Transcriptome-wide association analysis identifies candidate susceptibility genes for prostate-specific antigen levels in men without prostate cancer

Summary

Deciphering the genetic basis of prostate-specific antigen (PSA) levels may improve their utility for prostate cancer (PCa) screening. Using genome-wide association study (GWAS) summary statistics from 95,768 PCa-free men, we conducted a transcriptome-wide association study (TWAS) to examine impacts of genetically predicted gene expression on PSA. Analyses identified 41 statistically significant (p < 0.05/12,192 = 4.10 × 10⁻⁶) associations in whole blood and 39 statistically significant (p < 0.05/13,844 = 3.61 × 10⁻⁶) associations in prostate tissue, with 18 genes associated in both tissues. Cross-tissue analyses identified 155 statistically significantly (p < 0.05/22,249 = 2.25 × 10⁻⁶) genes. Out of 173 unique PSA-associated genes across analyses, we replicated 151 (87.3%) in a TWAS of 209,318 PCa-free individuals from the Million Veteran Program. Based on conditional analyses, we found 20 genes (11 single tissue, nine cross-tissue) that were associated with PSA levels in the discovery TWAS that were not attributable to a lead variant from a GWAS. Ten of these 20 genes replicated, and two of the replicated genes had colocalization probability of >0.5: CCNA2 and HIST1H2BN. Six of the 20 identified genes are not known to impact PCa risk. Fine-mapping based on whole blood and prostate tissue revealed five protein-coding genes with evidence of causal relationships with PSA levels. Of these five genes, four exhibited evidence of colocalization and one was conditionally independent of previous GWAS findings. These results yield hypotheses that should be further explored to improve understanding of genetic factors underlying PSA levels.

PSA levels, commonly used to screen for PCa risk, may be refined by identifying genes that influence normal PSA variation. Chen and Dong et al. investigated genes whose expression levels are associated with PSA levels in cancer-free men, yielding new hypotheses for genetic factors underlying PSA.

Introduction

Prostate-specific antigen (PSA) is a serine protease of the human tissue kallikrein-related (KLK) peptidase family.¹ Serum levels are commonly used as a biomarker for detection, monitoring, and risk stratification of prostate cancer (PCa).^2,3,4,5 A small portion of the highly abundant PSA in the prostate is released into blood and elevated in men with PCa. This process is incompletely understood, and disruption of prostate gland architecture by neoplastic transformation has been speculated as a possible mechanism.^2,4,6 PSA levels can additionally be influenced by age, ethnicity, body mass index, infection, prostate volume, benign prostatic hyperplasia (BPH), and germline genetic variation.^7,8

PSA screening has been used for more than 30 years for the detection of PCa, which is the second leading cause of cancer death among men in the United States.^2,9,10,11,12 However, low test specificity and discrimination have complicated the use and interpretation of PSA as a screening tool.^11,13 Long-term outcomes data from large, population-based, randomized PSA screening trials show that PSA screening significantly reduces deaths from PCa and also results in the considerable overdiagnosis of low-risk disease.^14,15 PSA screening might be improved were it to account for variation in PSA levels that is attributable to germline genetics rather than PCa.^16,17,18 Twin studies and genome-wide association studies (GWASs) have estimated that 30%–45% of serum PSA variation is influenced by hereditary factors.^19,20 The most recent GWAS meta-analysis from our group identified 318 independent PSA-associated variants, which explained approximately 9% of genetic variation in PSA levels.¹⁸

Additional genetic variation in PSA levels may be determined by analyses of gene expression, as opposed to individual variants. Such work could lead to the identification of regions or biological pathways that affect PSA levels and point to mechanisms underlying constitutional increases in PSA levels in the absence of carcinogenesis. Transcriptome-wide association studies (TWASs) allow for the identification of genes whose cis-regulated expression is associated with complex polygenic traits.^21,22,23,24 To identify genes associated with PSA levels and prioritize them for functional investigation, we performed a TWAS of PSA levels based on summary statistics from a GWAS meta-analysis of 95,768 men without PCa.^16,25 In doing so, we aimed to identify genes located at loci where prior GWAS signals fell below genome-wide significance. In further implementing colocalization analyses, we were able to identify variants that influence both complex traits and gene expression.²⁶ We additionally implemented conditional analyses, fine-mapping, colocalization analyses, and pathway analyses to yield comprehensive insights into genes’ relationships with PSA levels.

Subjects and methods

Ethical Considerations

Informed consent was obtained from all study participants. The UK Biobank received ethics approval from the Research Ethics Committee (reference: 11/NW/0382) in accordance with the UK Biobank Ethics and Governance Framework. The research was conducted with approved access to UK Biobank data under application number 14105. We used PSA GWAS results from Kaiser Permanente's Genetic Epidemiology Research on Adult Health and Aging (GERA) cohort; these results were previously published by Hoffmann et al.¹⁷ The original study was approved by the Kaiser Permanente Northern California Institutional Review Board and the University of California, San Francisco Human Research Protection Program Committee on Human Research. The Prostate, Lung, Colorectal, and Ovarian (PLCO) Cancer Screening Trial was approved by the Institutional Review Board at each participating center and the National Cancer Institute. The informed consent document signed by PLCO study participants allows use of these data by investigators for discovery and hypothesis generation in the investigation of the genetic contributions to cancer and other adult diseases. Our study includes publicly posted genomic summary results from the PLCO Atlas. (add reference: Machiela MJ, et al. GWAS Explorer: an open-source tool to explore, visualize, and access GWAS summary statistics in the PLCO Atlas. Sci. Data. 2023;10:25.) No institutional review board review is required for PLCO summary data use. The Vanderbilt University Medical Center Institutional Review Board approved the BioVU study. The Malmö Diet and Cancer Study (MDCS) was approved by the local ethics committee. The Veteran Affairs Central Institutional Review Board approved the Million Veteran Program (MVP).

Discovery populations

Existing GWAS summary statistics¹⁶ from 95,768 PCa-free men from the following five study populations were included: UK Biobank, GERA, PLCO, BioVU, and MDCS. The men included in these analyses were restricted to PCa-free individuals (i.e., without a PCa diagnosis or history of prostate resection, where information was available) with at least one PSA measurement. Analyses were restricted to men with PSA values between 0.01 ng/mL and 10 ng/mL to capture variations in PSA levels, while minimizing potential for including undiagnosed PCa cases. The upper threshold of 10 ng/mL corresponds to pre-treatment PSA levels for low-risk PCa under the validated D’Amico risk classification system for PCa.^27,28 The lower bound of 0.01 ng/mL ensured that all participants had a functioning prostate and corresponds to the analytical limit of detection for early generations of PSA assays, many of which were used in the studies included in this analysis.^29,30 Median PSA levels were used for individuals with multiple PSA measurements available, with the exception of PLCO, which used PSA levels at the start of follow-up of the trial. Additional details about the discovery cohorts are described in detail in Kachuri et al.¹⁶ and Table S1.

