Incorporating epistasis interaction of genetic susceptibility single nucleotide polymorphisms in a lung cancer risk prediction model.

Incorporation of genetic variants such as single nucleotide polymorphisms (SNPs) into risk prediction models may account for a substantial fraction of attributable disease risk. Genetic data, from 2385 subjects recruited into the Liverpool Lung Project (LLP) between 2000 and 2008, consisting of 20 SNPs independently validated in a candidate-gene discovery study was used. Multifactor dimensionality reduction (MDR) and random forest (RF) were used to explore evidence of epistasis among 20 replicated SNPs. Multivariable logistic regression was used to identify similar risk predictors for lung cancer in the LLP risk model for the epidemiological model and extended model with SNPs. Both models were internally validated using the bootstrap method and model performance was assessed using area under the curve (AUC) and net reclassification improvement (NRI). Using MDR and RF, the overall best classifier of lung cancer status were SNPs rs1799732 (DRD2), rs5744256 (IL-18), rs2306022 (ITGA11) with training accuracy of 0.6592 and a testing accuracy of 0.6572 and a cross-validation consistency of 10/10 with permutation testing P<0.0001. The apparent AUC of the epidemiological model was 0.75 (95% CI 0.73-0.77). When epistatic data were incorporated in the extended model, the AUC increased to 0.81 (95% CI 0.79-0.83) which corresponds to 8% increase in AUC (DeLong's test P=2.2e-16); 17.5% by NRI. After correction for optimism, the AUC was 0.73 for the epidemiological model and 0.79 for the extended model. Our results showed modest improvement in lung cancer risk prediction when the SNP epistasis factor was added.

Introduction

Lung cancer risk prediction models provide an estimate of individual's risk of developing lung cancer such that ‘at-risk’ subjects can be targeted for preventive and treatment interventions (1). Risk models hold promise for improving patient care by aiding the clinicians decision making process regarding choice of interventions and/or treatments. Risk models can also guide selection of individuals at the population level, for screening: this ensures limited resources are focussed on those individuals who are most likely to benefit. This risk guiding strategy ensures minimisation of unnecessary, invasive and potentially harmful interventions. Existing lung cancer absolute risk prediction models are mostly based on traditional epidemiological and/or clinical risk factors (2–7), limiting their predictive and discriminative abilities. For an improved precision, incorporation of genetic and molecular markers of disease in risk models has been advocated (8) and aided by recent proliferation of genetic/genomic research which has led to the identification of susceptibility genes and biological markers in many diseases (9–12).

Common gene variants involved in lung cancer have been recently identified through a number of large, collaborative, genome-wide association studies. Susceptibility genes identified to date include those on chromosomes 5p15.33, 6p21, and 15q24-25.1 (13–15). Apart from these, other genetic loci have also been identified in candidate gene association studies targeting specific molecular pathways; such as genes encoding proteins in cell cycle control, oxidant response, apoptosis, DNA repair, cell adhesion and airways inflammatory response (16,17).

While genomics research has been very fruitful in identifying these common, low-risk allelic variants, there is a growing scepticism regarding their usefulness in risk prediction. It has been shown that risk profiles generated by common low-moderate susceptibility loci, in a simple additive model, provides limited discrimination (18,19). The limited contribution of single nucleotide polymorphisms (SNPs) to risk profiling has been partly blamed on restriction to a limited number of significant alleles, methodological limitations regarding assessment of model performance and statistical approaches for incorporating the variants (19). Whilst the usual approach has been to utilise only the significant variants for risk profiling, an improved disease prediction may be attained by accounting for a large ensemble of markers (20). For the relatively few markers arising from candidate-gene studies, incorporation of the interactive effect of these genes, through epistasis modelling, may provide better predictions beyond that afforded by the limited effect of multiple loci using additive effects (21). Models including epistatic interactions take into account the complex biological relationships among the loci and extend the traditional method that focuses only on additive score using a weighted or unweighted number of risk alleles, which assume independence between the markers (22).

In the past three decades, improvements in risk prediction models brought about by the inclusion of markers and genetic factors were quantified using changes in the area under the receiving-operating characteristic curve (AUC) (23). Recently, an increasing popular measure of evaluating improvements in risk predictions, the net reclassification improvement was introduced (24). This measure involves cross-tabulating categories of predicted risk for 2 models, usually one with the new marker under study and the other without it, to see how persons are classified differently when these models are used (25).

