial and virus strains
Biological samples
Human gestational samples, E53-E67	University of Washington Birth Defects Research Laboratory	NA
Chemicals, peptides, and recombinant proteins
Collagenase II	VWR	80056-222
Nextera Tn5 transposase	Illumina	FC-121-1030
Critical commercial assays
Direct-zol RNA Microprep Kits	Zymo Research	R2062
KAPA mRNA HyperPrep Kit	KAPA	KK8580
KAPA Library Quantification Complete Kit	KAPA	KK4824
NEBNext High-Fidelity PCR Master Mix	NEB	M0541L
Deposited data
Human Appendicular Skeleton ATAC-seq	This paper	GSE252289
Human Appendicular Skeleton RNA-seq	This paper	GSE252288
Mouse reference genome NCBI build GRCm38.p6	Genome Reference Consortium	https://www.ncbi.nlm.nih.gov/assembly/GCF_000001635.26
Human reference genome NCBI build 37, GRCh37	Genome Reference Consortium	https://www.ncbi.nlm.nih.gov/projects/genome/assembly/grc/human/
Human E54 brain ATAC	Young et al. 202212	GSE165938
GIANT Consortium Datasets	Yengo et al. 20222	https://portals.broadinstitute.org/collaboration/giant/index.php/GIANT_consortium_data_files
Human fetal brain RNA-seq data	Davis et al. 201827	ENCODE: ENCFF654GII
Human chondrocyte promoter-capture Hi-C	Thulson et al. 2022111	GSE200345
LINSIGHT scores	Huang et al. 2017117	https://github.com/CshlSiepelLab/LINSIGHT
RegulomeDB scores	Boyle et al. 201230	regulomedb.org
PathwayCommons	Rodchenkov et al. 202056	https://www.pathwaycommons.org/archives/PC2/v12/
hTFtarget	Zhang et al. 202057	http://bioinfo.life.hust.edu.cn/hTFtarget#!/download
HOCOMOCO v11	Kulakovskiy et al.2018137	https://hocomoco11.autosome.org/
FOXP1 targets (CHEA database)	Lachmann et al. 201055	https://maayanlab.cloud/Harmonizome/gene_set/FOXP1/CHEA+Transcription+Factor+Targets
Adult pancreas RNA-seq	GTEx Consortium 201334	https://www.gtexportal.org/home/downloads/adult-gtex/bulk_tissue_expression
T2D GWAS summary statistics	Suzuki et al. 202460	http://www.diagram-consortium.org/downloads.html
Adult pancreas ATAC-seq peaks	Zhang et al. 2021140	http://catlas.org/catlas_downloads/humantissues/
gnomAD constraint scores v4.1	Karczewski et al. 2020142	https://gnomad.broadinstitute.org/news/2024-04-gnomad-v4-1/
Human developing limb bud scRNA-seq dataset	He et al.18	GSE143753
Experimental models: Cell lines
Experimental models: Organisms/strains
Oligonucleotides
Recombinant DNA
Software and algorithms
FastQC version 11.5	Babraham Institute	https://www.bioinformatics.babraham.ac.uk/projects/fastqc/
TrimGalore! version 0.6.5	Babraham Institute	https://github.com/FelixKrueger/TrimGalore
Salmon version 1.4	Patro et al. 201776	https://combine-lab.github.io/salmon/
R version 3.6.1	R Development Team77	https://www.r-project.org/
tximport version 1.14.0	Soneson et al. 201678	https://bioconductor.org/packages/release/bioc/html/tximport.html
DESeq2 version 1.26.0	Love et al. 201479	https://bioconductor.org/packages/release/bioc/html/DESeq2.html
ComplexHeatmap version 2.18	Gu et al., 201682	https://github.com/jokergoo/ComplexHeatmap
clusterProfiler version 3.13.1	Yu et al.83	https://bioconductor.org/packages/release/bioc/html/clusterProfiler.html
STAR version 2.7.0	Dobin et al. 201384	https://github.com/alexdobin/STAR
Bowtie2 version 2.3.2	Langmead and Salzberg 201288	https://bowtie-bio.sourceforge.net/bowtie2/index.shtml
Picard version 2.9.0	Broad Institute	http://broadinstitute.github.io/picard
MACS2 version 2.1.1.2	Zhang et al. 200889	https://github.com/taoliu/MACS
IDR version 2.0.3	Li et al. 201190	https://www.encodeproject.org/software/idr/
bedtools version 2.27.1	Quinlan and Hall 201091	https://bedtools.readthedocs.io/en/latest/
GREAT version 4.0.4	McLean et al. 201013	http://great.stanford.edu/public/html/
UpsetR version 1.4	Conway et al 201793	https://github.com/hms-dbmi/UpSetR
LDSC	Finucane et al 201523	https://github.com/bulik/ldsc
RobustRankAggreg version 1.1	CRAN	https://cran.r-project.org/web/packages/RobustRankAggreg/index.html
PAINTOR v3.0	Kichaev et al. 2017116	https://github.com/gkichaev/PAINTOR_V3.0
GATK version 4.3.0	DePristo et al. 2011125	https://gatk.broadinstitute.org/hc/en-us
1000Genomes phase 3	Gibbs et al. 2015126	https://www.internationalgenome.org/
BEAGLE version 4.1	Browning et al.2018127	https://faculty.washington.edu/browning/beagle/beagle.html
plink version 1.90b7	Chang et al. 2015128	https://www.cog-genomics.org/plink/
QTLtools	Delaneau et al.2017129	https://qtltools.github.io/qtltools/
WGCNA version 1.72	Langfelder et al.2018131	https://cran.r-project.org/web/packages/WGCNA/index.html
gprofiler2 version 0.2.2	Kolberg et al. 2020132	https://cran.r-project.org/web/packages/gprofiler2/index.html
HESS version 0.5.4	Shi et al. 201638	http://bogdan.bioinformatics.ucla.edu/software/hess
MEME	Bailey et al. 200998	https://meme-suite.org/meme/
SLDP	Reshef et al. 2018136	https://github.com/yakirr/sldp
rrvgo version 1.10	Sayols 2023138	https://bioconductor.org/packages/release/bioc/html/rrvgo.html
Other

STAR Methods

Experimental model and study participant details

Human samples: Human products of conception at E53-E67 were collected from the first-trimester termination through the Laboratory of Developmental Biology at the University of Washington in Seattle, the University of Washington in full compliance with the ethical guidelines of the National Institutes of Health (NIH) and with the approval of the University of Washington Institutional Review Boards for the collection and distribution of human tissues for research, and Harvard University for the receipt and use of such materials. The Laboratory of Developmental Biology obtained written consent from all tissue donors. The University of Washington Birth Defects Research Laboratory was supported by NIH award number 5R24HD000836 from the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD). Harvard University IRB determined this sample constitutes Non-Human Subjects Determination Status (Capellini: IRB16–1504). PI Capellini received no federal funds (e.g., NIH) to acquire, receive, process, or utilize these samples. The human samples were briefly washed in Hank’s Balanced Salt Solution (HBSS) and transported at 4 °C during shipment. Upon arrival the samples were dissected under a light dissection microscope and directly subjected to RNA-seq or ATAC-seq protocols described below, following approved Harvard University IRB (Capellini: IRB16–1504) and COMS (Capellini: 18–103) protocols. For this study, all limb elements were intact but then separated from one another manually, allowing to study each cartilaginous structure in isolation from bony diaphyses. Personal, racial, and ethnic information on human samples was not provided in agreement with ethical policies, nor was whole-genome sequencing performed for such considerations.

Method Details:

Human RNA-seq data generation:

Fresh long-bone samples were microdissected under a light microscope in 1× PBS on ice, epiphyses were separated, stripped clear of soft tissue, and collected in 2-ml tubes containing 200 μl of TRIzol and one 5-mm stainless steel bead. The right and left bone-ends were pooled from each donor sample. Each sample was then homogenized at 50 Hz for 2 min, followed by 1 min on ice, and a final 50-Hz homogenization for 2 min. Samples were then placed at −80°C until RNA extraction was performed. For each RNA extraction, samples were first subjected to a phenol-chloroform reaction in which samples were moved to a new microcentrifuge Eppendorf tube with 200 μl of chloroform per 1 ml of TRIzol in the sample. They were then vigorously vortexed and incubated at room temperature for 2 min before centrifugation for 5 min at 12,000g at 4°C. Following centrifugation, the aqueous phase was removed and transferred to a new microcentrifuge Eppendorf tube. After completion of this phenol-chloroform step, the RNA extraction continued using the Zymo Direct-zol RNA MicroPrep kit following the manufacturer’s protocols and eluted in 15 μl of nuclease-free water. Samples were then quantified using a Qubit and following the manufacturer’s protocols. Samples were also nano-dropped to determine 260/230 and 260/280 values and were then stored at −80°C. Aliquots of samples were also run on a TapeStation to determine their RNA integrity number (RIN) value. Only samples with RIN scores of 7 or higher were used for subsequent steps.

For cDNA library generation and sequencing, samples were normalized to a single concentration, and libraries were prepared using KAPA mRNA Directional Library Preparation methods following the manufacturer’s protocols. Twenty libraries were then run as quality control on a TapeStation to verify their quality. Quantitative polymerase chain reaction (PCR) was then performed on the library pool before sequencing of the library on NextSeq High 2 × 38. The library was sequenced repeatedly using paired-end sequencing on three lanes to obtain a minimum of 10 million reads for each sample, with some samples requiring only one or two lanes. On average, samples were sequenced at 16.5 million reads per sample. See Table S2 for sequencing sample, read count, and other information.

