Genetic polymorphisms associated with pancreatic cancer survival: a genome-wide association study

Abstract

Previous findings on the association of genetic factors and pancreatic cancer survival are limited and inconsistent. In a two-stage study, we analyzed the existing genome-wide association study dataset of 868 pancreatic cancer patients from MD Anderson Cancer Center in relation to overall survival using Cox regression. Top hits were selected for replication in another 820 patients from the same institution using the Taqman genotyping method. Functional annotation, pathway analysis, and gene expression analysis were conducted using existing software and databases. We discovered genome-wide significant associations of patient survival with three imputed SNPs which, in complete LD (r²=1), were intronic SNPs of the PAIP2B (rs113988120) and DYSF genes (rs112493246 and rs138529893) located on chromosome 2. The variant alleles were associated with a 3.06-fold higher risk of death (95% confidence interval [CI]=2.10-4.47, P = 6.4 × 10⁻⁹) after adjusting for clinical factors. Eleven SNPs were tested in the replication study and the association of rs113988120 with survival was confirmed (HR: 1.57, 95%CI: 1.13-2.20, P = 0.008). In silico analysis found rs1139988120 might lead to altered motif. This locus is in LD (D′=0.77) with 3 eQTL SNPs near or belong to the NAGK and MCEE genes. According to The Cancer Genome Atlas data and our previous RNA-sequencing data, the mRNA expression level of PAIP2B but not NAGK, MCEE or DYSF was significantly lower in pancreatic tumors than in normal adjacent tissues. Additional validation efforts and functional studies are warranted to demonstrate whether PAIP2B is a novel tumor suppressor gene and a potential therapeutic target for pancreatic cancer.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) is the most common type of pancreatic cancer which is the fourth leading cause of cancer death in the United States. PDAC has a dismal prognosis; the average 5-year survival rate was 6.7% according to cumulative statistics from 2004 to 2010 ¹. The poor outcome of this disease is due mainly to late diagnosis and the high invasiveness and profoundly drug-resistant nature of the tumor. The recent advances in the understanding of the molecular and genetic alterations in pancreatic tumor have not yet been translated into significant improvement in patient survival or reduced mortality. There is a critical need for novel therapeutic targets and molecular markers in personalized PDAC management to achieve better treatment efficacy.

Emerging evidence suggests that germline genetic variations influence survival in PDAC. Studies using the candidate-gene approach focusing on genes involved in DNA repair ², cell cycle regulation ³, drug metabolism ⁴, and signaling pathways ^{5, 6} have observed many nominally significant associations. Genome-wide association studies (GWAS) on overall survival (OS) in PDAC have also identified a number of interesting loci, but the findings are inconsistent and none of these loci have reached genome-wide significance^7-10. Identifying genetic variants associated with PDAC survival may lead to the discovery of novel therapeutic targets for the development of new strategies in patient treatment.

Many of the previous studies on genetic factors in PDAC survival suffered the limitations of insufficient study power or heterogeneous study population. Taking advantage of the existing GWAS data and the large patient population recruited in a case-control study on PDAC conducted at The University of Texas MD Anderson Cancer Center, we performed a two-stage study to identify genetic variants that are associated with survival in 1,688 patients with pathologically diagnosed PDAC.

Materials and Methods

Study population, GWAS dataset, and discovery study

The study population was drawn from a hospital-based case-control study of patients with pathologically confirmed PDAC conducted at The University of Texas MD Anderson Cancer Center from June 2002 through May 2009¹¹. All patients were recruited consecutively with the exclusion criteria of non-US resident or having a prior history of cancer. All patients signed an informed consent for interview and a blood sample. DNA was extracted from peripheral lymphocytes and GWAS was conducted as previously described by the Pancreatic Cancer Cohort Consortium (PanScan) and the Pancreatic Cancer Case-Control Consortium (PanC4)^{12, 13}. A total of 903 MD Anderson patients with sufficient amount of DNA samples were genotyped in the PanScan II and PamC4 GWASs.