In this study, we investigated the presence of epistasis among a panel of SNPs previously validated individually in lung cancer (26) and used both area under the receiver operating characteristic (AUC) analysis and net reclassification improvement (NRI) to assess the contribution of adding an interactive epistatic effect to an extensively validated clinical-based risk model for lung cancer.

Materials and methods

Study population

This study was performed as part of the Liverpool Lung Project (LLP). Details of recruitment procedure, study design and validation have been previously reported (3,27). Briefly, incident cases of histologically or cytologically confirmed lung cancer, ages between 20 and 80 years, were included. Lung cancer included any of topographical subcategories of code C34 of the International Classification of Disease for Oncology 9th revision. Two population controls per case, matched on year of birth (±2 years) and gender, were selected from registers of general practitioners in Liverpool area. All participants were Caucasians, residents in the Liverpool area. The study protocol was approved by the Liverpool Research Ethics Committee, and all research participants provided written informed consent in accordance with the Declaration of Helsinki.

In this study, we utilised complete genotype data on individuals included in the independent validation of SNPs identified in a candidate-gene genetic association study (26). The data comprises of 2385 subjects (cases=718, controls=1667) selected from individuals recruited into the LLP between 2000 and 2008. Of this number, 1362 (cases=418 and controls=914) were included in LLP case-control data used to develop the LLP risk model (3). Data on epidemiological, clinical and lifestyle factors were collected using a standardised questionnaire supplemented with hospital case note reviews conducted by trained LLP research nurses. Information documented includes: patients smoking status (smoking duration), previous history of pulmonary diseases (pneumonia, COPD and bronchitis), previous history of malignant diseases excluding skin melanoma, occupational exposure to asbestos, family history of lung cancer with age at onset, and case diagnosis details (date of diagnosis, histological subtype and staging).

Genetic data consist of 20 SNPs independently validated from 157 SNPs screened in a candidate-gene discovery study; details of selection and genotyping have been described elsewhere (26). Briefly, 157 candidate SNPs were screened in a discovery cohort of 439 subjects (200 controls and 239 lung cancer cases), which identified 30 SNPs associated with either the healthy smokers (protective) or lung cancer (susceptibility) phenotype. After genotyping this 30 SNP panel in a validation cohort of 491 subjects (248 controls and 207 lung cancers) and, using the same protective and susceptibility genotypes from the discovery cohort, a 20 SNP panel were selected based on replication of SNP associations in the validation cohort that includes variants in the metabolism of smoking-derived carcinogens (NAT2 and CYP2E1), inflammatory cytokines [interleukins 1(IL1B), 8(IL8), and 18(IL18), tissue necrosis factor α1 receptor (TNFR1), toll-like receptor 9 (TLR9)], smoking addiction [dopamine D2 receptor (DRD2) and Dopamine transporter 1(DAT1)], nicotine dependency [α5-nAChR (CHRNA3)], antioxidant response to smoking [α1 anti-chymotrypsin (SERPINA3) and extracellular super-oxide dismutase (SOD3)], cell cycle control, DNA repair and apoptosis (XPD, TP73, Bcl-2, FasL, Cerb1, and REV1) and integrins (ITGA11, ITGB3) implicated in apoptosis. Genomic DNA was extracted from whole blood samples by standard salt-based methods and purified genomic DNA was aliquoted (10 ng/μl concentration) into 96-well plates. Genotyping was performed on a Sequenom™ system (Sequenom Autoflex Mass Spectrometer and Samsung 24 pin nanodispenser) (26).

Statistical analysis

Characteristics of the subjects in the cases and controls were compared using t-test for continuous variables and χ2 test or Fisher's exact test for discrete variables as appropriate. Genotype and allele frequencies were checked for each SNP for Hardy-Weinberg equilibrium (HWE).