RNA-seq data outputs were screened with FastQC version 11.5 to confirm data quality in terms of per-base sequence quality, GC content, etc. Reads were trimmed with TrimGalore! Version 0.6.5 (https://github.com/FelixKrueger/TrimGalore) with the following parameters: --length 50 -q 20 –stringency 1 -e 0.1. Salmon version 1.4⁷⁶ was used to map reads to the human transcriptome, hg19, using the following parameters: quant --numBootstraps 100 --gcBias --validateMappings. All other parameters were left to defaults. Salmon quantification files were imported into R (version 3.6.1)⁷⁷ using the tximport library (version 1.14.0)⁷⁸ with the “type” option set to “‘salmon.” Transcript counts were summarized at the gene level using the corresponding transcriptome GTF file mappings obtained from ENSEMBL. Count data were subsequently loaded into DESeq2 (version 1.26.0)⁷⁹ using the “DESeqDataSetFromTximport” function, while accounting in the initial DESEQ2 model for both tissue and individual effects. In order to determine time/treatment dependent expression of genes, count data were normalized using the “estimateSizeFactors” function using the default “median ratio method” and output using the “counts” function with the “normalized” option. For subsequent differential-expression analysis, a low-count filter was applied prior to normalization, wherein a gene must have count data in at least 10% of samples, and further have a count greater than 5 in greater than three samples in order to be retained. Principle-components analysis (PCA) of samples across genes was performed using the “vst” function in DESeq2 (default settings) and was subsequently plotted with the ggplot2⁸⁰ package in R. Distance heatmap of all samples was generated using the ‘dist’ function in base R using the RNA read counts matrix generated via DESeq2, and subsequently visualized using the pheatmap package version 1.0.12.

We found that, when comparing expression patterns across developing long-bone-ends, there was less clear clustering of expression across tissues as compared to our ATAC-seq results (Figure S3A). This may be driven by a generally elevated homogeneity in expression across the skeleton, as evidenced by the substantial degree of similarity between any two samples (Figure S3B). However, performing differential-expression analysis followed by hierarchical clustering (see below), indicated that, for those genes behaving most differently across tissues, individual biological samples cluster by tissue-type, and further by limb-type (fore/hind limb) (as shown in Figure 2A).

Differential expression analyses were performed using the “DESeq” function of DESeq2 using all samples, with results subsequently summarized using the “results” function with the “alpha” parameter set to 0.05; p-values were adjusted using the Benjamini-Hochberg FDR method⁸¹, with differentially expressed genes filtered for those falling below an adjusted p-value of 0.05. Importantly, for all analyses we included an ‘individual’ term in the DESeq2 experimental design to control for the effects of variations across individuals in expression. We defined differentially-expressed genes at decreasing levels of granularity. To identify genes differentially-expressed in a given tissue, we compared the expression of a single tissue to the average of all other tissues using the ‘contrast’ option in the results function of DESeq2. To generate an expression heatmap (as shown in Figure 2A), the top 200 differentially-expressed genes in each tissue were aggregated. For each gene, expression values were z-transformed and plotted with ComplexHeatmap⁸² version 2.18, with row/column hierarchical clustering performed with default settings.

To identify genes changing expression between early and late time-points, we directly compared early/late samples for a given tissue (e.g. distal femur). To compare differential-expression across joints, we first grouped tissue samples across joints (e.g. proximal tibia and distal femur in the knee joint), then compared the expression of a particular joint relative to the average of the three other joints. We similarly considered the differential expression of genes across time within a given joint. Finally, we grouped tissue samples from forelimb and hind limb tissues to compare differential-expression across limb types. Gene-set enrichment analyses of differentially-expressed gene sets were performed with clusterProfiler⁸³ version 3.13.1 using default options.

In addition, for the eQTL analysis described below, we also mapped RNA-seq reads to the human genome using STAR⁸⁴ version 2.7.0 and default settings. We performed a first-pass mapping, subsequently using mapped junction information to perform a second-pass alignment via STAR. Given these alignments, we also performed a second differential-expression analysis using DESeq2, observing similar results to those observed in our initial Salmon-based analysis.

Human ATAC-seq data collection and analysis:

Fresh long-bone samples were microdissected under a microscope in 1× PBS on ice, and bone epiphyses were separated, stripped clear of soft tissue, and collected in microcentrifuge Eppendorf tubes containing 200 μl of 5% fetal bovine serum/Dulbecco’s modified Eagle’s medium (FBS/DMEM). The right and left bone-ends were pooled from each donor sample. For spinal-column samples (thoracic and lumbar vertebrae), individual vertebra were dissected, stripped clear of soft-tissue, and similarly collected in FBS-containing microcentrifuge tubes. Each sample was then subjected to 1% collagenase II (VWR, 80056–222, Radnor, Pennsylvania) digestion for 2 hours at 37°C rocking, mixing every 30 min. After placing on ice, samples were next filtered using a microcentrifuge filter setup by gently mashing the residual tissues through the filter followed by rinsing with 5% FBS/DMEM. Samples were then spun down at 500g at 4°C for 5 min. All cell counting methods were performed using trypan blue and a hemocytometer under an inverted microscope, and subsequent ATAC-seq steps were performed on those samples that had cell death rates well below 10%. Replicates consistently yielded several hundred thousand cells. Next, cells were resuspended in concentrations of 50,000 cells in 1× PBS. Cell samples were then subjected to the ATAC-seq protocol as described previously (113, 114), modifying the protocol by using 2 μl of transposase per reaction (see Richard et al.⁸⁵ and Guo et al.⁸). The transposase reaction product was then purified using the Omega MicroElute DNA Clean Up Kit following the manufacturer’s protocols, eluted in 10 μl of warmed double-distilled water (ddH2O), and stored at −20°C.

All samples were next subjected to PCR amplification and barcoding following Buenrostro et al. (2013, 2015)^86,87. Ten microliters of transposed DNA was placed in a reaction containing the NEBNext High-Fidelity PCR Master Mix, ddH2O, and barcoding primers. Following amplification, samples were transferred to new tubes and treated for fragment size selection using the OMEGA Bead Purification Protocol following the manufacturer’s instructions. The samples were eluted in 15 μl of tris EDTA (TE), nano-dropped, diluted to 5 ng/μl, and run on a Bioanalyzer for quality control and visualization of fragment size distribution. Before sequencing, sample concentrations were determined using the KAPA Library Quantification Complete Kit (KK4824). Samples were then sent out to the Harvard University Bauer Core Facility for paired-end sequencing using the Illumina NextSeq 500 platform. Each biological replicate was sequenced on average to 59 million reads per sample. Quality-control statistics are presented in Supplementary Table S1.

Computational analysis of ATAC-seq data began with checking read quality via FastQC. Adapters were then removed using NGmerge version 0.2. Reads were subsequently aligned to human reference hg19 genome assembly with Bowtie2⁸⁸ v2.3.2 using default parameters for paired-end alignment. Duplicate reads were removed using Picard’s MarkDuplicates function (version 2.9.0) (http://broadinstitute.github.io/picard), and mitochondrial reads were also filtered out. The resulting .bam files were then used for peak calling via MACS2 software (version 2.1.1.2)⁸⁹, using BAMPE and the following flags: --nolambda −bdg --verbose. Peaks reproducible across biological replicates were screened using a conservative irreproducible discovery rate (IDR) threshold of <0.01, as defined by the IDR statistical test (version 2.0.3)⁹⁰. The IDR method identifies overlaps in peak calls across pairs of sample replicates by comparing ranked peak lists (using MACS2 q value) to define a reproducibility score curve. These paired peaks are then assigned a pointwise score based on this curve. They are then sorted, and all peaks that fall below an IDR threshold (here defined as 0.01) are taken as our final reproducible peak set across replicates. This yielded peak calls, including those called using normal stringency IDR thresholds (0.05) within ENCODE guidelines of greater than 50,000 peaks²⁷. Following the identification of IDR-filtered peak sets for each bone-end, peaks were further filtered for epigenetic activity in the fetal brain. ATAC-seq was also performed on stage-matched E54 brain samples as part of a previous study¹². Called peaks were filtered so as to remove putative chondrocyte regulatory sequences overlapping with regulatory activity in the brain, and thus likely involved in regulating more general cellular processes. Intersections between bone and brain peaks were done using bedtools⁹¹ version 2.27.1 and was also used to define the sets of mutually-exclusive peaks described in the Results text. Ontology enrichments for non-coding sequences was performed using GREAT¹³ version 4.04.

Note about GREAT: GREAT takes an input set of genomic regions along with a defined ontology of gene annotations; first, it defines regulatory domains for all genes genome-wide and then measures the fraction of the genome covered by the regulatory domains of genes associated with a particular annotation (e.g., “cartilage development”). These fractions are used as the expectation in a binomial test, counting the number of input genomic regions falling within a given set of regulatory domains, which results in the reported significance of association between an input region set and a particular GO term. GREAT also performs a more traditional gene-based hypergeometric test to test for significance of region set–ontology association. The program returns a set of enriched ontologies sorted by the joint rankings of false discovery rate–corrected binomial and hypergeometric tests, as reported here in Table S1. For this study, we chose terms in the GREAT output with relevance to chondrocyte biology—bone-specific anatomy is not well annotated, particularly at the genetic level, and it is well known that there is a bias toward GO annotations for other phenotypes (such as immune function), particularly for phenotypes involving better-studied genes that tend to have increased GO term representation⁹².

Comparisons of epigenetic sets:

Overlaps in peaks between different tissues were performed using the UpsetR⁹³ package version 1.4 as well as bedtools version 2.30.0. All-by-all similarities between peak sets were calculated using the Jaccard index via bedtools and plotted with ComplexHeatmap version 2.10.0. For comparisons of ontology enrichments across early/late samples, peak sets were intersected per-tissue (e.g. early-vs-late distal femur), then pooled and used with rGREAT version 1.26.0.

Gene-set enrichment analyses:

Gene-set enrichment analyses were performed using GO Biological Processes via the ‘enrichGO’ function from the clusterProfiler⁸³ package version 3.13.1. The background gene set used was the set of all transcribed genes captured in our RNA-seq datasets. Semantically similar enriched GO terms were subsequently collapsed using the ‘simplify’ function from clusterProfiler, using default settings.

