TL;DR
This paper introduces GATE, a new computational method for analyzing genetic associations with time-to-event (TTE) phenotypes in large biobanks. GATE efficiently handles complex genetic relationships and heavily censored data, enabling the discovery of genetic variants influencing disease onset and lifespan that were previously missed.
What Problem Does This Paper Address?
Large biobanks, like the UK Biobank and FinnGen, link electronic health records with genetic data, offering a powerful resource to understand the genetic basis of complex diseases and lifespan. Analyzing time-to-event (TTE) phenotypes (e.g., age of disease onset, lifespan) is crucial for this, but existing genome-wide association study (GWAS) methods struggle with two major challenges:
- Computational Scalability: Analyzing millions of genetic variants across hundreds of thousands of individuals requires immense computational resources (time and memory).
- Data Characteristics: TTE phenotypes in biobanks are often "heavily censored," meaning most individuals haven't experienced the event of interest by the end of the study. This, combined with complex population structures and relatedness within biobanks, can lead to inaccurate results (inflated Type I errors) with standard statistical approaches, especially for low-frequency genetic variants. Existing methods like COXMEG, while accounting for relatedness, are too slow for biobank-scale data and can produce inflated p-values with heavy censoring.
Methods
The core innovation is the GATE (Genetic Analysis of Time-to-Event phenotypes) method. It tackles the challenges by:
- Frailty Model Formulation: GATE adapts a multivariate Gaussian frailty model, a statistical framework well-suited for TTE data with complex dependencies, by transforming its likelihood into a modified Poisson generalized linear mixed model (GLMM). This allows leveraging existing efficient GLMM fitting techniques.
- Efficient Optimization: It employs state-of-the-art optimization strategies, including Average Information Restricted Maximum Likelihood (AI-REML) for variance components and Pre-conditioned Conjugate Gradient (PCG) for solving linear systems. This avoids computationally expensive matrix inversions.
- On-the-Fly GRM Calculation: To save memory, GATE computes the Genetic Relatedness Matrix (GRM) elements as needed, rather than storing a massive matrix upfront.
- Saddlepoint Approximation (SPA): For accurate p-value calibration, especially with heavily censored data and low-frequency variants, GATE uses SPA. This is a more robust method than the traditional normal approximation for extreme tail probabilities.
- Variance Ratio Approximation: To speed up score statistic calculations, GATE uses a variance ratio approximation that has been shown to be stable for variants with minor allele counts of 20 or more.
Datasets
- UK Biobank: Approximately 408,582 participants of White British ancestry were used for analyzing five TTE phenotypes, including lifespan, and a phenome-wide scan of 871 TTE phenotypes.
- FinnGen: Approximately 180,000 individuals were used for analyzing lifespan and replicating findings from the UK Biobank.
Key Insights and Ideas
- Frailty Models are Crucial for TTE: The paper strongly advocates for frailty models over simpler logistic regression for TTE phenotypes, especially in biobanks. They better capture the nuances of time-to-event data and can uncover associations missed by binary trait analyses.
- SPA is Essential for Accuracy: The use of Saddlepoint Approximation (SPA) is critical for controlling Type I error rates, particularly in scenarios with high censoring rates and low-frequency variants, where standard methods fail.
- Computational Efficiency is Achievable: By combining advanced optimization techniques and clever data handling (like on-the-fly GRM calculation), GATE demonstrates that biobank-scale TTE GWAS is computationally feasible.
- Phenome-Wide Scans are Now Possible: The scalability of GATE opens the door to performing comprehensive phenome-wide association studies (PheWAS) for TTE traits, providing a much broader understanding of genetic influences on health trajectories.
Key Results
- Computational Performance: GATE is significantly faster and uses less memory than existing methods like COXMEG. For a biobank-scale GWAS (408,582 samples, 46 million variants), GATE took ~14.5 hours on 30 CPUs with peak memory under 11 GB, whereas COXMEG would take days and require much more memory.
- Type I Error Control: Simulations and real data analyses show that GATE effectively controls Type I error rates, even with censoring rates up to 99.8%, a significant improvement over methods that don't use SPA.
- Discovery of Novel Loci: GATE identified 114 loci for four exemplary TTE phenotypes (ischemic heart disease, female breast cancer, glaucoma, Alzheimer's disease) in the UK Biobank. Many of these were not detected by standard logistic regression models (e.g., SAIGE).
- Replication: 78 out of 81 identified loci were replicated in the FinnGen study, with 69 showing the same effect direction and being successfully replicated with p < 0.05.
- Lifespan Analysis: GATE successfully identified the known APOE locus for lifespan in both UK Biobank and FinnGen, a locus missed by a logistic mixed model analysis.
Limitations and Future Work
- No Direct Hazard Ratio Estimates: Similar to other score-test based methods, GATE does not directly provide hazard ratio estimates from the null model. However, it offers workarounds for this.
- Single-Variant Focus: The current version performs single-variant association analysis, which may have limited power for very rare variants. Future work could extend GATE to mask-based or region-based rare variant set association tests.
- No Left-Truncated Data: GATE currently does not accommodate left-truncated data, which could be relevant for early-onset phenotypes in older biobanks. Future work aims to extend GATE to handle this.
- Age-of-Diagnosis vs. Age-of-Onset: The analysis uses age-of-diagnosis from EHRs, which is a proxy for age-of-onset. While generally reasonable, this distinction is important for left-truncated phenotypes.
- Competing Risks: The model assumes independent censoring, which might not hold for all TTE phenotypes (e.g., IHD with competing risks of death from other causes). Future work plans to incorporate competing risk models.
Summary updated