Identification of SNPs epistasis

The multifactor dimensionality reduction (MDR) and random forest (RF) were used to investigate gene-gene interactions by identifying SNP combinations that provide the best discrimination of the status of the subjects. MDR is a non-parametric, model-free method that utilises a constructive induction technique to collapse high-dimensional genetic data into a single dimension (28,29). It pools multi-locus genotypes into high and low risk groups using an exhaustive search to identify optimal combination of polymorphisms, which can then be evaluated for its ability to classify or predict disease status. In our implementation of MDR, three separate genotypes were analysed for each SNP. The Relief-F algorithm as implemented in the MDR was used as a first approach to select among the 20 SNPs that are most likely to interact. An exhaustive search of all possible 1–5 loci were then explored using 10-fold cross validation as described by Hahn et al (28). Cross-validation allows estimation of the prediction error by leaving out a portion of the data as an independent test set. With 10-fold cross-validation, the data are divided into 10 equal parts, the model was developed on 9/10 of the data (i.e. the training data) and then evaluated on the remaining 1/10 of the data (i.e. the independent testing data). This is repeated for each possible 9/10 and 1/10 of the data and the resulting ten prediction errors are averaged (29). MDR, then, seeks to find the single-locus or multi-locus predictor(s) for explaining the outcome (based on a balanced accuracy measure - the arithmetic mean of sensitivity and specificity), based on the available genomic information (30). The prediction accuracy and cross-validation consistency defined as the number of cross-validation replicates (partitions) in which that same n-locus predictor(s) was chosen as the best predictor of lung cancer status i.e. the number of replicates in which it minimised the classification error were used to select the best SNPs in each 1 to 5-locus combination (31). The overall best SNP classifier of lung cancer status was selected as the one with the maximum prediction accuracy and cross-validation consistency and evaluated statistically using 1000-fold permutation test.

For comparison, we used the freely available Willows software package for generating RF (32). RF ranks variables by a variable importance index, a measure which reflects the ‘importance’ of a variable on the basis of the classification accuracy, while considering the interaction among variables (33). A classification tree was built by the recursive partitioning method; each tree is constructed using a different cohort of bootstrap samples from the original cohort. Approximately one-third of the samples are left out of the bootstrap (oob) samples and hence not used in the construction of the tree. The number of trees was set to 10,000 and the default values of the other parameters as provided by the program were used. Several classification trees were created with replacement from the original data imput into the program. To determine the importance of an SNP, first the values of the SNP in the oob samples are randomly permuted; then both the original oob samples and the permuted oob samples are classified by the corresponding tree. The difference in the correct classification rates between the original and permuted oob samples determines the importance of the SNP, and the variable importance is obtained by averaging the differences over all trees in the random forest (32,34).

Risk model predictions and incorporation of SNPs epistasis

Risk prediction was performed using the same risk factors included in the LLP risk model (3). Multivariable logistic regression was employed to generate estimates of predicted 5-year absolute risk of lung cancer in i) a model with epidemiological data and ii) an extended model with both genetic and epidemiological data. The baseline risk (α, the constant term in the regression model) for the prediction of 5-year absolute risk using the extended model with both genetic and epidemiological data was recalculated. The method for calculating the baseline α from age- and gender-specific lung cancer incidence rates from the Liverpool area has been described (3). The only difference is that the probability model now includes information on rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11).

The area under the receiver-operating characteristics (AUC) was used to i) assess the discriminatory ability of the models, and ii) compare the models with and without SNPs. The increase in AUC was evaluated and tested for significance using DeLong test (35). Furthermore, the net reclassification improvement (NRI) was used to assess the added discrimination offered by the addition of SNPs to the risk model (24). Bootstrapping techniques were utilised for internal validation of the models (36). Bootstrap samples were drawn 1000 times to adjust model parameters for overfitting. Improvement in model calibration was assessed using Akaike information criteria (AIC) and Bayesian information criteria (BIC). Unless otherwise stated, all analyses were performed using R version 3.1.1 and STATA® version 13.1 (StataCorp LP, College Station, TX, USA).

Results

Seven hundred and eighteen cases and 1667 population controls were successfully genotyped for 20 SNPs, which had been independently validated from 157 SNPs screened in a candidate-gene discovery study (26). Table I presents the general demographic and clinical characteristics of the study population. Men constituted the majority of the study population cases (57.7%) and (58.1%) controls. The proportion of ever smokers was significantly higher in cases (93.2%) compared with controls (65.2%). Significant differences were observed in other risk factors including smoking duration, prior diagnosis of pneumonia, occupational exposure to asbestos, and prior diagnosis of tumour (P<0.001).

Table I

Epidemiology, clinical and lifestyle characteristics of the subjects by case-control status.

