Complex regulatory interactions at GDF5 shape joint morphology and osteoarthritis disease risk

Abstract

Our ability to pinpoint causal variants using GWAS is dependent on understanding the dynamic epigenomic and epistatic context of each associated locus. Being the best studied skeletal locus, GDF5 associates with many diseases and has a complex cis-regulatory architecture. We interrogate GDF5 regulatory interactions and model disease variants in vitro and in vivo. For all regulatory regions we see that local epigenetic activation/repression impacts patterns of joint-specific expression and disease risk. By modeling the most cited risk variant in mice we found that it had no impact on expression, joint morphology, or disease. Yet, we identified significant epistatic expression interactions between this risk variant and others lying within regulatory regions subject to repression or activation. These findings are important lessons on how regulatory interactions and local epistasis work in the etiology of disease risk, and that assessment of individual variants of high GWAS significance need not alone be considered causal.

Introduction

Osteoarthritis (OA) is a serious aging disease and the leading cause of disability worldwide, affecting ~7.6% of the global population (1). Knee and hip OA are highly heritable (40–60%) and for both joints, morphology and mechanical injury are additional risk factors (2, 3, 4, 5, 6). Indeed, ~50% of hip OA cases result from undiagnosed developmental dysplasia of the hip (DDH) (7, 8) while recently, knee shape has been linked to osteoarthritic disease (5, 9). For both joints, developmental genetic programs underlying cartilage formation prior to and during endochondral ossification aid in determining joint shape (10). The Growth Differentiation Factor Five (GDF5) gene, expressed during embryonic joint formation (11, 12, 13), is critical in this regard; its loss in the brachypodism mouse model (bp/bp) results in serious knee and hip malformations (14). When challenged with collagenase, bp/bp mice go on to develop OA (15).

Of the numerous OA Genome-wide Association Study (GWAS) loci, GDF5 is also the most replicable. Hip/knee OA and DDH GWAS consistently reveal ~100 variants on a common 130 kb risk haplotype spanning the GDF5 regulatory locus. We previously whittled down this haplotype to two causal variants for knee OA and DHH, revealing their unique impacts on joint shape and disease. We first created a mouse model harboring a knee OA risk variant, rs6060369, in the downstream GDF5 regulatory region (enhancer/repressor), R4, and saw it caused statistically significant morphological changes to femoral condyles and tibial plateaus, and a 30% increase in knee OA risk(9). This finding mirrored the morphological changes observed in OA patients harboring the risk allele. We then generated a mouse model, harboring a different risk variant, rs4911178, present in the GDF5 growth-plate enhancer, GROW1. This variant resulted in alterations to acetabular and femoral neck shape but not the knee proper; with DDH patients stratified for this risk allele showing the same directions of effect (9). Despite this work, on this haplotype there still exist the most associated risk variants, rs143383 and rs143384 (16, 17, 18, 19, 20, 21), located in 5’UTR of GDF5. While reporter gene studies (22) and allele-specific expression (ASE) studies on patient knee tissue (23) have functionally tested these variants, their functional interrogation in vivo is lacking, and thus it is unclear if they are relevant to joint disease.

Importantly, ours and others research (14, 18, 19, 21, 24, 25, 26) have led to the GDF5 regulatory locus as one of the most well-annotated loci in the genome. Besides R4 and GROW1, four other identified regulatory regions (enhancers/repressors) (Figure 1) reside upstream or downstream of GDF5. All six enhancers drive different GDF5 expression patterns in vivo as assessed via lacZ mouse transgenesis (at embryonic day (E)14.5) and epigenomic studies on developing (E50’s-E60’s) human joints; findings which demonstrate that the mouse and human regulatory loci are functionally orthologous. Yet, as extensive as this work is, a major issue is that no regulatory region alone is able to reproduce the entire endogenous expression pattern (of GDF5), indicating that regulatory regions must work together to either enhance, repress, or restrict expression in a modular and tissue specific fashion. Elucidating this regulatory activity for GDF5, and how locus-specific activation and repression mechanisms operate, should give us insights into how disease risk variants function. Here, we set out to further explicate and characterize the interactions of GDF5 regulatory regions, interrogate the function of the leading OA GWAS risk variants rs143383/rs143384, and explore potential variant interactions across the GDF5 locus.

Figure 1. Depiction of the upstream and downstream regulatory regions of GDF5.

Modified UCSC Genome browser view (hg19) depicting GDF5 and UQCC1 genes accompanied by Human Long Bone (proximal tibia and femur, distal femur) ATAC-Seq data (E54/E67), GDF5 regulatory elements (R1-R9, R18–20, GROW1), followed by UCSC Gene locations and peaks of PhyloP100ways conservation. Locations of three variants, rs143384 in the 5’UTR of GDF5 rs4911178 in GROW1, and rs6060369 in R4, overlap with regulatory sequences in embryonic human tissues and mouse are indicated in red. Above, images of transgenic embryos collected at E14.5 depict location of expression for each regulatory element tested including an upstream 1kb region encompassing the R2 regulatory region (green), a downstream 37kb (purple) region encompassing GROW1, R18–20, R7, R8 and R9 and a 41kb (blue) region encompassing R3, R4 and R5. Endogenous Gdf5 expression in histological tissue of the embryonic knee joint, E14.5 (grey).

