Genome-wide analyses implicate 33 loci in heritable dog osteosarcoma, including regulatory variants near CDKN2A/B

Abstract

Background: Canine osteosarcoma is clinically nearly identical to the human disease, but is common and highly heritable, making genetic dissection feasible.

Results: Through genome-wide association analyses in three breeds (greyhounds, Rottweilers, and Irish wolfhounds), we identify 33 inherited risk loci explaining 55% to 85% of phenotype variance in each breed. The greyhound locus exhibiting the strongest association, located 150 kilobases upstream of the genes CDKN2A/B, is also the most rearranged locus in canine osteosarcoma tumors. The top germline candidate variant is found at a >90% frequency in Rottweilers and Irish wolfhounds, and alters an evolutionarily constrained element that we show has strong enhancer activity in human osteosarcoma cells. In all three breeds, osteosarcoma-associated loci and regions of reduced heterozygosity are enriched for genes in pathways connected to bone differentiation and growth. Several pathways, including one of genes regulated by miR124, are also enriched for somatic copy-number changes in tumors.

Conclusions: Mapping a complex cancer in multiple dog breeds reveals a polygenic spectrum of germline risk factors pointing to specific pathways as drivers of disease.

Figure 1

The three GWAS breeds (high risk of OS) and the AKC greyhounds (low risk of OS) are genetically distinct and highly inbred populations. (a) Along the top two principal components of variation, the three breeds (Rottweilers (n = 135, red), Irish wolfhounds (IWH, n = 141, purple), and racing greyhounds (n = 267, dark blue) ) are each tight clusters equidistant apart, with the AKC subpopulation of greyhounds (n = 19, light blue) near but not overlapping the racing dogs. (b) Estimate of the inbreeding coefficient (IC) shows the racing greyhounds as the least inbred, on average, while the other three populations fall within the range seen in an earlier survey of 28 dog breeds (grey bars mark average IC in breed; two outliers at 0.53 and 0.62 not shown) [17]. (c) The extent of linkage disequilibrium, measured as average pair-wise r², drops below 0.2 at 196 kb, 632 kb, and 1,533 kb in the greyhounds, Rottweilers, and IWH, respectively.

Figure 2

Mixed-model GWAS corrects for population structure and identifies 33 OSA associated loci explaining a large fraction of phenotype variance. In each breed, the QQ plots show no evidence of stratification relative to the expected distribution, identifying nominal significance at -log₁₀p of 3.5 and the 95% empirically determined confidence intervals (dashed grey line) at -log₁₀p of (a) 5 in the greyhounds, (b) 4 in the Rottweilers, and (c) 3.7 in IWH. In the IWH, a plateau of SNPs at P = 6.6 × 10^-5 corresponds to a 1.65 Mb haplotype on chromosome 18 peaking at the gene GRB10. (d) In greyhounds, 14 loci have P<0.0005, with one locus, on chromosome 11, exceeding 95% confidence intervals (dashed lines). (e) In Rottweilers, 15 and 6 loci are identified, (f) while only four and two loci are identified in IWHs. (g) For each breed, the phenotype variance explained by the associated loci, broadly defined by SNPs with r² >0.2 within 5 Mb of the peak SNP, exceeds 50%. In greyhounds, the 14 regions explain 56.9% +/− 12.5%, in Rottweilers, 15 regions explain 85.3 +/− 13.6%, and in IWH, four regions explain 53.1 +/− 15.5%. (h) For each affected dog (red circles) and unaffected dog (black circles), we estimated their relative risk based on the genotypes and odds ratio of the top SNP from each region for the breed (Table 1), showing that even using a small number of SNPs we see clear differences between the cases and the controls.

Figure 3

Identification of the top associated variant and functional analysis on chromosome 11. (a) We targeted 2.5 Mb around the greyhound GWAS peak on chromosome 11 for dense sequencing (15 dogs) and finemapping (180 cases and 115 controls). Imputation and association testing of sequenced variants narrowed the peak of association in greyhounds dramatically to a 15 kb risk haplotype (chr11:44390633–44406002), telomeric of the genes CDKN2A and CDKN2B, that is nearly fixed in both the Rottweilers (98% in cases and 96% in controls) and IWH (95% in cases and 92% in controls). (b) The top haplotype (blue vertical lines) maps to a locus downstream of the non-coding gene ANRIL on human chromosome 9 (hg19). (c) We tiled the human chromosome 9 region with luciferase assays and assayed the function in osteosarcoma cell lines compared to renilla. Potential markers of function in the region include H3K27 acetylation in osteoblasts and DNase hypersensitivity clusters (assayed from 125 cell types), most notably in regions that align between the dog and human genomes in a Multiz alignment of 46 species and are constrained across mammals as measured by Genomic Evolutionary Rate Profiling (GERP) [21-24]. (d) Of the seven non-control luciferase assays, four (B, C, E, and G) showed a significant increase compared to empty vector. Construct G showed by far the strongest increase with an approximate 32-fold increased activity suggesting a strong enhancer. (e) This fragment contains one of the top SNPs (Canfam2.0 chr11:44405676) which has a constrained reference allele C corresponding to a predicted transcription factor binding site, while the OS associated allele, A, is not found among 29 mammals or the wolf.

Figure 4

Connectivity and enrichment analysis identifies pathways linked to OS in multiple breeds. (a) Most of the associated regions in each breed contain one or more genes (black text). Genic regions are shown for the greyhounds (blue), IWH (purple), and Rottweilers (red), with hue alternating between light and dark to distinguish regions, numbered as in Table 1. GRAIL [37] analysis identified non-random connectivity (P<0.05) between associated genes (bold text), both within breeds (blue, purple, and red arched lines for greyhounds, IWH, and Rottweilers) and between breeds (grey arched lines). Twelve regions contain genes (blue dots) connected to the key word ‘bone’, one of the top terms identified by GRAIL (Table S4). (b, c, d) When gene set enrichment P values for the associated regions and regions of reduced variability are combined, most sets shows no inflation compared to background (grey circles; RRV P values from 28 other breeds). However a small number are inflated, including (b) five from the NCI pathway interaction database [47], (c) two from the Molecular Signatures database of shared cis-regulatory motif based sets [48], and (d) 0 from the Gene Ontology database. (e) Of the seven gene sets (colors match discovery breed), five are significantly enriched (bold numbers) in regions that are aberrant in all greyhound (blue), all Rottweiler (red), or all dogs (black) for which we compared normal and tumor DNA using comparative genomic hybridization. Boxed numbers show number of genes in gene set overlapping CGH regions; ‘up’ and ‘down’ indicates gain or loss in all samples; ‘any’ indicates all samples are either amplified or deleted.