DNArCdb: A Database of Cancer Biomarkers in DNA Repair Genes that Includes Variants Related to Multiple Cancer Phenotypes

Abstract

Functioning DNA repair capabilities are vital for organisms to ensure that the biological information is preserved and correctly propagated. Disruptions in DNA repair pathways can result in the accumulation of DNA mutations, which may lead to onset of complex disease such as cancer. The discovery and characterization of cancer-related biomarkers may allow early diagnosis and targeted treatment, which could significantly contribute to the survival rates of cancer patients. To this end, we have applied a hypothesis driven bioinformatics approach to identify biomarkers related to 25 different DNA repair enzymes, in combination with structural analysis of six selected missense mutations of newly discovered SNPs that are associated with cancer phenotypes. Our search on 8 distinct cancer databases uncovered 43 missense SNPs that statistically significantly associated at least one phenotype. Moreover, nine of these missense SNPs are statistically significantly associated with two or more cancers. In addition, we have performed classical molecular dynamics to characterize the impact of rs10018786 on POLN, which results in the M310L Pol ν variant, and rs3218784 on POLI, which results in the I236M Polι. Our results suggest that both of these cancer-associated variants result in noticeable structural and dynamical changes compared with their respective wild-type proteins.

Introduction

Cancer is a major healthcare challenge as the complexity and diversity of the disease inhibits easy solutions in developing effective treatments. Currently, cancer is the 2nd leading cause of death in the USA¹. In 2016, the number of diagnosed cancer cases in the USA was estimated at 1,685,210, while the number of deaths from cancer was reported to be 595,690². Cancers with high rates of incidence include prostate for men (21% of new cases), and breast for women (29%), while lung and bronchus cancers result in the largest number of cancer caused deaths for both men (27% of all cancer deaths) and women (26%)². While the treatment for some types of cancers has a relatively high success rate, e.g. prostate cancer has 98.6% survival rate, others remain a challenge, e.g. pancreatic cancer has 8.2% survival rate³.

The search for genes and mutations that are involved in carcinogenesis is an ongoing endeavor. As cancer has its roots in genetic information, finding the genetic variants that are responsible for carcinogenesis would be beneficial for cancer prevention, diagnostics, and treatment. Cancer causing genetic mutations can be of both germline and somatic origin⁴. Of particular interest are the genes involved in cell proliferation, differentiation, death, and DNA repair as their disruption is associated with carcinogenesis⁴. Cancer population genetic information is accumulating thanks to a number of initiatives, such as the Cancer Genome Atlas (TCGA) Research Network⁵, and the NCBI database of genotypes and phenotypes (dbGAP)⁶.

Taking into consideration personal genetic data in identifying patients’ cancer treatment can significantly contribute to finding the best interventions at the right time and thus increase survival rates. Personalized medicine, the term that has been used in the literature, can allow to foresee the possibility of development of a particular disease and to implement preventive measures, as well as to find the most effective medications tailored individually to a patient⁷. In addition to individual genomic profile, transcriptomics, proteomics, and metabolomics data may also be used for diagnostic decisions⁷.

Nevertheless, challenges still remain, although biomarkers’ development have been facilitated by the progress in genotyping and data processing technologies. The search for biomarkers is complicated by the fact that out of many possible mutations in cancer linked genes only a few are actually cancer driving mutations, while most are passengers⁸. Also, genes confer their role as constituents of various pathways and thus when analyzed individually may not represent the full picture of biological processes. Therefore, methods such as gene sets analysis have been developed to account for biological complexity and interplay of various genes and pathways⁹. Furthermore, genetics composition of neoplasms are heterogeneous, complicating pinpointing a causative mutation. Methods have been developed to detect low frequency somatic mutations¹⁰. More recently, methods that analyze a small to moderate number of SNPs on a several phenotypes have been developed to investigate the possibility of these SNPs for a common underlying genetic predisposition^{5, 11}.