Quality control and GWAS meta-analysis

Genotyping, imputation, study-specific quality control of genetic data, and the GWAS meta-analysis have been previously described in Kachuri et al.¹⁶ Briefly, ancestry and study-specific GWAS analyses used linear regression of log(PSA) on genetic variants, age, and the first 10 genetic ancestry principal components. After results across studies were meta-analyzed within ancestral groups, ancestry-specific summary statistics from European ancestry (n = 85,824), African ancestry (n = 3,509), East Asian ancestry (n = 3,337), and Hispanic/Latino (n = 3,098) individuals were meta-analyzed to generate multi-ancestry summary statistics.

MetaXcan transcriptome-wide gene-based analysis

We undertook a TWAS using the MetaXcan approach, which directly estimates Z scores for associations between gene expression levels and PSA levels using meta-analyzed GWAS summary statistics with tissue-specific prediction models. Multivariate adaptive shrinkage (mashr) prediction models were trained on GTEx (version 8) expression quantitative trait loci (eQTL) data for whole blood and, separately, prostate tissue. The covariances and weights for models trained on individual tissue types were from the PredictDB Data Repository (https://predictdb.org/).²⁵ We also undertook a cross-tissue analysis across 45 GTEx tissues to identify potential associations with genes that may have been missed in single-tissue analyses. Tissue types found primarily or exclusively in females (mammary breast, ovary, uterus, and vagina) were excluded from the multiple tissue analysis. A Bonferroni correction was implemented based on the number of genes tested (whole blood: 0.05/12,192 = 4.10 × 10⁻⁶; prostate: 0.05/13,844 = 3.61 × 10⁻⁶; cross-tissue: 0.05/22,249 = 2.25 × 10⁻⁶). Primary analyses were based on the multi-ancestry GWAS summary statistics. Sensitivity analyses restricted to men of European ancestry.

Replication analyses

We performed replication analyses of the findings from our discovery TWAS by implementing the MetaXcan approach in 209,318 PCa-free men of European ancestry enrolled in the MVP who had at least one PSA measurement.^18,31 The MVP has been previously described. Briefly, it is an expansive and diverse biobank that was initiated in 2011 and contains linkages to electronic health record data from more than 50 veterans affairs medical centers. Based on GWAS summary statistics, we estimated gene expression level associations using the aforementioned GTEx v8 mashr prediction models.³² Replication analyses were conducted based on whole blood, prostate tissue, and cross-tissue. We used the following replication significance thresholds based on the number of significant genes from our discovery TWAS analyses: 0.05/41 = 1.21 × 10⁻³ for whole blood, 0.05/39 = 1.29 × 10⁻³ for prostate tissue, and 0.05/155 = 3.22 × 10⁻⁴ for cross-tissue.

Conditional analyses ascertaining independence of TWAS and previous GWAS findings

Methods that assess the conditional fit of eQTLs on GWAS hits are lacking. Thus, to identify the set of TWAS genes for PSA that were not previously identified by GWAS, we first limited consideration to significant TWAS genes that did not contain genome-wide significant (p < 5 × 10⁻⁸) variants from the Kachuri et al.¹⁶ GWAS (i.e., within exact gene boundaries). Second, for the remaining genes, we performed conditional analyses using genome-wide complex trait analysis conditional and joint multiple-SNP analysis (COJO)³³ that simultaneously modeled eQTLs for a given TWAS-identified gene and GWAS results. For individual tissue analyses, eQTLs used in the prediction models of transcriptome-wide significant genes were selected from the tissue-specific PredictDB TWAS eQTL weight files for each gene. For the cross-tissue analyses, the maximum noncollinear eQTLs (variance inflation factor of <8 within a window size of 50 kb) were extracted from all 45 GTEx tissues (i.e., to remove variants in linkage disequilibrium [LD]). COJO was performed using LD reference panels from 10,000 European ancestry UK Biobank participants (as 89.6% of our population was European ancestry).¹⁶ GWAS summary statistics were then conditioned on the eQTLs for one gene at a time. If the set of genome-wide significant variants was reduced after conditioning on eQTLs used to predict gene expression, then the eQTLs and genome-wide significant variants were assumed to be in LD and thus not independent. However, if the set of genome-wide significant variants remained unchanged after conditional analyses, then we considered the TWAS gene to be independent of GWAS variants from Kachuri et al.¹⁶

Probabilistic fine-mapping

Similar to GWASs, marginal TWAS associations cannot be interpreted as causal. Identifying the most likely causal genes requires accounting for LD and eQTL sharing among genes in cis-regions. We performed genome-wide statistical fine-mapping for whole blood and prostate tissue using causal TWAS (cTWAS),³⁴ which jointly models the effects of imputed gene expression and genetic variants to derive posterior inclusion probabilities (PIPs). Cross-tissue fine-mapping is currently unavailable within the cTWAS analysis framework. Standard cTWAS pre-processing included restriction to protein-coding genes. LD was estimated from UKB reference data in individuals of European ancestry. Prior parameters were estimated for gene and variant effects in each genomic block using an empirical Bayesian approach, and Sum of Single Effects regression with summary statistics fine-mapping was applied to all variables after parameter estimation.^35,36 A random subset of 10% of all variants was used during initial parameter estimation assuming at most five causal effects for each region, but regions with PIPs of 0.8 or greater were recomputed. PIPs were calculated for all genes, all variants in regions with strong gene signals, and 10% of variants in other regions. A standard threshold of PIP of more than 0.8 was used for declaring statistically significant genes.³⁴

Colocalization analyses

To further investigate whether genes and variants exhibited shared genetic signals, we used fastENLOC software (version 2) to undertake colocalization analyses of eQTLs across multiple tissues for all genes with GTEx v8 prediction models and all variants investigated in the Kachuri et al. GWAS.^16,26,37 Colocalization was performed across tissues rather than complete individual-tissue colocalization analyses. Variant-specific enrichment priors were directly estimated by applying an established multiple imputation procedure to GWAS summary statistics and eQTL data.^37,38 Unlike traditional window-based approaches, fastENLOC uses Bayesian credible sets—clusters of variants in LD that share common underlying association signals—as regional LD blocks for colocalization analyses. These independent association signals, which were inferred from multi-variant fine-mapping analyses, were used to compute a colocalization probability for each LD region. For these analyses, LD blocks were based on 1000 Genome Project European ancestry individuals.³⁹ The GWAS summary statistics were colocalized for 49 GTEx version 8 tissues using pre-computed GTEx multiple-tissue annotations. Signal-level results returned regional-level colocalization probabilities (RCPs) between eQTL and GWAS signals, which sum up the variant-level colocalization probabilities of correlated variants within an LD block that harbors a single GWAS association signal. Gene results were based on gene-level RCPs (GRCPs), which represent the probability that a candidate gene contains at least one colocalized variant.²⁶ RCP and GRCP values of greater than 0.5 indicate strong evidence of shared genetic signals between eQTLs and GWAS variants.²⁶