Prior to analysis, we carried out further quality control of the GWAS data. We extracted 11 duplicate samples that were intentionally preset to check the genotyping concordance between PanScan II and PanC4. Comparison of the duplicate samples showed 100% concordance. Then, we excluded 15 patients who had later been diagnosed with diseases other than PC or who did not have complete clinical data. Identity-by-descent analyses did not show evident familial relationships between any of the patients. A total of 877 samples remained for population structure analysis. Using the International HapMap Project genotype data (phase 3 release #3, National Center for Biotechnology Information [NCBI] build 36, SNP Database (dbSNP) b126, 2010-05-28, minor allele frequency [MAF]>5%) for CEU, JPT/CHB, and YRI ¹⁴, we seeded 10,195 high-quality markers from our dataset (r² < 0.004) in STRUCTURE analysis ¹⁵. A total of 868 individuals of European descent (0.82-1.00 similarity to CEU) were identified for the survival analysis in the discovery study. We also derived five principal components for population substructure based on the 868 patients using Genome-wide Complex Trait Analysis software¹⁶.

Replication study

The replication study was conducted in 820 MD Anderson patients who were mostly enrolled in the case-control study after the conduction of the GWAS in 2009. The only selection criteria we applied are: 1) have not been involved in GWAS and 2) being self-reported non-Hispanic whites. We selected the highest-ranked single-nucleotide polymorphisms (SNPs), which had the smallest P values in Cox regression analysis, high imputation quality scores (≥0.7), and MAFs of >0.15. We also preferentially considered the nonsynonymous SNPs or SNPs of known PDAC-related genes. Genotyping was performed using the TaqMan method in an Applied Biosystems 7900HT Fast Real-time PCR System (Foster City, CA). Five percent of the samples were analyzed in duplicates, and a 99% agreement rate was achieved. The inconsistency was resolved by further genotyping.

Statistical methods

Imputation

Because PanScan II and PanC4 used different genotyping platforms, imputation was conducted to generate a common dataset for both studies. The GWAS data were first pre-phased in SHAPEIT2 ¹⁷, then imputed in IMPUTE2^{18, 19} with the 1000 Genomes Project (phase1_release_v3.20101123) as the reference ²⁰. Because SNPs in PanScan II were originally mapped to an older genome assembly (NCBI build 36 [hg18]), we systematically converted their genome positions to genome assembly NCBI build 37 (hg19) using the liftOver tool by the University of California, Santa Cruz, (http://genome.ucsc.edu). SNPs not listed in NCBI build 37 were removed, with the RsMergeArch.bcp.gz database as the reference (ftp://ftp.ncbi.nih.gov/snp/organisms/human_9606/database/organism_data/). Moreover, we removed SNPs that had a MAF of less than 1%, deviated from Hardy–Weinberg equilibrium (P < 0.0004), or had a high number of missing genotypes (>2%). Because the PanScan II and PanC4 datasets had only ∼320,000 SNPs in common, imputation was performed in separate dataset. After imputation, we pruned out SNPs with quality scores of <0.3, MAFs of <0.01, missing rates of >0.02, or deviating from Hardy–Weinberg equilibrium (P < 0.0001), resulting in 7,738,399 bi-allelic SNPs for further analysis.

Survival analysis

OS was calculated from the date of cancer diagnosis to the date of death or to the last follow-up date. The maximum follow-up time was 5 years. Patients who were alive at the last follow-up date were censored. Median survival times were estimated and compared using the Kaplan–Meier plot and log-rank test. A Cox proportional hazards model was fitted to assess risk of death using hazard ratios (HRs) and 95% confidence intervals (95% CIs). Genotype data were analyzed using an additive inheritance model with adjustment for demographics such as age (continuous) and sex and for clinical predictors such as tumor stage (localized, locally advanced, or metastatic), tumor resection (performed or not performed) and chemotherapy (given or not given). In the discovery study, five principal components accounting for population substructure were also adjusted.

