Germline Variants in DNA Repair Genes, Diagnostic Radiation, and Risk of Thyroid Cancer

The link between ionizing radiation and thyroid cancer, particularly papillary thyroid carcinoma (PTC), has been well characterized in the literature. Given the rapidly increasing use of diagnostic radiation in health care settings, iatrogenic PTC is a potential sequela of routine medical workups (1) that should be mitigated as much as possible through primary prevention efforts.

DNA repair pathways work to protect the body from DNA damage caused by ionizing radiation and other mutagenic sources in the environment. These pathways include nonhomologous end joining, homologous recombination, nucleotide excision repair, and base excision repair, as well as direct reversal of DNA damage. Mutations in any of these pathways, whether acquired during life or inherited at birth, might alter an individual's lifetime risk of carcinogenesis, especially for individuals who have been exposed to higher lifetime doses of ionizing radiation.

The literature of recent years has started to examine the role of somatic mutations in PTC carcinogenesis, such as RET/PTC and other chromosomal arrangements (2, 3). However, the role of germline mutations in PTC development remains poorly documented. Furthermore, no studies to date have examined the role that such germline mutations might have in modifying the risk of PTC due to ionizing radiation. Here, we test the hypothesis that germline variations in DNA repair genes are associated with risk of PTC and, furthermore, that these variations further modify the effects of ionizing radiation on PTC risk. We do so by evaluating 299 SNPs in 52 genes related to DNA repair, using data from the Connecticut population–based case–control study.

Details of the population-based case–control study were described in previous publications (1, 4). We identified cases of histologically confirmed, incident thyroid cancer via the Yale Cancer Center's Rapid Case Ascertainment Shared Resource. A total of 462 of the eligible cases (65.9% of all eligible cases) completed in-person interviews and were included in our study. Connecticut residents who had no lifetime history of cancer of any type were recruited via random-digit dialing as population-based controls. A total of 498 individuals participated as controls in the study. Cases and controls were frequency matched by age (±5 years). The study was approved by the Human Investigations Committee at Yale and the Connecticut Department of Public Health. Written informed consent was obtained from each participant.

After undergoing the standardized interview process described previously, a total of 448 thyroid cancer cases and 465 controls donated samples of venipuncture whole blood. Peripheral blood leukocyte DNA was extracted using the Qiagen Phenol-Chloroform Extraction Kit (Qiagen, N.V.) according to standard manufacturer protocol. DNA was then genotyped using a custom-made Golden Gate Illumina assay. Genotyping data were successfully obtained for 440 thyroid carcinoma cases and 465 controls. The GoldenGate assay included analysis of 299 SNPs in 52 gene regions involved in DNA repair, based on statistical significance previously demonstrated at the SNP and gene levels in the 2011 analysis by Neta and colleagues (included in our analysis were all individual SNPs that had demonstrated a significance of P_snp < 0.1 in the study by Neta and colleagues, as well as all additional SNPs associated with gene regions with P_gene < 0.1 in the same study; ref. 5). Quality control duplicate samples were also included in the genotyping platform. All duplicate samples yielded a concordance rate of ≥99%. Hardy–Weinberg equilibrium (HWE) was assessed in controls for each SNP using a χ² test. SNPs with a P > 0.00001 from the χ² test were considered to be in HWE. Of the 299 SNPs tested, 3 SNPs were not in HWE and were excluded from the final analyses.

Unconditional logistic regression models were employed to estimate risk and to calculate ORs and 95% confidence intervals (CIs), adjusting for age, gender, and race. For each SNP of interest, the common homozygote was treated as the reference. We tested for linear trend using an additive model that assigned a value of 0 to common heterozygotes, 1 to heterozygotes (single variant), and 2 to rare homozygotes (double variant). We also compared the sum of heterozygotes and rare homozygotes with the common allele homozygote for each SNP to test for the collective significance of variation compared with the common allele. To control for race, subgroup analysis was performed using only Caucasian participants as the cases and controls. Further subgroup analysis was performed for PTC, PTC microcarcinoma (n = 163), and PTC large tumor size (n = 168). A final subgroup analysis used “PTC microcarcinoma” as cases and “PTC large tumor size” as controls.