Pathway enrichment analysis

To explore the potential biological relevance of PSA-associated genes, we applied the Enrichr tool to all significant genes identified in the whole blood, prostate, and cross-tissue analyses to assess enrichment against three gene set libraries: Kyoto Encyclopedia of Genes and Genomes (KEGG) 2021 human; Gene Ontology (GO) biological process 2021; and GO molecular function 2021.^40,41 Enrichment was assessed by multiplying the p value from a Fisher’s exact test with the Z score of the deviation from the expected rank.^42,43 Pathways with a Benjamini-Hochberg-corrected p value of <0.05 were considered statistically significantly enriched.

Results

Using MetaXcan and GWAS summary statistics based on 95,768 individuals (Table S1), TWASs were conducted for whole blood, prostate, and cross-tissue matrices. The sample was primarily European ancestry (n = 85,824), although African ancestry (n = 3,509), East Asian ancestry (n = 3,337), and Hispanic/Latino (n = 3,098) individuals were also included. The median PSA value across cohorts was 2.35 ng/mL.

Overall, 93% of the GTEx variants used to predict gene expression in individual tissue models were available in our multi-ancestry GWAS summary statistics that included 20,060,440 variants. GTEx gene expression models for whole blood were derived from n = 670 participants with genotype data, and samples from n = 221 individuals were used to train prostate tissue prediction models.⁴⁴ In the TWAS based on whole blood gene expression models, 41 of 12,192 genes were associated with PSA levels at the Bonferroni-corrected threshold (Figure 1 and Table S2). In the TWAS based on prostate tissue gene expression models, 39 of 13,844 genes were statistically significantly associated with PSA levels (Figure 2 and Table S3). Cross-tissue analyses identified 155 of 22,249 genes whose expression was associated with elevated PSA levels (Figure 3 and Table S4). In total, we replicated 151 of 173 unique associations across the three individual tissue analyses in the MVP cohort. Sensitivity analyses restricting to European ancestry individuals were largely comparable with our primary TWAS results (Tables S2–S4).

Figure 1.

Miami plot for the discovery TWAS of PSA levels based on predicted gene expression in whole blood

Each dot represents a gene, and the red dashed lines represent the significance threshold after Bonferroni correction (4.10 × 10⁻⁶). All statistically-significant genes are annotated; the genes independent of results from the largest prior GWAS are highlighted in yellow. Genes on the top half of the plot are positively associated with PSA levels, and the genes on the lower half of the plot are inversely associated with PSA levels.

Figure 2.

Miami plot for the discovery TWAS of PSA levels based on predicted gene expression in prostate tissue

Each dot represents a gene, and the red dashed lines represent the significance threshold after Bonferroni correction (3.61 × 10⁻⁶). All statistically significant genes are annotated; the genes independent of results from the largest prior GWAS are highlighted in yellow. Genes on the top half of the plot are positively associated with PSA levels, and the genes on the lower half of the plot are inversely associated with PSA levels.

Figure 3.

Manhattan plot for the discovery TWAS of PSA levels based on cross-tissue predicted gene expression

Each dot represents a gene, and the red dashed line represents the significance threshold after Bonferroni correction (2.25 × 10⁻⁶). The top 30 genes with the smallest association p values are annotated and colored blue. Directions of association between genes and PSA levels are not available for cross-tissue analyses due to the joint tissue inference methodology used in S-MultiXcan.

Whole blood TWAS

Of the 41 genes whose whole blood expression was significantly associated with PSA levels (Figure 1 and Table S2), the expression of 20 genes was positively associated with elevated PSA and the expression of 21 genes was inversely associated. Replication analyses in the MVP cohort confirmed 33 of 41 genes in whole blood, and all genes demonstrated consistent direction of effect. Two of the significant (and replicated) genes are located at 19q13.33, which contains KLK3, the gene most strongly linked to serum PSA levels.⁴⁵ While a prediction model for KLK3 was not available in whole blood, a member of the same gene family at 19q13.33, KLK2, was the gene most strongly associated with PSA levels (p = 1.28 × 10⁻⁶²).⁴⁵ Notably, KLK2 and KLK3 are located at the same genomic locus and are known to have high LD.⁷ Twenty-four of the 41 genes did not contain genome-wide significant variants and had not been annotated in previous PSA GWASs. Conditional analyses identified 7 of the 41 TWAS genes to be independent of previously published PSA GWAS findings (Table 1).¹⁶ Of these, increased expression of six genes was associated with elevated PSA levels in PCa-free men: GPBP1L1 (1p34.1, p = 2.21 × 10⁻⁷); TMEM69 (1p34.1, p = 2.21 × 10⁻⁷); UQCRH (1p33, p = 4.76 × 10⁻⁷); ACTRT3 (3q26.2, p = 5.90 × 10⁻⁷); EXOSC9 (4q27, p = 1.80 × 10⁻⁶); and CCNA2 (4q27, p = 7.80 × 10⁻⁷). Decreased expression of ITH4 was associated with increased PSA levels (3p21.1, p = 3.65 × 10⁻⁶).

Table 1.

Conditionally independent genes associated with prostate-specific antigen levels in transcriptome-wide association studies based on whole blood, prostate tissue, and cross-tissue predicted gene expression