Statistical analysis used the R package version 3.1.0. We took P < 5.0 × 10⁻⁸ as genome-wide significant in the GWAS analysis and took P < 5.0 × 10⁻² as nominally significant and P < 4.5 × 10⁻³ (P/the number of SNP tested) as Bonferroni-corrected significant in analyses of the validation and the combined GWAS and validation dataset.

Functional annotation and network analysis

We conducted functional annotation and linkage disequilibrium (LD) analysis of the top hit SNP using HaploReg version 4 ²¹, RegulomeDB²² and PolyPhen-2²³. The association of selected SNPs and gene expression levels was evaluated using Genotype-Tissue Expression (GTEx) and NCBI eQTL database and the RNASeq level 3 data from TCGA. A previously conducted RNA-Seq study in paired normal and tumor tissues from 10 patients with resected PDAC tumors²⁴ was also considered.

To explore functionally enriched biological pathways, we conducted network analysis on SNPs with a P value of <0.0005 using Ingenuity Pathway Analysis (Ingenuity Systems, Redwood City, CA [www.ingenuity.com]). SNPs were assigned to relevant genes using the University of California Santa Cruz (UCSC) Table Browser data retrieval tool²⁵. For each gene region, SNPs within 20 kb upstream or downstream of the gene were included.

Replication analysis of SNPs reported in association with OS of PDAC

As a validation effort, we examined SNPs that were associated with PDAC survival in previous GWASs (P<10⁻⁵)^7-10 in association with survival in the current MD Anderson GWAS dataset.

Results

Patient characteristics and OS

The distributions of sex, age, and tumor stage were similar between patients in the discovery and validation sets (P > 0.15 for each) (Table 1). However, more patients in the discovery set received chemotherapy than did those in the validation set (73.73% vs. 66.34%, P = 0.002), and more patients underwent tumor resection in the validation set than did those in the discovery set (34.15% vs. 29.26%, P = 0.031). As expected, disease stage and tumor resection status were strong predictors for OS (Table 1).

Table 1. Patient characteristics and overall survival.

Variable	Discovery set			Validation set			Combined set
	No. (%)	MSTa	P valueb	No. (%)	MST	P value	No. (%)	MST	P value
Vital Status
Dead	800 (92.2)			630 (76.8)			1430 (84.7)
Alive	68 (7.8)			190 (23.27)			258 (15.3)
Sex
Male	517 (59.6)	14.7		492 (60.0)	17.6		1009 (59.8)	15.4
Female	351 (40.4)	14.5	0.84	328 (40.0)	17.9	0.70	679 (40.2)	16.2	0.72
Age (years)c
≤63	467 (53.8)	15.2		407 (49.6)	18.9		874 (51.8)	15.9
>63	401 (46.2)	13.8	0.55	413 (50.4)	16.6	0.18	814 (48.2)	15.7	0.69
Stage
Local	266 (30.7)	32.1		277 (33.8)	32.9		543 (32.2)	32.1
Locally advanced	215 (24.8)	13.9		214 (26.1)	14.8		429 (25.4)	14.7
Metastatic	387 (44.6)	9.5	<0.0001	329 (40.1)	11.1	<0.0001	716 (42.4)	10.0	<0.0001
Chemotherapy
Yes	640 (73.73)	14.9		545 (66.95)	18		1185 (70.45)	16.2
No	228 (26.27)	13.9	0.74	269 (33.05)	17.2	0.74	497 (29.55)	14.8	0.28
Resection
Yes	254 (29.26)	33.7		280 (34.15)	33.7		534 (31.64)	33.7
No	614 (70.74)	11.2	<0.0001	540 (65.85)	12.8	<0.0001	1154 (68.36)	12.0	<0.0001

^aMST: median survival time in months.

^bP Values were derived from log-rank test.

^cMedian age for all patients.

