Correlating observed odds ratios from lung cancer case-control studies to SNP functional scores predicted by bioinformatic tools

Abstract

Bioinformatic tools are widely utilized to predict functional single nucleotide polymorphisms (SNPs) for genotyping in molecular epidemiological studies. However, the extent to which these approaches are mirrored by epidemiological findings has not been fully explored. In this study, we first surveyed SNPs examined in case-control studies of lung cancer, the most extensively-studied cancer type. We then computed SNP functional scores using four popular bioinformatics tools: SIFT, PolyPhen, SNPs3D, and PMut, and determined their predictive potential using the odds ratios (ORs) reported. Spearman’s correlation coefficient (r) for the association with SNP score from SIFT, PolyPhen, SNPs3D, and PMut, and the summary ORs were r = −0.36 (p = 0.007), r = 0.25 (p = 0.068), r = −0.20 (p = 0.205), and r = −0.12 (p = 0.370) respectively. By creating a combined score using information from all four tools we were able to achieve a correlation coefficient of r = 0.51 (p < 0.001). These results indicate that scores of predicted functionality could explain a certain fraction of the lung cancer risk detected in genetic association studies and more accurate predictions may be obtained by combining information from a variety of tools. Our findings suggest that bioinformatic tools are useful in predicting SNP functionality and may facilitate future genetic epidemiological studies.

1. Introduction

SNP-disease association research in the emerging field of genetic and molecular epidemiology has been driven by the candidate gene and genome-wide approaches. In the candidate gene approach, an interesting SNP will typically be investigated in a population-based study if a plausible biological mechanism, which relates the gene harboring the SNP to disease etiology or progression, has been proposed. Our understanding of disease etiology is continuing to improve and with it a greater number of disease candidate genes are being discovered. Concurrent with this are numerous SNP-finding efforts that are generating millions of putative SNPs for potential association studies. The entirety of this multitude cannot be immediately assessed and therefore, the current obstacle lies in extracting useful and informative SNPs from the public databases [1]. Approaches to prioritizing these functional SNPs will tremendously enhance SNP-based association research in molecular epidemiology.

Combining modern molecular phenotypic assays with polymorphisms (i.e. SNP data) will provide an ultimate assessment of the effect that genetic variations have on the network of interacting molecular and physiologic systems under normal and pathological human conditions. This, however, is not an easy task and it currently poses one of the greatest challenges to modern medical scientists. In addition to this, functional assays are currently not available for most candidate SNPs [2].

Taking advantage of recent developments in evolutionary biology, protein structural genomics, and transcriptomics, several computational methods may be utilized to discriminate between neutral SNPs, which constitute the majority of genetic variations, and the small portion of SNPs of likely functional importance. The most straightforward approach to predicting SNP functionality involves determining the importance of the SNP’s location. For example, exonic SNPs may alter amino acid sequences and consequently influence the kinetic parameters of enzymes, the DNA-binding properties of proteins that regulate transcription, the signal transduction activities of transmembrane receptors, and the architectural roles of structural proteins [3]. SNPs may also change the sequence of DNA splicing sites in both exons and introns, and in turn, affect post-transcriptional processes [4]. Moreover, from an evolutionary perspective, SNPs altering a conserved amino acid site are more likely to have functional importance [5].

Many SNPs in candidate disease-related genes have been genotyped in molecular epidemiological studies, especially in the field of cancer research. This provides a great opportunity to validate these bioinformatic tools by correlating predicted SNP functional scores to findings from case-control studies. In this study, we first surveyed previous publications which genotyped SNPs in case-control studies of lung cancer, the most extensively examined cancer type. We then used four major bioinformatic tools, SIFT [5], PolyPhen [3], SNPs3D [6], and PMut [7] to predict the functional impact of these SNPs. Finally, we tested the hypothesis that SNPs predicted to have a significant impact on protein function are more likely to be associated with cancer risk in terms of odds ratios by correlating the functional scores to ORs obtained from actual published case-control studies.

