Overview of LDAK-KVIK

LDAK-KVIK is a method for performing mixed-model association analysis (MMAA) for GWAS. LDAK-KVIK uses a three-step process to test SNPs and genes for association with the phenotype:

  1. Compute a Leave-One-Chromosome-Out (LOCO) PRS using an Elastic Net model
  2. Use the PRS as offset in single-SNP analysis
  3. Use the summary statistics of single-SNP analysis to run a gene-based association analysis LDAK-GBAT.

A summary of the each step is included in the LDAK-KVIK steps, and a detailed description of LDAK-KVIK is included in the LDAK-KVIK publication (see preprint). Below, we list some innovations of LDAK-KVIK.

Key innovations of LDAK-KVIK

LDAK-KVIK never needs to store genotypes for more than 512 SNPs, and thus has low memory demands. LDAK-KVIK constructs PRS in Step 1 using a chunk-based variational Bayes solver that requires 5-20 times fewer updates than conventional variational Bayes solvers. This step is also used to efficiently compute a test statistic scaling factor.

LDAK-KVIK increases statistical power, relative to existing MMAA tools, by using more realistic models for how SNP effect sizes vary across the genome (specifically, the dependency between SNP heritability and MAF). An Elastic Net prior distribution is used for SNP effect sizes (i.e., a mixture of a Gaussian and a Laplace distribution), whereas existing MMAA tools generally restrict to mixtures of Gaussian distributions.

LDAK-KVIK includes a fast empirical algorithm for computing the saddlepoint approximation (SPA), which is used for controlling type 1 errors when testing binary traits with case:control imbalance.

LDAK-KVIK includes a novel test for structure, based on the average pairwise correlation between 512 SNPs randomly picked from the genome (the correlation is calculated after regressing out covariates). The outcome of this test determines how LDAK-KVIK calculates the test statistic scaling factor.

In addition to testing SNPs individually for association with the phenotype, LDAK-KVIK can also perform gene-based association analysis. This is achieved by providing the results of the single-SNP analysis to our existing software LDAK-GBAT.