OSCA: a tool for omic-data-based complex trait analysis

Abstract

The rapid increase of omic data has greatly facilitated the investigation of associations between omic profiles such as DNA methylation (DNAm) and complex traits in large cohorts. Here, we propose a mixed-linear-model-based method called MOMENT that tests for association between a DNAm probe and trait with all other distal probes fitted in multiple random-effect components to account for unobserved confounders. We demonstrate by simulations that MOMENT shows a lower false positive rate and more robustness than existing methods. MOMENT has been implemented in a versatile software package called OSCA together with a number of other implementations for omic-data-based analyses.

Fig. 1

Power and false positive rate for the MWAS methods in simulation scenario 1. The phenotypes were simulated based on the effects from 100 causal probes but no direct effects from the CTCs. a Mean genomic inflation factor from a method across 100 simulation replicates with an error bar representing ± SE of the mean. The dashed line at 1 shows the expected value if there is no inflation. b Box plot of AUCs for each method from 100 simulation replicates

Fig. 2

Genomic inflation factor and family-wise error rate for the MWAS methods in simulation scenario 2 (effects from CTCs but no causal probes). Shown on the horizontal axis are the $R_{\text{CTCs}}^2$ values used to simulate the phenotype. a Each dot represents the mean λ value from 1000 simulation replicates given a specified $R_{\text{CTCs}}^2$ value for a method with an error bar representing ± SE of the mean. b Each dot represents the family-wise error rate, calculated as the proportion of simulation replicates with one or more null probes detected at a methylome-wide significance level

Fig. 3

QQ plot of p values from MWAS analysis for 4 quantitative traits in the LBC data. The DNAm measures were adjusted for age, sex, and batches. The phenotypes were stratified into groups by sex and cohort and were adjusted for age and standardized to z-scores by rank-based inverse normal transformation in each group. The phenotypes are a BMI, b height, c lung function, and d walking speed