Ensembl gene IDa	Gene name	Chr	Bandb	Analysisc	Z	Beta	Disc pvalue	Rep pvalue
ENSG00000132128	LRRC41	1	p33	prostate	5.04	0.29	4.76 × 10−7	9.39 × 10−3
ENSG00000159592	GPBP1L1	1	p34.1	whole blood	5.18	17.92	2.21 × 10−7	1.81 × 10−3
				cross-tissued	−1.92	N/A	2.08 × 10−8	4.28 × 10−8
ENSG00000159596	TMEM69	1	p34.1	whole blood	5.18	0.28	2.21 × 10−7	1.81 × 10−3
ENSG00000173660	UQCRH	1	p33	whole blood	5.04	8.04	4.76 × 10−7	9.39 × 10−3
ENSG00000070831	CDC42	1	p36.12	cross-tissue	−1.55	N/A	6.24 × 10−7	1.01 × 10−3
ENSG00000184378	ACTRT3	3	q26.2	whole bloodd	4.99	0.25	5.90 × 10−7	7.81 × 10−11
ENSG00000055955	ITIH4	3	p21.1	whole bloodd	−4.63	−0.03	3.65 × 10−6	2.75 × 10−8
				prostated	−4.69	−0.03	2.69 × 10−6	8.25 × 10−8
ENSG00000163950	SLBP	4	p16.3	cross-tissue	−2.15	N/A	8.40 × 10−8	4.23 × 10−3
ENSG00000145386	CCNA2	4	q27	whole bloodd	4.94	0.61	7.80 × 10−7	3.93 × 10−7
				cross-tissued	−1.81	N/A	1.02 × 10−6	6.89 × 10−6
ENSG00000123737	EXOSC9	4	q27	whole blood	4.77	0.09	1.80 × 10−6	2.07 × 10−2
				prostate	−4.86	−0.42	1.18 × 10−6	1.87 × 10−2
				cross-tissue	0.36	N/A	8.60 × 10−7	4.04 × 10−5
ENSG00000233822	HIST1H2BN	6	p22.1	prostated	5.35	0.24	8.86 × 10−8	3.58 × 10−18
ENSG00000272447	RP11-182L21.6	10	q22.3	cross-tissue	1.25	N/A	1.54 × 10−7	5.19 × 10−2
ENSG00000264230	ANXA8L1	10	q11.22	cross-tissue	−0.46	N/A	2.17 × 10−7	1.50 × 10−2
ENSG00000165891	E2F7	12	q21.2	cross-tissued	1.45	N/A	8.54 × 10−9	1.05 × 10−13
ENSG00000188277	C15orf62	15	q15.1	cross-tissue	−0.85	N/A	2.01 × 10−7	3.56 × 10−1
ENSG00000259359	RP11-327J17.2	15	q26.2	cross-tissue	−0.16	N/A	5.82 × 10−7	3.71 × 10−2
ENSG00000155719	OTOA	16	p12.2	prostate	−4.82	−5.36	1.44 × 10−6	2.01 × 10−1
ENSG00000197445	C16orf47	16	q22.3	cross-tissued	0.44	N/A	1.11 × 10−7	1.24 × 10−12
ENSG00000141052	MYOCD	17	p12	cross-tissued	−3.03	N/A	1.20 × 10−7	8.40 × 10−7
ENSG00000181513	ACBD4	17	q21.31	cross-tissued	−1.05	N/A	2.17 × 10−6	6.57 × 10−7
ENSG00000186834	HEXIM1	17	q21.31	prostated	5.03	0.12	4.95 × 10−7	3.80 × 10−3

Chr, chromosome; Disc, discovery; Rep, replication.

^aThe genes included in this table were significant (Bonferroni-adjusted p < 0.05) in at least one TWAS and, based on conditional analyses, independent from the largest prior GWAS of PSA levels.

^bGRCh38.

^cIndicates the tissue(s) in which predicted gene expression was significantly associated with PSA levels.

^dReached Bonferroni significance in tissue-specific replication analyses.

Fine-mapping analyses in whole blood were based on 10,106 protein-coding genes. We identified 14 genes with a PIP of 0.8 or greater (Table S5), of which two overlapped with Bonferroni-significant genes in the whole blood TWAS discovery analyses: PMVK (1q21.3) and A2M (12p13.31).

Prostate tissue TWAS

In prostate tissue, increased expression of 18 and decreased expression of 21 genes was associated with elevated PSA levels (Figure 2 and Table S3). Thirty-one of 39 genes in prostate tissue replicated in the MVP cohort, and effect sizes for all were in the same direction as those reported in the discovery analyses. No gene prediction models were available for KLK3 or any other genes in the KLK family in prostate tissue. Instead, MMP7 on 11q22 was most strongly associated with PSA levels (p = 2.78 × 10⁻¹⁸). We identified 18 PSA-associated TWAS genes containing known genome-wide significant variants. Of the remaining 21 significant genes, six were conditionally independent of previous GWAS results after conditioning on the eQTL variants used in gene prediction models (Table 1). Of these six, increased expression of three genes was associated with elevated PSA levels: LRRC41 (1p33, p = 4.76 × 10⁻⁷), HIST1H2BN (6p22.1, p = 8.86 × 10⁻⁸), and HEXIM1 (17q21.31, p = 4.95 × 10⁻⁷). Decreased expression of the remaining three genes was associated with increased PSA levels: ITIH4 (3p21.1, p = 2.69 × 10⁻⁶), EXOSC9 (4q27 p = 1.18 × 10⁻⁶), and OTOA (16p12.2, p = 1.44 × 10⁻⁶).

Using 10,806 protein-coding gene expression models in prostate tissue, cTWAS identified 16 genes with a PIP of 0.8 or greater (Table S6), three of which were Bonferroni-significant in main analyses: ADD3 (10q25.1), MMP7 (11q22.2), and OTOA (16p12.2). As specified above, OTOA was also identified by conditional analyses.

Cross-tissue TWAS

In cross-tissue analyses, 136 of 155 genes had expression levels that were associated with elevated PSA levels in our replication cohort. One additional gene, TMEM69, which was discovered in the whole blood analyses and did not reach statistical significance in the cross-tissue discovery analyses, did not replicate in whole blood but was significantly associated with elevated PSA levels in cross-tissue replication analyses (p = 1.948 × 10⁻⁸). Cross-tissue analyses do not yield directionality of effects, so we were unable to compare effect directions between discovery and replication analyses. Among the 155 statistically significant genes from cross-tissue analyses, eight, all of which replicated, are located on 19q13.33–19q13.41 and are part of the expanded human KLK gene family (Figure 3 and Table S4).

We found 71 genes that do not contain genome-wide significant PSA variants within their gene boundaries. Conditional analyses identified 13 genes from this set that were independent of known genome-wide significant variants: CDC42 (1p36.12, p = 6.24 × 10⁻⁷), GPBP1L1 (1p34.1, p = 2.08 × 10⁻⁸), SLBP (4p16.3, p = 8.40 × 10⁻⁸), EXOSC9 (4q27, p = 8.60 × 10⁻⁷), CCNA2 (4q27, p = 1.02 × 10⁻⁶), ANXA8L1 (10q11.22, p = 2.17 × 10⁻⁷), RP11-182L21.6 (10q22.3, p = 1.54 × 10⁻⁷), E2F7 (12q21.2, p = 8.54 × 10⁻⁹), C15orf62 (15q15.1, p = 2.01 × 10⁻⁷), RP11-327J17.2 (15q26.2, p = 5.82 × 10⁻⁷), C16orf47 (16q22.3, p = 1.11 × 10⁻⁷), MYOCD (17p12, p = 1.20 × 10⁻⁷), and ABCD4 (17q21.31, p = 2.17 × 10⁻⁶) (Table 1).