Characteristics	Case (n=718)	Control (n=1667)	All subjects (n=2385)
Age (yrs.)
<60	162 (22.6)	457 (27.41)	619 (25.9)
60–70	264 (36.8)	647 (38.8)	911 (38.2)
70+	292 (40.7)	563 (33.8)	855 (35.9
Gender
Male	414 (57.7)	969 (58.1)	1383 (58.0)
Female	304 (42.3)	698 (41.9)	1002 (42.0)
Smoking statusa
Never	43 (6.0)	575 (34.5)	618 (25.9)
Former	316 (44.0)	820 (49.2)	1136 (47.6)
Current	353 (49.2)	267 (16.0)	620 (26.0)
Smoking duration (yrs.)a
Never	43 (6.0)	575 (34.5)	618 (25.9)
1–20	38 (5.3)	341 (20.5)	379 (15.9)
21–40	175 (24.4)	440 (26.4)	615 (25.8)
41–60	399 (55.6)	278 (16.7)	677 (28.4)
>60	51 (7.1)	27 (1.6)	78 (3.3)
Previous pneumoniaa
Yes	105 (14.6)	243 (14.6)	348 (14.6)
No	590 (82.2)	1420 (85.2)	2010 (84.3)
Previous malignant
Yes	183 (26.3)	38 (2.3)	221 (9.4)
No	512 (73.7)	1625 (97.7)	2136 (90.6)
Asbestos exposurea
Yes	134 (18.7)	158 (9.5)	292 (12.2)
No	395 (55.0)	1505 (90.3)	1900 (79.7)
Family lung CA
No history	566 (78.8)	1348 (80.9)	1914 (80.3)
Early onset	74 (10.3)	101 (6.1)	175 (7.3)
Late onset	78 (10.9)	218 (13.0)	296 (12.4)
Histology
Squamous cell carcinoma	239 (33.3)	-
Adenocarcinoma	228 (31.8)	-
Small cell	87 (12.1)	-
NSCLC	77 (10.7)	-
Other	87 (12.1)	-

Numbers do not add up to total due to missing data; NSCLC, non-small cell lung cancer.

Table II presents the results of additive gene-dosage model for all SNPs. Heterozygosity for rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11) conferred an increased risk for lung cancer in reference to the wild-type genotype [OR 1.74 (95% CI 1.31–2.32); 1.42 (95% CI 1.16–1.72) and 2.53 (95% CI 2.06–3.12), respectively]. The homozygote genotype for rs16969968 (CHRNA3/5), rs13181 (ERCC2), rs5744256 (IL-18) and rs2306022 (ITGA11) increased the risk of developing lung cancer with reference to the wild-type [OR 1.46 (95% CI 1.11–1.93); 1.46 (95% CI 1.13–1.89); 4.07 (95% CI 3.11–5.31) 4.09 (95% CI 2.21–7.56), respectively].

Table II

Univariable analysis of associations between 20 candidate SNPs and lung cancer (33).

SNP	Chromosome	Gene	Genotype					Additive model assumption
			Wilda	Heterozygote		Homozygote
			ca/co (%)	ca/co (%)	OR (95% CI)	ca/co (%)	OR (95% CI)	P-valuetrend
rs2279115	18q21.3	Bcl-2	30.1/29.0	49.0/50.4	0.91 (0.75, 1.11)	20.1/20.6	0.91 (0.71, 1.17)	0.91
rs10115703	9p22.3	Cerb1	86.2/84.7	12.7/14.6	0.85 (0.66, 1.10)	1.1/0.7	1.66 (0.66, 4.15)	0.21
rs16969968	15q25.1	α5-nAChR	40.1/44.9	45.7/44.1	1.16 (0.96, 1.40)	14.2/10.9	1.46 (1.11, 1.93)	0.012
rs2031920	10q26.3	CYP2E1	94.7/94.7	5.2/5.2	0.99 (0.67, 1.47)	0.1/0.1	1.16 (0.11, 12.8)	0.71
rs6413429	5p15.33	DAT1	87.2/86.4	12.5/13.3	0.93 (0.72, 1.21)	0.3/0.3	0.92 (0.18, 4.76)	0.74
rs1799732	11q23.2	DRD2	79.5/79.4	13.0/7.4	1.74 (1.31, 2.32)	7.5/13.1	0.57 (0.42, 0.78)	0.30
rs13181	19q13.32	XPD(ERCC2)	38.6/39.9	43.3/47.3	0.95 (0.78, 1.15)	18.1/12.8	1.46 (1.13, 1.89)	0.10
rs763110	1q24.3	FasL	42.5/40.2	43.4/46.6	0.88 (0.73, 1.07)	14.1/13.2	1.00 (0.77, 1.32)	0.27
rs5744256	11q23.1	IL18	32.7/47.2	43.7/44.5	1.42 (1.16, 1.72)	23.6/8.3	4.07 (3.11, 5.31)	<0.0001
rs16944	2q13	IL1B	42.9/46.1	44.7/43.2	1.11 (0.92, 1.34)	12.4/10.7	1.24 (0.93, 1.65)	0.24
rs4073	4q13.3	IL8	27.6/29.9	51.3/47.4	1.17 (0.96, 1.44)	21.7/22.7	1.01 (0.79, 1.30)	0.50
rs2306022	15q23	ITGA11	65.9/83.6	30.6/15.4	2.53 (2.06, 3.12)	3.5/1.0	4.09 (2.21, 7.56)	<0.0001
rs2317676	17q21.32	ITGB3	87.9/87.5	11.6/12.2	0.95 (0.72, 1.24)	0.6/0.3	1.54 (0.43, 5.48)	0.88
rs1799930	8p22	NAT2	50.3/48.4	39.4/42.7	0.89 (0.74, 1.07)	10.3/8.9	1.12 (0.82, 1.52)	0.95
rs3087386	2q11.2	REV1	31.6/31.4	49.7/49.4	0.99 (0.82, 1.22)	18.7/19.3	0.96 (0.75, 1.24)	0.63
rs4934	14q32.13	SERPINA3	26.9/27.3	50.3/49.2	1.04 (0.84, 1.28)	22.8/23.5	0.99 (0.77, 1.26)	0.99
rs1799895	4p15.2	SOD3	96.7/97.2	3.3/2.7	1.25 (0.75, 2.06)	0.0/0.1	-	0.44
rs5743836	3p21.2	TLR9	71.2/69.0	25.4/28.1	0.88 (0.72, 1.07)	3.5/2.9	1.15 (0.70, 1.88)	0.24
rs1139417	12p13.31	TNFR1	32.0/31.5	49.3/50.8	0.96 (0.78, 1.16)	18.7/17.8	1.03 (0.80, 1.34)	0.96
rs2273953	1p36.33	TP73	58.5/62.8	35.5/31.7	1.20 (0.99, 1.45)	6.0/5.5	1.17 (0.80, 1.70)	0.11