Materials & Methods

Ethics:

All experiments performed on adult or embryonic mice, including euthanasia, have been approved by Stanford University and Harvard University’s respective Institutional Animal Care and Use Committees (IACUC) with protocols (SU, 10665; HU, 13–04-161–2). No human subjects were used. As we are exploring the functional effects of previously published OA GWAS variants, such experiments are not performed in living humans, so no patients were used or involved in this study.

Animal models:

The 5’UTR^{rs143384-T/ rs143384-+} single allelic replacement mouse line contains a single “T” base-pair replacement of the orthologous human rs143384 variant in the 5’UTR at mm10 position (chr2:155,945,103–155,945,103). The R2de mouse model contains a specific 428bp deletion of the R2d+e regulatory region plus adjacent sequence at mm10 position (chr2:155,945,327–155,945,755). Both were CRISPR-Cas9 generated on C57BL/6J Mus musculus backgrounds by Applied StemCell as previously described(9).

Transgenic mice:

Transgenic mice were generated as previously described(19). Briefly, transgenic mice were generated by pronuclear injection into FVB or C57BL6/CBA F1 fertilized oocytes (27, 28). Embryos were collected at E14.5 for X-gal staining. For each construct, multiple transgenic embryos derived from independent integration events were analyzed and only consistent patterns are reported (Supplementary Table 1).

X-gal staining/in situ hybridization:

Whole mount β-galactosidase activity was performed as described with minor modifications (28). Embryos were fixed in ice cold 4% PFA in PBS, hemisected then fixed for an additional 15 minutes at 4°C. Following 3 washes in wash buffer, embryos were stained for 16–24 hours in the dark with 1 mg/ml X-gal in staining buffer at room temperature. Embryos were then washed and fixed in 4% PFA for 5 hours. For sectioning, X-gal stained embryos were placed in sucrose solution before being embedded in gelatin, then cryosectioned at 25um. Nuclear Fast Red was used to counterstain sections. In situ hybridization: Antisense and sense digoxigenin-labeled probes for in situ hybridization were generated for Gdf5 (29).

BARX binding site analysis:

To find upstream transcription factors predicted to bind to R4 regulatory sequences, we used UNIPROBE (30, 31). At the recommended enrichment threshold (0.4), UNIPROBE identified over 1,000 sites (i.e., specific 8-mer sequences bound by a transcription factor) in mouse/human R4 sequences. Predicted factors were intersected with expression/phenotypic data (Eurexpress, http://www.eurexpress.org/ee/; Genepaint, http://www.genepaint.org/Frameset.html; Mouse Genome Informatics, http://www.informatics.jax.org) to narrow down to factors expressed or required in limbs and joints; Barx1–2 displayed overlap with known Gdf5 expression patterns, specifically at gestational days when R4 enhancer was active (32, 33, 34). We next engineered R4 sequences carrying site-specific BARX mutations and these were synthesized by GenScript. Supplementary document shows the sequence changes per BARX site as calculated using UNIPROBE enrichment score analysis. This technique allowed us to identify those sequences where BARX could bind strongly (i.e. wild type, as above) versus those where BARX could no long bind (mutated sites). Each wild-type (WT) or mutated R4 element was cloned between R3 and R5 within the Hsp68 lacZ reporter (Fig. 3b). Each resulting concatenated construct was then used to generate multiple independent E14.5 transgenic mouse embryos for lacZ expression analysis (Supplementary Table 1).

Figure 3. Repressive interactions between nearby regulatory elements affects expression patterns.

(a-b) Transgenic embryos collected at E14.5 of (a) the 41kb construct encompassing R3, R4 and R5, and (b) the PHC19 construct concatenating regulatory elements R3, R4, and R5 (1898bp), both of which drive expression in the same limb domains. Cartoon depiction of construct containing concatenated regions. (c) Transgenic embryos collected at E14.5 of the R3 region, driving lacZ expression in the autopod, and forelimb and hind limb joints, the R4 region, driving expression in the forelimb, hind limb, and digit synovial joints, and the R5 region, driving expression predominantly in the digits. (d) Transgenic embryos collected at E14.5 showing how R3+R4 drive expression in the forelimb, hind limb and digit joints with some expression in the digit periphery; R4+R5 drive expression in the forelimb, hind limb and digit joints; and R3+R5 drive expression only in the webbing of the digits. Red arrows indicate areas of expression patterns that change between constructs. See text for details.

Allele-specific expression:

Pyrosequencing was performed to calculate the allelic ratio of C57BL/6J R2de (or 5’UTR rs143384 mouse ‘T’ allele) in the heterozygous state to 129×1/SVJ (wild-type). Heterozygous ratios determined from cDNA products were then normalized by the ratio of wild-type C57BL/6J to 129×1/SVJ genomic products, amplified from known 1:1 mixtures of each sequence. A non-parametric permutation test was used to assess the significance of affect between the wild-type and heterozygous allelic expression allele in R.