DNA repair genes serve a genome maintenance function and are important for preventing carcinogenesis^{12, 13}. Mutations that disrupt the function of DNA repair genes can result in increased rates of somatic mutations in other genes and thus significantly increase susceptibility to cancer. Without maintenance and repair, cells retain the damaged DNA, continue their growth cycle, divide, and thus propagate and accumulate the damage, eventually turning into cancerous cells. Some of us developed a new method called hypothesis driven-SNP search (HyDn-SNP-S), and used it to uncover SNPs on DNA polymerases ^{14, 15}. Our method uncovered over 75 cancer—associated SNPs on various DNA polymerases. In addition, we characterized the structural/functional impact on the structure of a specific mutation on DNA Polymerase λ resulting from a breast cancer-associated SNP¹⁵.

In this work we extend our investigation to uncover and characterize SNPs related to other cancer phenotypes on DNA repair genes (including DNA polymerases) based on cancer case-control studies available in the dbGAP database (project access request #12236)¹⁴. Following the HyDn-SNP-S method¹⁵, we apply a hypothesis driven approach and uncover SNPs in the genes of interest, i.e. DNA repair genes. By narrowing the gene pool tested, the error of finding false positive associations and the multiple testing penalty are reduced when compared to GWAS. Therefore, a targeted approach provides useful SNP leads for further investigation. Subsequently, the uncovered SNPs are classified as intronic or exonic. SNPs resulting in missense mutations in the protein product are further investigated by all-atom classical molecular dynamcis (MD) methods if high-resolution crystal structures of that particular protein are available. This subsequent characterization allows the determination of whether and/or how the mutation arising from the cancer-related SNP affects the protein structure and/or function.

Methods

SNP Search and Statistical Analysis

Data for analysis was obtained from the NCBI dbGAP database (project access #12236)¹⁴. Disease phenotypes analyzed include breast cancer (study reference id: phs000147^{16, 17}; cases: 1145; controls: 1142), prostate cancer (study reference id: phs000207¹⁸; cases:1172; controls:1157), lung cancer (study reference id: phs000336¹⁹; cases: 5739; controls: 5848), pancreatic cancer (study reference id: phs000206^20–22; cases: 5533; controls: 3904), chronic lymphocytic leukemia (CLL) (study reference id: phs000802²³ and phs000818²³; cases: 2178; controls: 2685), diffuse large B-cell lymphoma (DLBCL) (study reference id: phs000889²⁴ and phs000818²³; cases: 2661; controls: 2685), follicular lymphoma (FL) (study reference id: phs000890²⁵ and phs000818²³; cases: 2142; controls: 2685), and marginal zone lymphoma (MZL) (study reference id: phs000891²⁶ and phs000818²³; cases: 825; controls: 2685).

SNPs in the DNA repair genes were selected utilizing previous compilations^27–29, in particular the following gene families were included in the present analysis: Alkb human homologs (ALKBH1, ALKBH2, ALKBH3, ALKBH4, ALKBH5, ALKBH6, ALKBH7, ALBH8, FTO), mismatch repair (EXO1, MLH1, MLH3, MSH2, MSH6, PMS1, PMS2, MSH4, MSH5), DNA polymerases (DNTT, MAD2L2, POLA1, POLA2, POLB, POLD1, POLD2, POLD3, POLD4, POLE, POLE2, POLE3, POLE4, POLG, POLG2, POLH, POLK, POLL, POLM, POLN, POLQ, REV1, REV3L), ten-eleven translocation enzymes (TET1, TET2, TET3, APO1, APO2, APO3A, APO3B, APO3C, APO3D, APO3F, APO3G, APO3H, APO4), base excision repair (UNG, SMUG1, MBD4, TDG, OGG1, MUTYH, MPG, NEIL1, NEIL2, NEIL3, APEX1, APEX2, LIG3, XRCC1, PNKP, APLF), and poly ADP-ribose polymerases (PARP1, PARP2, PARP3). Statistical analysis to detect associations between the SNPs of interest and cancer phenotypes was performed with the logistic regression model. Four inheritance models were considered: multiplicative, additive, dominant, and recessive³⁰. Haplotype analysis for a selected group of SNPs was performed with the haplo.stats package³¹.