For the gene–environment interaction analysis, exposure to diagnostic radiation was defined as exposure to any of the following procedures: (i) upper gastrointestinal series; (ii) lower gastrointestinal series; (iii) chest X-rays; (iv) head and neck CT scans; (v) chest CT scans; (vi) abdominal CT scans; (vii) pelvic CT scans; (viii) nuclear cardiology tests; (ix) thyroid uptake studies using I-131 or another radioactive agent; (x) nuclear medicine tests, including bone, brain, liver scans, or other studies that utilize pretest injection of a radioactive agent; (xi) kidney X-rays involving dye injection into a vein or artery; and (xii) mammograms. Nonexposure was defined as lack of exposure to any of these 12 procedures. An unconditional logistic regression model was used to estimate the ORs and 95% CIs for association between exposure to diagnostic X-rays and risk of thyroid cancer and its subtypes in different genotypic strata. To increase statistical power, the heterozygous and homozygous variant genotypes were combined for each SNP and compared with the “common homozygous” genotype. The significance of gene–exposure interaction was assessed by adding an interaction term in the logistic models. Results were adjusted for age (continuous), gender, race, and body mass index (BMI).

We used an a priori significance level of 0.01 for each test rather than a Bonferroni correction because the Bonferroni correction is overly conservative when hypothesis tests are correlated (6, 7). A significance level of 0.01 will increase the type I error rate to the point of certainly identifying some false positive findings. However, it will also reduce the number of false negative findings and, thus, the actual α is likely to be substantially less than the nominal α in this case. All P values presented were two-sided. All analyses were performed using Statistical Analysis Software, version 9.3 (SAS Institute, Cary, NC).

Table 1 displays selected demographic characteristics of cases and controls. Distribution of these characteristics was similar to that obtained for the original population (1). Supplementary Table S1 lists the genes represented in our analysis, grouped by category and displaying the number of analyzed SNPs per gene.

The results of the additive model SNP analysis were similar for thyroid cancer and for PTC. HUS1 rs2708896 genotypes (P_trend = 0.0057), HUS1 rs10951937 genotypes (P_trend = 0.0070), and MGMT rs12769288 genotypes (P_trend = 0.0023) were associated with PTC risk (Table 2). A number of other SNPs yielded statistically significant ORs for the heterozygous or the homozygous rare genotypes, but did not meet statistical significance in the additive model. These SNPs include EME2 rs2076431, MGMT rs10764901, RAD54B rs2046666, MBD4 rs4273365, and ATR rs10804682.

In the PTC subtype analysis, the variations in HUS1 rs2708896, HUS1 rs10951937, and MGMT rs12769288 all demonstrated statistically significant associations with PTC microcarcinoma in the additive model (Table 3). MGMT rs10764901, MBD4 rs4273365, and ATR rs10804682, along with RECQL rs12312710, yielded statistically significant ORs for the heterozygous or the homozygous rare genotypes, but did not meet statistical significance in the additive model. No SNPs displayed statistically significant associations with PTC large tumor size in the additive model. However, a number of SNPs did yield statistically significant ORs for the less common genotypes. These SNPs include EME2 rs2076431, OGG1 rs159154, OGG1 rs159153, XAB2 rs1674034, XAB2 rs794078, and ATR rs10804682. None of the SNPs listed above met statistical significance in the case–control analysis of microcarcinoma versus large tumor size.

We performed a post hoc analysis of the results for PTC large tumor size stratified by tumor size (11–15 mm, 16–30 mm, and >30 mm). This analysis revealed one SNP, rs1674034 in XAB2, that was significantly associated with PTC 16 to 30 mm (P_trend = 0.0009 for all races, P_trend = 0.0059 for Caucasians). No other SNPs remained significantly associated with any of the subgroups of PTC large tumor size after adjusting for race.