2. Methods

2.1. Literature search

We performed a PubMed search using the keywords “lung cancer polymorphism case control,” while setting the limits to case-control studies examining SNPs and lung cancer risk, published in English from January, 1994 to December, 2006. The time restriction was imposed as few studies were identified prior to 1994, and these studies might be of questionable methodological quality. The bioinformatic tools we used make predictions about the impact of a given SNP based on the resulting amino acid substitution. As such, we imposed a further restriction on the studies to include only those which investigated non-synonymous SNPs (nsSNPs), i.e. those resulting in an amino acid change. Studies matching this additional criteria were manually identified from the larger pool of studies returned from the initial keyword search. We chose the overall odds ratio (OR) based on the dominant disease model, comparing carriers of combined heterozygous and homozygous variant allele with the homozygous wild type allele carriers.

2.2 Assessment of nsSNP functionality

The impact of nsSNPs can be assessed by evaluating the importance of the amino acids they affect. Four widely-used computational tools for determining the functional significance of nsSNPs were employed for this study. The SIFT software (blocks.fhcrc.org/~pauline/SIFT.html) was used to determine the conservation level of a particular amino acid position in a protein, which leads to a tolerance index (from 0 to 1) for SNP functionality [5]. The higher a tolerance index, the less functional impact a particular amino acid substitution is likely to have, as a higher tolerance index indicates that the position is less conserved across species. We also used the PolyPhen tool (http://coot.embl.de/PolyPhen/) to estimate the structural and functional impact of an amino acid substitution, which returns a “Position-Specific Independent Count” (PSIC) score [3]. Large values of this PSIC score indicate that the substitution is rarely or never observed in the protein family, suggesting likelihood that the amino acid replacement will be deleterious. SNPs3D (http://www.snps3d.org) uses two methods for determining whether a SNP will be deleterious to protein function[6]. The first makes predictions based on the estimated impact of the nsSNP on protein stability[8], while the second considers conservation of the given amino acid within a protein family[9]. In both cases, a negative score indicates a deleterious substitution, while a positive number corresponds to a neutral change, with greater absolute values indicating stronger confidence in the prediction. In cases where both scores were available for a given SNP, we used the more confident prediction. No score was available from either model for 12 of the 54 SNPs under study. The final bioinformatic tool we employed was PMut (http://mmb2.pcb.ub.es:8080/PMut/). This tool uses neural networks trained using a large database of disease-associated and neutral SNPs to predict the impact of a given amino acid substitution[7,10]. We used the prediction tool which returns a neural network output value from 0 to 1, with 0 corresponding to a “highly reliable” neutral prediction and 1 corresponding to a “highly reliable” deleterious prediction. As values approach 0.5 they become less robust. All searches performed on each of the four tools were performed using the respective default parameters.

2.3 Calculation of summary score

In order to incorporate data from all four tools, a summary score was created which combined the scores returned from each individual algorithm. In order to provide a meaningful summary value, an equation had to be generated which took into consideration the direction corresponding to a deleterious SNP, along with the range of possible output values for a particular tool. Since larger values for the scores generated by Polyphen and PMut indicate deleterious SNPs, these were included together in the numerator. Furthermore, since PMut scores take on a value between 0 and 1, while PolyPhen scores can be greater than 2, Polyphen scores were divided by the PMut score, to create one value in the numerator which would increase as confidence in a deleterious prediction increased. Since larger SIFT and SNPs3D values correspond to neutral SNPs, these were placed in the denominator of the final equation and multiplied by one another. In order to generate comparable values from the numerator and denominator portions of the summary equation, we took the square root of the numerator and the square of the denominator. The resulting range of values for the numerator piece of the equation was 0.39 to 6.46, while the range for the denominator was 0 to 5.85, indicating fairly equal contributions from the two halves of the equation. The net effect of this final equation was to create one summary value which drew on information from each of the 4 tools, and would take on larger values with increasing confidence in a deleterious prediction for each SNP. The final equation was as follows:

$$\frac{\sqrt{(\text{PolyPhen}/\text{PMut})}}{{(\text{SIFT}\ast \text{SNPs}3\text{D})}^2}$$ (1)