Discovery study

In the discovery study, with the use of the GWAS data, 528,088 SNPs were nominally associated with OS, and three SNPs in complete linkage disequilibrium reached genome-wide significance (Figure 1). Detailed information on the top 214 SNPs with a P value of less than or equal to 1 × 10⁻⁵ is given in STable 1. The top ten SNPs with the smallest P values in this analysis were all imputed, and their quality scores ranged from 0.556 to 0.995 (STable 1). To check the quality of imputation, we re-genotyped the top SNP, rs113988120, in the 868 GWAS samples using the Taqman method, and a concordance rate of 99.7% was observed. We estimated the genomic control inflation factor λ using the imputed genotype dosage after adjusting for the leading five principal components in the survival analysis. The quantile-quantile plot showed a slightly inflated λ value of 1.063 (SFigure 2). It is unlikely the inflation is due to population stratification because we have adjusted the five principal components accounting for population substructure in the analysis. It's more likely due to the survival analysis itself.

Fig. 1.

Manhattan plot for P values in survival analysis of the GWAS data.

The top three SNPs that showed a genome-wide significant association with survival are intronic SNPs (rs113988120) of the PAIP2B (poly(A) binding protein interacting protein 2B) and rs112493246 and rs138529893 of the DYSF (dysferlin) gene located on chromosome 2. The minor alleles of these three SNPs were significantly associated with shorter survival by a margin of 5.2 months (log-rank P = 1.6 × 10⁻⁵) (Figure 2A). Cox regression analysis with adjustment for demographic and clinical factors demonstrated a genome-wide significant (P ≤ 6.4 × 10⁻⁹) association of the three SNPs with increased risk of death (HR: 3.06, 95% CI: 2.10-4.47).

Fig. 2.

Kaplan-Meier plot for overall survival of patients with pancreatic cancer by rs113988120 genotype. Panels A, B, and C present survival curve in the discovery, validation, and combined datasets, respectively. The median survival time was 14.87 vs 9.67 months, 19.7 vs 13.3 months, and 16.9 vs 10.6 months for TT vs TA/AA genotype carriers and the number of patients with the at-risk (TA/AA) genotypes are 29, 31 and 60 in the discovery, replication and combined datasets, respectively. P values are from log-rank test.

Replication study and combined dataset

For the replication study, we selected 11 SNPs located in eight genes coding for PAIP2B, RAB6B, ZBTB2O, ROBO1, B4GALT4, UPK1B, LRRC15, and SLIT3; in a non-coding RNA (LOC101927026); and in an intergenic region (Table 2). The top hit rs113988120 was included in the replication study even though the MAF (0.015) was quite low (0.015). The allele frequencies of the 11 SNPs tested in the replication study were similar to those in the discovery dataset (Table 2). The variant allele of rs113988120 was associated with a 6.4 months and 6.3 months shorter survival in the replication and combined discovery and replication datasets, respectively (Figure 2B and 2C). A HR (95% CI) of 1.57 (1.13-2.20, P = 0.08) in the replication dataset and 1.86 (1.46-2.37, P = 4.6 × 10⁻⁷) in the combined dataset was observed for this variant allele in Cox proportion hazards regression analysis with adjustment for demographics, tumor stage, and resection status, and chemotherapy (Table 3). Other SNPs in the replication study were not significantly associated with OS (Table 3).

Table 2. Eleven SNPs tested in the validation and combined datasets.