Colocalization analysis of cross-tissue association signals

Cross-tissue colocalization analyses identified 372 signals with an RCP of more than 0.5 at the variant level (Table S7) and 307 genes with a GRCP of more than 0.5 at the gene level (Table S8). The top 10 genes with the largest GRCP were OTX1 (2p15), MAFF (22q13.1), FUT2 (19q13.33), EMSY (11q13.5), IFT80 (3q25.33), EXOSC9 (4q27), SMC4 (3q25.33), RCAN3 (1p36.11), PBXIP1 (1q21.3), and PMVK (1q21.3).

Among 155 significant genes from the cross-tissue TWAS, locus-level results identified 50 genes with strong evidence of colocalization (Table 2). Signal-level results identified 66 regions in 51 genes (Table S9). Colocalization analyses identified 50 genes from cross-tissue analyses that shared genetic signals with GWAS variants, five of which were significantly associated with PSA levels in cross-tissue TWAS analyses.

Table 2.

Gene level colocalization results for significant genes from the transcriptome-wide association study of prostate-specific antigen levels based on cross-tissue predicted gene expression

Ensembl gene ID	Gene name	Chr	Cytobanda	GRCP	GLCP
ENSG00000069275	NUCKS1	1	q32.1	0.97	2.55
ENSG00000117280	RAB29	1	q32.1	1.13	0.52
ENSG00000117602	RCAN3	1	p36.11	2.13	−24.54
ENSG00000143603	KCNN3	1	q21.3	1.42	−136.50
ENSG00000158715	SLC45A3	1	q32.1	0.59	15.17
ENSG00000163344	PMVK	1	q21.3	1.97	−50.77
ENSG00000163346	PBXIP1	1	q21.3	1.99	−14.21
ENSG00000270361	RP11-307C12.13	1	q21.3	1.08	−0.32
ENSG00000115507	OTX1	2	p15	4.91	−7.33
ENSG00000143869	GDF7	2	p24.1	1.00	620.30
ENSG00000068885	IFT80	3	q25.33	2.41	24.05
ENSG00000072858	SIDT1	3	q13.2	0.62	−211.30
ENSG00000113810	SMC4	3	q25.33	2.18	−52.64
ENSG00000132394	EEFSEC	3	q21.3	1.34	−160.40
ENSG00000175792	RUVBL1	3	q21.3	1.24	−72.12
ENSG00000177311	ZBTB38	3	q23	1.22	−39.66
ENSG00000186432	KPNA4	3	q25.33	1.27	−19.18
ENSG00000213186	TRIM59	3	q25.33	1.54	−38.26
ENSG00000123737	EXOSC9b	4	q27	2.24	−108.20
ENSG00000145386	CCNA2b	4	q27	1.39	−54.29
ENSG00000049656	CLPTM1L	5	p15.33	0.71	−27.67
ENSG00000164362	TERT	5	p15.33	0.72	4.54
ENSG00000184357	HIST1H1B	6	p22.1	0.92	2.70
ENSG00000189298	ZKSCAN3	6	p22.1	0.84	1.28
ENSG00000204713	TRIM27	6	p22.1	0.76	−12.84
ENSG00000233822	HIST1H2BNb	6	p22.1	1.08	−9.88
ENSG00000281831	HCP5B	6	p22.1	0.88	1.25
ENSG00000164684	ZNF704	8	q21.13	1.00	−94.45
ENSG00000167034	NKX3-1	8	p21.2	0.96	−293.10
ENSG00000056558	TRAF1	9	q33.2	0.78	5,137.00
ENSG00000095261	PSMD5	9	q33.2	0.61	−14.25
ENSG00000119403	PHF19	9	q33.2	0.97	−23.73
ENSG00000130956	HABP4	9	q22.32	0.82	30.20
ENSG00000130958	SLC35D2	9	q22.32	1.22	−11.44
ENSG00000165244	ZNF367	9	q22.32	0.60	−10.13
ENSG00000148700	ADD3	10	q25.1	1.21	−118.60
ENSG00000272447	RP11-182L21.6b	10	q22.3	0.99	5.94
ENSG00000137673	MMP7	11	q22.2	0.96	183.70
ENSG00000158636	EMSY	11	q13.5	2.64	−9.56
ENSG00000177951	BET1L	11	p15.5	0.78	−63.92
ENSG00000111707	SUDS3	12	q24.23	1.33	−225.60
ENSG00000175899	A2M	12	p13.31	0.64	−156.60
ENSG00000172766	NAA16	13	q14.11	1.25	−14.31
ENSG00000259359	RP11-327J17.2b	15	q26.2	1.10	2.93
ENSG00000141052	MYOCD	17	p12	0.62	−1.94
ENSG00000275410	HNF1B	17	q12	0.81	2.32
ENSG00000176920	FUT2	19	q13.33	3.04	320.20
ENSG00000177045	SIX5	19	q13.32	1.87	−15.39
ENSG00000185800	DMWD	19	q13.32	1.00	461.40
ENSG00000185022	MAFF	22	q13.1	3.63	3.06

Chr, chromosome; GLCP, gene-level colocalization probability; GRCP, gene-level variant colocalization probability.

^aGRCh38.

^bThese genes were independent from the largest prior GWAS of PSA levels based on COJO analyses.

When comparing the overlap between transcriptome-wide significant genes and results from colocalization analyses, we identified five genes (EXOSC9, CCNA2, HIST1H2BN, RP11-182L21.6, and RP11-327J17.2) with GRCP and RCP values of greater than 0.5 that were not attributable to a known lead variant from GWAS. Two of these genes are long non-coding RNA, and the remaining three are protein-coding genes. CCNA2 and HIST1H2BN had expression levels that were associated with PSA levels in replication analyses. Four genes identified through tissue-specific fine-mapping analyses also exhibited strong evidence of colocalization: A2M and PVMK in whole blood and MMP7 and ADD3 in prostate tissue.

Significant genes overlapping across TWAS

Of the 41 genes that were significantly associated with PSA levels in the whole blood TWAS, 34 were also imputed in the prostate tissue TWAS. Thirty of 39 significantly associated genes from the prostate tissue TWAS were also evaluated in whole blood analyses. Of the 155 significantly associated genes detected in the cross-tissue analysis, 89 and 105 genes had transcriptome prediction models available for whole blood and prostate tissue, respectively.

We identified 13 genes whose predicted expression was associated with elevated PSA levels across all three TWASs: RAB29 (1q32.1), LDAH (2p24.1), ZFP36L2 (2p21), SIDT1 (3q13.2), TRIM59 (3q25.33), EXOSC9 (4q27), HIBADH (7p15.2), TAX1BP1 (7p15.2), MEGF9 (9q33.2), ADD3 (10q25.1), APIP (11p13), LINC01569 (16p13.3), and EIF2S2 (20q11.22) (Figure 4).