2.4 Statistical analysis

All statistical analyses were performed using the STATA statistical software (StataCorp.; College Station, TX) unless otherwise specified. Meta-analysis was first performed to estimate summary ORs for SNPs examined in multiple studies in order to generate one data point for each SNP in the subsequent correlation analysis. Overall ORs were calculated using a random-effects model [11] which includes both within- and between-study variations. Briefly, ORs, standard errors, and 95% Confidence Intervals (CIs) were calculated for all studies using published frequencies for cases and controls of the genotypes of interest. Weighting was applied in the calculation by using the inverse of an individual ORs’ variance in order to take into account the quality of information available (e.g. the sample size of the study and the precision of the point estimate). If a SNP had been found to be protective (OR < 1) in a study, we re-expressed the odds ratio in terms of the risk genotypes (OR > 1) in order to facilitate the comparisons, which were based on the strength of the association, rather than the direction. Neither the odds ratios nor their natural logarithms followed a Gaussian distribution, so correlations were determined by nonparametric methods (Spearman’s rank correlation) using the re-expressed ORs. In addition to correlating the scores from each bioinformatic tool to the observed odds ratios for all SNPs, Spearman’s rank correlation coefficients were also calculated to assess the correlation among the various functional scores obtained for each SNP.

Results

Our PubMed search identified 51 case-control studies examining a total of 54 nsSNPs in 37 different genes for risk estimates of lung cancer (Table 1). All of these SNPs are located in the coding regions of cancer related-genes, such as those involved in DNA repair, metabolism, and cell cycle checkpoints.

Table 1.

List of SNPs investigated in lung cancer case-control studies.