SNP	Gene	Quality	RegulomeDB	MAF	Discovery		Validation		Combined
(chromosome)		scorea	scoreb		N (Dead|Alive)	MSTd	N (Dead|Alive)	MST	N (Dead|Alive)	MST
rs113988120 (2)	PAIP2B	0.9	6	0.015
TT					721|118	14.9	505|248	19.7	1226|366	16.3
TA					29|0	9.7	20|5	14.7	49|5	10.3
AA					0|0	-	4|1	13.3	4|1	13.7
rs4780973 (16)	LOC101927026	0.99	5	0.31
CC					377|53	12.9	267|109	17.6	644|162	14.7
CT					315|50	15.07	228|119	17.6	543|169	16.2
TT					58|15	21.7	44|27	22.4	102|42	21.93
rs9815265 (2)	RAB6B	0.9	3	0.403
TT					279|36	13.4	165|86	19.5	444|122	14.8
TC					362|55	14.6	264|108	19.1	626|163	15.6
CC					109|27	22	126|65	18.7	235|92	18.8
rs6438217 (3)	ZBTB20	1	6	0.164
TT					492|93	15.7	367|173	19.5	859|266	16.6
TC					234|22	13.9	166|80	19.5	400|102	14.8
CC					24|3	10.5	19|5	13.3	43|8	12.4
rs12638565 (3)	ROBO1	0.93	No data	0.233
CC					448|72	16	310|143	20.7	758|215	17.3
CT					264|43	13.1	180|87	17.6	444|130	14.4
TT					38|3	9.5	60|27	19.7	98|30	13.7
rs4568126 (3)	B4GALT4	0.99	6	0.421
CC					266|35	13.4	120|59	18.4	386|94	14.5
CA					361|53	14.3	249|123	18.8	610|176	15.4
AA					123|30	18.6	187|77	20	310|107	18.9
rs12186112 (3)	UPK1B	0.99	6	0.405
AA					286|40	12.9	199|82	19.5	485|122	14.7
AC					342|52	14.8	250|122	19.5	592|174	16.3
CC					122|26	18.3	93|49	18.1	215|75	17.4
rs13060627 (3)	LRRC15	1	V264I	0.264
CC					423|58	13.1	304|144	19	727|202	14.7
CT					281|46	15.7	207|95	20.3	488|141	17.2
TT					46|14	20	37|18	15	83|32	18.4
rs923936 (3)	LRRC15	1	L133L	0.234
GG					458|62	13.5	29|13	18.5	487|75	13.8
GA					254|43	15.2	203|92	19.5	457|135	16.9
AA					38|13	20.7	313|150	19	351|163	17.6
rs13173842 (5)	SLIT3	0.79	4	0.321
CC					363|59	15.1	246|116	18.9	609|175	16
CT					305|52	14.8	237|112	20.1	542|164	16.3
TT					82|7	12.4	60|27	17.3	142|34	13.7
rs10926274 (1)	NA	0.72	4	0.483
AA					189|22	12.4	127|50	19.1	316|72	14.2
AG					393|57	14.8	262|131	19.9	655|188	15.8
GG					168|39	17	162|72	18.4	330|111	17

^aImputation quality score.

^bMinor allele frequency based on imputation of the GWAS data.

^cScore 3a, 4,5, and 6 indicate TF binding + any motif + DNase peak, TF binding + DNase peak, TF binding or DNase peak, or other, respectively.

^dMedian survival time in months.

Table 3. Cox regression on 11 SNPs in the discovery, validation and combined dataset.

SNP	Discovery		Validation		Combined
	HR (95%CI)	P value	HR (95%CI)	P value	HR (95%CI)	P value
rs113988120	3.06 (2.10-4.47)	6.40E-09	1.44 (1.02-2.02)	0.037	1.79 (1.40-2.29)	2.70E-06
rs4780973	0.77 (0.68-0.86)	2.60E-06	0.90 (0.79-1.02)	0.11	0.82 (0.76-0.89)	4.40E-06
rs9815265	0.75 (0.68-0.83)	6.70E-08	1.02 (0.92-1.14)	0.70	0.87 (0.81-0.94)	2.70E-04
rs6438217	1.40 (1.23-1.59)	4.40E-07	1.01 (0.86-1.17)	0.94	1.19 (1.08-1.31)	5.80E-04
rs12638565	1.35 (1.19-1.52)	1.80E-06	1.04 (0.92-1.17)	0.57	1.14 (1.05-1.24)	2.10E-03
rs4568126	0.77 (0.70-0.85)	3.10E-07	0.97 (0.87-1.09)	0.64	0.85 (0.79-0.91)	6.10E-06
rs12186112	0.78 (0.70-0.86)	6.00E-07	0.99 (0.88-1.12)	0.92	0.87 (0.81-0.94)	3.60E-04
rs13060627	0.79 (0.70-0.88)	4.50E-05	0.94 (0.82-1.07)	0.33	0.85 (0.78-0.93)	2.40E-04
rs923936	0.79 (0.70-0.88)	5.90E-05	1.08 (0.94-1.24)	0.27	0.88 (0.82-0.94)	1.30E-04
rs13173842	1.27 (1.14-1.42)	1.30E-05	0.97 (0.86-1.10)	0.61	1.11 (1.02-1.20)	1.40E-02
rs10926274	0.83 (0.75-0.91)	2.20E-04	1.02 (0.91-1.14)	0.72	0.91 (0.84-0.98)	1.50E-02