Figure 4.

Statistically significant genes overlapping across transcriptome-wide association studies based on whole blood, prostate tissue, and cross-tissue predicted gene expression

Of the significant gene sets from whole blood (41 genes), prostate tissue (39 genes), and cross-tissue analyses (155 genes), 13 genes were significant across all three analyses, 36 genes were significant in exactly two analyses, and 124 genes were significant in a single analysis.

Of these genes, only EXOSC9 was conditionally independent after accounting for prior loci and demonstrated evidence of colocalization (described above). Five additional associations were not statistically significant in the cross-tissue analysis, but demonstrated significant signals in whole blood and prostate tissue. Seventeen significant genes were shared by only the whole blood and cross-tissue TWASs, six of which did not have prediction models in prostate tissue. Similarly, the prostate tissue and cross-tissue TWAS shared 14 significant genes, of which seven did not have prediction models in whole blood.

Pathway enrichment analysis

Using significantly associated signals from our main analyses, TWAS genes were significantly enriched in 18 pathways underlying molecular activities and biological processes present in KEGG and GO gene set databases (Table 3; Figure 5). Within the whole blood gene set, analyses detected nominally significant enrichment of the renin-angiotensin-aldosterone system in KEGG, which works to regulate arterial blood pressure. The observed signal was primarily driven by two genes in the KLK family and was only nominally significant. Transcriptional dysregulation in cancer was significantly enriched in KEGG for cross-tissue results. No pathways were significantly enriched in prostate gene sets in the KEGG catalog.

Table 3.

Results from pathway analyses based on all genes that were analyzed in the transcriptome-wide association studies of prostate-specific antigen levels based on whole blood, prostate tissue, and cross-tissue predicted gene expression

Database	Pathway	Overlapa	p value	Genes
Whole blood
KEGG	renin-angiotensin system	2/18	4.56 × 10−2	KLK1, KLK2
Cross-tissue
KEGG	transcriptional misregulation in cancer	8/48	1.95 × 10−3	SLC45A3, CDKN2C, TRAF1, TP53, HOXA10, HOXA9, H3C10, CCNA2
GO molecular function	histone deacetylase binding	8/120	4.48 × 10−4	NKX3-1, TFAP4, SUDS3, TP53, HOXA10, MYOCD, ZBTB7B, H1-5
GO molecular function	serine-type endopeptidase activity	9/120	6.58 × 10−4	KLK3, KLK2, KLK1, KLK11, KLK4, KLK5, KLK15, MMP7, KLK8
GO molecular function	serine-type peptidase activity	9/120	8.65 × 10−4	KLK3, KLK2, KLK1, KLK11, KLK4, KLK5, KLK15, MMP7, KLK8
GO molecular function	serine hydrolase activity	9/120	8.65 × 10−4	KLK3, KLK2, KLK1, KLK11, KLK4, KLK5, KLK15, MMP7, KLK8
GO molecular function	DNA-binding transcription activator activity, RNA polymerase II specific	12/120	5.69 × 10−3	NKX2-6, TBX3, EHF, TFAP4, OTX1, TP53, HOXA10, HOXA9, GLIS2, SIX5, ZBTB7B, MAFF
GO molecular function	DNA-binding transcription activator activity	12/120	5.69 × 10−3	NKX2-6, TBX3, EHF, TFAP4, OTX1, TP53, HOXA10, HOXA9, GLIS2, SIX5, ZBTB7B, MAFF
GO molecular function	DNA-binding transcription factor binding	11/120	2.18 × 10−3	NKX3-1, TBX3, NUCKS1, FAF1, ETS2, TP53, SDR16C5, MYOCD, PBXIP1, HMGA1, EXOSC9
GO molecular function	endopeptidase activity	10/120	2.18 × 10−2	KLK3, KLK2, KLK1, KLK11, KLK4, KLK5, KLK15, MMP7, ADAMTS18, KLK8
GO molecular function	RNA polymerase II-specific DNA-binding transcription factor binding	9/120	3.17 × 10−2	NKX3-1, TBX3, FAF1, ETS2, TP53, SDR16C5, MYOCD, HMGA1, EXOSC9
GO molecular function	structural constituent of chromatin	4/120	3.20 × 10−2	H3C10, H1-5, H2BC15, HMGA1
GO biological process	stem cell differentiation	11/121	5.06 × 10−3	TBX3, HNF1B, ZFP36L2, PHF19, DMRTA2, JARID2, FGFR2, TP53, MYOCD, CDC42, A2M
GO biological process	regulation of DNA replication	7/121	3.34 × 10−2	ZBTB38, NUCKS1, FAF1, RUVBL1, TP53, E2F7, CCNA2
GO biological process	regionalization	11/121	3.34 × 10−2	NKX3-1, TBX3, HNF1B, OTX1, DMRTA2, FGFR2, ETS2, TP53, HOXA10, HOXA9, SETDB2
GO biological process	negative regulation of cyclin-dependent protein serine/threonine kinase activity	4/121	3.34 × 10−2	CDK5RAP1, TFAP4, CDKN2C, MYOCD
GO biological process	negative regulation of cyclin-dependent protein kinase activity	4/121	3.34 × 10−2	CDK5RAP1, TFAP4, CDKN2C, MYOCD
GO biological process	positive regulation of histone modification	6/121	3.34 × 10−2	PHF19, JARID2, TP53, SDR16C5, ZBTB7B, H1-5

^aNumber of imputed genes in our analysis / number of genes in pathway.

Figure 5.

GO term enrichment results from the transcriptome-wide association study of prostate-specific antigen levels based on cross-tissue predicted gene expression

The GO molecular function and biological process data repositories were queried for pathway enrichment analyses across a background set of all genes used in our main analyses. The color gradient represents the magnitude of p values, with darker colors corresponding to smaller values. The size of the circles represents the number of genes in the pathway. "p.adjust" represents the p value adjusted for multiple testing.

Enrichment analyses within the GO biological process and molecular function gene set libraries identified significant enrichments only in the cross-tissue gene set, with 16 pathways. No pathways were significantly enriched in GO for whole blood or prostate tissue TWAS sets. We also observed enrichment in four pathways involving both single and double stranded DNA binding mechanisms and regulation by transcription factors, including RNA polymerase II, a core component of the DNA transcription machinery. Moreover, two pathways underlying extracellular matrix and cellular component disassembly were enriched in the cross-tissue gene set. We detected significant enrichment in GO biological pathways underlying the regulation of systemic arterial blood pressure, which was also observed for whole blood TWAS using KEGG.