Putative regulatory element motif-enrichment analyses:

For all motif analyses, we fixed all peaks to a static width of 1250bp (the median peak size across all IDR-filtered sets) and retrieved human reference sequences for hg19 using the UCSC ‘twoBitToFa’ utility⁹⁴.

For un-biased motif-enrichment analysis we obtained the JASPAR database (2022 version)⁹⁵ and combined this with the PWM database provided with the HOMER⁹⁶ software. In instances where a factor had more than one defined PWM (e.g. due to overlap between JASPAR and HOMER databases), the information content (IC) of each matrix was calculated, with the matrix having the highest IC retained using TFBSTools⁹⁷ version 1.40. This combination yielded a final set of 787 unique PWMs. We also took advantage of our transcriptomic datasets to define sets of transcription factors which were differentially-expressed in at least one tissue, and for which PWMs were available (n = 209).

To define background nucleotide frequencies of our region sets, we took the set of DNA sequences for a given bone-end and generated a Markov background model using the ‘fasta-get-markov’ function from the MEME suite⁹⁸ version 5.4.1. Next, we took our sets of sequences and tested for enrichments of TF motifs using the ‘meme-chip’ function from the MEME suite with default options. MEME-ChIP results for each sequence set were aggregated; those motifs reaching significance (E-value < 0.05) in at least one set were retained to produce a final, filtered motif database comprising 267 unique motifs.

To test tissue-specificitty of TF enrichments, we took our filtered motif database and tested for enrichments in each individual sequence set using the ‘ame’ function from the MEME suite. As before, a Markov background model for each sequence set was used, along with the ‘--control --shuffle--’ options. To force AME to return enrichment results for all TFs (rather than just those passing a significance threshold), we used ‘--evalue-report-threshold 1000000’. For reporting enrichment E-values in Supplemental Table 1, we took the −log10(E-value) for each score. In instances of very small E-values (E << 1e-100), we fixed the −log10(E-value) at 1000.

The AME results were next aggregated across tissues, defining enriched motifs as AME E-value < 0.001. For the majority of PWMs we tested (182 of 267), significant enrichments were shared across all tissues, while 12 PWMs were enriched in seven or fewer tissues (73 motifs were not significant in any tissue after running AME and using E-cutoff). Using PWMs defined using our transcriptomic datasets, 17 (of 209) PWMs showed some degree of tissue specificity. To visualize these differential-enrichments, we took the E-values of each motif for a given tissue and plotted them on a log10 scale using ggplot2⁸⁰.

The above analyses were re-iterated using ‘specific’ region sets (exclusive from one another) (treated upstream in the same way above), which contain smaller numbers of DNA sequences. After aggregating significant MEME-ChIP results (n = 121 unique PWMs) and running AME, no motifs were enriched across all eight tissues, with the remaining factors demonstrating some degree of tissue-specificity (41 motifs were not significant in any tissue after running AME and using E-cutoff). Using transcriptome-defined PWMs, 76 (of 209) PWMs showed some degree of tissue specificity.

Some TFs showed strong bias in motif occurrence specific to one or a handful of tissues - for example, motifs for the homeobox factor BARX2 were enriched specifically in the proximal and distal humerus sets, which might reflect an increased activity of this TF in the more developed forelimb at this timepoint⁹⁹. PITX1, a known regulator of hindlimb development^16,100,101, was enriched only when considering specific distal-femur regions (Table S1) – possibly suggesting tissue-specific PITX1 regulation.

TF motif enrichments were also tested at the level of skeletal joints (Figure 3). For each joint, the relevant sets of ATAC-seq peaks (e.g. distal femur, proximal tibia of the knee joint) were pooled. The superset of all genes differentially-expressed in any individual joint (see Table S2) was intersected with the aggregated PWM database. This reduced set of PWMs was then used to test for significant enrichments of TF motifs across our joint sequence sets using AME as described above. Transcription factors were identified which were both up-regulated in the given joint, and whose motifs were enriched for joint-specific enhancers. For example, TWIST1 motifs were strongly enriched in knee and hip sites, but significantly up-regulated only at the hip; TWIST1 is a known regulator of pelvic development¹⁰², acting downstream of the key axial and hip joint regulators PBX1/2^100,103. Similarly, BARX2 motifs were strongly enriched specifically in elbow sites; this transcription factor is associated with chondrogenesis during limb bud development and is controlled by joint regulators such as GDF5^99,104.

Single-cell analysis and deconvolution:

To confirm the homogeneity of cell-types within our developing cartilage datasets, we utilized data from a previous scRNA-seq study of the developing limb bud¹⁸ to define signatures of cell-type-specific expression during early human limb formation from GEO Datasets (accession GSE143753). Pre-processed data were obtained as ‘loom’ files and converted to Seurat objects using the loomR library. This dataset comprised three different sets of experiments from developing long-bone tissues. We integrated these samples using the SCT-based pipeline developed for Seurat¹⁰⁵. Individual datasets were first transformed using the SCTransform function, then the ‘SelectIntegrationFeatures’ function was used with the ‘nfeatures=3000’ flag. Subsequent steps of the pipeline were performed using default settings. Clusters were defined using a resolution of 0.5, and significant marker genes were subsequently defined using the ‘FindAllMarkers’ function with the following parameters: only.pos = TRUE, min.pct = 0.25, logfc.thresold = 0.25. This resulted in a total of twenty different clusters. Marker gene sets were used to manually annotated using known marker genes for generic chondrocytes, articular chondrocytes, hypertrophic chondrocytes, tendon cells, etc. We noted that of these clusters, twelve were relatively rare in the limb single-cell dataset (occurring at a frequency of less than 5%). For our visualizations of estimated cell-type proportions (below, Figure S2) we constrained our graphical representation to those clusters representing at least 5% of the processed single-cell dataset.

Once this reference cell-type dataset was established, we next used the MuSiC2 package⁷⁵ to deconvolute our bulk RNA-seq datasets based on these cell-type signatures (Figure S2, Table S2). We generated an FPKM matrix from our aggregated transcriptomic datasets using Salmon, and subsequently used this, along with our single-cell reference, for the ‘music_prop’ function of MuSiC2, using default settings. The resulting estimated fractions of cell-type proportions, aggregated by tissue-type, were subsequently visualized using ggplot2. The majority of deconvoluted signals in our samples could be attributed to four main cell clusters expressing chondrocyte-lineage markers, with little signal emanating from non-skeletal cell types. Furthermore, there was little deviation (< 10%) of estimated cell-type proportions across different anatomical sites (Figure S2, Table S2), suggesting that the transcriptomic differences we observe across the skeleton are driven by changes in cell-intrinsic behaviors, rather than substantial differences in cell-type proportions between sites.

Height variant enrichments:

To test for enriched intersections between our ATAC-seq peak sets and significant height-associated variants from the GIANT dataset, we took each individual peak set (e.g. distal-femur-specific peaks) and generated n = 1000 randomized background sets using bedtools ‘shuffle’, matching both set and peak size. These background sets were subsequently intersected with sets of height variants using bedtools ‘intersect’ and counted. We calculated significance by standardizing the distribution of background intersect values and calculating p-values using a CDF of the normal distribution via the ‘pnorm’ function in base R. Multiple-hypothesis testing was corrected for using Benjamini-Hochberg FDR for the number of sequence sets tested for enrichment. As an additional significance test, we also calculated a permutation-based p-value as the number of randomized background sets with an equal or greater number of intersections than the target set. Both sets of results are reported in Table S3.

MAGMA analysis:

We performed MAGMA analysis³⁷ according to the detailed workflow previously described¹⁰⁶. We initially mapped variants to genes using MAGMA version 1.0.9a with the EUR subset of 1000Genomes, DBSNP151 information, and hg37 gene locations available from the MAGMA website (https://ctg.cncr.nl/software/magma). For MAGMA we prepared an FPKM matrix based on our Salmon RNA-seq analyses and combined this matrix with the per-gene metrics of tissue-specific accessibility described in the STAR Methods, along with the average cross-tissue expression of genes (AVG_EXP), average cross-tissue accessibility of gene loci (AVG_ACC), and the total number of peaks present around a given gene locus (NUMPEAKS). We then performed the initial MAGMA analysis using default GO BP, MF and CC ontologies, and applied a post-hoc Bonferroni correction for each ontology. This yielded 24 GO BP terms, 2 GO MF terms, 1 GO CC term, 11 expression-based metrics, and 1 accessibility metric (NUMPEAKS). Significant co-variants, along with GO terms, were subsequently used in a MAGMA model conditioned on both AVG_EXP and NUMPEAKS. This conditioned model yielded insignificant terms for all expression-based metrics, indicating that the original expression signal was likely driven largely by average skeletal expression. We then took the set of significant GO enrichments from this conditioned model and generated pairwise conditional models for each pair of enrichments – testing for independence of heritability signal. GO terms for which the pairwise conditional model was not significant (p > 0.05) were collapsed. Finally, we tested for interaction effects between these significant GO terms and our confounding variables (AVG_EXP and NUMPEAKS), identifying three terms for which significant associations were found.

Partitioned heritability analysis with LDSC:

Summary statistics for the entire GIANT dataset were analyzed using the LDSC software²³. Summary data was converted to an appropriate format using the ‘munge_numstats.py’ script. The ‘baseline-LD’ model available from the Alkes group¹⁰⁷ (https://data.broadinstitute.org/alkesgroup/LDSCORE) was downloaded. Initial heritability analysis (using the ‘ldsc.py’ script, ‘–h2′ mode) was performed using 1000 Genomes reference allele frequencies provided by the Alkes group, along with weights extracted from the baseline-LD model using a custom script. For the partitioned heritability analysis, custom scripts were used to generate a modified version of the baseline-LD model with a reduced set of features, including our epigenetic datasets (see Table S3 for list of features); these models were subsequently used to re-calculate LD scores with the ‘ldsc.py’ script with the following options: plink files for the EUR subset of 1000G (obtained from the Alkes group, see above link),–ld-wind-cm 1, constraining to the same set of SNPs used in the original baseline LD model. Partitioned heritability of features was done using the ‘ldsc.py–h2′ mode, using re-calculated LD scores, extracted weights, reference 1KG3 frequencies, the ‘–overlap-annot’ flag, and all other settings left to defaults. P values for significant proportion of heritability captured by particular features were corrected using BH correction. Results are shown in Table S3 with statistical significance defined as adjusted p-value < 0.05.