Histomorphology:

The left hind limbs of a minimum of 4 animals/genotype/sex/time point were obtained for histological analysis to assess cartilage integrity. Post sacrifice, skin and excess muscle was removed, and each limb was fixed in 10% neutral buffered formalin for 24 hours at room temperature before being decalcified in 14% Ethylene-diaminetetraacetic acid (EDTA) at pH 7.5 for 8 days. Formalin-fixed tissues were sent to the Massachusetts General Hospital Center for Skeletal Research (CSR) Histology & Histomorphometry Services core facility and processed in one batch for proper sectioning and histological staining using Fast Green and Safranin O (Saf O) staining. Embedding, sectioning, and staining were performed without knowledge of genotype. A minimum of 10 coronal sections taken throughout the joint (60–80um levels) were generated per knee joint. Sections were blind scored by two readers (CC and AK/BP) using the summed Osteoarthritis Research Society International (OARSI) scoring method of OA in the mouse. Briefly, per slide, each quadrant (medial and lateral tibial plateaus, medial and lateral femoral condyles), was assigned a score and summed, with 0 representing intact cartilage, and 6 representing cartilage erosion down to the bone extending >75% of the articular surface. The top 3 summed slides per joint were added to generate one final score per joint. One way ANOVA (R2de) was used to compare OARSI scores between genotypes whereas, Two-way ANOVA (5’UTR^{rs143384/rs143384}) was used to compare OARSI scores between the sex and genotype.

Functional Analysis of GROW1/R4 + 5’ UTR risk and non-risk variants using luciferase reporter assays:

CHON-002 cells, derived from human fetal femoral growth plate chondrocytes at 18 weeks of gestation (female), were sourced from ATCC (CRL-2847), NIH/3T3 (Hoekstra Lab, Harvard University) cells and T/C28a2 cells were cultured at 37°C in a 5% CO₂ environment using ATCC’s complete growth medium, which consists of Dulbecco’s Modified Eagle’s Medium (DMEM) supplemented with 10% fetal bovine serum (FBS), 50 μg/mL penicillin-streptomycin, and 0.1 mg/mL G-418.

For transfection experiments, CHON-002 (GROW1 constructs) cells, T/C-28a2 (R4 constructs and R4 single variant constructs previously described(26)) or NIH/3T3 (R4 single variant constructs) cells were seeded in 96-well plates at a density of 1–4 × 10⁴ cells per well and cultured in DMEM supplemented with 10% FBS for 24 hours. Transient transfections were performed using Lipofectamine^™ 3000 (Invitrogen L3000015) in serum-free DMEM, according to the manufacturer’s protocol. The transfection complex was prepared by diluting Lipofectamine 3000 in Opti-MEM Medium (5 μL Opti-MEM + 0.3 μL Lipofectamine 3000 per well of a 96-well plate). Separately, a DNA-P3000 mix was prepared (5 μL Opti-MEM + 100 ng (or 200ng for NIH/3T3) DNA + 0.2 μL P3000 Reagent per well). The diluted DNA solution was then added to the diluted Lipofectamine 3000 in a 1:1 ratio and incubated at room temperature for 10–15 minutes. Cells were transfected with a firefly luciferase reporter vector containing a GROW1 and 5’UTR or a R4 and 5’UTR and fusion sequence with various combinations of GROW1 (rs4911178) and 5’UTR (rs143383 and rs143384) variants and R4 (rs6060369) and 5’UTR (rs143383 and rs143384) (100 ng total). An empty pGL4.23 luciferase vector (100 ng) was used as a control. To normalize transfection efficiency, the pRL-CMV Renilla luciferase vector (Promega; E226A) was co-transfected. The epistatic activity of the fusion variant combinations was measured 24 hours post-transfection using the Dual-Luciferase Reporter Assay System (Promega, TM040) on an Agilent BioTek Synergy Neo2 Hybrid Multimode Reader (Thermo Fisher Scientific, USA), following the manufacturer’s instructions. Standard and variant synthesis of either R4 or GROW1 sequences concatenated to the 5’UTR including added 5’ KpnI and 3’ HindIII sequences with sequence verification and custom cloning into pGL4.23 custom (Ampicillin) via 5’ KpnI and 3’ HindIII (by recombination) were generated by Genewiz and delivered as a mini-scale DNA sample (see Supplementary Materials for sequences of each construct). The constructs used for the functional analysis of GROW1 and 5’UTR variants in the Dual-Luciferase Reporter Assay are shown in Supplementary Materials.

CRISPR methods

CRISPR targeting of regulatory regions R9, and R2de in vitro. All sgRNAs flanking (1) the R9 regulatory region (2) and the R2de region were designed using MIT CRISPR Tools (http://crispr.mit.edu) and synthesized by Integrated DNA Technologies, Inc (Coralville, Iowa), and cloned into a PX458 vector as previously described in published protocols(35). See supplementary materials for sequences and chromosomal locations of sgRNAs. All guide RNAs, were tested for deletion efficiency of respective human elements in cultured T/C-28a2 cells (n=3 biological replicates per assay). T/C-28a2 cells were maintained as described above and seeded in a 6 well plate 24 hours before transfection. Transfection efficiency was measured using a fluorescence microscope (>70% of cells were GFP positive). Extraction of DNA was performed using the E.Z.N.A Tissue DNA Kit (Omega Bio-Tek, Norcross, GA), and respective regulatory elements were amplified using PCR with primers flanking each sgRNA location (Supplementary Table 2), followed by purification from 1% agarose gel (E.Z.N.A Gel Extraction Kit). Sanger sequencing was used to verify each successful targeting event

Impacts of each modification on GDF5, UQCC, and CEP250 expression were assessed by extracting RNA from control and CRISPR-Cas9 targeted T/C-28a2 or NIH/3T3 cells (n = 3 biological replicates, with three technical replicates per experiment per condition) using Trizol Reagent (Thermo Fisher Scientific, Springfield Township, New Jersey) and Direct-zol^™ RNA Miniprep kit (ZYMO). SuperScript III First-Strand Synthesis System (Thermo Fisher Scientific) was used to prepare cDNA. qRT-PCR analysis was then performed with gene specific primers (9, 26) and Applied Biosystems Power SYBR master mix (Thermo Fisher Scientific) with GAPDH house-keeping gene as an internal control.