Reference genotype; ca, cases; co, controls.

Table III summarises the result obtained from the MDR analysis investigating epistatic effects among the SNPs. The best candidate classifiers of lung cancer status based on five SNP loci selected using the cross-validation consistency, training and testing accuracy were as follows: Single locus: rs2306022 (ITGA11); 2 loci: rs5744256 (IL-18), rs2306022 (ITGA11); 3 loci: rs1799732 (DRD2), rs5744256(IL-18), rs2306022 (ITGA11); 4 loci: rs1696998 (CHRNA3/5), rs1799732 (DRD2), rs5744256 (IL-18), rs2306022 (ITGA11); 5 loci: rs1799732 (DRD2), rs763110 (FasL), rs5744256 (IL-18), rs4073 (IL-8), rs2306022 (ITGA11). The 3 loci consisting of SNPs rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11) appears to be the overall best classifier of lung cancer status. These loci had training and testing accuracy of 0.6592 and 0.6572 respectively, and the cross validation consistency of 10/10 (model selected as the best of 3 in 10 CV) P<0.0001 (permutation test).

Table III

Comparison of different Multi-locus SNP combinations using MDR.

Model of inheritance	No. of loci	Selected SNPs in selected best model	Cross Validation consistency (CV)	Balanced training accuracy	Balanced testing accuracy
Additive effect	1	ITGA11_rs2306022	10/10	0.5886	0.5886
	2	IL18_rs5744256ITGA11_rs2306022	10/10	0.6418	0.6418
	3	DRD2_rs1799732IL18_rs5744256ITGA11_rs2306022	10/10	0.6575	0.6538
	4	CHRNA3_A5_rs16969968DRD2_rs1799732IL18_rs5744256ITGA11_rs2306022	4/10	0.6652	0.6321
	5	DRD2_rs1799732 FASL_rs763110IL18_rs5744256 IL8_rs4073ITGA11_rs2306022	6/10	0.6869	0.6178

Table IV shows the importance score results in the RF. RF ranks variables by a variable importance index, which is an indication of the importance of a variable on the basis of classification accuracy while considering interaction among variables. The three SNPs [rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11)] selected as the overall best classifier of lung cancer status in MDR were also ranked top 3 by RF using variable importance index.

Table IV

Importance score results in the random forest.

SNP	Gene name	Variable importance
rs5744256a	IL18	18.0783
rs2306022a	ITGA11	14.2703
rs1799732a	DRD2	4.4401
rs4934	SERPINA3	2.8533
rs13181	XPD(ERCC2)	2.7543
rs16969968	α5-nAChR	2.4906
rs16944	IL1B	2.1737
rs1139417	TNFR1	1.5054
rs2273953	TP73	1.4667
rs3087386	REV1	1.4185
rs1799930	NAT2	1.1701
rs10115703	Cerb1	0.9366
rs2279115	Bcl-2	0.8465
rs5743836	TLR9	0.7407
rs4073	IL8	0.6093
rs763110	FasL	0.4508
rs2317676	ITGB3	0.0477
rs2031920	CYP2E1	−0.048
rs1799895	SOD3	−0.1922
rs6413429	DAT1	−0.3696

Top 3 ranked SNPs using variable importance.