Tissue-specific gene expression heritability analysis with LDSC:

To perform a tissue-specific expression heritability analysis, we followed the suggested analysis described in Finucane 2018¹⁰⁸. We first used our Salmon-based expression datasets, calculated differential-expression across tissues (as described above) and subsequently extracted the top 10% most differentially-expressed (based on test-statistic) genes in each tissue. As a control set, we also considered all coding genes in our Salmon dataset. We then used the ‘make_annot.py’ function from LDSC to generate different models for each set of genes, using the EUR subset of the 1000G dataset and a --windowsize of 100000. Taking these sets of models, we performed LD-score regression for each tissue-specific model using ‘--ld-wind-cm 1 and --thin-annot’. We then aggregated results and adjusted p-values via BH correction (Table S3).

Gene-wise association ranking method:

We developed a simple metric with which to compare the enrichment of a set of regions for associations with height-related variants. First, we took all protein-coding genes in the genome and, for each gene, generated a 200kb window centered on the TSS. These regions were subsequently intersected with the GIANT genome-wide association results using bedtools ‘map’ – we considered all variants falling within all gene windows and assigned the strongest-associated variant to each window. All windows were accordingly ranked. Only those genes for which at least one variant was detected in the window were retained for ranking. To test for significant-enrichment of ranks for expressed genes (relative to those not detected in our skeletal RNA-seq datasets), we took the sets of ranks from both target and background sets and compared them using the Student’s t-test in base R. As a more stringent test, we also used a sub-sampling approach, wherein we randomly sample 10,000 sets of genes from the global genes and calculated the average set-wise rank of each set. This background was standardized and subsequently compared to the average rank of target genes using a CDF of the normal distribution in base R. Both sets of results are presented in Table S3. We used a similar t-test and subsampling-based method to compare the ranks of genes expressed in fetal brain RNA-seq data, obtained from ENCODE²⁷ (accession ENCFF654GII), to those not-detected in brain – using a reported TPM cutoff of > 1 to binarize expression. We also considered those brain-expressed genes which shared expression with fetal skeletal tissue, as well as those expressed exclusively in brain (Table S3).

We also compared the rank-enrichments of genes differentially-expressed in one tissue relative to those either (a) not differentially-expressed in any tissue, (b) differentially-expressed in hindlimb (relative to forelimb), or (c) differentially-expressed in forelimb (relative to hindlimb). For this comparative analysis, we aggregated set of rankings and compared them using the ‘glm’ function in base R, while correcting for the effects of average cross-tissue gene expression, using the following formula: rank ~ prefix + mean + prefix*mean. Given that these are rankings, negative t-statistics indicate that a given class (e.g. hindlimb DEGs) results in lower rankings (i.e. stronger associations with height).

In addition to the testing for module-wise enrichment for height associations using MAGMA³⁷, we also performed enrichment tests using this gene-wise ranking method, for the permutation testing again using 10,000 sets of randomized genes (set-size matched) (Table S5). (

Axial ATAC-seq datasets:

Raw ATAC-seq datasets were processed similarly to that described above for the appendicular data. IDR-filtered peaks for each tissue (vertebral and thoracic vertebrae) were subsequently intersected using bedtools to define exclusive/shared sites. We subsequently intersected these region sets with the pooled set of appendicular peaks (obtained at the E67 timepoint) to define sets of appendicular-filtered axial peaks, as well as axial-filtered appendicular peaks. Re-analyzed UKB height metrics were obtained from the recent GIANT publication authors², and used with the enrichment-testing method described above in section ‘Height variant enrichments’. Enrichment p-values were adjusted for each height metric tested using BH correction (see Table S3).

Tissue-specific Loci Metrics:

To prioritize important loci based on our different datasets, we generated multiple different metrics, defined at the gene level, as follows. These methods are largely similar to those implemented in a previous publication¹²:

(1). Tissue-specific accessibility.

Similar to our RNA-seq analyses, we implemented a differential accessibility methodology to define regions as being more/less accessible in a given pelvic tissue. All IDR-filtered, brain-filtered peaks from each bone-end were pooled, with overlapping peaks merged using bedtools “merge.” We next defined 2.5-kb regions centered on each pooled peak (1.25 kb upstream/downstream); along these regions, a 1.25-kb sliding window (chosen based on the average called peak size across individual pelvic elements) was slid in 50-bp increments to generate a set of overlapping 1.25-kb regions. ATAC-seq read coverage within these regions was calculated using the “bedcov” function of samtools version 1.5 for each mapped .bam file corresponding to individual ATAC-seq samples. This sliding-window approach is used to identify the area of strongest cross-sample signal (i.e., ATAC-seq read coverage) around a pooled ATAC-seq peak, improving our confidence in defining differential accessibility (i.e., avoiding regions around an ATAC-seq peak with lower read coverage, which may make our analysis more sensitive to noise). For sets of overlapping windows corresponding to a single ATAC-seq peak, we applied a smoothening algorithm to eliminate extreme values occurring in overlapping windows for individual samples. Windows whose read coverage fell outside 1 SD of the set (of sliding windows for a single peak) were assigned the average read coverage of the two windows adjacent. This adjustment does not affect the later differential accessibility analysis but will affect the choice of sliding window assigned to represent this given peak and is done to avoid consistently using the most extreme windows to define differential accessibility (which may otherwise bias our results). Following this smoothening, for each window, we calculated the 75th percentile read coverage value across all skeletal subelements and samples (this was found to be more robust than mean read coverages, even after smoothening adjustment). The window with the greatest 75th percentile coverage was then selected as the representative region for that pooled ATAC-seq peak. Subsequently, raw read coverages for all optimized windows across all bone samples were imported as a matrix into DESeq2 version 1.26.0 using the “DESeqDataSetFromMatrix” function, with differential accessibility calculated using the “DESeq” function with tissue type as the main variable. For each individual tissue, we then calculated the log fold change (logFC) increase/decrease in accessibility for windows relative to the average of all other tissues (e.g., distal-femur-specific logFC relative to all other bones) using contrast vectors and the “results” DESeq2 function. We also calculated per-tissue read counts (FPKM) as an absolute (instead of relative) metric of accessibility strength, finding that these absolute values were more informative than relative logFC values.

Given these continuous ATAC-seq data, which we calculate across all peaks genome-wide, we next implemented a ‘tissue-specific regulatory score’ for all genes in our RNA-seq dataset, inspired by previous integrative approaches¹⁰⁹. For each gene, we captured all chromatin accessibility regions (i.e., the optimized windows above) within 100 kb upstream/downstream of the TSS. For each accessible region, we took the calculated tissue-specific read-counts (FPKM) and scaled it by distance to TSS using the following formula:

$$\text{Regulatory}\text{score}=\left(\text{tissue-specific}\text{FPKM}\right)*\left({\text{e}}^{-\left(0.5+4*\left(\text{distance}\text{to}\text{TSS}/100\text{kb}\right)\right)}\right)$$

Where the second distance-scaling term is adapted from Tang et al. 2011¹¹⁰. These per-region regulatory scores were summed across all regions to give a single tissue-specific regulatory score. Thus, a single gene will have multiple scores for each tissue.

(2). Height associations of nearby variants.

For each 200kb window centered on each protein-coding gene, we aggregated all variants for which height-associations were available. For each variant, we took the −log10(p-value) of association and scaled this by distance using the above distance-scaling term. For each gene, all these scaled values were summed.

(3). Tissue-specific height associations of nearby variants.

For each 200kb window centered on each protein-coding gene, we aggregated all variants for which height-associations were available, but conditioning on variants falling within ATAC-seq peaks of a given tissue. For each variant, we took the −log10(p-value) of association and scaled this by distance using the above distance-scaling term. Subsequently, we took these distance-scaled values and multiplied them by the tissue-specific accessibility (FPKM) of their overlying peaks, such that variants which are more accessible in a given tissue have stronger scoring. For each gene, all these scaled values were summed, such that for each gene there are multiple tissue-specific association scores.

(4). Tissue-specific expression.

We took the results of the differential-expression analysis we performed for individual tissues, assigning to each gene the calculate logFC expression change in each tissue (relative to all others).

To increase the stringency of this method, we also integrated a dataset of promoter-capture Hi-C interactions generated on adult primary chondrocyte samples¹¹¹. For metrics 1–3 above, we further required that all variants and peaks fall within a region contacted a given gene promoter. To incorporate the strength of the interaction, we further scaled the ‘APScoreAvg’ values from the Hi-C calling pipeline for all fragments contacting gene promoters using the ‘scales’ package version 1.2.1 to the range of [0,1]. For each variant/peak, we substituted the distance-dependent scaling factor for this scaled promoter-interaction value.

The global set of all coding genes (in our RNA-seq dataset) were ranked according to each of these scores, such that we generated a large matrix of gene rankings across multiple metrics. To look for consistent trends in ranking, we aggregated ranks using the RobustRankAggreg library version 1.1 in R, which seeks to identify genes that are consistently ranked higher across different ranking sets, assigning a significance value per gene for their rankings relative to a null hypothesis of uncorrelated inputs. Using the full model (all metrics across all tissues), we took the top 200 top-ranked genes (based on rank aggregation) and performed gene-set enrichment analysis using ClusterProfiler version 4.6.0. This set of genes was significantly enriched for GO terms relevant to skeletogenesis, including ‘skeletal system development’, ‘extracellular matrix organization’, ‘chondrocyte differentiation’, among others. The results from our models either independent of, or conditioned on, interaction data are shown in Supplemental Figure 3 and Supplemental Table 3.