Micro-CT and anatomical measurements:

Acetabular joint, femur, and tibia from right hind limbs were collected using high-resolution Micro-Computed Tomography (μCT40, SCANCO Medical AG, Brüttisellen, Switzerland). Scan parameters were 12 μm3 isotropic voxel size, 70 kVp peak X-ray tube intensity, 114 mA X-ray tube current, and 200 ms integration time. Resultant DICOM images were exported for measurements following anatomical features in Osirix MD v7.5 (Pixemo SARL, Bernex, Switzerland). First, using previously described methods (9), measurements were taken of the acetabulum (depth, diameter, inclination), proximal femur (valgus cut angle, neck-shaft angle, neck length, neck diameter, head offset, head diameter), distal femur (bicondylar width, notch width, condylar width (medial and lateral), condylar curvature (medial and lateral), trochlear width (medial, central, and lateral), trochlear groove depth, trochlear angle), and proximal tibia (plateau width, posterior tibial slope (medial and lateral), tibial spine height (medial and lateral)). All measurements displayed strong inter- and intra-examiner reliability (ICC > 0.78). Second, MicroCT-DICOM images were segmented to generate 3D models of each bone using image processing software (Mimics v17.0, Materialise). 3D models were imported to 3-matic software package (v9.0, Materialise) and co-registered together using a global n=point registration method. 3D models of wild type and homozygous mice were used to generate 3D heatmaps indicating the geometrical differences between genotypes for each mouse line. The heatmaps were generated by calculating the distance between the corresponding points in co-registered models, where dark blue indicates the maximum deviation in the negative direction and red indicates the maximum deviation in the positive direction.

Statistical analysis:

Expression data for GDF5, CEP250, and UQCC1 were normalized relative to GAPDH house-keeping gene expression and compared between control and R2de enhancer deletion. All data are presented as the mean ± SEM unless otherwise stated. Individual pairwise comparisons between control and experimental condition were analyzed by two-sample, two-tailed Student’s t-test, with p < 0.05 regarded as significant. N = 4 technical replicates per biological replicate (3 biological replicates).

Results

Activation/repression domains within a 37 kb sequence containing GROW1

We previously reported on a broad 37kb human sequence downstream of GDF5 (purple region, Fig. 1, Fig. 2a) that drives GDF5 expression only in the sub-perichondral region of long-bone growth plates, identical to that of GROW1, a 2.54kb enhancer containing a key DDH risk variant (rs4911178). Interestingly, this 37kb sequence consists of a ~12.17kb subregion PHC17 (Fig. 1, light grey) that drives expression within the entire growth plate chondrocytes (i.e., not only in the sub-perichondral region) and a ~10kb subregion PHC18 (Fig. 1, dark grey) that drives expression in the shoulder joint proper (Fig.2b-c). Since Gdf5 is not endogenously expressed throughout the growth plate (Fig. 1), this means that built-in or adjacent to PHC17 there are repressors that restrict expression to the sub-perichondral space. We therefore first tested two conserved PHC17 sub-regions, one called R18–20 (~2.7kb from GROW1), and another called R7 (~6.9kb from GROW1). We found that neither drive growth plate expression (Fig. 2b-c), but rather at other joint and digit sites. As noted, the adjacent PHC18 (Fig. 2c) drives expression in the shoulder, which is not a pattern controlled by the larger 37kb sequence. We next tested two conserved regions within PHC18 termed R8 and R9. Surprisingly, R8 alone drives expression within many more hind limb and forelimb joints, indicating that it is being repressed in PHC18 and the 37kb constructs. Upon testing R9 we found no expression in limbs indicating it might act as a repressor, especially when considered in PHC18 and 37kb parental constructs. To test this, we deleted R9 in human T/C-28a2 chondrocytes and observed a significant increase in GDF5 expression, thus revealing a normal repressor role (Fig. 2d). We note that in this R9 repressor there exists two OA GWAS risk variants (rs2378349 and rs2248393) that may modulate repressor and thus GDF5 expression levels. We performed a luciferase reporter test of these two variants in T/C-28a2 chondrocytes and observe some impacts of the risk allele on GDF5 expression (see Supplementary Figure 1c). Thus, in considering both PHC17 and PHC18 within the 37kb region, we note that the endogenous growth plate sub-perichondral expression, recapitulated by both 37kb and GROW1 regions (Fig. 2a) results from two potential repression systems: (1) a localized system within PHC17 via R18–20 and/or R7 and (2) a long-range system within PHC18 via the actions of the R9 repressor.

Figure. 2. Restriction of GROW1 to the perichondrium exerted by downstream regulatory regions.

Transgenic embryos driving lacZ were collected at E14.5 of (a) 37kb fosmid driving growth plate expression and (b) regulatory region PHC17 (GROW1, R18–20, and R7) with two histological panels of the hind limb showing PHC17-driven LacZ expression throughout the entire growth plate. (c) Transgenic embryos collected at E14.5 of region PHC18 (which includes R8 and R9). Red arrows indicate locations of differential expression patterns between regulatory elements. (d) Relative Gdf5 expression following CRISPR knockout of R9 in T/C-28a2 cells.

