TL;DR
This paper introduces fastSPA, a computationally efficient algorithm for analyzing large-scale Phenome-Wide Association Studies (PheWAS). It significantly speeds up the process of identifying genetic variants associated with thousands of binary health traits, even when case-control data is highly imbalanced, outperforming existing methods like Firth's test by orders of magnitude.
What Problem Does This Paper Address?
Genome-Wide Association Studies (GWAS) have revolutionized our understanding of genetics, but they typically focus on a single trait. Phenome-Wide Association Studies (PheWAS) offer a powerful alternative by examining the association of genetic variants across a vast landscape of thousands of health phenotypes, leveraging electronic health records (EHRs). However, a major challenge in PheWAS is the prevalence of binary phenotypes with highly unbalanced case-control ratios (e.g., 1:10 or even 1:600). Standard statistical tests, while effective in balanced scenarios, can suffer from inflated Type I error rates (false positives) or become computationally prohibitive when dealing with the sheer scale of PheWAS data. Existing methods like Firth's test, which addresses some of these issues, are too slow for large-scale PheWAS.
Methods
The authors propose a novel score test-based method, fastSPA (fast Saddlepoint Approximation), to address these challenges.
- Core Idea: Instead of relying on traditional normal approximations for the score test's null distribution,
fastSPAuses the saddlepoint approximation. This method provides a more accurate approximation, especially for skewed distributions common in unbalanced case-control data, and is more robust at the tails of the distribution. - Computational Efficiency: A key innovation is the
fastSPAimprovement, which significantly reduces computation time by leveraging a partially normal approximation approach. This optimization makes the calculation of the cumulant generating function and its derivatives dependent on the number of individuals with at least one minor allele (carriers) rather than the total sample size, which is particularly beneficial for rare variants. - Covariate Adjustment: The method can also adjust for non-genetic covariates, a crucial feature for real-world genetic association studies.
- Thresholding for Speed: To further accelerate computation,
fastSPAincorporates a hybrid approach that uses normal approximation for p-values close to the mean (where it's accurate and faster) and the saddlepoint approximation for more extreme p-values. Different thresholds (e.g.,fastSPA-2with a threshold of 2,fastSPA-0.1with a threshold of 0.1) are explored to balance speed and accuracy.
Datasets
The primary dataset used for evaluation and demonstration is the Michigan Genomics Initiative (MGI) data. This dataset includes genetic data from over 20,000 surgical patients, genotyped on the Illumina HumanCoreExome v12.1 array and imputed using the Haplotype Reference Consortium. Phenotypes were derived from 8,940 ICD-9 billing codes, resulting in 1,815 PheWAS disease states. For the analysis, four specific phenotypes were selected: Skin Cancer, Type-2 diabetes, Primary Hypercoagulable state, and Cystic Fibrosis.
Key Insights and Ideas
- Saddlepoint Approximation for Skewed Data: The saddlepoint approximation is a powerful tool for improving the accuracy of statistical tests when dealing with highly skewed data, such as unbalanced case-control ratios in PheWAS. It offers a significant improvement over traditional normal approximations.
- Carrier-Based Optimization is Key: The
fastSPAimprovement, which makes computation dependent on the number of carriers rather than the total sample size, is a critical breakthrough for handling rare variants efficiently. This is a major bottleneck in current genetic analysis. - Speed vs. Accuracy Trade-off: The use of thresholding (e.g.,
fastSPA-2vs.fastSPA-0.1) demonstrates a practical way to balance computational speed with statistical accuracy, allowing researchers to choose based on their resource constraints and desired precision. - Replication of Known Signals: The method not only controls Type I errors but also successfully replicates known genetic associations, validating its utility in real-world PheWAS applications.
Key Results
- Computational Speed:
fastSPAis dramatically faster than Firth's test, performing 100-300 times faster. For example, analyzing 10 million SNPs across 1500 phenotypes took 1.09 CPU-years withfastSPA-2compared to 117 CPU-years for Firth's test in an unbalanced scenario. - Type I Error Control:
fastSPA-2effectively controls Type I error rates across various case-control imbalances and MAFs, outperforming the traditional score test, which shows severely inflated error rates in unbalanced scenarios. Firth's test also performs well but is much slower. - Power:
fastSPA-2and Firth's test demonstrate comparable statistical power, suggesting no loss in detection ability compared to the gold standard. The score test shows lower power in unbalanced scenarios due to its inflated Type I errors. - MGI Data Application: In the MGI data analysis,
fastSPA-2identified significant associations that were consistent with previous findings, while the traditional score test produced many spurious associations.
Limitations and Future Work
- Threshold Selection: The choice of threshold for switching between normal and saddlepoint approximations (e.g.,
r=0.1vs.r=2) impacts computation time and the accuracy of genomic control estimates. The paper suggestsfastSPA-0.1offers better inflation factor estimates but at a higher computational cost. - Rare Variant Association: While
fastSPAis efficient for rare variants, the paper notes that rare variant association analysis, especially with multiple variants (gene/region-based tests), is an area for future research. Existing rare variant tests can also suffer from inflated Type I errors, andfastSPAcould potentially be adapted to address this. - Computational Resources: For extremely large datasets or when computational resources are severely limited, researchers might opt for larger thresholds (like
fastSPA-2) or the Berry-Esseen bound for faster results. - Future Research: The authors suggest future work could involve adapting
fastSPAfor gene or region-based multiple variant tests, particularly for rare variants, and exploring its use in whole-exome or whole-genome sequencing data.
Summary updated