New Paper: Simultaneous SNP Selection and Adjustment for Population Structure in High Dimensional Prediction Models

Congratulations to our Platform 5 Lead Dr. Celia Greenwood on this recent publication!

Table 1. Simulation study results. Mean (standard deviation) from 200 simulations stratified by the number of causal SNPs (null, 1%), the overlap between causal SNPs and kinship matrix (no overlap, all causal SNPs in kinship), and true heritability (10%, 30%). For all simulations, sample size is n = 1000, the number of covariates is p = 5000, and the number of SNPs used to estimate the kinship matrix is k = 10000. TPR at FPR = 5% is the true positive rate at a fixed false positive rate of 5%. Model Size () is the number of selected variables in the training set using the high-dimensional BIC for ggmix and 10-fold cross validation for lasso and twostep. RMSE is the root mean squared error on the test set. Estimation error is the squared distance between the estimated and true effect sizes. Error variance (σ2) for twostep is estimated from an intercept only LMM with a single random effect and is modeled explicitly in ggmix. For the lasso we use [28] as an estimator for σ2. Heritability (η) for twostep is estimated as from an intercept only LMM with a single random effect where and are the variance components for the random effect and error term, respectively. η is explictly modeled in ggmix. There is no positive way to calculate η for the lasso since we are using a PC adjustment. show less
Table 1. Simulation study results. Mean (standard deviation) from 200 simulations stratified by the number of causal SNPs (null, 1%), the overlap between causal SNPs and kinship matrix (no overlap, all causal SNPs in kinship), and true heritability (10%, 30%). For all simulations, sample size is n = 1000, the number of covariates is p = 5000, and the number of SNPs used to estimate the kinship matrix is k = 10000. TPR at FPR = 5% is the true positive rate at a fixed false positive rate of 5%. Model Size () is the number of selected variables in the training set using the high-dimensional BIC for ggmix and 10-fold cross validation for lasso and twostep. RMSE is the root mean squared error on the test set. Estimation error is the squared distance between the estimated and true effect sizes. Error variance (σ2) for twostep is estimated from an intercept only LMM with a single random effect and is modeled explicitly in ggmix. For the lasso we use (cf. paper for formula) as an estimator for σ2. Heritability (η) for twostep is estimated as (cf. paper for formula) from an intercept only LMM with a single random effect where and are the variance components for the random effect and error term, respectively. η is explictly modeled in ggmix. There is no positive way to calculate η for the lasso since we are using a PC adjustment.

Abstract

Complex traits are known to be influenced by a combination of environmental factors and rare and common genetic variants. However, detection of such multivariate associations can be compromised by low statistical power and confounding by population structure. Linear mixed effects models (LMM) can account for correlations due to relatedness but have not been applicable in high-dimensional (HD) settings where the number of fixed effect predictors greatly exceeds the number of samples. False positives or false negatives can result from two-stage approaches, where the residuals estimated from a null model adjusted for the subjects’ relationship structure are subsequently used as the response in a standard penalized regression model. To overcome these challenges, we develop a general penalized LMM with a single random effect called ggmix for simultaneous SNP selection and adjustment for population structure in high dimensional prediction models. We develop a blockwise coordinate descent algorithm with automatic tuning parameter selection which is highly scalable, computationally efficient and has theoretical guarantees of convergence. Through simulations and three real data examples, we show that ggmix leads to more parsimonious models compared to the two-stage approach or principal component adjustment with better prediction accuracy. Our method performs well even in the presence of highly correlated markers, and when the causal SNPs are included in the kinship matrix. ggmix can be used to construct polygenic risk scores and select instrumental variables in Mendelian randomization studies. Our algorithms are available in an R package available on CRAN.

Author summary

This work addresses a recurring challenge in the analysis and interpretation of genetic association studies: which genetic variants can best predict and are independently associated with a given phenotype in the presence of population structure? Not controlling confounding due to geographic population structure, family and/or cryptic relatedness can lead to spurious associations. Much of the existing research has therefore focused on modeling the association between a phenotype and a single genetic variant in a linear mixed model with a random effect. However, this univariate approach may miss true associations due to the stringent significance thresholds required to reduce the number of false positives and also ignores the correlations between markers. We propose an alternative method for fitting high-dimensional multivariable models, which selects SNPs that are independently associated with the phenotype while also accounting for population structure. We provide an efficient implementation of our algorithm and show through simulation studies and real data examples that our method outperforms existing methods in terms of prediction accuracy and controlling the false discovery rate.

Publication: Simultaneous SNP selection and adjustment for population structure in high dimensional prediction models. Bhatnagar SR., et al. PLoS Genet. 2020 May 4.

Share This Post

Twitter
LinkedIn
Email

Recent Posts