Supplementary Table S2 displays associations between diagnostic radiation exposure and risk of thyroid cancer among genotyped cases and controls. The results were similar to those obtained in the original 2015 analysis (1). The initial GxE analysis of thyroid cancer (Table 4) yielded 23 total SNPs with P_interaction < 0.05. Only 3 of these SNPs reached a priori significance of P_interaction < 0.01: ALKBH3 rs10768994 (P_interaction = 0.0086), LIG1 rs2163619 (P_interaction = 0.0060), and LIG1 rs10421339 (P_interaction = 0.0081). Three more SNPs only achieved P < 0.05 but yielded large ORs: MGMT rs4750763 (OR = 3.79; CI, 1.44–9.98; P_interaction = 0.015), MGMT rs1762444 (OR = 3.36; CI, 1.37–8.27; P_interaction = 0.024), and RPA3 rs4720751 (OR = 2.91; CI, 1.39–6.09; P_interaction = 0.034). When cases were restricted to PTC (Table 5), only LIG1 rs2163619 and LIG1 rs10421339 reached P_interaction < 0.01. In total, the SNPs with P_interaction < 0.05 for thyroid cancer included mutations in ALKBH3 (7), LIG1 (7), TOPBP1 (3), RPA3 (2), MGMT (2), PARP4 (1), and UBE2A (1). The thyroid microcarcinoma subanalysis (Table 6) yielded 12 SNPs with P_interaction < 0.05, but only 1 SNP that reached a priori significance, XRCC2 rs10234749 (OR = 7.82; CI, 2.20–27.78; P_interaction = 0.0041). Of the other 11 SNPs with P_interaction < 0.05, 4 were associated with ALKBH3, 4 were associated with ERCC5, and one was associated with PARP4. The subanalysis of PTC microcarcinoma (Supplementary Table S3) yielded similar results. The subanalysis of thyroid cancer with large tumor size yielded 5 SNPs with P_interaction < 0.05 (Supplementary Table S4). Three of these SNPs reached a priori significance: LIG1 rs251693 (OR = 2.20; CI, 1.02–4.73; P_interaction = 0.0056), LIG1 rs2288878 (OR = 2.19; CI, 1.02–4.70; P_interaction = 0.0040), and LIG1 rs274897 (OR = 2.26; CI, 1.05–4.87; P_interaction = 0.0038). When cases were restricted to PTC, none of the SNPs remained significant.

The role of germline mutations in thyroid cancer and PTC has thus far received limited investigation. Gudmundsson and colleagues (8) performed a genome-wide association study (GWAS) to search for PTC susceptibility loci and uncovered two candidate SNPs. Individuals homozygous for both variant SNPs carried an estimated thyroid cancer risk more than five times greater than that of noncarriers. The authors replicated these results in two populations of European descent. Further studies have identified additional susceptibility loci (9, 10)

Most extant candidate gene studies of PTC have only examined up to 5 to 10 SNPs each (11–15). However, the study performed by Neta and colleagues in 2011 tested 5,077 SNPs from 340 candidate genes involved in genomic integrity (5). This study revealed 9 genomic integrity SNPs associated with PTC risk with P_trend < 0.0005, as well as 7 gene regions associated with PTC risk with P_trend < 0.01, although none of these SNPs remained statistically significant after adjustment for FDR. Three of the identified SNPs (HUS1 rs2708906, ALKBH3 rs10838192, and MGMT rs4751109), and two of the identified gene regions (HUS1 and ALKBH3), correspond to genes involved in DNA repair. These promising results have merited validation.

Our study identifies a number of SNPs in DNA repair genes statistically significant associated with PTC. Among these genes are the three identified by Neta and colleagues: HUS1, MGMT, and ALKBH3. The HUS1 protein forms a complex with two other proteins, RAD1 and RAD9, and deposits in regions of damaged DNA. This activates the ATR kinase signaling cascade and thus the overall cellular response to DNA damage (16). MGMT repairs mutagenic methylguanine lesions generated by alkylating agents; decreased expression has been linked to increased incidence of gliomas (17) and testicular germ cell tumors (18). ALKBH3 plays a similar role by removing 1-methyladenine and 3-methylcytosine lesions from DNA (19).

Interestingly, our subanalysis reveals SNPs being differentially associated with PTC microcarcinoma and PTC larger tumor. For example, significant mutations in HUS1 and MGMT were identified in microcarcinoma but not in larger tumor. Other mutations, such as those in XAB2 and OGG1, only showed association with larger tumor. None of these SNPs yielded significant values in the head-to-head comparison of microcarcinoma versus larger tumor (one possible explanation could be that some “microcarcinoma” cases in fact represent a misclassification of larger tumors in their early stages of development). The differential association of SNPs with tumor sizes suggests that there might be underlying biological differences between microcarcinomas and larger tumors, which would merit further genomic investigation.