Structural and Dynamics Analysis of Missense Mutations

Two enzymes with SNPs resulting in missense mutations were selected to characterize the impact of the cancer variant on the structure/function using classical molecular dynamics (MD) simulations. We have previously reported the characterization of the ALKBH7 variant resulting from rs6540³². Here we present the characterization of the cancer variants for two DNA polymerases: Pol ν (SNPs rs9328764, rs10011549, and rs10018786) and Pol ι (SNP rs3218784). The corresponding mutations for each variant were introduced with the tleap tool of AMBER16³³.

Molecular dynamics simulations were performed in triplicate for the wild type and all mutant structures. All systems were prepared as apo- or holo- binary structures, that is, the holo-structures include the DNA substrate without the incoming nucleotide triphosphate. Crystal structures for Pol ν (PDB ID: 4XVK³⁴) and for Pol ι (PDB ID: 3GV8³⁵) were obtained from Protein Data Bank³⁶. Coordinates of missing loops in the crystal structures were modeled by MODELLER³⁷. For Pol ν, a magnesium ion was added to the active site of the crystal structure; the structures with DNA (holo-) and without DNA (apo-) were made by either leaving or deleting DNA atoms from crystal structures.

Hydrogen atoms were added to the structures using Molprobity^{38, 39}. The pmemd.cuda⁴⁰ program of AMBER16³³ with protein.ff14SB and DNA.OL15 force fields^{41, 42} was used for molecular dynamics simulations. All systems were subjected to MD for 200 ns for each replicate trajectory, with a 1 fs time step. Water molecules were modeled with the TIP3P force field using periodic boundary conditions. All systems were created with a minimum distance from the surface of the protein to the edge of the box of 15 Å⁴³. The smooth particle mesh Ewald (sPME) method with an 8 Å real space cutoff was used for modeling the electrostatic interactions⁴⁴. The water density was brought to 1 g/ml by a restrained MD in the NPT ensemble. Subsequently, calculations were performed in the NVT ensemble at 300K temperature with a Langevin thermostat. The cpptraj tool of AMBER16³³ was used for trajectory and hydrogen bonds analyses. Chimera software⁴⁵ was used for visualization and structure analysis and editing.

Results and Discussion

Statistical Analysis

Five hundred and sixty two SNPs on DNA repair genes are found to have a statistically significant (p<0.05) association with cancer phenotypes for at least one of the inheritance models (see Tables S.1 to S.8). Thirty one of these SNPs are associated with the breast cancer phenotype, 26 with prostate, 134 with lung, 64 with pancreatic, 116 with CLL, 82 with DLBL, 69 with FL, and 40 with MZL. Out of these 564 SNPs only 43, i.e. around 8%, correspond to missense mutations (see Table 1).

TABLE 1.

Missense mutation SNPs linked to cancer phenotypes (p-values of at least one inheritance model < 0.05). Data for polymerase genes for breast and prostate cancers are from Swett et al¹⁵.

	breast	prostate	lung	pancreatic	CLL	DLBCL	FL	MZL
ABH7		rs7540
MSH3			rs184967 rs26279
POLE					rs4883543 rs4883544
POLA2								rs487989
POLD1							rs1726801
POLG						rs2307441		rs3087374
POLG2		rs1427463
POLI		rs8305
POLL	rs3730463				rs3730476	rs3730463 rs3730476	rs3730476
POLN	rs9328764 rs10011549 rs1001878							rs11937432 rs2353552
POLQ	rs3218651	rs3218651		rs3218651	rs3218642
REV3L			rs462779
TET1	rs12221107		rs12221107 rs16925541		rs3998860			rs3998860
APO2						rs2076472	rs2076472
APO3G			rs8177832
APO3H			rs139298		rs139297	rs139297
APO4				rs1174657 rs1174658		rs16861394
MBD4				rs140693		rs2307293
MUTYH						rs3219484	rs3219484	rs3219484
NEIL2					rs8191613 rs8191664
NEIL3		rs13112390 rs13112358	rs13112358	rs1876268	rs1876268	rs13112358 rs13112390 rs1876268
APEX1							rs1130409
APEX2						rs2301416
PARP1			rs1136410				rs1136410
PARP2				rs3093921				rs3093921

The POLN gene was found to have three missense mutations linked to breast cancer (Table 1). Haplotype analysis of POLN and POLI genes based on the patients’ data used in this study has been previously published and did not detect haplotypes of genetic variants from these genes¹⁵. To check if the combinations of the missense mutations in the POLN gene constitute haplotypes linked to breast cancer all double, triple and quadruple combinations of the POLN genes that were individually found to be linked to breast cancer were analyzed using the haplo.stats package, (Table 2). Each of the haplotypes contained only one missense causing variant.