Cox regression analysis was under additive genetic model with adjustment for age, sex, tumor stage, resection, and chemotherapy in validation and combined datasets and with additional adjustment for five principal components accounting for population structure in the discovery set.

Next we assessed the impact of rs113988120 on survival by disease stage. The median survival time was 34.3 vs 14.9 months (P_log–rank <0.001), 15.7 vs 10.3 months (P_log–rank <0.001), and 10.6 vs 8.0 months (P_log–rank =0.318) for TT vs TA/AA genotype carriers for patients with localized, locally advanced and metastatic disease, respectively (SFigure 2).

Functional annotation and gene expression analysis

RegulomeDB analysis suggest that it may have an undefined altered mortif with a score of 6, and it is located at a DNAseI hypersensitive site and an enhancer region of the chromatin, which make it more accessible for the binding of proteins such as transcription factors. HaploReg analysis demonstrated that rs113988120 was in LD (r²=0.49, D′=0.77) with three eQTL SNPs (rs34634781, rs35098046 and rs34022557) near or belong to the NAGK (N-acetylglucosamine kinase) and MCEE (methylmalonyl-CoA epimerase) gene on chromosome 2, in addition to the two SNPs of the DYSF gene (STable 2). Cox regression analysis of the GWAS data showed that rs34634781 had an HR of 2.44 (95% CI: 1.70 -3.49, P = 1.10 × 10^-6) and s35098046/rs34022557 had an HR of 2.43 (95% CI: 1.71 -3.46 P = 8.50 × 10^-7) (STable 1).

Using TCGA RNA-seq level 3 data from four PDAC patients, we found that PAIP2B expression was downregulated in tumors (mean=90.58) compared with paired normal tissues from the same patient (mean = 150.22) (P = 0.23, paired t-test). This observation agreed with data from our previous RNA-seq study on paired tumor-normal tissues of 10 PDAC patients, which showed a 7.8-fold lower expression of PAIP2B in tumors than in normal adjacent tissues (P = 0.00032)²⁴. Notably, the expression of PAIP2B was significantly reduced in tumors compared with that in normal adjacent tissues of other gastrointestinal cancers according to TCGA RNA-seq data (Ps ≤ 8 × 10⁻⁵) (Table 4). We did not detect a significant association of PAIP2B expression with patient survival in the entire TCGA dataset of 178 tumor samples (P = 0.30). Among patients with stage II or earlier disease (n=48), those with lower (below median) PAIP2B expression exhibited a non-significant shorter survival (by 39 months, P_log–rank = 0.10) and increased risk of death than those with higher (at or above median) PAIP2B expression (HR=2.45, 95%CI: 0.90-6.80, P = 0.081). The mRNA expression levels of NAGK, MCEE and DYSF genes were not significantly different between tumor and normal adjacent tissues and were not related to survival (data not shown).

Table 4. Expression of PAIP2B in gastrointestinal cancers in TCGA databases.