Gene	Amino Acid change	SNP ID	OR*	SIFT Score	PSIC Score	SNPs3D Score	PMut Score	Reference
ADRB2	Gly16Arg	rs1042713	1.01	0.40	1.353	1.96	0.44	[16]
APE1/APEX1	Gln51His	rs1048945	1.04	0.12	0.093	0.53	0.0341	[17]
APE1/APEX1	Ile64Val	rs2307486	0.60	0.33	0.911	0.82	0.0226	[17]
APE1/APEX1	Asp148Glu	rs1130409	0.92	1	0.467	1.42	0.0876	[17–19]
ATR	Thr211Met	rs2227928	0.64	0.21	0.980	-	0.3969	[17]
CDKN1A	Ser31Arg	rs1801270	0.92	0.80	0.488	2.58	0.7492	[20]
CYP1A1	Ile462Val	rs1048943	1.35	0.05	0.785	−0.64	0.0485	[21–28]
CYP1B1	Leu432Val	rs1056836	1.78	0.13	0.277	−0.25	0.2893	[27,29]
EPHX1	His113Tyr	rs1051740	0.69	0.01	2.261	−0.78	0.2091	[30,31]
EPHX1	His139Arg	rs2234922	0.74	0.49	0.141	0.07	0.166	[30,31]
ERCC2/XPD	His201Tyr	rs1799792	1.05	1	0.350	1.95	0.1941	[17]
ERCC2/XPD	Asp312Asn	rs1799793	1.03	0.58	0.333	2.03	0.345	[17,19,32,33]
ERCC2/XPD	Lys751Gln	rs13181	1.29	0.84	0.472	2.88	0.697	[17,19,32–34]
ERCC4/XPF	Arg415Gln	rs1800067	1.20	0.28	1.892	−0.91	0.6268	[17]
ERCC5/XPG	Cys529Ser	rs2227869	1.32	0.80	0.660	1.63	0.2999	[32]
ERCC5/XPG	His1104Asp	rs17655	1.11	0	1.628	−0.98	0.6505	[32,35,36]
EXO1	Glu589Lys	rs1047840	0.81	0.72	0.252	-	0.5013	[17]
GSTP1	Ile105Val	rs1695	1.07	1	0.166	1.36	0.0276	[29,37–41]
GSTP1	Ala114Val	rs1138272	1.40	0.12	1.455	1.22	0.498	[40]
HDAC5	Asp593Glu	rs228757	1.19	1	0.114	1.04	0.0274	[42]
LIG4	Ile658Val	rs2232641	1.20	0.60	0.432	1.13	0.0402	[35]
MYCL1	Ser362Thr	rs3134614	0.68	0.04	1.235	-	0.0296	[43]
MBD1	Pro345Ala	rs125555	1.29	0.40	0.866	−0.7	0.8374	[44]
MGMT	Leu84Phe	rs12917	1.12	0.71	0.024	−1.16	0.1301	[17,45]
MGMT	Ile143Val	rs2308321	1.35	0.70	0.193	0.78	0.0332	[17,46]
MGMT	Lys178Arg	rs2308327	1.24	0.52	0.112	0.64	0.0701	[17,47] [47]
MMP9	Arg279Gln	rs17576	1.29	0.35	1.537	2.28	0.2577	[48]
MMP9	Pro574Arg	rs2250889	1.61	0.50	0.764	2.4	0.8465	[48]
MMP9	Arg668Gln	rs2274756	1.04	1	0.084	1.9	0.5389	[48]
MTHFR	Ala222Val	rs1801133	0.91	0.04	1.266	−2.01	0.3939	[49]
MTHFR	Ala429Glu	rs1801131	1.12	0.28	0.786	-	0.6345	[49]
MTHFR	Arg594Gln	rs2274976	1.23	0.06	1.729		0.3811	[49]
MTR	Asp919Gly	rs1805087	0.93	0.33	0.390	1.29	0.7541	[49]
MTRR	Leu22Met	rs1801394	1.36	0.03	1.267	−	0.1041	[49]
NBS1	Glu185Gln	rs1805794	1.27	0.64	0.624	1.78	0.0689	[17,50,51]
NQO1	Pro187Ser	rs1800566	1.1	0.40	1.499	−1.29	0.1403	[52–55]
OGG1	Ser326Cys	rs1052133	1.19	0.22	0.397	−0.46	0.707	[17,29,56]
p53	Arg72Pro	rs1042522	1.20	0.54	0.521	2.63	0.7672	[20,35,57–59]
PARP1	Val762Ala	rs1136410	1.26	0.66	1.357	0.54	0.2393	[60]
PARP1	Lys940Arg	rs3219145	1.30	0.21	1.678	−	0.0431	[35]
PDZ-RhoGEF	Ser1416Gly	rs868188	0.90	0.12	1.323	-	0.2736	[61]
POLB	Pro242Arg	rs3136797	2.31	0.07	2.587	0.97	0.4716	[17]
POLI	Thr706Ala	rs8305	1.17	0.67	0.227	-	0.3617	[17,35]
RAD23B	Ala249Val	rs1805329	1.81	0.29	1.060	0.11	0.6915	[32]
REV1	Phe257Ser	rs3087386	1.00	0.90	0.342	2.25	0.0736	[35]
SOD2	Ala16Val	rs4880	1.24	0.53	0.173	-	0.5608	[58]
SULT1A1	Arg213His	rs9282861	1.36	0.07	1.725	0.71	0.465	[62,63]
XPC	Ala499Val	rs2228000	1.11	0.73	0.346	-	0.3404	[32,64,65]
XPC	Lys939Gln	rs2228001	0.99	0	2.068	-	0.2929	[32,64,65]
XRCC1	Arg194Trp	rs1799782	0.99	0.11	2.495	1.47	0.7963	[17,66]
XRCC1	Arg280His	rs25489	1.11	0.14	1.595	2.1	0.5637	[17,66]
XRCC1	Arg399Gln	rs25487	0.97	0.80	0.201	1.25	0.8085	[17–19,51,58,60,66]
XRCC2	Arg188His	rs3218536	2.82	0.20	0.631	2.3	0.1764	[17]
XRCC3	Thr241Met	rs861539	0.86	0.61	1.068	1.79	0.3058	[17,19,51]

^*Only summary OR is listed for a SNP investigated in multiple studies.