TABLE 2.

Haplotypes of POLN gene associated with the breast cancer phenotype. SNP reference IDs marked with * are missense mutations.

SNPs: rs9328764* rs2353552					SNPs rs10011549* rs11937432
Haplotype Effect Model: additive					Haplotype Effect Model: additive
global-stat = 6.83406, df = 2, p-val = 0.03281					global-stat = 6.3784, df = 2, p-val = 0.0412
Haplotype-specific Scores					Haplotype-specific Scores
loc-1	loc-2	Hap-Freq	Hap-Score	p-val	loc-1	loc-2	Hap-Freq	Hap-Score	p-val
A	C	0.12	−2.36	0.02	A	T	0.07	−2.37	0.02
G	A	0.12	−0.83	0.41	G	T	0.05	−0.73	0.46
G	C	0.76	2.38	0.02	A	C	0.88	2.3	0.02
SNPs: rs9328764* rs11937432					SNPs rs10011549* rs11937432
Haplotype Effect Model: additive					Haplotype Effect Model: recessive
global-stat = 7.92266, df = 3, p-val = 0.04764					global-stat = 7.99672, df = 3, p-val = 0.04608
Haplotype-specific Scores					Haplotype-specific Scores
loc-1	loc-2	Hap-Freq	Hap-Score	p-val	loc-1	loc-2	Hap-Freq	Hap-Score	p-val
A	A	0.06	−2.45	0.01	G	T	0.05	−1.66	0.1
G	A	0.05	−0.83	0.41	A	T	0.07	−1.48	0.14
A	G	0.88	2.39	0.02	A	C	0.88	2.15	0.03
SNPs rs9328764* rs11937432					SNPs rs10018786* rs2353552
Haplotype Effect Model: recessive					Haplotype Effect Model: additive
global-stat = 8.03298, df = 3, p-val = 0.04534					global-stat = 6.21471, df = 2, p-val = 0.04472
Haplotype-specific Scores					Haplotype-specific Scores
loc-1	loc-2	Hap-Freq	Hap-Score	p-val	loc-1	loc-2	Hap-Freq	Hap-Score	p-val
G	A	0.05	−1.74	0.08	G	C	0.12	−2.23	0.03
A	A	0.06	−1.1	0.27	T	A	0.12	−0.83	0.41
A	G	0.88	2.27	0.02	T	C	0.76	2.29	0.02

A significant finding from our statistical analysis is the relatively large number of missense SNPs on DNA repair enzymes that are found to be associated with more than one cancer phenotype. A total of 109 SNPs were found to be statistically significantly associated with multiple cancer types (TABLE S.9), with 43 of those corresponding to exonic-nonsynonymous mutations. Out of these, SNPs on POLL, POLQ, TET1, APO2, APO3H, MUTYH, NEIL3, PARP1 and PARP2 all have at least one exonic-nonsynonymous SNP that is statistically significantly associated with two or more cancer phenotypes. That is, over 30% of the genes that have a cancer-associated missense mutation, share this mutation with more than one cancer phenotype. Moreover, our results indicate that there are four different enzymes, DNA polymerase θ, APOBEC3H, MUTY and NEIL1, with the same SNP statistically significantly associated with three different cancer phenotypes (see Table 1). It should be noted that our search procedure produces results that are consistent with recently-developed approaches such as phenome-wide association studies, which have been developed to evaluate the impact of one or a small number of genetic variants with phenotypic data⁴⁶.