To confirm that leveraging these multiple metrics yields more meaningful insights than single-metrics (e.g. just GWAS statistics), we performed rank aggregation on different subsets of the metric sets. These include the full model conditioned on contacts (“Contacts – Full”), GWAS only (condition on contacts – ‘Contact-GWAS’), full model not conditioned on contacts (‘FULL – no Contacts’), GWAS only (‘GWAS Only’), the full model conditioned on contacts – but lacking GWAS (‘Contacts – no GWAS’), the full model not conditioned on contacts and lacking GWAS (‘No Contacts – no GWAS’). We subsequently performed GO enrichment analyses on the top 200 ranked genes from each of these analyses. The GO enrichments were subsequently aggregated, and shared terms across analyses were compared in base R (see Supplemental Table 3).

We next sough to prioritize candidate loci based on tissue-specificity. For each tissue (e.g., distal femur), we reverse the gene-wise ranks of all tissue-specific metrics, looking for genes which are consistently ranked higher in that target tissue, and consistently ranked lower across all non-target tissues, and again applied rank aggregation. For each tissue, we further required that the top-ranked genes were also significantly-upregulated in expression for said tissue, identifying several genes per-tissue (Table S3). For example, LHX9 was identified in the proximal-tibia set – this factor plays a known role in limb outgrowth¹¹² and is expressed in fetal epiphyseal cartilage¹¹³. Additionally, CREB5 was identified in the distal-femur set – this factor promotes formation of synovial joints¹¹⁴ and is specifically expressed in the developing mouse knee¹¹⁵.

We also combined this integrated approach with the above MAGMA³⁷ analysis, with these integrated metrics as additional model parameters. None of the continuous metrics for (tissue-specific) local accessibility were significantly associated with height heritability. However, the absolute number of putative regulatory regions in a given locus (across all tissues) was significantly linked with height association at that locus (p < 0.05, Table S3). Conversely, all metrics for tissue-specific expression were significantly associated with height (p < 0.05), though these were largely explained by average skeletal gene expression (Table S3).

We next performed a gene-set association analysis using GO ontology terms (see Table S3 for full lists) and built models that condition on the number of regulatory regions in a gene locus, as well as average skeletal expression of each gene, as covariates (see above). This approach yielded a set of 12 aggregated GO terms (p < 0.05) (Table S3). Significant (p < 0.05) interaction effects were observed between three different GO terms and average skeletal gene expression: embryo development (GO:0009792), skeletal system development (GO:0001501), and chromatin components (GO:0000785).

SNP-wise set comparisons via ANOVA:

To test whether SNPs, genome-wide, have different per-variant height associations when conditioned on their accessibility patterns (i.e. whether their activity is limited to one part of the skeleton, or more pleiotropic), we first considered all variants in the GIANT dataset. Initially, we binned variants based on whether their accessibility was (1) exclusive to fetal brain ATAC, (2) broadly active across the entire skeleton as well as brain, (3) active in part, but not all, of the skeleton, or (4) active across the skeleton but not brain. We subsequently compared the log10 p-values of variants across these groups using the ‘aov’ function in base R, followed by Tukey post-hoc corrections with the ‘TukeyHSD’ function from the stats package. We visualized these results using the ‘plotmeans’ function from gplots version 3.1.3 (Supplemental Figure S3I). For each variant, we then counted the number of distinct tissues for which accessibility was detected (i.e. overlapped a peak). Converting these counts to an ordinal variable, we generated a linear model between association and counts with the ‘lm’ function, followed by ANOVA analysis with the ‘anova’ function.

Extending GWAS results using functional annotations:

To incorporate our functional annotations to extend our resolution to detect possibly causal variants at a given fine-mapped locus, we used the PAINTOR v3.0¹¹⁶ software. For all COJO fine-mapped loci, we generated 250kb windows centered at the lead variant. We then intersected these windows with the GIANT GWAS dataset, using a relaxed p-value threshold of 5e-8. For each window, we then calculated LD between variants using the ‘CalcLD_1KG_VCF.py’ function from PAINTOR, based on the EUR subset of 1000G reference panel, subsequently running ‘AnnotateLocus.py’ for each window. Given the large number of loci involved, we shuffled all loci (to avoid potential chromosomal-biases) then split them into sets of 100. Each set was then used as input to PAINTOR with the following parameters: ‘-mcmc -enumerate 3’, all others left to default. To gage the additional information gained from our functional annotations, for each locus we ran PAINTOR both with and without the use of functional annotations. Conditioning on developing chondrocyte regulatory activity, in greater than 50% of fine-mapped loci additional/alternate significant variants (i.e., not the principal COJO-prioritized variant in the locus) were detected with posterior probabilities >= 99% (Figure S4, Table S4).

To compare the predicted functional potential of lead variants to those detected with PAINTOR, we incorporated information from LINSIGHT¹¹⁷ (which aggregates information on inter- and intra-species sequence conservation, functional annotations, predicted TFBS, etc.), RegulomeDB³⁰ probability scores, and RegulomeDB rank assignments. For all windows for which predicted regulatory information was available for multiple variants (including the lead), we compared these three metrics between leads and variants detected with PAINTOR, considering any variant which scored higher by any metric as ‘improved’ relative to the lead. Over a quarter of loci with significant additional variants were predicted to be of increased regulatory effect compared to the original fine-mapped variant (Figure S4B, Table S4).

To putatively tie these non-coding variants to their cognate genes, we incorporated chondrocyte interaction promoter-capture data¹¹¹, intersecting looping anchors with our variant sets (Figure S4C). We grouped genes into those contacted by both fine-mapped variants and PAINTOR-detected variants (based on our ATAC-seq annotations) and tested for gene-set enrichment using ClusterProfiler version 4.6.0, finding similar biological enrichments (Table S4). We also tested for enrichment of OMIM⁷² genes associated with skeletal defects using a chi-sq test, comparing the number of OMIM skeletal genes contacted by each set of variants relative to the proportion of all contacted genes which were OMIM skeletal genes, with the strongest enrichment observed for genes shared in contacted fine-mapped and PAINTOR variants. Figure S4F illustrates an example at the IGF2 locus, wherein both COJO fine-mapped and PAINTOR alternative variants both intersect enhancers described in chondrocyte Hi-C datasets, which also show evidence of activity in our ATAC-seq data. IGF2 is a growth signaling molecule that induces anabolic metabolic activity in human¹¹⁸ and mouse¹¹⁹ chondrocytes and plays a known role in skeletogenesis¹²⁰. Furthermore, variants in the IGF2 locus have been associated with adult height and birth length^121–123.

Variant-calling pipeline from expression datasets:

We aligned all our RNA-seq datasets to the human genome (hg38) via STAR version 2.7.0 using a two-pass mapping strategy. Subsequently, these alignments were imported into R and normalized using ‘estimateSizeFactors’ and default parameters. The normalized read-count matrix was converted to an expression co-variance matrix using the ‘prcomp’ function in base R, as suggested in the original GTEX Consortium publication¹²⁴.

For genotyping, duplicates in aligned BAM files were removed using picardtools ‘MarkDuplicates’. We subsequently made use of the GATK pipeline¹²⁵ version 4.3.0. BaseRecalibrator was run on individual duplicate-filtered bam files using known sites from the 1000Genomes¹²⁶ phase 3 dataset, followed by ‘ApplyBQSR’. Subsequently, ‘HaplotypeCaller’ was used with the following flags: --dont-use-soft-clipped-bases -stand-call-conf 20. Finally, ‘VariantFiltration’ was used to clean generated VCF results with the following flags: -window 35 -cluster 3 --filter-name “FS” --filter-expression “FS > 30.0” --filter-name “QD” --filter-expression “QD < 2.0”. Following this variant filtration, VCF files were combined using bcftools.

Next, we used the BEAGLE software¹²⁷ version 4.1 to impute variants based on the EUR subset of 1000Genomes. Firstly, we confirmed the genetic heterogeneity of our sample set by generating a genetic eigenvector matrix using plink¹²⁸ version v1.90b7, and visualizing the resulting points in PCA space using base R – observing no obvious stratification of samples (i.e. as would be expected with mixed ancestry, e.g. EUR/AFR). We used the conform-gt.jar utility from BEAGLE to first align our VCF files to the 1000Genomes datasets. We then imputed using BEAGLE and default settings. Subsequently, we made use of plink to generate another eigenvector matrix, which we reformatted as a genetic co-variance matrix for use with QTLTools using base R. Additionally, we used the ‘identify_outliers’ function from the rstatix package version 0.7.2 to identify extreme outliers in the genetic PCA distribution, removing the following samples from subsequent analysis: E03_ProxHumerus_59_S21, F01_Ilium_59_S102, F03_Acetabulum_59_S22. Similarly, we also removed samples based on their sparsity in the expression matrix (> 50% genes with missing data), removing the following samples from subsequent analysis: A03_Femur2_57A_S113, C01_Femur1_57A_S3, G12_Acetabulum_57C_S95.

We then used QTLtools¹²⁹ to calculate statistical significance of links between gene expression and imputed genetic variants. For this analysis, we conditioned on the following: (1) that variants fell within chromatin regions accessible in the developing skeleton – as detected using our ATAC-seq datasets. The epigenetic activity of variants increases the likelihood that they are functionally linked to expression of genes expressed during skeletal development (implicitly suggested by our findings that epigenetic activity of variants is linked to greater genetic associations with height). (2) That variants were detected as significant QTLs in the GTEX cross-tissue dataset (using a significance threshold of 1e-6) – the logic being that if a given variant acts to regulate expression across multiple tissues in GTEX, that it is (by definition) less likely to be tissue-specific, and thus possibly also acts to regulate expression in developing chondrocytes. As described in the original QTLtools paper, we first used an iteration algorithm to progressively add co-expression principal-components to our QTLtools cis models, calculating the number of significant hits at q-value < 0.05 (with a permutation parameter set to 100000). Plotting the subsequent curve, we identified the optimal number of PCs to be n = 100.

