Large-scale transcriptome-wide association study identifies new prostate cancer risk regions

Abstract

Although genome-wide association studies (GWAS) for prostate cancer (PrCa) have identified more than 100 risk regions, most of the risk genes at these regions remain largely unknown. Here we integrate the largest PrCa GWAS (N = 142,392) with gene expression measured in 45 tissues (N = 4458), including normal and tumor prostate, to perform a multi-tissue transcriptome-wide association study (TWAS) for PrCa. We identify 217 genes at 84 independent 1 Mb regions associated with PrCa risk, 9 of which are regions with no genome-wide significant SNP within 2 Mb. 23 genes are significant in TWAS only for alternative splicing models in prostate tumor thus supporting the hypothesis of splicing driving risk for continued oncogenesis. Finally, we use a Bayesian probabilistic approach to estimate credible sets of genes containing the causal gene at a pre-defined level; this reduced the list of 217 associations to 109 genes in the 90% credible set. Overall, our findings highlight the power of integrating expression with PrCa GWAS to identify novel risk loci and prioritize putative causal genes at known risk loci.

Fig. 1

Tissue-specific predictive models for gene expression. a Cross-validation prediction accuracy of cis-regulated expression and splicing events (R²) for all 109,170 tissue-specific models. b Normalized prediction accuracy ($R^2∕\mathrm{c}is-h_g^2$) for all 109,170 tissue-specific models. c Histogram of the number of reference panels per gene. The majority of genes were heritable in a small number of tissues, but many genes exhibited heritable levels across many tissues

Fig. 2

OncoArray PrCa TWAS and GWAS. The top figure is the TWAS Manhattan plot. Each point corresponds to an association test between predicted gene expression with PrCa risk. The red line represents the boundary for transcriptome-wide significance (4.58 × 10⁻⁷). The bottom figure is the GWAS Manhattan plot where each point is the result of a SNP association test with PrCa risk. The red line corresponds to the traditional genome-wide significant boundary (5 × 10⁻⁸)

Fig. 3

Predicted expression of MLPH explains majority of GWAS signal at its genomic region. Each point corresponds to the association between SNP and PrCa status. Gray points indicate the marginal association of a SNP with PrCa status (i.e., GWAS association). Green points indicate the association of the same SNPs with PrCa after conditioning on predicted expression of MLPH using models trained from normal prostate (GTEx) and tumor prostate (TCGA). The dashed gray line corresponds to the genome-wide significant threshold (i.e., P = 5 × 10⁻⁸). MLPH was discussed in previous works as a possible susceptibility gene for PrCa. Association between total expression of MLPH and PrCa risk was transcriptome-wide significant in normal and tumor prostate tissue

Fig. 4

Average TWAS association statistics for genes predicted in each expression panel. Each bar plot corresponds to the average TWAS association statistic using all gene models from a given expression reference panel. Lines represent 1 standard-deviation estimated using the median absolute deviation under normality assumptions. Normal and tumor prostate tissues are marked in green. All other tissues are marked in gray