Some of these SNPs have been previously reported to be statistically significantly associated with one of a subset of the cancer phenotypes (breast, lung, prostate or melanoma) ¹⁵ including two SNPs on the POLL gene, which codes for DNA polymerase λ (Pol λ). One of these SNPs rs3730463, translated to T221P on Pol λ, was previously linked to breast cancer. Our new analysis reveals that this same SNP is also linked with DLBCL (TABLE 1). A haplotype of this SNP coupled with rs3730477, translated to R438W, was previously identified to be associated with breast cancer¹⁵. Moreover, SNP rs3730477 has been experimentally shown to be present in significantly higher ratios in germline DNA of breast cancer patients and has been confirmed to be a risk factor for estrogen-driven breast cancer ⁴⁷.

Structural and Dynamics Analysis of Missense Mutations on Polymerase ν Gene

The SNPs found to be linked to cancer phenotypes in the POLN, and POLI genes were selected for further investigation of the impact of the mutation on the protein structure by computational simulations based on deposited crystal structures available in the protein data bank³⁶. This analysis has been applied previously to cancer variants of Pol λ and ALKBH7.^15,32 Our previously reported computational simulations on the ALKBH7 variant resulting from the prostate cancer associated SNP predicted that the structural changes would preclude co-substrate binding, which was confirmed experimentally in the same work as discussed in Ref³².

Three SNPs that result in missense mutation on the POLN gene, which codes for DNA Polymerase ν (Pol ν), were found to be statistically linked to breast cancer¹⁵ (Table 1). These missense SNPs, rs10018786, rs10011549, and rs9328764, are located within the available crystal structures of Pol v. These three SNPs result in Pol ν variants M310L, G336S, and R425C respectively. (Figure 1).

Figure 1.

Missense mutations resulting from rs9328764, rs10011549, and rs10018786 mapped onto the structure of Pol ν with DNA bound (pdbid 4XVK). DNA is shown in blue.

The molecular dynamics simulations of the holo and apo structures for each of the three mutants revealed changes in the fluctuation of specific regions with respect to the wild type (WT) Pol ν (Figure 2). In particular, a significant increase in the fluctuation of residues 493-509 is observed in the holo structures of R425C and M310L mutants (Figure 2a). This is an area adjacent to the DNA (Figure 2c). In the same region of the apo structures the wild type and R425C mutant show higher fluctuations with larger change in the magnitude of the fluctuations, as well as the number of the residues involved (Figure 2, b and d). There is also a difference in fluctuation for both the holo and apo structures in the loop comprising residues 256-270. Residues in this region in the holo structure of wild type Pol ν and apo structures of R425C and M310L mutants have higher RMSF. Higher fluctuation was also revealed in the 593-603 region of the holo structures of the wild type and R425C mutant.

Figure 2.

Root mean square fluctuation (RMSF) per residue for the a) holo and b) apo systems of WT Pol ν and three cancer variants. c) holo and d) apo structures with regions exhibiting altered fluctuations highlighted in red.

Interestingly, the regions with altered fluctuations compared with WT Pol ν are located on opposite ends of the protein relative to the sites of the cancer-related mutations, and include loop structures. The area comprising residues 482-515 is located in proximity to the DNA, with T514 forming hydrogen bonds to the DNA backbone in the holo structure. The M310L and G336C mutants having decreased fluctuations in this region in the apo structures, suggesting that these variants may have hindered or incorrect DNA binding. The altered fluctuations in the loops could be indicative of tertiary structure changes caused by the mutations. The change in fluctuations is not limited to the protein structure, indeed, an increase in the mobility of the DNA is observed for the M310L variant compared to the WT structure (Figure S2).

Although all three variants show changes with respect to WT, the M310L mutant shows larger changes in RMSF and RMSD for certain regions compared with the other variants. In particular, residue-wise correlation analysis suggests a change in the motion of several residues in the M310L variant compared with the WT in both the holo and apo structures, consistent with the RMSF results. As evident from the brighter red hue in the difference correlation matrix for the holo structure (Figure 3a), the area near residue 493, which also exhibits higher fluctuation in the M310L and R425 variants (Figure 2a), has a change in the dynamics. In addition, a region comprising residues 678-753 (Figure 3) shows a change in correlation relative to the residues 468-518 with respect to WT (Figure 3b). The two regions with altered correlations correspond to areas in the fingers and thumb domains, which contact the DNA in the holo structure (Figure 4).