We then took these optimized QTLtools results and ran the included ‘qtltools_runFDR_cis.R’ script with an fdr parameter of 0.05. We then re-ran the QTLtools cis function using the ‘--mapping’ flag with this FDR output. This yielded a set of 9,582 significant SNP-eGene pairs (FDR < 0.05, Table S4), representing 694 unique variants across 498 genes. More than half of these SNP-eGene pairs recapitulated the known relationship found in the original GTEX study³⁴, and many of these validated variants also had predicted effects on additional target genes. We next integrated these variants with height GWAS information and identified 53 unique eQTLs which were also significant height variants, significantly linking to 197 different genes (Table S4). Of these, 18 had strong evidence for predicted regulatory effects via RegulomeDB variant ranks, including evidence for transcription factor (TF) motifs, binding activity, and DNase-accessibility (Table S4).

Locus Modularity Analysis:

We have previously found that musculoskeletal GWAS disease-associated loci exhibited complex regulatory architecture during early skeletal development¹³⁰. We thus asked whether the strength of height association for a given was related to the behavior of adjacent putative regulatory regions. For all genes detected in our RNA-seq datasets, we defined a 200kb window centered on the gene TSS and intersected different sets of accessible regions detected in each tissue. We defined a locus as being ‘non-modular’ if, at each detected putative regulatory element, the exact same set of tissues displayed an open chromatin site. This also includes loci for which only a single adjacent chromatin site was detected. Loci were considered ‘modular’ if different adjacent regulatory elements were represented in different sets of tissues. Genome-wide, 18898 gene loci were defined as ‘modular’, versus 5337 which were non-modular. Of these, 2621 had complete pleiotropy (i.e. all elements in the locus were active in all skeletal tissues), while 2716 exhibited consistent overlap (i.e. shared patterns) of tissue activity across all elements.

To test the height associations of different sets of modular/non-modular loci, we used the gene-ranking method described above. Modular genes tended to have significantly-greater associations with height relative to both the global background and randomly-sampled subsets (Table S3). Conversely, genes which were neither modular nor exhibiting complete pleiotropy tended to not be more associated with height, while non-modular genes exhibiting complete pleiotropy did have significantly better associations, but to a smaller extent than modular genes.

As an additional test to confirm this pattern (that modular gene loci tend to be more associated with height), we took the set of COJO fine-mapped loci and asked whether these loci tended to be modular. Indeed, we found that COJO loci tended to be modular (chi-sq statistic: 62.01, p = 3.42e-15). Finally, we compared the COJO loci which were defined as modular/non-modular using ANOVA, finding that modular loci tended to have more significant association p-values, larger absolute effect sizes, and explained a greater per-locus proportion of heritability (data not shown).

Defining Gene Expression Modules:

To define modules of genes based on co-expression patterns, we used the WGCNA package¹³¹ version 1.72, following the recommended pipeline from the authors. Briefly, we took the processed RNA-seq dataset from our initial processing in R and applied a variance-stabilization transformation (via ‘vst’), using this to select a soft threshold with the ‘pickSoftThreshold’ function WGCNA, for which we selected a power of 9. We then defined modules using the ‘blockwiseModules’ function in WGCNA with the following parameters: power = 9, maxBlockSize = 26000, networkType = “signed”, corType = “bicor”, mergeCutHeight = 0.1, minModuleSize = 40, pamStage = T, reassignThreshold = 1e-6, deepSplit = 2, minKMEtoStay = 0.3). Subsequently, cutreeHybrid was used to group genes into modules. Once the final set of gene modules were defined, we ran GO term enrichments using the ‘gost’ function from gprofiler2¹³² version 0.2.2 on each set, using the set of all coding genes used in the initial WGCNA analysis as the background set (see Table S5). As a confirmation of the co-expression patterns within modules, we confirmed that the pairwise expression correlation of genes within modules were significantly greater than randomly-scrambled gene modules (data not shown).

We then calculated the tissue expression of genes in each module across all tissues and used ANOVA to test whether individual modules had significant skews towards expression in certain tissues, finding no significant biases for any set (data not shown). We then calculated the tissue-average expression of genes in each module and tested whether different modules were expressed at significantly-higher levels than expected. For each module, we generated 10,000 randomized sets of genes (of equal size), calculating the average expression for each randomized set. This background distribution was then standardized and the average expression of a given set was tested using a CDF of the normal distribution with the ‘pnorm’ function in base R. As an additional test, we also calculated a p-value as the number of randomized gene-sets with equal or greater average expression to the target set. Both test results are reported in Table S5. Finally, we took those modules with significantly-increased expression and compared their expression distributions across modules using ANOVA and Tukey Post-hoc Corrections (Figure S11, Table S5).

To assess the enrichment for genetic height associations for individual modules, we adapted a script and methodology from Gandal 2018³⁶. Briefly, we generate customized annotations for each module set, which are then fed into the MAGMA software along with the genome-wide height genetic association data. We generated gene annotations for MAGMA using either gene bodies or 100kb windows centered on gene start sides to compare the effects of proximate and more distal genetic variation. The results of the MAGMA program are then processed in R, where the gene-wise association p-values are log-transformed, then correlated with the eigen vector values of gene assignments to a given module using the Spearman’s correlation test. Test p-values were adjusted using Benjamini-Hochberg correction (Table S5).

Module-Module Correlation Analyses:

Taking inspiration from Gandal 2018³⁶, which considered the expression correlation between their sets of defined modules, we calculated the pairwise Sperman correlations between all genes in a given pair of modules and took the median correlation value for each pairwise comparison. This all-by-all matrix was converted to a graph wherein each node was a module, and each edge weighted by the correlation coefficient. We generated multiple such graphs for the median correlation (positive and negatives), the median positive correlations, median negative correlations, and the percent of pairwise comparisons (between two modules) which had significant positive or negative correlations. To look for ‘centrality’ of modules – i.e. if there was an organized structure or hierarchy relationship between our module sets, we used the ‘calculate_centralities’ function from the CINNA¹³³ package version 1.2.2. However, we found that across multiple different types of centrality calculations that no clear, robust structure was apparent in any of our graphs – suggesting that there is no well-organized series of relationships between our different modules (data not shown). We subsequently looked at the behaviors of individual modules relative to others (to look for expression-program outliers, rather than program ‘master sets’) by visualizing our pairwise-correlation matrices (as the percent of pairwise comparisons with significant positive/negative correlations) both in PCA space, as well as using the ComplexHeatmap⁸² package (as shown in Figure S5D–E). No visually-obvious organization of relationships was apparent; however, when considering negative correlations between modules, one module (CD1) behaved differently – tending to have a greater degree of negative correlation (both average and summed across modules) with expression of genes across other module sets.

Genome-wide blocked heritability:

To assess the contribution of a given module to overall height heritability, we made use of previously-defined LD blocks from HESS³⁸. Using the hess software version 0.5.4, we partitioned the GIANT height associations in these LD blocks genomewide, requiring a minimum MAF of 0.01 and using the EUR subset of 1KG3. We subsequently used the ‘estimate_lambdagc.py’ script from hess to specify the ‘--reinflate-lambda-gc’ parameter for the second step in the HESS analysis. Once obtaining the final set of per-block assigned heritabilities, we plotted the cumulative distribution of heritability across the genome using base R, highlighting those blocks in the top 99% percentile of all heritabilities. To associated module genes with blocks, we looked for the presence of at least one SNP with an assigned p-value of < 1e-6 within 100kb of a gene. We then intersected these SNPs with our set of blocks, following the methodology described in Sinnott-Armstrong 2021¹³⁴. We subsequently intersected the set of CD1 module genes with these top 1% blocks, as shown in Figure S14. To calculate the fold-enrichment for heritability signals captured by the set of blocks hitting core genes, we calculated the fraction of all blocks captured by our module set and compared this to the fraction of global heritability represented in this set of blocks (Table S5).

In addition to the HESS-block based heritability partitioning, we also considered the presence of COJO-defined windows falling within individual blocks captured by our module set. For all blocks across the genome, we aggregated all COJO windows falling within, summing the per-window assigned h². We then calculated the fraction of blocks captured by our module set and compared this to the fraction of global COJO-based heritability captured by these blocks, calculating a fold-enrichment for heritability. Similarly, we also calculated the fold-enrichment of COJO windows captured by our module-block set (relative to the global distribution of COJO windows, see Table S5).

For the set of enriched GO terms associated with CD1, we extracted the genes associated with each individual GO term. Subsequently, we aggregated the per-gene FPKM values for each GO term across all skeletal samples and compared the per-term FPKM distributions using ANOVA in base R, followed by Tukey post-hoc correction. We also aggregated the per-gene associated blockwise-heritability estimates by GO-term, again comparing these distributions across terms using ANOVA.

Module-wise TF Analysis:

For a given gene-set module, we took all the genes and defined a 200kb window (centered on the gene TSS) and subsequently intersected these windows with the pooled set of E54 ATAC-seq peaks. Aggregating these sets of peaks across all genes in a module, we defined a modulewise set of putative regulatory sequences. We subsequently scan this set of putative regulatory sequences with ‘meme-chip’ program from MEME⁹⁸ using an aggregated TF database and defining a background model with the ‘fasta-get-markov’ program from MEME. We subsequently aggregated the MEME results generated from each module set, retaining those TFs with a reported E-value of < 0.05. This reduced motif database was subsequently used with the ‘ame’ program from MEME to test for enrichments of these TFs across all module sets. The subsequent AME results were aggregated in base R, with significant enrichment for a given TF in a given module set defined as a reported E-value < 1e-3. To visualize the patterns of motif enrichments across all module sets, we took the −log10(E-value) values and plotted them using the ‘ComplexHeatmap’ package (Figure S5).