Activation/repression domains within a 41 kb sequence containing R4

We had characterized a human 41 kb sequence (blue region, Fig. 1, Fig. 3a), further downstream of GDF5, showing it recapitulated the classic Gdf5 limb joint-specific expression pattern. This regions three-known enhancers drive unique patterns: R3 in the interdigital space, digital transverse stripes, and knee/elbow joints (R3, Fig. 3c); R4 in the elbow/knee, and digit joints, with weaker expression in hip/shoulder (R4, Fig. 3c); R5 in the prechondrogenic phalangeal mesenchyme and weakly in elbow/knee (R5, Fig. 3c)(19). Alone, R3, R4, or R5 cannot drive the broader 41 kb pattern. Moreover, R3 and R5 patterns are not observed by the broader 41 kb sequence, nor are they known Gdf5 expression territories. Interestingly, when concatenated (in construct PHC19, Fig. 3b) R3(586bp) + R4(975bp) + R5(337bp) (i.e.,1898bp out of the 41kb) strikingly generated the 41kb pattern. To understand this complex interaction further, we first concatenated R3+R4 and found together they drive strong expression in the knee/elbow, and digit joints including transverse stripes and some interdigital webbing (Fig. 3d). Here, R3 acts to restrict knee, elbow, and shoulder expression of R4, whilst R4 represses R3 expression in the interdigital webbing. We next tested R4+R5 and found together they drive weak expression in the elbow, but strong expression in the knee, transverse stripes of the digits, and weak expression of the metacarpals (Fig. 3d). Here, R4 acts to suppress R5 mesenchyme expression whilst R5 acts to restrict (and weaken) R4 expression in large joints. Finally, we tested R3+R5 and found together they cause the strong digital mesenchyme expression of R5 to be lost, while R3 expression becomes further restricted in the interdigital webbing (Fig. 3d). These studies reveal that in addition to enhancer activities, each regulatory element can act as a repressor.

As we have shown, the ability of a regulatory region to restrict expression appears to be joint and tissue dependent, emphasizing the importance of location for gene regulatory activity to recapitulate endogenous Gdf5 expression. To emphasize the importance of this finding to causal disease biology, we tested the previously reported R4 “T” OA risk allele at rs6060369 in two contexts. First, by testing the risk “T” allele (compared to the non-risk “C” allele) in human T/C-28a2 chondrocytes (a cell type in the developing knee), we find that the R4 element acts as an enhancer, and the “T” risk allele decreases its activity (Supplementary Fig. 1a). However, when the “T” risk allele is tested in a fibroblast line (NIH/3T3 cells; another cell-type in the developing knee), we find it serves to derepress the activity (and thus increase expression towards baseline) (Supplementary Fig. 1b). These findings reveal that as in our in vivo LacZ experiments, R4 in different human cellular contexts can repress or activate, and importantly that its risk variant effects depend on the epigenomic context.

Predicted transcription factor binding sites in R4 are required for expression in the forelimb and hind limb joints

While variant roles in disease biology are dependent on the cis-regulatory, epigenomic context for which they are located, the transcription factor (TF) or trans-environment is also important to consider. We next deeply interrogated the R4 regulatory element given its importance to knee OA risk. Using published methods (Methods) we identified BARX, a homeodomain protein with known roles in chondrogenesis, as predicted to strongly bind to five locations across R4 (Fig. 4a)(32). BARX1 and BARX2 are known to be expressed across developing joints, exhibiting strong overlap with GDF5 expression (32, 33, 34, 36). We next sought to experimentally test each BARX binding site using in vivo lacZ approaches (Methods). We first tested all 5 sites within R4 by mutating each to destroy (through reshuffling) their predicted binding sequence. When all 5 sites were mutated (MUTB1-MUTB5,) lacZ expression was restricted only to the digits, reducing expression in digit webbing and eliminating large limb joint expression (Fig. 4 b and 4c and Supplementary Figure 2a and 2b). We next generated separate lacZ constructs with each TF binding site mutated independently and found no impact on expression patterns as a result of MUTB1, MUTB4 or MUTB5 (Fig. 4d, and Supplementary Figure 2c), but strong repression of knee and shoulder expression as a result of MUTB3 (Fig. 4e). MUTB2 (Supplementary Figure 2c) also reduced expression within the limb joints but not as strongly as MUTB3. Interestingly, the OA GWAS associated R4 variant (rs6060369)(9) lies between MUTB2 (92bp away) and MUTB3 (66bp away), respectively. Overall, these data reveal the importance of transcription factor functioning on R4 activity and that individual TF binding sites (and variants nearby) can have sizable impacts of the regulatory activity of important enhancers.

Figure 4. Predicted BARX binding sites are required for R4 hind limb and forelimb joint expression.

(a) Five predicted (BARX) homeodomain binding sites were identified within R4 and mutated at each individual site, or all were mutated simultaneously. (b) WT transgenic E14.5 embryos of construct PHC19 (R3+R4+R5) driving expression of lacZ. (c) Transgenic E14.5 embryos of construct PHC19 with all 5 BARX sites mutated, showing loss of expression in the forelimb and hind limb joints. (d) Transgenic E14.5 embryos of construct PHC19 with only one BARX site mutated (MUTB1) where there is no impact on lacZ expression. (e) Transgenic E14.5 embryos of construct PHC19 with mutated (MUTB3) where there is a complete loss of forelimb and hind limb joint expression. Red arrows indicate areas of expression patterns that change between constructs.