Discussion

This TWAS of 95,768 PCa-free men identified 173 unique genetically predicted transcripts whose expression levels were significantly associated with PSA. We replicated 151 unique gene expression signals in analyses across all three tissues, and all associations had effect estimates that were in the same direction as those in the discovery analyses. Conditional analyses identified 20 unique susceptibility genes for PSA that did not correspond with a known GWAS signal, 10 of which replicated in the original tissue. Fine-mapping analyses for whole blood and prostate tissue identified five protein-coding genes for PSA levels, four of which had strong evidence of colocalization across multiple tissues: MMP7, ADD3, A2M, and PVMK. Colocalization analyses further highlighted five of the 20 genes from conditional analyses: EXOSC9 and CCNA2 at 4q27, HIST1H2BN at 6p22.1, RP11-182L21.6 at 10q22.3, and RP11-327J17.2 at 15q26.2. Two of these genes (CCNA2 and HIST1H2BN) replicated in the same tissue matrix as the discovery analysis in which they were originally identified. Pathway enrichment analyses across three catalogs implicated regulatory pathways related to transcription, cell signaling, and disassembly of cellular and noncellular components.

The colocalized genes introduce interesting hypotheses regarding the genetic mechanisms regulating PSA production. Of particular interest are the signals at the 4q27 locus (CCNA2 and EXOSC9), a genomic region that was not detected in previous GWASs of PSA.^16,17 Researchers have previously investigated an autoimmune-related block on the 4q27 locus and did not find an association with PCa risk.^46,47,48,49 Signals at CCNA2, which were Bonferroni-significant in whole blood and cross-tissue matrices, replicated in additional TWAS analyses for both tissues. Literature documenting a relationship between CCNA2 and PSA is limited, although CCNA2 was previously tested as a candidate autoantibody signature marker for distinguishing PCa from BPH in patients with elevated serum PSA;⁵⁰ it was not determined to be a top antibody signature of any specific PCa targets. Gene co-expression network analyses have linked CCNA2 to biochemical recurrence and survival in men with PCa.^51,52 Further exploration of CCNA2 expression in individuals without PCa may unveil biological pathways that influence PSA levels. EXOSC9, which encodes a core protein involved in the RNA degradation machinery in humans, was also identified in conditional and colocalization analyses as a putative gene candidate for PSA levels. EXOSC9 was identified across all three discovery TWAS analyses, though replication analyses did not yield any significant associations in individual tissue or cross-tissue matrices. Currently, there is no literature supporting an association between EXOSC9 and PSA levels or PCa risk. Genetically predicted expression of EXOSC9 in whole blood was positively associated with PSA levels, while the opposite direction of effect was observed for expression in prostate tissue. There exists limited tissue-specific knowledge of the physiological and pathological function of EXOSC9 in non-cancer cell lines.^48,49 It is worth noting that fine-mapping yielded PIP values lower than 0.8 at the 4q27 locus, suggesting that gene expression of CCNA2 and EXOSC9 in individual tissues may not be causally linked to PSA levels; the association between pooled gene expression and PSA levels is unknown.

Colocalized gene HIST1H2BN was identified in prostate tissue discovery and replication analyses, but not in fine-mapping results. HIST1H2BN encodes a component of a core nucleosome histone and has been linked to PCa cell growth and epithelial-mesenchymal transition through upregulated nuclear factor κB (NF-κB)/Rel expression.⁵³ Activation of the NF-κB pathway can induce activation of the PSA promoter enhancer, even in the absence of androgens, and NF-κB can directly bind to the PSA enhancer in PCa cell lines.⁵⁴ To our knowledge, no experiments have been conducted to investigate the relationship between NF-κB and PSA in PCa-free populations.

PSA is used for both PCa detection and monitoring of PCa progression, making it difficult to disentangle the mechanisms underlying our observed associations. PSA-related screening bias may account for a portion of the observed relationship between gene expression and PSA levels. Genes may also exert pleiotropic effects on PCa and PSA through overlapping biological mechanisms.^2,11 Nevertheless, our restriction to PCa-free men for all analyses minimizes the potential for reverse causation and bolsters confidence that the observed genetic signals inform constitutive PSA levels. In addition, there was no overlap between genes reported in a recent PCa TWAS⁵⁵ that used prostate tissue gene expression models and the 20 genes we identified in our conditional analyses, nor for the five genes identified in our colocalization analyses. A comparison between the results of the PCa TWAS and our prostate tissue discovery analyses revealed that 14 of the 39 PSA-associated genes were also significantly associated with PCa risk. None of these genes were determined to be independent from previous GWAS findings identified in Kachuri et al.¹⁶

Fourteen of the 20 PSA-related genes identified from conditional analyses, spanning 12 genomic regions not implicated by prior PSA GWAS, have been associated with prostatic malignancies in prior literature: CDC42 (1p36.12), GPBP1L1 and TMEM69 (1p34.1), LRRC41 (1p33), ITIH4 (3p21.1), SLBP (4p16.3), CCNA2 (4q27), HIST1H2BN (6p22.1), ANXA8L1 (10q11.2), E2F7 (12q21.2), C15orf62 (15q15.1), OTOA (16p12.2), C16orf47 (16q22.3), and HEXIM1 (17q21.31).^{53,56,57,58,59,60,61,62,63,64,65} Prostate tissue fine-mapping also highlighted OTOA as a gene whose expression levels may be causal for PSA. OTOA has been identified as a testis-selective gene that shares sequence homology with cancer antigen mesothelin. The remaining six genes identified from conditional analyses have not been implicated in PCa susceptibility in gene function, experimental, or human population research: UQCRH (1p33), ACTRT3 (3q26.2), EXOSC9 (4q27), RP11-182L21.6 (10q22.3), RP11-327J17.2 (15q26.2), and ACBD4 (17q21.31). ACTRT3 and ACBD4 were significantly associated with PSA levels in replication analyses, but were not causally associated with PSA level in fine-mapping analyses. ACTRT3 is of particular interest, as it is critical in regulating sperm nucleus cytomorphology upstream of the processing of spermatid into mature motile sperm.⁶⁶ Increased expression of ACTRT3, which forms a testis-specific profilin III-ACTRT3 complex that facilitates male germ cell head cytomorphology and maintains sperm motility in animal models, was associated with elevated PSA levels in the TWAS based on whole blood.^66,67 Though there is no documented link between ACTRT3 and PSA levels or PCa risk, the pronounced role of ACTRT3 in mediating conformational changes in sperm nuclei suggests possible shared biological pathways between PSA and the production and processing of male germ cells.^66,68 Loci that map to the multigenic region that contains ACTRT3 have been linked to colorectal and bladder cancer susceptibility.^69,70

The function of ABCD4 is not well understood, although the literature suggests that ATPase activity of the protein it encodes plays an essential role in vitamin B12 (cobalamin) metabolism.⁷¹ ABCD4 is a member of the ATP-binding cassette transporter family that contributes to intracellular transport of B12 through interactions with LMBD1.^71,72 Elevated plasma levels of circulating B12 have been associated with an up to three-fold increased PCa risk,^73,74 although an analysis of B12 and changes in post-diagnostic PSA levels was not suggestive of an association.⁷⁵ Associations between serum levels of B12 and PSA levels have not been evaluated in PCa-free men. It must be noted that 17q21.31 includes a ∼900 kB chromosomal inversion polymorphism that is prevalent in approximately 20% of European ancestry individuals.⁷⁶ This chromosomal inversion is known to have a tissue-specific effect on the expression of multiple genes and is linked to gene expression differences in neurological disease and breast carcinogenesis.^77,78,79 As such, our results at this locus should be interpreted with caution.