To test for biases in association p-values for SNPs falling within different motifs, we took the set of motifs with significant AME enrichments and used the ‘fimo’ program from MEME to generate motif hits across all peaks in our pooled E54 ATAC-seq dataset, using a p-value cutoff for retained hits of p < 1e-4 (as recommended in the FIMO documentation). We subsequently intersected these sets of motif hits with the genome-wide height association statistics. To test for significant shifts in the distribution of per-SNP association p-values for SNPs falling within motifs of a given transcription factor, we randomly sampled all SNPs not falling within motifs for a given TF (but still falling within motifs), generating 1000 sets of randomized SNPs (matching the number of SNPs intersecting motifs for the target TF). We subsequently took the average −log10(p-value), beta and absolute-beta coefficients from each set of randomized SNPs to generate a background distribution. This distribution was subsequently standardized and compared to the average of the target distribution using the normal CDF using the ‘pnorm’ function in base R. Resulting p-values across all TFs tested were adjusted with Benjamini-Hochberg correction.

To test for biases in association p-values for SNPs falling within different motifs, conditioning on the presence of these motifs around module genes, we performed a similar analysis – but this time conditioning on SNPs falling within a 200kb window of any gene. The distribution of SNPs falling within a given motif around CD1 genes was compared to the distribution of SNPs falling within any motif (from the AME enrichment set). We first considered the set of factors which were enriched either (a) across all modules, (b) across some modules (i.e. showing some degree of specificity), or (c) exclusively around the CD1 module. As a more stringent analysis, we further conditioned on SNPs to only consider those falling around motifs adjacent to CD1 genes to generate a more stringent background distribution for testing. Subsequently, we tested individual factors, testing all motifs which were enriched in the CD1 AME results (n = 50). For the nine factors which were enriched relative to the general background, we tested their grouped enrichment for height-association using the ranking-based method described above.

Cis-regulatory factor analyses:

Reasoning that cis-QTLs that impact upstream factors (e.g., TFs) may act as trans-QTLs for downstream genes¹³⁵, we mined the GTEX cross-tissue dataset for sets of significant QTLs for these upstream factors. We leveraged existing databases of TF-target pairs - to consider predicted upstream regulators of defined gene modules, we first combined two databases of known/predicted TF-target relationships – PathwayCommons⁵⁶ (https://www.pathwaycommons.org/archives/PC2/v12/) and hTFtarget⁵⁷ (http://bioinfo.life.hust.edu.cn/hTFtarget#!/download) for the former extracting all relationships with an annotated INTERACTION_TYPE as ‘controls-expression-of’. We then combined these two datasets in base R. We then took the GTEX cross-tissue QTL dataset (using a p-value cutoff of 1e-6) and obtained all QTL variants which showed evidence of regulating the set of regulatory factors in our combined database, using our pooled epigenetic datasets to filter for those variants showing some evidence of epigenetic activity in skeletal development. For an additional higher-stringency analysis, we further subset this set of variants to those falling within COJO-defined height-associated windows. Finally, for all regulatory analyses (both for target and randomized modules) we only considered those transcription factors expressed in skeletal tissue (based on our RNA-seq datasets).

We then linked QTL-regulated transcription factors with putative downstream target genes and collected sets of TFs (and their associated QTLs) based on the presence of their targets within different modules. For a given gene module, we took the set of genes in said module and looked for all annotated upstream regulators in the combined dataset, then extracted the set of QTLs associated with these regulators. For the entire set of variants (not conditioned on COJO windows), we summed the per-SNP heritabilities (H²), effect size (beta), and p-values (log-transformed) by gene. In instances where QTLs are shared by multiple possible upstream regulators for a given gene, only the unique set of variants were summed. Conditioned on COJO windows, we took the set of all windows containing at least one associated QTL and summed the heritability and effect-size assigned to a given window as reported in the original data table – only summing the unique set of windows in the event of recurrent hits. To compare our module sets, which were derived in an unbiased, data-driven manner, to a manually-curated gene set, we also performed the above variant-aggregation for the set of genes associated with developmental skeletal abnormalities (from OMIM). To establish a background set for comparison, we pooled the set of module-assigned genes (excluding module CD0, which were unassigned genes from the WGCNA analysis). Then, for a given module, we generated 100,000 randomized modules of equal set-size for comparison. For each randomized module, we performed the same variant-assignment method as described above, establishing background distributions for per-gene summed heritabilities and effect-sizes. For each randomized module, we first compared the distribution of per-gene heritabilities and effect-sizes to the distributions observed for the actual gene module with a one-sided student’s T-test (with an alternative hypothesis of a greater distribution mean in the randomized set). Performing this over all 100K randomized sets, we calculated a false-discovery rate of a random set of genes yielding an improved distribution of per-gene assigned heritabilities/effect-sizes (as described above) relative to the target module set. We calculated a similar p-value for the probability of detecting a distribution average greater than the distribution average of a given gene module.

We next took the distribution of heritabilities/effect-sizes calculated for OMIM skeletal genes and compared this to the distributions of our different module sets, using the Student’s T-test. Given these repeat comparisons to the same distribution of OMIM genes, we adjusted t-test p-values using Benjamini-Hochberg correction. Finally, for each module, we took the set of 100,000 randomized gene sets and compared these to the distribution of OMIM skeletal genes, as described above.

Signed LD Profile Regression:

To test for significant skews in the effects of variants hitting motifs in our epigenetic datasets, we made use of the SLDP software¹³⁶. For this analysis, we used the HOCOMOCO¹³⁷ motif PWM database (v11). Given our epigenetic datasets, we adapted the SLDP software to allow us to only consider motifs and variants falling within our putative regulatory regions. We first ran the software with our set of TFs, combining the three TFs with greater associations across all module genes, and those nine enriched around CD1 module genes – though only five of these were represented in the HOCOMOCO dataset. Globally, we saw no significant biases for any of our tested TFs. Given the identification of TFs with significantly-greater associations across all module genes, and particularly CD1 module genes, we tested for similar biases in direction-of-effect of variants by conditioning variants on proximity to module genes (Table S5, Figure S5). Only one TF, FOXP1, had significant biases in direction were observed around module genes, including CD1. Variants intersecting putative FOXP1 sites tend to be more associated with height (Table S5) and, around CD1 gene loci, variants improving FOXP1 motifs also tended to be height-increasing. We obtained a set of FOXP1 target genes from the CHEA⁵⁵ database via: https://maayanlab.cloud/Harmonizome/gene_set/FOXP1/CHEA+Transcription+Factor+Targets. For each of the five module sets which yielded significant adjusted p-values from the SLDP FOXP1 analysis, we tested whether module genes were significantly over-represented in FOXP1 targets using the GeneOverlap library version 1.34, using the set of all transcribed genes as a global background, adjusting enrichment p-values with BH correction (Table S5).

Finally, we then took a broader gene set-based approach testing for signed-variant effects with a set of GO Biological Processes (C5), subsequently correcting p-values for the total number of gene sets tested. The top 10 enriched module results for each TF are shown in Table S5. For FOXP1, we plotted the per-module enrichments as calculated effect-size (calculated as log2(mean_in_wgt/mean_out_wgt) from the SLDP output). For the GO enrichments, we filtered for those with an adjusted p-value of < 0.05, then subsequently ran semantic clustering to collapse similar terms using the ‘calculateSimMatrix’ and ‘reduceSimMatrix’ functions from rrvgo¹³⁸ version 1.10 with a similarity threshold of 0.8. The resulting clustered terms were subsequently plotted with ggplot2, scaling the size of points by the calculated effect-size for each term, and plotted as a PCA of the GO term similarity matrix.

Methods for the Genetic Architecture of Type 2 Diabetes:

To demonstrate the efficacy of our framework for examining the omnigenic model as it applies to the genetics of other complex traits, we chose type II diabetes⁵⁸ with the pancreas as a key organ involved in its biology⁵⁹. Taking an approach similar to that we took in our skeletal biology and height analyses, we first sought to define putative gene regulatory modules active in the pancreas. We obtained a dataset of processed bulk RNA-seq read counts from adult pancreas samples via the V8 data release of the GTEx consortium³⁴. We included only those genes with more than 10 mapped reads across at least three samples and excluded any samples missing data in more than 10% of filtered genes (final n = 275). To further exclude lowly-expressed genes from our analysis, we further retained genes with a calculate CPM > 1 in at least 10% of samples (n = 16462).

We performed a similar WGCNA analysis to that described above. The filtered gene expression counts matrix was transformed using ‘vst’, and we then select a soft threshold with the ‘pickSoftThreshold’ function WGCNA, for which we selected a power of 28. We then defined modules using the ‘blockwiseModules’ function in WGCNA with the following parameters: power = 28, maxBlockSize = 26000, networkType = “signed”, corType = “bicor”, mergeCutHeight = 0.1, minModuleSize = 40, pamStage = T, reassignThreshold = 1e-6, deepSplit = 2, minKMEtoStay = 0.3). Subsequently, cutreeHybrid was used to group genes into modules. Once the final set of gene modules were defined, we ran GO term enrichments using the ‘gost’ function from gprofiler2¹³² version 0.2.2 on each set, using the set of all coding genes used in the initial WGCNA analysis as the background set (see Table S6). While a small number of modules had more specific associations (e.g. ‘immune response’), the majority of modules were associated with terms related to protein synthesis. This follows with the important role of the pancreas in both exocrine and endocrine secretion¹³⁹.

We next integrated genetic-association information for T2D. We obtained GWAS summary statistics from a recent GWAS meta-analysis on type 2 diabetes⁶⁰ (http://www.diagram-consortium.org/downloads.html). These variants were aligned to DBSNP 151 to obtain RSID accessions. We assessed the enrichments of these modules for associations with T2D using the gene-ranking method described above, relative to 10,000 randomly-generated gene sets (sampling from the set of all genes used in the pancreas WGCNA analysis) (Table S6). Finding significant enrichments among several co-expression modules for T2D signals, we next used the above-described MAGMA-based approach to confirm the correlations between module eigen-vectors and locus T2D associations (Table S6).