Table V summarises reclassifications for cases and controls using epidemiological model and models with SNPs [rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11)]. Subjects were categorised into three different thresholds; low-risk (<0.91), intermediate risk (0.91 to 5.12), and high-risk (>5.12) groups. The threshold values were defined from the predicted 5-year absolute risks for the original LLP control samples (n=1,272), assuming the risk distribution in this group is similar to that of the general Liverpool population. The upper threshold (5.12) corresponds to the value for the top 20% of predicted absolute risks in the population; individuals whose 5-year predicted absolute risk is equal to or above this value are designated as ‘high risk’ group. The lower threshold value of 0.91 corresponds to the bottom 40% of absolute risks in the control population and represents the ‘low risk’ group. This definition of high risk and low risk groups was used in an earlier study (13). Overall, 42.7% of cases (311/727) and 35.7% of controls (592/1657) had their predicted risks re-classified into other risk groups when SNPs were incorporated into risk prediction model. This reclassification showed improvement (upward shift) in approximately 25% of cases and became worse (downward shift) for 18% resulting in a net gain of ~6%. The net gain was higher for controls (10%) with overall improvement in risk (downward shift) for 23% and worse performance (upward shift) for 13%. The NRI was estimated at 13.5% (P<0.001).

Table V

Reclassification of predicted risk for cases and controls using the epidemiological model and extended model with rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11).

Epidemiological model	Extended model with rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11)				Total
	<0.91%	0.91 to 2.5%	>2.5 to 5.12%	>5.12%
Cases
<0.91%	69 (57.5)	43 (35.8)	8 (6.7)	0 (0)	120
0.91 to 2.5%	15 (12.4)	46 (38.0)	46 (38.0)	14 (11.6)	121
>2.5 to 5.12%	0 (0)	43 (26.7)	49 (30.4)	69 (42.9)	161
>5.12	2 (0.6)	9 (2.8)	62 (19.1)	252 (77.5)	325
Total	86	141	165	335	727
Controls
<0.91%	726 (89.9)	77 (9.5)	4 (0.5)	1 (0.1)	808
0.91 to 2.5%	180 (45.0)	147 (36.7)	58 (14.5)	15 (3.8)	400
>2.5 to 5.12%	20 (8.8)	85 (37.4)	70 (30.8)	52 (22.9)	227
>5.12%	3 (1.3)	29 (13.1)	68 (30.6)	122 (55.0)	222
Total	929	338	200	190	1657

Table VI depicts the odds ratios (OR) and 95% confidence intervals (95% CI) of the multivariate logistic regression models for the epidemiological model and the extended model with SNPs. The ORs and 95% CI for both models were comparable which suggests the absence of any serious confounding effects of SNPs on the relationship between each of the other clinical and epidemiological risk factors and lung cancer risk. Model fit was assessed using Akaike information criterion (AIC) and Bayes information criteria (BIC). There was an improvement in model fit as indicated by the reduction of the AIC from 2098.42 from the epidemiological model to 1930.14 for the extended model with SNPs. Likewise, a similar reduction was observed in BIC from 2167.75 from the epidemiological model to 2016.80 for the extended model with SNPs. Fig. 1 shows the AUC of the epidemiological model and extended model with SNPs. The apparent AUC of the epidemiological model without SNPs was 0.75 (95% CI 0.73–0.77). When epistatic data were incorporated in the extended model, the AUC increased to 0.81 (95% CI 0.79–0.83) which corresponds to 8% increase in AUC for the model with SNPs (DeLong's test P=2.2e-16). After correction for optimism, the AUC was 0.73 for the epidemiological model and 0.79 for the extended model.

Table VI

Summary of multivariable risk model for the epidemiological model and the extended model with rs1799732 (DRD2), rs5744256 (IL-18) and rs2306022 (ITGA11).