Activating/repression domains within the upstream R2 joint enhancer

We had reported (19) a ~1 kb enhancer, R2 (green region, Fig 1.), that drives strong expression within limb joints and recapitulates a subset of the entire endogenous Gdf5 expression pattern. This element resides near the GDF5 promoter and 5’UTR. There are 5 highly conserved regions(19) across the R2 element denoted a-e (Fig. 5a). Interestingly, whilst a, b, and c do not reproduce expression either independently or together (a+b+c) (Fig.5b), sub-elements d and e can recapitulate expression in the hip and knee or shoulder and elbow, respectively (19). We identified a shorter 112bp sequence within d that suppresses elbow and metapodial joint expression (Supplementary Figure 3). Most strikingly, by concatenating a+b+c with either d or e (i.e., a+b+c+d or a+b+c+e), abc suppresses expression in both the hind limb and forelimb. This reveals built-in repression domains within individual enhancers. Strikingly, under any combination of sub-element concatenation, digit expression could not be recapitulated, revealing that digit expression requires the full R2 regulatory element to be intact for expression.

Figure 5. Sub-regions of R2 repress regulatory region activity.

a) Cartoon representation of conservation of 5 sub-regions (a-e) within the R2 element across different species (19). (b) Transgenic E14.5 embryos of the entire R2a-e sequence driving expression in the forelimb, hind limb and digit joints; sub-region R2abc is unable to drive expression in any joint, whilst subregions de drive expression in hind limb and forelimb joints, but not in digit joints. (c) Transgenic E14.5 embryos show sub-regions R2d and R2e are able to drive expression in hind limb and forelimb joints respectively. (d) Transgenic E14.5 embryos showing that sub-regions abc when placed adjacent to sub-region d or e represses hind limb or forelimb expression, respectively. Red arrows indicate areas of expression patterns that change between constructs.

Deletion of R2de results in morphological changes not associated with OA development

Enhancer R2 is adjacent to GDF5 5’UTR, in which resides the most associated OA GWAS variant, rs143384. To understand the role of R2 in large joint (e.g., knee) biology and OA risk, we generated a mouse line with a deletion of the limb-joint-specific R2de region. Using allele-specific expression analysis (ASE) on 7 separate E15.5 limb joint sites, we observe a consistent ~40% decrease in Gdf5 expression at each site (Supplementary Figure 4a). When deleted in T/C-28a2 human chondrocytes, GDF5 (but not nearby UQCC1 and CEP250) expression is similarly reduced (Supplementary Figure 4b, 4c and 4d respectively). Given the expression reduction, we performed MicroCT morphological analyses on wild-type (WT), R2de^−/+ and R2de^−/− mice at P56 and found statistically significant shape changes across the knee’s femoral and tibial plateau and hip’s acetabulum (Fig. 6a and Supplementary Figure 4f). Strikingly, by 1.5 years, only knee bicondylar width remained statistically impacted (Fig. 6b and Supplementary Figure 4g). By 3D analysis, these changes predominantly locate to the medial femoral condyles at P56 and 1.5years, while at 1.5years hip alterations are ameliorated though some changes remain at the anterior tibial plateau (Fig. 6c). Apart from two WT cases of spontaneous OA at 1.5 years, there are no changes to cartilage integrity measured by OARSI scoring at P56 or 1.5 years of age, and separation of scores by plateau and condyle does not reveal regionally-specific effects (Fig. 6d, 6e and Supplementary Figure. 4e). Likewise, no statistically significant changes are observed in tibial or femoral articular cartilage thickness at either time point (Fig. 6e and Supplementary Figure. 4e). Thus, the loss of R2de, which has a marked effect on Gdf5 gene expression, has no long-term observable impacts on joint morphology and no increased risk of spontaneous OA.

Figure 6. Morphological characterization of the Gdf5 R2de enhancer mouse model.

(a) MicroCT measurements of significantly different anatomical features in R2de deletion mice at P56 (WT n = 4, Het n = 5, Hom n = 6) and 1.5 years (WT n= 8, Het n= 7, Hom n = 7). One-way ANOVA with Tukey-kramer test was used for comparisons between all groups (*p<0.05, **p<0.01, ***p<0.001, ****p<0.0001, bars indicate medians and 95% Confidence Intervals) at both P56 and (b) 1.5 years of age. (c) 3D morphological comparative analysis indicating locations of largest anatomical differences between wild type (WT) and homozygous R2de null hind limbs at P56 and 1.5 years. (d) Coronal histological sections of the medial compartment stained with Saf O for representative OA score per timepoint. (e) Tibial articular cartilage thickness measurements of the lateral and medial plateaus at P56 and 1.5 years alongside respective OARSI scores for each time point (P56: WT n = 4, Het = 5 HOMO n = 6; 1.5 years: WT n= 8, Het n= 7, Hom n = 7).

rs143384 risk allelic mice lack joint alterations and OA disease

Within the GDF5 5’UTR is the OA GWAS risk variant, rs143384, a ‘G’ to ‘A’ risk mutation. While others have shown that the variant impacts reporter gene expression in vitro or GDF5 expression in patients (23), no one has tested its role in joint development and homeostasis in vivo. We generated a humanized 5’UTR^{rs143384-A/rs143384-G} allelic replacement line. Using ASE at E15.5 on 7 joint sites, we observe no impact of the risk ‘A’ allele on Gdf5 expression (Supplementary Figure 5a). Using MicroCT imaging on P56 wild-type, 5’UTR^{rs143384-G/rs143384-A}, and 5’UTR^{rs143384-A/rs143384-A} mice, we find the “A” allele causes no significant changes to the width and curvature of the knee’s femoral condyles or to the size and slopes of the tibial plateau at P56 (Fig. 7a, 7b, and Supplementary Fig 5c). We observe only a modest impact on notch height in female mice, albeit GDF5 does not exhibit sex effects on OA risk in any GWAS to date. Histologically, we observe no increase in disease risk between each genotype, or sex by OARSI scoring at P56 (Fig. 7c and Supplementary Figure 5b). We conclude that modeling the variant alone reveals little-to-no impact on knee OA biology and disease risk.

