TL;DR
This paper introduces LLARRMA, a novel statistical method that uses LASSO regression and resampling to better identify true genetic signals within "hit regions" of genome-wide association studies (GWAS). By accounting for uncertainty in model selection and linkage disequilibrium, LLARRMA outperforms traditional single-SNP analysis and other advanced methods in pinpointing causal variants, especially in complex genetic regions.
What Problem Does This Paper Address?
Genome-wide association studies (GWAS) are powerful for identifying genomic regions associated with diseases. However, once a "hit region" is found, pinpointing the exact causal genetic variants becomes challenging. This is primarily due to linkage disequilibrium (LD), where nearby genetic markers (SNPs) are inherited together, making it difficult to distinguish the true signal from correlated noise. Traditional single-SNP analysis, while effective for initial discovery, falters in these complex regions. More advanced methods like simultaneous multi-locus modeling are often applied in an ad hoc manner, lacking a robust way to assess the reliability of their findings against sampling variability. This paper aims to address this gap by providing a principled statistical framework to re-prioritize genetic associations within these challenging hit regions.
Methods
The core of the proposed method is LASSO local automatic regularization resample model averaging (LLARRMA). It combines several key techniques:
- LASSO (Least Absolute Shrinkage and Selection Operator): A penalized regression method that can shrink the coefficients of less important predictors (SNPs) to zero, effectively performing variable selection and producing "sparse" models.
- Resample Model Averaging (RMA): Instead of fitting a single model, LLARRMA repeatedly fits LASSO models to different subsamples (resamples) of the data. This allows for the estimation of a Model Inclusion Probability (MIP) for each SNP, representing the likelihood that it would be selected in a model across various data realizations. This helps characterize the uncertainty in model choice.
- Local Automatic Regularization: A crucial innovation is the automatic selection of the LASSO penalty parameter ($\lambda$) locally for each subsample. This is achieved through two criteria:
- Complement Deviance Selection: Optimizes for out-of-sample prediction performance.
- Permutation Selection: A more conservative approach aimed at identifying SNPs that are robustly associated even under random permutations of the phenotype, helping to control for false positives.
- Multiple Imputation: The method is designed to seamlessly incorporate uncertainty arising from missing genotype data, which is common in GWAS. It can handle hard, dosage, or multiple imputations of missing genotypes.
LLARRMA contrasts with Stability Selection (SS), another resampling-based method, by using locally chosen $\lambda$ values for each subsample, whereas SS uses a single global $\lambda$.
Datasets
The paper uses two real-world genotype datasets for evaluation:
- Cancer Data: Genotype data from a case-control GWAS for colorectal cancer, specifically focusing on a hit region previously identified on chromosome 18q21. Both complete and incomplete (with missing genotypes) versions of this dataset were used.
- '58 Data: Genotype data from a subset of the WTCCC (Wellcome Trust Case Control Consortium) study, focusing on a region on chromosome 22 with a mixture of high and low LD.
These datasets were used to simulate phenotypes and evaluate the performance of LLARRMA and its competitors under various scenarios of LD, effect sizes, and missing data.
Key Insights and Ideas
- The problem of LD in hit regions is critical: While single-SNP analysis is good for initial discovery, it's insufficient for fine-mapping causal variants in regions with high LD. The paper clearly demonstrates how correlated SNPs can obscure true signals.
- Local regularization is key: LLARRMA's ability to select the LASSO penalty parameter ($\lambda$) locally for each subsample appears to be a significant advantage over methods like Stability Selection that use a global $\lambda$. This allows the model to adapt better to the specific characteristics of each resampled dataset.
- Model inclusion probabilities (MIPs) are more informative than binary model selections: By providing a probability of inclusion, LLARRMA offers a nuanced view of SNP importance, reflecting the uncertainty inherent in statistical modeling, rather than a definitive "in" or "out" decision.
- Resampling and model averaging are powerful tools for uncertainty quantification: The paper effectively leverages these techniques to provide a more robust assessment of SNP associations, especially in the face of sampling variability and imputation uncertainty.
- Multiple imputation integrates seamlessly: The framework's ability to incorporate uncertainty from genotype imputation without significant computational overhead is a practical advantage.
Key Results
- LLARRMA consistently outperforms single-locus regression: In simulations, especially in scenarios with multiple causal loci and strong LD, LLARRMA's identified candidate sets were significantly enriched for true signals compared to single-locus analysis.
- LLARRMA outperforms Stability Selection (SS): Even when SS was "oracle-tuned" (given foreknowledge of the true answer to select its global $\lambda$), LLARRMA generally performed better, particularly in challenging scenarios with small effects and strong LD. This highlights the advantage of LLARRMA's local regularization.
- Permutation selection for $\lambda$ showed a slight advantage: In the simulations, the permutation-based criterion for selecting $\lambda$ tended to perform slightly better than the complement deviance criterion, suggesting its discovery-oriented nature aligns well with the evaluation goals.
- Imputation type has less impact than expected: While multiple imputation is theoretically superior for accounting for imputation uncertainty, the study found that simpler hard or dosage imputations performed comparably well when combined with LLARRMA, especially when imputation accuracy was reasonably high.
- LLARRMA is computationally feasible: The method's runtime was comparable to or better than Stability Selection, making it practical for use on large GWAS datasets.
Limitations and Future Work
- Focus on hit regions: LLARRMA is designed for focused analysis within pre-identified hit regions, not for initial genome-wide screening.
- Computational cost: While feasible, LLARRMA involves resampling and fitting multiple LASSO models, which is more computationally intensive than single-locus regression.
- Additive effects assumption: The current implementation assumes additive genetic effects. The authors suggest future work could extend this to more complex models, such as group LASSO, to account for non-additive effects or epistasis.
- Choice of $\lambda$ criteria: While permutation selection showed a slight edge, the authors suggest complement deviance might be superior in predictive settings, and further investigation into optimal $\lambda$ selection criteria could be beneficial.
- Generalizability to quantitative traits: The paper focuses on case-control data (binary outcomes) but notes that the method can be extended to quantitative traits.
The authors plan to release an R-package implementation of LLARRMA, which will facilitate its adoption by the research community.
Summary updated