Our study is the first to examine interaction between SNPs and lifetime exposure to ionizing radiation. The GxE analyses revealed significant SNPs linked to MGMT and ALKBH3 (described previously) as well as in four additional genes: ERCC5, PARP1, XRCC2, and LIG1. The ERCC5 endonuclease makes the 3′ incision in DNA excision repair following UV-induced damage (20). Genetic variation in ERCC5 has been associated with risk of lung cancer (21), gastric cancer (22), and xeroderma pigmentosum (23). XRCC2 and PARP1 are both involved in homologous recombination. XRCC2-deficient cells appear more sensitive to PARP1 inhibitors than XRCC2-expressing cells, suggesting that XRCC2 and PARP1 share a DNA repair pathway (24). The LIG1 ligase is involved in DNA replication, recombination, and base excision repair (25); LIG1 germline polymorphisms have been associated with non–small cell lung cancer (26, 27). SNPs in LIG1 appeared to be most strongly associated with larger tumor in this study, whereas the SNPs in ERCC5, PARP1, and XRCC2 displayed association with microcarcinoma.

None of our significant SNPs were in protein-coding regions. SNPs within introns might affect RNA splicing patterns and thus upregulate or downregulate key DNA repair protein products (28). In particular, SNPs that alter the usual pattern of exonic splicing enhancers (ESE) could affect spliceosome assembly and lead to exon skipping (29). To investigate this hypothesis, we used ESEFinder (Cold Spring Harbor Laboratory) to test 43 of our most significant SNP “hits” for changes in their pattern of ESEs. Thirty-three SNPs demonstrated the potential to modify at least one ESE (Supplementary Table S5). SNPs located upstream of genes could be involved with promotor or regulator sequences, influencing the amount of RNA transcribed in various biological circumstances (30). Mounting evidence suggests that a majority of gene–environment interaction is determined by distant regulatory sequences (31). We did not directly test for cellular RNA or protein content, but further biological experiments could test for evidence differential gene transcription and RNA modification.

An important limitation of the current study is the sample size, which, although larger than those of previous candidate SNP studies, still proved insufficient to detect SNPs with unequivocal statistical significance after Bonferroni correction for multiple comparison. SNP validation would require an even larger pool of cases and controls. A larger sample size would also increase confidence in certain “significant” results that are based on very small sample sizes. Finally, although our post hoc analysis of large tumor size subsets suggests that additional SNP associations might exist, the current study was not powered to detect such associations.

Recall bias must be considered in this study. Data on diagnostic radiography exposure relied upon self-reporting by cases and controls, rather than health record documentation. However, as described in our previous publication, other studies (32–34) have suggested nondifferential reporting error between thyroid cancer cases and controls. Even if differential recall bias were a possibility for subjects in a particular age range (35), our age-stratified analysis of the original data described previously was unable to uncover evidence of such bias in our own findings. Furthermore, our significant findings are limited to thyroid microcarcinoma. In addition, diverse types of diagnostic radiation exposure were combined, which might cause potential exposure misclassification. In gene–environment interaction studies, both differential and nondifferential misclassification of a binary environmental factor biases a multiplicative interaction effect toward the null (36).

As previously discussed, our candidate gene approach likely misses many potentially significant regions of the genome. A fruitful future direction for this work would be performing a GWAS using a randomly selected subset of our cases and controls, to identify novel regions of genomic significance.

This study provides further evidence that germline mutations in DNA repair genes are significantly associated with risk of thyroid cancer and PTC. It suggests that microcarcinoma, which is particularly associated with diagnostic radiation exposure, represents a distinct subset of thyroid cancer and PTC with its own biological signature. Finally, our study provides novel evidence to suggest a significant interaction between germline mutations in DNA repair genes and ionizing radiation in the pathogenesis of thyroid cancer. These results merit replication with larger sample sizes and alternative study methods to establish statistical significance, and to further explore the basis of the molecular biology behind them.

No potential conflicts of interest were disclosed.

The authors assume full responsibility for analyses and interpretation of these results.

Conception and design: R. Udelsman, Y. Zhang

Development of methodology: J.E. Sandler, Y. Zhang

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.E. Sandler, H. Huang, N. Zhao, W. Wu, F. Liu, S. Ma, R. Udelsman

Writing, review, and/or revision of the manuscript: J.E. Sandler, H. Huang, S. Ma, R. Udelsman, Y. Zhang

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): N. Zhao, F. Liu, Y. Zhang

Study supervision: Y. Zhang

This research was supported by the American Cancer Society grant RSGM-10-038-01-CCE (to Y. Zhang) and the NIH grant R01ES020361 (to Y. Zhang).

We gratefully acknowledge Dr. Alina Brenner, Dr. Gila Neta, and Dr. Alice Sigurdson of the NCI for consulting on this project.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.