Cancer type	No. of patients	Mean expression (tumor|normal)	P valuea
Pancreatic cancer	4	90.58|150.22	2.30E-01
Cholangiocarcinoma	9	107.93|950.49	1.40E-07
Colon adenocarcinoma	26	166.67|253.97	5.30E-05
Colorectal adenocarcinoma	32	171.43|264.07	2.40E-06
Hepatocellular carcinoma	50	569.4|1009.75	8.00E-05
Stomach adenocarcinoma	32	116.79|414.84	4.90E-08

^aDerived from paired t-test

Ingenuity Pathway Analysis (IPA)

Next, we conducted IPA on 162 genes whose tagging SNPs had P < 0.0005 in Cox regression analysis. The top four most significant pathways identified were G beta-gamma signaling (P = 1.77 × 10⁻⁵, q value = 0.012), CXCR4 signaling (P = 8.82 × 10⁻⁵, q value = 0.03), CREB signaling in neurons (P = 1.99 × 10⁻⁴, q value = 0.07), and CCR5 signaling in macrophages (P = 4.51 × 10⁻⁴, q value = 0.08) and Agrin interactions at neuromuscular junction (P = 4.47 × 10⁻⁴, q value = 0.08) (STable 3). When we reran IPA on genes selected at various P value thresholds ranging from 0.008 to 0.00001, the top four pathways remained almost unchanged (data not shown).

Replication analysis of SNPs previously associated with OS or risk of PC

We analyzed the top SNPs associated with PC survival in previous GWAS studies ^7-10 and found that 11 of 131 such SNPs (P<10⁻⁵) in the study by Wu et al. ⁷ were nominally significant in our study (P < 0.05) (STable 4).

Discussion

In this GWAS study, we discovered a significant association of a polymorphic variant (rs113988120) of the PAIP2B gene with shorter survival in patients with PADC and confirmed this association in a replication study. This is, to our knowledge, the first report of a genome-wide association of germline genetic variation and PDAC survival in a GWAS dataset.

The rs113988120 SNP is located in the intron region of the PAIP2B gene. In silico analysis suggested weak evidence for possible adverse functional consequences of this SNP but it is not related to gene expression. Furthermore, this SNP is located in a DNAseI hypersensitive and chromatin enhancer region, which suggests more accessibility for transcription factor binding. Furthermore, rs113988120 is in LD with five SNPs, including three eQTLs, near or located at the NAGK, MCEE, and DYSF genes and all five SNPs were among the top hits identified from the GWAS dataset with P values less than or equal to 1.1 × 10^-6. Thus, which gene or SNP is truly responsible for the observed association with PDAC survival is unclear at present. Further fine mapping and functional studies are required to identify the causal allele and to understand the mechanisms involved.

At the gene level, the gene expression data support a tumor suppressor role of PAIP2B and possible link with patient survival. While a significantly reduced expression of PAIP2B mRNA was observed in pancreatic cancer²⁴ and several other gastrointestinal cancers (TCGA data), the mRNA expression level of NAGK, MCEE, and DYSF genes were not significantly different between tumor and normal tissues. In addition, the reduced expression of PAIP2B was related to shorter survival in PADC patients with early-stage disease (N=48) even though the difference was not statistically significant. Because of the sample size is small, this association needs further investigation.

PAIP2B is a translational inhibitor²⁶ that regulates poly(A) binding protein activity. Poly(A) binding protein enhances translation by circularizing mRNA through its interaction with the translation initiation factor EIF4G1 and the poly(A) tail²⁷. PAIP2B regulates the translation of many genes that have important biological significance in cancer^{28, 29}. For example, it is a strong regulator of vascular endothelial growth factor³⁰ and is an anti-proliferative factor³¹. Although we have no evidence to link SNP rs113988120 with the expression of the PAIP2B gene, it is conceivable that the polymorphic variants may result in altered poly(A) binding protein binding activity, which in turn leads to upregulated expression of many pro-proliferative or metastatic genes that contribute to reduced patient survival. Further replication studies in other GWAS datasets and functional studies on PAIP2B in pancreatic carcinogenesis are warranted to confirm the observed association and to understand the mechanisms underlying the association.

