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.

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 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 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 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 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 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.

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.