Analyzing longitudinal trait trajectories using GWAS identifies genetic variants for kidney function decline
Abstract
Understanding the genetics of kidney function decline, or trait change in general, is hampered by scarce longitudinal data for GWAS (longGWAS) and uncertainty about how to analyze such data. We use longitudinal UK Biobank data for creatinine-based estimated glomerular filtration rate from 348,275 individuals to search for genetic variants associated with eGFR-decline. This search was performed both among 595 variants previously associated with eGFR in cross-sectional GWAS and genome-wide. We use seven statistical approaches to analyze the UK Biobank data and simulated data, finding that a linear mixed model is a powerful approach with unbiased effect estimates which is viable for longGWAS. The linear mixed model identifies 13 independent genetic variants associated with eGFR-decline, including 6 novel variants, and links them to age-dependent eGFR-genetics. We demonstrate that age-dependent and age-independent eGFR-genetics exhibit a differential pattern regarding clinical progression traits and kidney-specific gene expression regulation. Overall, our results provide insights into kidney aging and linear mixed model-based longGWAS generally.
What the paper shows and why it matters
(AI-generated)
Genome-wide association studies (GWAS) link genetic variants to traits. For kidney function (estimated glomerular filtration rate), single-time-point studies have found hundreds of loci. Few variants are known for how the trait changes over time, although worsening biomarkers mark disease progression. Earlier work used the difference between a person’s last and first measurement, or a fast two-stage approach that estimates each person’s slope first and tests it against genotype. Neither can include people measured only once; one-stage mixed models can, but were considered too demanding, and no systematic comparison existed. The problem is an approach that controls false positives, has good power, gives unbiased effects and is fast enough for genome-wide use.
The paper argues that, for GWAS of change over time, a one-stage linear mixed model of the trait over age is preferable when power and unbiased effect sizes both matter. The model gives each person a starting level and a slope, lets them correlate, and includes people measured once. In simulations and UK Biobank kidney data comparing seven approaches, the two-stage approach was most powerful among those controlling false positives, but shrank effects by 11–38%. Models that ignored person-specific slopes, or the correlation between starting level and slope, inflated false positives. The age-based model was unbiased and, including people measured only once, nearly as powerful as the two-stage approach. With the GMMAT/MAGEE software, it could be run genome-wide. For kidney function it identified 13 independent variants for decline, six of them novel, and five genome-wide significant loci. Of 12 known kidney-function variants associated with decline, all but one had near-zero effects at age 40 and large effects at 70; 11 others had large, age-independent effects.
The paper advises the age-based model for longitudinal GWAS of trait change when the start of follow-up is random, and a model of time since baseline when the start is informative, for example when analysing trait change in patients. It reads the two-stage approach as possibly acceptable for finding loci but not for interpreting effect sizes. For kidney research, the results suggest that decline variants act mainly from about age 40 onwards, and the enrichment of genes differentially expressed in kidney cortex in their loci points to kidney-intrinsic mechanisms, making the likely causal genes there candidates for studying kidney ageing.