Abstract
Genome-wide association studies (GWAS) have proven successful in predicting genetic risk of disease using single-locus models; however, identifying single nucleotide polymorphism (SNP) interactions at the genome-wide scale is limited due to computational and statistical challenges. We address the computational burden encountered when detecting SNP interactions for survival analysis, such as age of disease-onset. To confront this problem, we developed a novel algorithm, called the Efficient Survival Multifactor Dimensionality Reduction (ES-MDR) method, which uses Martingale Residuals as the outcome parameter to estimate survival outcomes, and implemented the Quantitative Multifactor Dimensionality Reduction method to identify significant interactions associated with age of disease-onset. To demonstrate efficacy, we evaluated this method on two simulation sets to estimate the type I error and power. Simulations show that ES-MDR identifies interactions using less computational workload and allows for adjustment of covariates. We applied ES-MDR on the OncoArray-TRICL Consortium data with 14,935 cases and 12,787 controls for lung cancer (SNPs = 108,254) to search over all two-way interactions to identify genetic interactions associated with lung cancer age-of-onset. Our experiment on the OncoArray-TRICL data identified many one-way and two-way models with a single-base deletion in the noncoding region of BRCA1 (HR = 1.24, P = 3.15 x 10-15), as the top marker to predict age of lung cancer onset. From the results of our extensive simulations and analysis of a large GWAS study, we demonstrate that our method is an efficient algorithm that identifies genetic interactions to include in our models to predict survival outcomes.
Introduction
Genome-Wide Association Studies (GWAS) are highly effective at identifying single genetic variants (SNPs) linked to disease. However, biological risk is often driven by the interaction between multiple variants—a phenomenon known as epistasis. Detecting these interactions at a genome-wide scale is a significant computational challenge; for instance, testing every possible pair among 100,000 SNPs results in billions of tests. This burden is particularly heavy for "survival data" (time-to-event data, such as age of cancer onset), which requires complex statistical modeling. Existing tools like Surv-MDR are computationally slow and struggle to adjust for confounding covariates like smoking. The ES-MDR algorithm was developed to make these searches both faster and more accurate.
Methods: The Martingale Residual Innovation
The core innovation of ES-MDR is a two-step process that bridges survival analysis and machine learning. First, researchers run a standard Cox proportional hazards model that includes known covariates (e.g., smoking status). From this, they calculate "Martingale Residuals," which represent the difference between observed and expected events.
As shown in the equation above, the residual $M_i(t)$ serves as a "purified" quantitative score of the genetic effect, stripped of noise from known risk factors. These residuals are then plugged into the Quantitative MDR (QMDR) framework, which categorizes genetic variant combinations into high- or low-risk groups. By transforming a survival problem into a quantitative one, ES-MDR bypasses the need for expensive log-rank tests for every genetic combination.
Datasets and Validation
The method was validated using two primary approaches:
- Extensive Simulations: Two sets of simulated data were used to ensure the model correctly identifies real interactions (power) without producing false positives (Type I error).
- OncoArray-TRICL Consortium: A real-world dataset of 14,935 lung cancer cases and 12,787 controls. The team filtered this to 108,254 high-quality SNPs to search for interactions related to the age of disease onset.
Key Results
ES-MDR demonstrated a massive computational speedup. In head-to-head testing on 100 simulated datasets, the previous standard (Surv-MDR) took over 12 hours, while ES-MDR completed the task in just 2 minutes and 14 seconds. Furthermore, ES-MDR showed improved power to detect true interactions as sample sizes increased.
In the OncoArray-TRICL analysis, ES-MDR identified a single-base deletion in a noncoding region of BRCA1 as the top marker for predicting the age of lung cancer onset (Hazard Ratio = 1.24). Models incorporating genetic interaction data alongside smoking history significantly improved predictive accuracy, reaching an AUC of 0.68 for specific age groups.
Conclusion
ES-MDR provides an efficient algorithm for identifying genetic interactions in survival outcomes. By utilizing Martingale Residuals as a surrogate, the method achieves significant speed gains and allows for the critical adjustment of covariates. While future work is needed to replicate findings like the BRCA1 variant in independent datasets, ES-MDR represents a major advancement in the computational feasibility of epistasis research in GWAS.
Summary updated