Figure 7. Morphological characterization of the Gdf5 5’UTR rs143384 variant in the mouse model.

(a) MicroCT measurements of anatomical features in 5’UTR rs143384 single base-pair replacement mice at P56 (Males: WT n = 10, Het n = 17, Hom n = 12. Females: WT n= 10, Het n= 15, Hom n= 12). Welch’s t test was used for comparisons between WT and Homozygous measurements (*p<0.05 bars indicate medians with 95% confidence intervals). (b) 3D morphological comparative analysis indicating locations of largest anatomical differences between WT and homozygous risk 5’UTR rs143384 hind limbs at P56. (c) Coronal histological sections of the medial compartment stained with Saf O of representative OARSI scoring (Males: WT n = 8, Het n = 8, Hom n = 8. Females: WT n= 7, Het n= 8, Hom n= 8).

Modeling variant epistasis across the complex GDF5 regulatory locus

The common 130 kb risk haplotype spanning GDF5 associates with knee OA, DDH, height, and other musculoskeletal disease/traits (18). We had revealed causal risk alleles at rs4911178 (“A”) in GROW1 and rs6060369 (“T”) in R4, uncoupling risk for DDH and knee OA, respectively. Yet above we revealed complex regulatory interactions within the vicinity of each variant and across the locus. Given that risk “A” rs143384 is in strong linkage disequilibrium with both risk “A” rs4911178 and “T” rs6060369 risk alleles, we hypothesized that there could be epistatic interactions between these variant positions (Fig. 1, variants in red). To first test for epistasis, we generated constructs containing different combinations of risk (R) and non-risk (N) alleles at three variants; rs143383 (5’UTR) (an additional associated variant), rs143384 (5’UTR), and rs6060369 (R4) (Fig. 8a) and transfected them into T/C-28a2 human chondrocytes (methods). We found that the R4 risk/non-risk alleles drive lower luciferase expression than 5’UTR risk/non-risk alleles. Strikingly, risk alleles at R4 and 5’UTR variants (Fig. 8a, blue values) drive nearly half the expression on non-risk alleles, indicating a strong epistatic interaction (Fig. 8a, red values). In the same fashion we tested for epistasis between or rs4911178 (in GROW1) and 5’UTR variants, but transfected constructs into the human growth plate chondrocyte cell line, CHON-002. Here, we observe much lower luciferase expression with constructs containing either GROW1 non-risk/risk alleles in comparison with 5’UTR non-risk/risk alleles, with risk alleles driving lower expression than the non-risk. Additionally, we observe that all three risk alleles at GROW1 and the 5’UTR (Fig. 8b, red values) result again in an approximate halving of activity compared to non-risk alleles at these 3 variant positions (Fig. 8b, blue values). Together, these studies reveal complex interactions between variants that impact GDF5 expression. We have included a diagram showing all interactions across the locus discovered to date (Fig. 8c).

Figure 8. Disease risk variants interact to impact expression.

(a) Relative normalized luciferase expression produced by plasmids containing different non-risk (N) or risk (R) variants of either the R4 variant rs6060369, or the 5’UTR variants rs143383 and rs143383 in T/C-28a2 cells. (b) Relative normalized luciferase expression produced by plasmids containing different non-risk (N) or risk (R) variants of either the GROW1 variant rs4911178, or the 5’UTR variants rs143383 and rs143383 in CHON-002 cells. (c) Cartoon model of identified interactions across the GDF5 locus including risk variants marked as red lines (depictions not to scale).

Discussion

Localized interactions of regulatory sequences act to finely pattern Gdf5 expression:

Prior to this study, what was known about Gdf5 regulatory activity was restricted to broad patterns across the locus or specific regulatory sequences, with no large or small regulatory region able to fully recapitulate endogenous Gdf5 expression. On interrogation of a downstream 37kb region encompassing the reported GROW1 sub-perichondral enhancer (Fig.1, purple region), we identified an overlapping conserved subregion PHC17 that drove lacZ expression throughout the growth plate, a pattern not observed for any larger or smaller construct nor for endogenous Gdf5 expression. Indeed, PHC17 individual components, GROW1, R18–20, and R7, were each unable to recapitulate this full growth plate signal. Interestingly, in the adjacent PHC18 subregion, we found that the R9 element acts to restrict R8 expression patterns to only the shoulder joint, with its repressor activity confirmed in human chondrocytes. Therefore, R9 is the strongest candidate to restrict GROW1 to the sub-perichondral region across the entire 37 kb window. Further downstream in the 41kb joint region, we then observed that different combinations of the R3, R4, and R5 enhancers regions drove very distinct patterns across limb joints with none alone able to recapitulate the complete 41kb, PHC19, or Gdf5 expression patterns. Here, we identified that these enhancers also act as repressors with small or large impacts on expression depending on the skeletal site. Finally, for the upstream joint region R2, we identified subregion abc as a repressor sequence of the d and e subregions. Moreover, we identified a shorter 112bp sequence within the d region itself that restricts d expression to the major hind limb joints only. From these findings we conclude that there are complex activator and repressor sequences that operate between and within regulatory regions that can be tissue/cell type dependent as well as joint specific. We argue that this complex regulatory landscape likely evolved to precisely shape joints and long-bones, a point evinced by the high dysmorphic joints of Gdf5 complete loss-of-function (bp/bp) mice.