Subsequently, we performed a variant-regulatory analysis on our modules, similar to that described above (Cis-regulatory factor analyses). We linked transcription factors associated with genes in a given module and, intersecting GTEX cross-tissue QTLs with T2D association statistics, extracted the set of regulatory variants associated with these transcriptional regulators. As done in our skeletal analyses, we limited our analyses (both for target and randomized modules) to those transcription factors expressed in the adult pancreas (based on the definition of expressed gene we used in the pancreas WGCNA analysis). Also similar to above, we conditioned our regulatory variant analysis on the epigenetic activity of variants. We obtained uniformly-defined ATAC-seq accessibility peaks from adult human tissues via CATLAS¹⁴⁰; specifically, we obtained peaks from the following categories: Pancreatic beta cell 1, Pancreatic alpha cell 1, Pancreatic beta cell 2, Pancreatic delta/gamma cell, Pancreatic alpha cell 2 – similar to a previous publication focusing on T2D genetics¹⁴¹. We pooled significant peak calls from all categories, taking the unique set of final peaks and lifting over from hg38 to h19. We required that regulatory variants associated with upstream regulators for a given module where active in pancreatic islet cells for inclusion in our analysis. As a point of comparison to our unbiased module sets, we also performed the above variant-aggregation for the set of genes associated with Type 2 diabetes via OMIM, which we obtained using the OMIM GeneMap and manual review. However, given the limited number of OMIM T2D genes we were able to obtain, we were unable to directly compare this set to either or defined modules nor their respective size-matched background distributions (as described below).

To establish a background set for comparison, we pooled the set of module-assigned genes (excluding module CD0, which were unassigned genes from the WGCNA analysis). Then, for a given module, we generated 100,000 randomized modules of equal set-size for comparison. For each randomized module, we performed the same variant-assignment method as described above, establishing background distributions for per-gene averaged log p-values and summed effect-sizes. For each randomized module, we first compared the distribution of per-gene p-values and effect-sizes to the distributions observed for the actual gene module with a one-sided student’s T-test (with an alternative hypothesis of a greater distribution mean in the randomized set). Performing this over all 100K randomized sets, we calculated a false-discovery rate of a random set of genes yielding an improved distribution of per-gene assigned heritabilities/effect-sizes as described above (Table S6). We calculated a similar p-value for the probability of detecting a distribution average greater than the distribution average of a given gene module. For visualization, we show both the distribution of average p-values (as well as effect sizes) for a given module relative to an example randomized ‘module’, as well as the distribution of randomized gene-set averages relative to the average of the target module (see Figure 6, Figure SF6). Across several modules, we observed aggregated per-variant association (p-value) distributions significantly increased relative to randomized module sets (Figure 6C - top/middle, Figure S6B, Table S6).

Similarly, we aggregated both signed and absolute effect-size (beta-coefficient) for a given module and compared to the randomized background distribution. Performing these analyses, we observed a mix of significant increases and decreases in effect size (Figure 6B - bottom, Figure S6B).

Unlike in our skeletal analysis, we did not find that the module most enriched in gene-loci associations (Figure 6B) was that which was most significantly enriched in the network-association analysis. The most strongly-enriched (by gene-locus) module (CD4) was not-significantly enriched relative to randomized modules – instead trending towards decreased aggregate variant associations (Figure S6B). This may suggest that, for this module, the effects of individual gene members matter more than the coordinated effects of these genes as dictated by upstream factors (in the context of T2D associations). Module CD14, having moderate enrichment at the gene-locus level, was strongly enriched for aggregate p-value in this network analysis, possibly suggesting the reverse scenario - where individual gene loci in a module matter less individually, but instead their upstream regulators have an outsized T2D association. It may be the case that genes in this module are themselves strongly discouraged from mutations impacting local regulation, limiting their individual contributions to T2D, but changes in regulation to their upstream TFs have an outsized impact on T2D associations. To explore this latter possibility, we obtained gene loss-of-function (LoF scores) from the gnomAD¹⁴² database (specifically, gnomad.v4.1.constraint_metrics.tsv from the gnomAD website) and intersected these with our module sets. Using the variables ‘lof.pNull’ and ‘lof.z_score’, we calculated the average likelihoods (log10 p-value and z-score) for genes in a given module being under sequence constraint, along with the proportion of a module being significantly (p < 0.05) constrained. These are reported in Table S6.

Compared to genes in other modules, CD14 genes had the greatest average LoF Z-scores (i.e., are more intolerant to variation, Table S6). This may suggest that constraint on the locus-level may limit the contributions of individual CD14 genes to T2D, while TF-mediated network effects may instead play an outsized role in this module’s T2D contributions. In contrast, T2D module CD1 – whose gene members were significantly depleted in locus-level associations (Figure 6B), had the strongest enrichment in the network-aggregation analysis (Figure 6C). This may not be totally explained by heightened constraint around CD1 genes, as CD14 and other modules had greater average LOF Z-scores (Table S6). It may be that CD1 genes, individually, have very weak impacts on T2D, and these weak associations are consistent across many genes. The transcriptional regulators upstream of CD1 may stand to regulate many such small-effect genes (CD1 being the largest-defined module), such that their network-aggregated effects on T2D are quite large. The two modules with the strongest enrichments in network-aggregated associations were CD1 and CD14, the largest and smallest defined modules (respectively). Thus, our analysis is likely not uniformly biased by module set-size effects, but rather may reflect different regulatory scenarios at play for different co-expression gene modules. That the individual behaviors of certain modules we observed in our T2D analysis were different than those observed for certain modules (e.g. CD1) in our height analysis suggests that omnigenicity (as formulated here) is not the same for each disease/trait. Instead, this genetic architecture emerges via the complex relationship the genome has with evolution and its effects on trait manifestation at the transcriptional network level.

Quantification and Statistical Analysis:

Differential-gene expression:

The statistical results from differential gene-expression analyses, performed with DESeq2⁷⁹, were corrected using the Benjamini-Hochberg⁸¹ (FDR) method for all analyses indicated.

Height variant enrichments:

To test for enrichments of height variants in our ATAC-seq peak sets, we generated custom background sets for each individual peak set, generating n = 1000 randomized sets matching both set and peak size. These background sets were subsequently intersected with sets of height variants and counted. Significance was calculated by standardizing the distribution of background intersection values and calculating p-values using a CDF of the normal distribution. Multiple-hypothesis testing correction was applied using FDR for the number of ATAC-seq peak sets tested for enrichment. As an additional significance test, we also calculated a permutation-based p-value as the number of randomized background sets with an equal or greater number of intersections than the target set.

Gene-wise association ranking:

To compare the enrichment for a set of genic regions for association with height GWAS, we generated gene-based windows, assigned association p-values, and ranked windows across the genome (see above). For a given target gene set, we randomly sample 10,000 sets of genes (matching set-size) from the global set of protein-coding genes and calculate the average set-wise rank of each set. This random background distribution is standardized and compared to the average rank of the target gene set using a CDF of the normal distribution and applying FDR correction. Using gene-window rankings, we also compared the ranks of genes differentially-expressed in tissues relative to one another by fitting generalized linear models, correcting for the effects of average cross-tissue gene expression.

SNP-wise set comparisons via ANOVA:

To test whether variants have different per-variant height associations when conditioned on ATAC accessibility patterns, we binned variants into different groups (see description above). We compared the log10 p-values of variants across groups using ANOVA, followed by Tukey post-hoc corrections. Similarly, to test the relationship between the pleiotropic activity of a variant (i.e. how many different tissues it is active in) and height association, we counted distinct tissues and generated a linear model between height association and tissue counts, assessing significance using ANOVA analysis in R.

Module network heritability analysis:

Leveraging transcription factor – target databases and cross-tissue QTL datasets, we considered the putative upstream regulators of gene modules. For a given gene module, we took the set of genes in said module and looked for all annotated upstream regulators in the combined dataset, then extracted the set of QTLs associated with these regulators, and further aggregated the height associations of this variant set. These variants are summed per-gene. To establish a background set for comparison, we pooled all module-assigned genes (excluding module CD0, which were unassigned genes from the WGCNA analysis). To test a given module, we generated 100,000 randomized modules of equal set-size for comparison (randomly pulling from the set of all module-assigned genes), performing the same upstream-regulator analysis for each randomized module. We then implemented two complementary statistical comparisons. (A) For each randomized module, we first compared the distribution of per-gene heritabilities and effect-sizes to the distributions observed for the actual gene module with a one-sided student’s T-test (with an alternative hypothesis of a greater distribution mean in the randomized set). Performing this over all 100K randomized sets, we calculated a false-discovery rate of a random set of genes yielding an improved distribution of per-gene assigned heritabilities/effect-sizes (as described above) relative to the target module set. (B) We calculated the average per-gene heritability for the target module, as well as the 100K randomized modules. A p-value is calculated as the probability of randomly detecting a distribution average greater than the distribution average of the target module.

Supplementary Material

1

Table S1: Epigenetic Datasets. Related to Figures 1, 3, 4 and 5.

2

Table S2: Transcriptomic Datasets. Related to Figures 2, 3, 4 and 5.

3

Table S3: Height Genetics Integration. Related to Figures 4 and 5.

4

Table S4: Extending GWAS with Functional Data. Related to Figure 4.

5

Table S5: Network Characterization. Related to Figure 5.

6

Table S6: Framework application to T2D. Related to Figure 6.

Highlights:

Combined RNA/ATACseq atlas of human skeletal development during critical cartilage stages

Identification of key regulators of bone-end & joint programs with relevance to height

Unbiased detection of cartilage expression modules strongly supports height as Omnigenic

New approach to testing omnigenicity is applicable to other traits like Type-2 diabetes