DYSF is a skeletal muscle protein, which plays a role in calcium-mediated membrane fusion events, suggesting that it may be involved in membrane regeneration and repair³². Mutation of these genes have been associated with sarcolemma in animal model^{33, 34}. The NAGK gene encodes N-acetylhexosamine kinase that catalyzes the conversion of N-acetyl-D-glucosamine to N-acetyl-D-glucosamine 6-phosphate, and is the major mammalian enzyme which recovers amino sugars. NAGK has been associated with speckle, paraspeckle and general transcription factor suggesting its regulatory roles in gene expression ³⁵. NAGK gene mutation has been related to inclusion body myopathy ³⁶. MCEE catalyzes the interconversion of D- and L-methylmalonyl-CoA during the degradation of branched chain amino acids, odd chain-length fatty acids, and other metabolites. Mutations in this gene result in methylmalonyl-CoA epimerase deficiency, which is presented as mild to moderate methylmalonic aciduria³⁷. Whether these two genes play any role in cancer is currently unknown.

Network analysis identified several signaling pathways that are significantly associated with survival, such as the G beta-gamma and CXCR4 signaling pathways. The CXCL12/CXCR4 axis may promote dissociation of the G beta-gamma complex (G_βγ), then activate PI3K-AKT and Rho-ROCK-MLC pathways to promote cancer cell survival and migration³⁸. A high level of CXCL12/CXCR4 expression was reported to be significantly associated with metastasis and low OS in many types of cancer, including PDAC^{39, 40}. Suppression of the CXCL12/CXCR4 axis contributed to immune control of pancreatic ductal cancer growth⁴¹. Our findings on the CXCL12/CXCR4 pathway genes may have potential value in future PDAC therapy targeting this particular pathway.

The strengths of the current study include a large sample size and a relatively homogeneous study population from the same institution. As part of a hospital-based study, the clinical data are relatively accurate and complete. The existing large databases such as TCGA and GTEx have facilitated the functional annotation of the GWAS top hits. However, the number of SNPs selected for validation was limited because of cost constraints. We may have missed some important ones. The low MAF of the identified SNP and the small number of samples available for PC gene expression analysis limited our ability to fully characterize this gene variant.

Overall, our study has reported a low-frequency SNP of the translation inhibitor gene PAIP2B with a significant association with PC survival. These data need further validation in other datasets. If confirmed, they may open a new research avenue in illustrating the molecular mechanisms underlying the clinical phenotypes and offer a potential tool in identifying therapeutic targets for future individualized cancer treatment.

Supplementary Material

SFigure 1

SFigure 1. Quantile-Quantile (QQ) plot of observed vs. expected P values for the discovery set. The genomic control inflation factor λ was calculated using the imputed genotype dosage after adjusting for the leading five principal components account for population structure in the survival analysis

SFigure 2

SFigure 2. Kaplan-Meier plot for overall survival of patients with pancreatic cancer by rs113988120 genotype. Panels A, B, and C present survival curve for patients with localized, locally advanced and metastatic disease, respectively. The median survival time was 34.3 vs 14.9 months, 15.7 vs 10.3 months, and 10.6 vs 8.0 months for TT vs TA/AA genotype carriers and the number of patients with the at-risk (TA/AA) genotypes are 15, 16 and 29 for patients with localized, locally advanced and metastatic disease, respectively. P values are from log-rank test.

STable 1

STable 1. Top SNPs (P < 0.00001) associated with overall survival of pancreatic cancer patients in PanScanII + PanC4

STable 2

STable 2. SNP: rs113988120 and variants with r2 ≥ 0.4

STable 3

STable 3. IPA analysis of significant canonical pathways in the GWAS dataset

STable 4

STable 4. Top SNPs associated with survival of pancreatic cancer by previous GWAS (P<1.0 × 10^-5).

Novelty and Significance.

We found a significant association of a single nucleotide polymorphic variant and pancreatic cancer survival in a two stage study. This finding may help to identify novel tumor suppressor gene and therapeutic target for pancreatic cancer.

Abbreviations

CI

confidence interval

eQTL

expression quantitative trait locus

GWAS

genome wide association study

GTEx

Genotype-Tissue Expression

HR

hazard ratio

LD

linkage disequilibrium

MAF

minor allele frequency

OS

overall survival

PDAC

pancreatic ductal adenocarcinoma

SNP

single nucleotide polymorphism, TCGA, The Cancer Genome Atlas