Abstract
Identifying relevant genetic variants in family-based studies is often hindered by computational complexity and the presence of shared environmental factors. Traditional single-marker association analyses frequently miss weak genetic signals that contribute to complex traits. We present a resampling-based fast variable selection technique for detecting relevant single nucleotide polymorphisms (SNPs) in multi-marker mixed-effect models. By utilizing an e-values framework and a scalable generalized bootstrap procedure, our method effectively identifies multiple SNPs simultaneously while accounting for familial dependency. Numerical studies and analysis of the Minnesota Center for Twin and Family Research (MCTFR) dataset demonstrate that this approach outperforms traditional single-marker and model selection methods in detecting SNPs associated with traits like alcohol consumption.
Introduction
Genome-Wide Association Studies (GWAS) are essential for linking genetic variations to specific diseases or behaviors. However, the standard practice of "single-SNP analysis"—testing one variant at a time—often fails to capture the subtle, joint contributions of multiple genetic variants. This limitation is particularly pronounced in family studies (involving twins or siblings), where shared genes and environments create complex dependencies in the data.
While joint modeling of genetic variants within a gene or pathway offers higher power to detect weak effects, the computational cost of traditional model selection in family data is often prohibitive. This paper introduces a computationally efficient approach based on the e-values framework to overcome these bottlenecks, allowing for the simultaneous detection of relevant SNPs in familial datasets.
Methodology
The Statistical Model and E-Values
Our approach utilizes a Linear Mixed Model (LMM), specifically the ACE model, which partitions phenotypic variance into additive genetic effects (A), common shared environment (C), and unique environmental error (E). To perform variable selection within this framework, we employ e-values rather than traditional p-values.
E-values quantify the "relevance" of a variable by measuring how much the model's predictive power changes when a specific SNP is excluded. Unlike p-values, which rely on rigid distributional assumptions and null hypothesis testing, e-values provide an "assumption-lean" metric that is more robust for complex biological data.
Computational Efficiency via Generalized Bootstrap (GBS)
To address the computational intensity of refitting complex models, we implement a Generalized Bootstrap (GBS) procedure. Instead of the traditional bottleneck of refitting the model thousands of times, GBS fits the model once and uses Monte-Carlo sampling to weight the results. This makes the selection process highly parallelizable and scalable to large datasets.
Quantile Selection Logic
A critical innovation in our method is the use of quantile e-values. We observed that for weak genetic signals, the average of the bootstrap distribution can often wash out the signal. By focusing on the tails of the distribution (e.g., the 90th percentile), we can more effectively separate true genetic signals from random noise.
Results and Validation
Simulation Studies
In synthetic environments with 250 families and 50 SNPs, our e-value method consistently outperformed standard methods such as mBIC2 and RFGLS+BH. Notably, at low heritability levels—where traditional methods often failed to detect any signal—our approach maintained a respectable detection rate, demonstrating its sensitivity to weak genetic effects.
Real-World Application: MCTFR Dataset
We applied the method to the Minnesota Center for Twin and Family Research (MCTFR) dataset, comprising 7,605 individuals across 2,151 nuclear families. Focusing on alcohol consumption phenotypes, our method identified significantly more SNPs in established "alcohol genes" (such as GABRA2, ADH, and ALDH2) than competing approaches. Many of the detected SNPs were located in or near exons, suggesting high biological and functional relevance.
Conclusion
Our proposed resampling-based variable selection method provides a powerful and efficient alternative to single-SNP analysis in family-based GWAS. By leveraging e-values and the generalized bootstrap, researchers can now identify multiple relevant SNPs simultaneously without the prohibitive computational costs of traditional joint modeling. Future work will focus on scaling this method to genome-wide levels through initial screening steps and extending the framework to handle more complex pedigree structures.
Summary updated