Correlations between each of the SNP functional scores obtained for each nsSNP were assessed by calculating Spearman’s rank correlation coefficients. Significant correlations were found between the SIFT tolerance index and PSIC score (r = −0.607, p < 0.001), as well as SIFT and SNPs3D scores (r = 0.348, p = 0.024). PMut scores did not significantly correlate with any other scores (Table 2).

Table 2.

Correlations among predicted SNP functionality scores and observed odds ratios.

	Sift	PolyPhen	SNPs3D	Pmut	Combined
OR	−0.361	0.250	−0.200	−0.124	0.513	Rho*
	0.007	0.068	0.205	0.370	<0.001	p-value*
	54	54	42	54	42	N
Sift		−0.607	0.348	−0.013	−0.828	Rho*
		<.001	0.024	0.923	<.001	p-value*
		54	42	54	42	N
PolyPhen			−0.154	0.181	0.488	Rho*
			0.330	0.189	0.001	p-value*
			42	54	42	N
SNPs3d				0.191	−0.636	Rho*
				0.225	<.001	p-value*
				42	42	N
PMut					−0.150	Rho*
					0.344	p-value*
					42	N

^*Rhos and p-values are for Spearman’s rank correlation coefficient.

SIFT, PSIC, SNPs3D, and PMut scores were also correlated to the observed ORs associated with the corresponding nsSNPs detected in molecular epidemiologic studies of lung cancer. Spearman’s rank correlation showed that of the four tools, SIFT scores were most strongly correlated with observed ORs (r = −0.361, p = 0.007). PolyPhen scores were modestly correlated (r = 0.250, p = 0.068), while SNPs3d and PMut demonstrated very weak associations (r = −0.200, p = 0.205, and r = −0.124, p = 0.370), respectively (Figure 1). The summary score, which considers the magnitude of the scores returned from each tool, as well as which direction corresponds to a deleterious substitution, was computed for each of the 42 nsSNPs for which scores were available from all four tools (N = 42). Spearman’s rank correlation coefficient for this summary score and the observed ORs was r = 0.513 (p < 0.001) (Figure 2).

Fig. 1.

Correlation between log of observed ORs and SNP scores from each of the four tools.

Fig. 2.

Correlation between log of observed ORs and log of combined SNP scores

Discussion

One unsolved issue of SNP genotyping in molecular epidemiological studies is how to choose target SNPs for investigation, since functionalities of most SNPs are unknown. Though haplotype tagging SNPs (htSNPs) generated from the HapMap project tremendously facilitates genotyping for genetic association studies, the ultimate goal of such studies is still to locate functional genetic changes linked to htSNPs in candidate genes. Given the lack of phenotypic data for most of the SNPs identified, results from our analyses suggest that bioinformatic tools based on recent findings from evolutionary biology, protein structure research, and computational biology may provide useful information in assessing the functional importance of SNPs.

Our findings are congruent with our previous observations, which demonstrated that SNPs altering conserved amino acids assessed by the SIFT tool are more likely to be associated with cancer risk [12]. The current study expanded on the previous analysis in three important ways, resulting in more meaningful and interpretable results. First, we included functional scores from SIFT, PolyPhen, SNPs3D, and PMut tools, each of which employs fundamentally different algorithms that can be used to assess the functionality of the same nsSNPs. Second, we restricted the analyses to case-control studies investigating lung cancer only, as the importance of a given gene in carcinogenesis may vary significantly across cancer types. Moreover, lung cancer is the most extensively studied cancer type, which allowed us to collect a reasonable number of studies for our analysis. Third, meta-analysis was performed to obtain summary ORs for SNPs examined in multiple studies in order to generate one data point for each SNP in the analysis.