Figure 3.

Difference correlation plots, comparing correlation between residues in M310L mutant and the wild type structures for a) holo and b) apo Pol ν.

Figure 4.

Regions with changes in correlation for M310L Pol ν. Residues 468-518 are shown in orange, 678-753 in green, and DNA is shown in blue.

Structural and Dynamics Analysis of Missense Mutations on Polymerase ι Gene

The POLI gene codes for DNA polymerase ι (Pol ι), and was found to have a SNP linked to melanoma, rs3218784, as previously reported in Ref. ¹⁵. This genetic variant results in a I236M missense mutation (Figure 5).

Figure 5.

Pol ι crystal structure 3GV8 with the mutation site, I236, highlighted in red.

Our MD simulations revealed changes in the dynamics of Pol ι I236M. As can be seen from Figure 6, several residues exhibit higher fluctuations in the cancer-related variant including regions comprising residues 142-152, 225-234, and 240-277 (highlighted in green in Figure 7). Similarly, the melanoma mutation results in decreased fluctuations in two regions, namely residues 350-357, and 369-377 (highlighted in purple in Figure 7). The residues with increased fluctuations in the I236M variant are located in the thumb domain and are close to the mutation site. RMSF analysis for the DNA substrate for the wild type and Pol ι I236M binary complexes indicates that the DNA has significantly smaller fluctuations in the cancer variant structure (Figure 8). Thus, the I236M Pol ι variant arising from the melanoma-associated mutation affects the dynamics of the protein in two regions, which in turn results in a decrease in the fluctuation of the DNA in the binary structure. This change in fluctuations could result in a negative effect on the protein’s function.

Figure 6.

RMSF plots for wild type and I236M mutant Pol ι.

Figure 7.

RMSF difference between WT and I236M Pol ι. Regions that show an increase(decrease) in RMSF in I236M Pol ι with respect to WT are shown in green(purple). The DNA substrate is shown in blue ribbons and the mutation site in orange sticks.

Figure 8.

RMSF for DNA for the wild type and mutant Pol ι complexes.

Taken together, our computational simulations for these two enzymes with four cancer variants suggest that these mutations affect the dynamics of the protein, which in turn could affect the function of these important enzymes. It would be interesting to experimentally confirm whether these predicted effects indeed result in a change in the function of the cancer variants.

Conclusions

We have used our previously developed HyDn-SNP-S method to search for cancer mutations on 25 different DNA repair genes related to 8 different cancer phenotypes. Our analysis uncovered 562 total SNPs on DNA repair genes that are statistically significantly associated with at least one of the tested cancer phenotypes, 43 of which result in missense mutations on the final protein product. Moreover, our analysis uncovered that out of the 562 total SNPs, 109 are associated with more than one cancer phenotype, and nine out of the 25 tested DNA repair genes involve a missense SNP that is associated with two or more cancer types, and four of these genes (Polθ, APOBEC3H, MUTY and NEIL1) have the same SNP associated with three cancer phenotypes. The subsequent all-atom MD characterization of four cancer variants on Pol ν and Pol ι provide insights into the possible effects of these cancer mutations on the structure and function of the respective enzymes. In particular, we investigated three breast cancer-related missense mutations on Pol ν: M310L, R425C, and G336S. In all cases, the mutations are observed to affect the fluctuations and correlation of various regions of the thumb and finger domains of Pol ν. One more melanoma-related missense mutation on Pol ι, resulting in the I236M variant, also exhibits altered dynamics. In particular, a decrease in fluctuations in the fingers domain, which in turn also results in a decrease in the observed motion of the DNA substrate in the binary structure. Overall, our results underscore the importance of DNA repair genes and their relation to cancer as observed by the large number of single mutations related to multiple cancer phenotypes, as well as the possible effects of the resulting mutations on the protein structure and function.