	Epidemiological model		Extended model with SNPs
Covariates	OR (95%CI)	P-values	OR (95%CI)	P-values
Age	1.01 (0.99–1.02)	0.312	1.00 (0.99–1.02)	0.610
Gender	1.24 (0.95–1.63)	0.107	1.14 (0.87–1.52)	0.340
Smoking duration (years)
None	1.00		1.00
1–19	1.41 (0.82–2.42)	0.209	1.23 (0.69–2.18)	0.476
20–39	4.30 (2.81–6.57)	<0.001	4.90 (3.10–7.73)	<0.001
40–59	11.12 (5.41–22.86)	<0.001	15.70 (7.22–34.14)	<0.001
≥60	13.91 (9.26–20.91)	<0.001	18.58 (11.90–29.01)	<0.001
Pneumonia	1.53 (1.12–2.09)	0.007	1.55 (1.11–2.15)	0.008
Asbestos	3.25 (2.34–4.52)	<0.001	3.10 (2.19–4.39)	<0.001
Previous tumour	16.97 (11.25–25.61)	<0.001	16.52 (10.79–25.31)	<0.001
Family history of lung cancer
None	1.00
Early onset (<60 years)	1.33 (0.84–2.09)	0.223	1.11 (0.69–1.80)	0.659
Late onset (≥60 years)	1.07 (0.76–1.54)	0.672	1.14 (0.78–1.66)	0.495
rs1799732			0.78 (0.63–0.97)	0.028
rs5744256			2.04 (1.69–2.46)	<0.001
Rs2306022			4.04 (3.10–5.26)	<0.001
Goodness of fit statistic
AIC	2098.42		1930.14
BIC	2167.75		2016.80

Figure 1

Performance of lung cancer risk model with and without the SNP epistatic effect.

Discussion

This study demonstrates the use of comprehensive analytical techniques for investigating the contribution of adding an interactive effect of a panel of genetic markers (SNPs) to the prediction of individual absolute risk of developing lung cancer, using a risk model similar to the LLP model (3). Using genotype data from 2385 individuals included in the independent validation of SNPs identified in a candidate-gene genetic association study from the LLP case-control study, we found the 3 loci genotype interaction rs2306022 (ITGA11), rs5744256 (IL-18) and rs1799732 (DRD2) provided the best classifier of disease status using both MDR and RF. Adding these SNPs to a clinically-based lung cancer risk model lead to an increase in AUC (0.75 to 0.81); and increase in net reclassification (NRI=17.5%).

We utilised two different approaches; discrimination and reclassification to evaluate the contribution of adding an interactive epistatic effect to a risk model for lung cancer. AUC is the most popular metric used for measuring the discriminatory power of a model to correctly classify subjects with or without a disease. Our result showed 8% increase in AUC (DeLong's test P=2.2e-16) for risk prediction in the extended model with SNPs (AUC=0.81) compared with the epidemiological model without SNPs (AUC=0.75), which is higher than that reported by Li et al (9). Li et al, in a Chinese case-control study, genotyped five SNPs identified in Genome Wide Association study of 5068 subjects. The genetic risk scores based on these SNPs were estimated by two approaches: a simple risk alleles count (cGRS) and a weighted method (wGRS). Their AUC in combination with the bootstrap resampling method was used to assess the predictive performance of the genetic risk score for lung cancer. Smoking history contributed significantly to lung cancer (P<0.001) risk [AUC=0.619 (0.603–0.634)], and incorporated with wGRS gave an AUC value of 0.639 (0.621–0.652) after adjustment for over-fitting (9). For clinical risk prediction, it is expedient that a new risk model correctly classify individuals into higher or lower risk categories (37). Pencina et al introduced a new metric, the NRI that assesses the improvement in model performance by quantifying the degree of correct classification (24). By applying the NRI, we demonstrated that the addition of SNPs lead to a 17.5% improvement in the risk classification of the subjects.