While previous studies have investigated the predictive power and accuracy of computational approaches[13,14], this analysis is unique in that it draws on risk estimates derived directly from human epidemiological studies. While correlations based on predicted protein functionality or mutagenesis studies are important, we believe that those obtained by observational studies of human risk are even more relevant. Our data suggests that individual tools correlate modestly with observed odds ratio, and that combining information from a variety of tools may significantly increase the predictive power for determining the functional impact of a given nsSNP.

Our results demonstrate a significant agreement between the SIFT and PolyPhen tools, as well as SIFT and SNPs3D in the prediction of SNP functionality. This finding supports results from a previous analysis, which showed that the predicted scores of SNP functionality from the two algorithms are highly associated, with concordance in the predicted impact observed for approximately 62% of the variants [15].

However, while the combined scores yielded a respectable correlation coefficient (r) with observed ORs, the correlations for the individual tools detected in our analyses were fairly modest. This implies that the functional impact of a SNP on a gene may not be predicted 100% correctly by computational tools. It might also indicate that molecular phenotype (e.g. functional SNPs) may not always penetrate through to clinical phenotype (e.g. ORs). Furthermore, lung cancer is a complex environment-related cancer type, and disease-associated SNPs may only trigger tumorigenesis in the presence of certain environmental exposures such as tobacco carcinogens.

Although the correlation we observed is readily apparent, the current study has several limitations. First, there is the concern that many of the SNPs under study may not in fact be singly causal, but may associate with disease in more complicated ways. While linkage disequilibrium between the SNPs under study and other causative SNPs cannot be ruled out, each of these tools make predictions for missense mutations only, and base their predictions on the impact of the amino acid change on protein function. In addition, the ORs reported for each of the SNPs included in the analysis were based on studies designed using a candidate gene approach, rather than an array-based method. As such, the SNPs under study lie within genes that have established relevance to carcinogenesis such as DNA repair and metabolic genes, and are therefore likely to affect cancer risk directly. Couple this with the fact that the tools make predictions based on the specific SNP’s effect on protein function, and the result should be a fairly specific relationship between the bioinformatic prediction and the epidemiologic finding. Furthermore, the utility of these tools and others like them, lies in their ability to make predictions on the functional impact of an amino acid change for which little or no direct experimental evidence is available, and then to use this information in further studies to relate genetic variations to disease etiology. It is possible that a SNP predicted harmful does not show an association with cancer because multiple factors, such as gene-gene and/or gene-environment interaction are involved in tumorigenesis. It is also possible that a SNP predicted benign is actually associated with cancer, if the SNP is genetically linked to adjacent causal variations. Second, the accuracy with which a SNP predicts cancer risk depends on the alignment obtained in some bioinformatic tools. For example, the SIFT method depends on homologous sequence alignments among different species. The number of sequences available from different species may be different from gene to gene, which in turn may affect the accuracy of the prediction. The PolyPhen and SNPs3D predictions depend on protein structure information available for a given gene product, which may be of varying quality across the human proteome.

Nevertheless, our data indicates that bioinformatic tools are indeed useful in predicting the functional impact of SNPs, as our findings could explain a respectable fraction of the lung cancer risk that has been detected. Although these algorithms have been developed based on empirical data, correlations between predictive scores and findings from human studies have not been explored, with the exception of SIFT. The results of our study have therefore provided novel evidence of the correlation using human data, which in turn facilitate genotyping efforts in future molecular epidemiological studies and provide targets for phenotypic analysis of genetic variants. These results can also be used to refine the bioinformatic algorithms. These findings warrant a more comprehensive approach that includes other cancer types and more available bioinformatic tools in future analyses.