GDF5 complex regulatory landscape impacts how, when, and where risk variants drive abnormal disease phenotypes:

By generating mice with a deletion of the R2de sequence adjacent to Gdf5 5’UTR we identified a 40% decrease per allele in Gdf5 expression at all large limb joint skeletal sites. Interestingly, complete loss-of-function (R2de^−/− mice) did lead to hip and knee morphological changes early on (P56), but by 1.5 years these differences were no longer significant and importantly did not lead to any increase in OA disease. This is in-line with what is observed in mice with Gdf5 coding mutations - i.e., bp/+ have a 50% drop in expression and bp/bp mice have no functional expression, and neither develop OA, even though bp/bp mice have massive joint disruptions throughout life. These observations are in stark contrast to our findings from functional tests of individual Gdf5 regulatory sequences and variants therein. Mice with a R4 enhancer deletion have decreases (of ~32%/allele) in Gdf5 expression only in the knee, but not in other joint sites. This joint-specific decrease led to morphological changes in R4^−/− mice only in the knee at P56 and 1 year, with resulting effects on OA. Humanized mice with a GWAS OA risk variant (rs6060369, C→ T) in R4 also have decreased Gdf5 expression (~16%/allele) within knee epiphyseal and articular cartilage only, and this reduction causes significant morphological changes in the same direction as OA patients, and a 30% increase in OA in homozygous “T/T” mice (9). Mice with a GROW1 enhancer deletion have decreases (16%/allele) in Gdf5 expression only in the proximal femur and acetabulum, but not at other joint sites. This led to femoral head and neck and acetabulum alterations (but not in knees), recapitulating those observed in DDH patients. Finally, humanized mice with a GWAS DDH risk variant (rs4911178, G→A) in GROW1 also have decreased Gdf5 expression (16%/allele) only within the proximal femur and acetabulum sub-perichondral chondrocytes, and this causes hip changes in mice in the same direction of effect as DDH patients (9).

Collectively, these functional studies reveal that location-specific decreases in Gdf5 expression are far more important mediators of complex disease risk (OA, DDH, etc.) than large effects of expression reduction observed across different joint sites and tissues, the latter which are typically not observed in human patients with complex (i.e., common variant mediated) joint disease (but in fact are more indicative of patients with syndromic disorders due to GDF5 coding mutations). We argue that this is the case because such location-specific decreases alter joint shape locally, changes that are not reciprocated on the opposing joint surface; resulting in a misregistration of joint articulation, which developmentally (DDH) or over time (OA), can predispose to disease. Furthermore, these effects on complex disease risk are in turn the product of the fine sculpting of GDF5 expression by complex interacting regulatory sequences.

Regulatory variant interactions drive complex disease risk at GDF5:

We have shown that across the locus, GWAS risk variants can reside in enhancers (e.g., rs4911178 in GROW1), repressors (rs2378349 and rs2248393 in R9), or in sequences that behave as both (e.g., rs6060369 in R4), which in turn markedly obscures one’s ability to pinpoint causality at the individual base-pair level and at cell-type resolution. For example, by generating humanized mice harboring the most highly cited OA SNP variant (rs143384, G→A) located in GDF5 5’UTR, we did not observe any significant expression reductions at any joint site, nor did we find any significant morphological changes, or evidence of knee OA disease in adult mice. And this reveals that alone, rs143384 “A” is likely not causal for OA disease risk. Yet, we would be remiss to have not considered this (or any) variant within GDF5 complex cis-regulatory architecture; a system that has evolved built-in activation/repression mechanisms locally (i.e., within a regulatory element) and further afield (across broad growth plate or joint regions) to generate endogenous Gdf5 expression territories at each joint site, and thus precisely sculpt joints. Indeed, by testing for epistasis, we did observe substantial interactions between 5’UTR rs143384 “A” and R4 rs6060369 “T” risk alleles, and with GROW1 rs4911178 “A” risk alleles. In each context, we argue that cis-regulatory variants provide locational and cell-type specificity, and by doing so they (i.e., R4 and GROW1) both reduce and direct how reductions caused by other (5’UTR) variants become channeled to specific tissues. This in turn results in local joint mis-registration causing developmentally driven site-specific musculoskeletal disease risk at GDF5. We posit that disease risk at this developmental locus and most others, are analogous to the risks for developing cancers due to the actions of double or triple hits, but here hits are the myriad cis-regulatory variants residing on risk haplotypes and within their built-in complex epigenetic interactions (37, 38). Therefore, understanding the epistatic context in which a disease risk variant lies will be fundamental to disentangling and prioritizing the close to 2000 OA associated GWAS (39, 40) variants that have been identified.

Teaser:

Genetic interactions at the most studied skeletal disease locus reveal hidden complexities in pinpointing causal mutations.

Data Availability Statement

All data produced in the present work are contained in the manuscript.