This study is the first to replicate the association between the ITGA11 locus and lung cancer described by Young and colleagues (26). ITGA11 (integrin α11) belongs to the family of transmembrane receptors that mediate physical interactions between cells and extracellular matrix protein collagens (38). ITGA11 is localised to stromal fibroblast and commonly overexpressed in non-small cell lung cancer (NSCLC) (38). Earlier studies have reported that the interactions of tumour cells with the stroma play a crucial role in tumour growth, invasion, metastases, angiogenesis, and chemoresistance (38–41). It has been shown that carcinoma-associated fibroblasts in NSCLC express higher levels of ITGA11. One of the factors which are affected by higher levels of ITGA11 during tumour growth is IGF2 (38,42). Higher levels of IGF2, in turn, can stimulate growth of tumour epithelial cells leading to tumour progression and metastasis (38). IL18 (Interleukin-18) is a multifunctional cytokine (an extracellular signalling molecule) that augments IFN-γ production and affects tumour immune response, leukocyte recruitment, cancer proliferation, and angiogenesis (43,44). An earlier study reported the presence of IL-18 in induced sputum of lung cancer patients (45). Farjadfar et al also reported an association between IL-18 and lung cancer in a case-control study including 73 lung cancer patients (53 squamous carcinoma and 20 small cell lung carcinoma), and 97 healthy regional aged-matched individuals (46). They suggested that their finding may be attributed to the disruption of the potential of cAMP responsive element-binding protein site and subsequent reduction in IL-18 production as observed in other cancer types (46). Reduced production of IL-18 can result in decreased IFN-γ synthesis, imbalanced Th1/Th2 differentiation, insufficient activation of natural killer cells and CD8+ lymphocytes (46,47) impairment of cancer cell apoptosis and efficient angiogenesis (47,48). DDR2 is a receptor tyrosine kinase that binds collagen as its endogenous ligand (49). It has been previously shown to promote cell migration, proliferation, and survival when activated by ligand binding and phosphorylation (49,50). Harmmerman et al reported that DDR2 mutations are present in 4% of small cell lung carcinomas; gain-of-function mutations in this gene are important oncogenic events and are amenable to therapy with dasatinib (49). However, the mechanism of this mutation is unknown.

Since epistasis is known to contribute to unexplained genetic variation of common diseases, some genetic variants may have a weak and insignificant independent effect, but strong epistatic effect (biological interaction) with other variants. The integration of genetic variants in risk prediction models beyond the traditional epidemiological covariates have been applauded as the way forward in lung cancer risk prediction modelling (8). The result presented in this study supports this notion. Genetic factors function primarily through complex mechanisms that involve interactions between multiple genes and environmental factors (21,22). However, the effect of interaction will be disregarded if the genetic effect is examined in isolation, without taking cognisance of potential interactions with other unknown factors (31). The inherent nonlinearity implies that epistasis can occur among polymorphisms even in the absence of independent effect of the components, presenting computational intensive difficulties and statistical challenges because an infinite number of combinations that needs to be evaluated (21,22). The use of nonparametric and genetic model-free machine learning algorithms such as MDR (28,29) and RF (32,33) have been proposed to overcome the caveat of the traditional parametric statistics and have proven to be useful in this study. Here we see that the addition of the three SNPs increases the AUC, indicating that the interaction of these loci may be important. There was an improvement in model fit, as indicated by the reduction of the AIC and BIC. Furthermore, the SNPs used in this study were internally validated using a two stage design as described by Young et al (26) and the use of HWE to minimise genotyping error are methodological advantages utilised to minimise false positive results.

To the best of our knowledge, this is the first study to evaluate the addition of these specific interactions of SNPs to a lung cancer risk model. However, the result of this study must be considered in the light of a number of limitations. First, our prediction model used covariates in the LLP risk model but did not include other risk factors for lung cancer such as chronic obstructive pulmonary diseases. However, the objective of this study is to evaluate the contribution of adding an interactive effect of a panel of genetic SNPs to the LLP risk model and the model has been validated in three independent external datasets with good discrimination and calibration (27). Second, our study demonstrates how a modest increase in AUC can lead to a substantial improvement in reclassification as quantified by the NRI. This finding supports a suggestion by Pencina et al that a small increase in AUC might still be suggestive of a meaningful improvement (24). Third, the LLP comprise predominantly Caucasians and therefore, the lack of ethnic diversity implies that this model may be less applicable in non-white population. Fourth, our approach to reclassification did not distinguish between persons with competing events and those without an event because both are classified as not having the event of interest. Fifth, the lack of validation of the epistatic model in an independent population is a limitation, however, the application of bootstrap correction for optimism addresses in part the lack of independent validation. Sixth, many of the 20 SNPs from Table II failed to replicate in the current study, particularly given the larger sample size (718 cases, 1667 controls) in the current study when compared with previous study (248 cases, 207 controls). A plausible explanation for this observation may be due to the fact that the non-significant SNPs play lesser or no role in epistatic interaction. Finally, our threshold values for risk classification was based on the predicted 5-year absolute risk for original LLP control samples but the appropriateness of these threshold values in other populations is uncertain. Using different values could have affected the results of our reclassification analyses and subsequent clinical implications.

In conclusion, our result shows in principle how an SNP epistatic factor can be incorporated into an epidemiological risk prediction model. In this study, inclusion of SNPs rs1799732 (DRD2), rs5744256 (IL-18), rs2306022 (ITGA11) resulted in a modest improvement in lung cancer risk prediction.