Individual-tissue fine-mapping identified five protein-coding genes from our discovery TWAS that were causal for PSA levels. None of these genes have been evaluated in relation to PSA levels in PCa-free men, but MMP7 and A2M have been previously linked to PSA overall. Serum levels of MMP7, which were the strongest signals in our discovery prostate tissue TWAS, have been linked to large PSA declines in men with metastatic hormone-sensitive PCa who were undergoing chemotherapy treatment.⁸⁰ MMP7 was also significantly associated with PCa risk in a recent TWAS.⁵⁵ A2M, which was significantly and causally associated with PSA in whole blood, is also inversely correlated with PSA levels in men with PCa.⁸¹ The alpha-2-macroglobulin (A2M) protein functions as an enzymatic inhibitor of PSA in blood; PSA interacts with A2M and forms inactive PSA-A2M complexes.⁸² The remaining three genes identified by fine-mapping have not been linked to PSA or PCa. ADD3 encodes the cytoskeleton protein adducin and is located in a known tumor suppressor region for glioblastomas.⁸³ PMVK encodes the phosphomevalonate kinase protein, which catalyzes a key step in the mevalonate pathway, but its function with respect to cancer development and progression is unknown.^84,85 OTOA is primarily implicated in normal hearing and sound transduction.^86,87 Fusion gene annotations using FusionGDB 2.0⁸⁸ have found that OTOA can form a fusion gene with NCOA4 (FusionGDB2 ID: 61935). NCOA4 encodes an androgen receptor co-activator that potentially forms a tripartite complex with PSA and the androgen receptor, highlighting a potential mechanism through which OTOA can be linked to PSA levels.^89,90

In general, fine-mapping analyses for whole blood and prostate tissue yielded fewer signals than our primary TWAS, and only five of 30 genes with a PIP of greater than 0.8 overlapped with TWAS hits. cTWAS does not currently allow for evaluation of results from cross-tissue analyses, whereas the majority of signals from our discovery TWAS were from cross-tissue analyses. It is also possible that pooled gene expression across multiple tissue matrices exerts an effect on PSA levels in PCa-free men. Moreover, cTWAS has comparatively low power to detect associations compared to other methods for integrating GWAS data and eQTL results. Associations that are causal may not be captured by cTWAS due to the conservative significance threshold, particularly in gene sets with a low proportion of variance explained.³⁴ Additionally, preprocessing using the cTWAS analytical pipeline restricted gene expression models to protein-coding genes. Regardless, our fine-mapping results suggest that results from TWAS must be interpreted with caution.

Our study had several limitations. First, nearly 90% of men in the study population were of primarily European ancestry,¹⁶ and 85% of the tissue samples used to train the TWAS models were derived from European populations.⁴⁴ Investigations into ancestry differences in genetic regulatory effects suggest that for variants that are common (minor-allele frequency of >1%) across populations, the majority of eQTL signals tend to be shared with little evidence of heterogeneity in cross-population effect sizes.^44,91,92 However, our findings may not be generalizable to broader ancestral populations for rarer variants and further investigation in multi-ethnic populations are still necessary to equitably improve PSA screening. Second, our study assessed only cis-eQTLs, so any trans-eQTL effects were not incorporated.^93,94 Third, probabilistic colocalization methods have a high type two error rate and may be underpowered to detect shared association signals.²⁶ As such, a lack of colocalization signal at a particular locus does not rule out the presence of a true signal and may instead reflect analytical factors like LD mismatch.³⁸ Fourth, although we restricted our analyses to men who had not been diagnosed with PCa, we cannot rule out the possibility of latent, undiagnosed disease or disease diagnosed at a later time point. However, the prevalence of undiagnosed PCa in our population was likely to be low on account of increased monitoring and surveillance in longitudinal cohorts.^9,11 Fifth, we were unable to impute and run analyses conditional on the chromosomal inversion at 17q21.31 due to the summary-level nature of our available GWAS data. Finally, many of the genes that we identified were clustered at multi-gene loci, in part due to co-regulation by a shared set of eQTLs. Sentinel genes at these loci should be interpreted with caution, as there may be correlated predicted expression.⁹⁵

Our study also had several key strengths. The use of GWAS summary statistics from a cohort of 95,768 men provided us with high statistical power to detect PSA-associated genes. In addition to conducting tissue-specific association analyses in tissues that are biologically meaningful for PSA, we integrated association signals across 45 GTEx tissues to improve power for genes with similar regulatory mechanisms across tissues. TWAS is one of several approaches for connecting GWAS variants to target genes. Although other approaches like locus-to-gene fine-mapping⁹⁶ and combined variant-to-gene heritability frameworks⁹⁷ are powerful for systematically integrating functional data for gene prioritization, they do not restrict to disease-specific tissues and cell types. Our analysis better captures gene expression levels for PSA-relevant tissue matrices and incorporates similar gene-prioritization methods via colocalization. Our study also used COJO analysis to find genes conditionally independent from known GWAS variants. Last, the study results were validated in an independent cohort of more than 200,000 MVP participants.

In summary, our TWASs identified gene expression profiles associated with PSA levels in men without PCa. These findings provide several hypotheses for genes that affect constitutive PSA. Further exploration of these results, including functional analyses of these genes in in vivo settings, will augment our understanding of the genetic etiology of PSA variation. Transcriptomic analyses might also be vertically integrated with downstream -omic approaches to uncover complete mechanisms through which genetics influence circulating PSA levels. In addition, TWAS findings may be used to develop polygenic transcriptome risk scores⁹⁸ for PSA levels, which could be leveraged for improving PSA as a screening tool for PCa.

Data and code availability

The GWAS summary statistics from Kachuri et al.¹⁶ used in this analysis are available for download from the following Zenodo repository: https://doi.org/10.5281/zenodo.7460134. Transcriptome prediction weights and models used for MetaXcan analyses are available from: https://predictdb.org.

Web resources

https://doi.org/10.5281/zenodo.7460134.

https://predictdb.org.