mixWAS

This repository contains code for the paper:

Li R., Benz L., Duan R., Denny J., Hakonarson H., Mosley J., Smoller J., Wei WQ., Ritchie M., Lumley, T., Moore J., and Chen Y. A One-Shot Lossless Algorithm for Cross-Cohort Learning in Mixed-Outcomes Analysis. (2024) (Pre-Print)

An R package for the mixWAS algorithm is available for download as follows

# install.packages('devtools')
devtools::install_github('lbenz730/mixWAS', build_vignettes = TRUE)

Example Usage

A detailed tutorial for using mixWAS can be explored via

vignette('mixwas_tutorial') 

Functions

mixWAS exports the following 4 functions, with the following arguments

• mixWAS: mixWAS algorithm for all sites (if all data can be supplied at once). Returns either p-value for SNP or score/variance components.

  • snps: list of snps (one for each site), each a vector of SNPs $\in {0,1,2}$
  • phenotypes: list of phenotypes, each a matrix of phenotypes (one per column), with names of phenotypes specified as column names
  • covariates: list of covariates, each a matrix or data frame of covariates
  • covariate_map: Default = NULL. If NULL, function assumes by all covariates are to be used for each phenotype. If this is not desired behavior, user can supply a data frame with two columns one called variable and a second called phenotype. In the variable column is the name of covariates, with phenotypes being specified as 'all' (to use the variable for all phenotypes) or the name of a phenotype. Variables can be entered multiple times if they go to multiple phenotypes (but not all). A phenotype specific data set of covariates with use 'all' covariate + phenotype specific covariates.
  • phenotype_index: list of vectors giving the index (numeric) of which phenotypes are in each site's matrix. If NULL (default), will be inferred from matrix colnames.
  • types: optional vector specifying data types e.g. ('continuous', 'binary', 'count'). Default = NULL (phenotype data types will be inferred), Note that 'count' will never be inferred, only 'binary' or 'continuous'.
  • parallel_sites: logical, if score/variance component computations should be parallelized over sites. Default = FALSE.
  • return_p: logical, if TRUE return P-values, else return components like score/variance. Default = TRUE.

• mixWAS_single_site: Compute score vector and covariance matrix for a single site via mixWAS algorithm.

  • snps: matrix of phenotypes (one per column), with names of phenotypes specified as column names.
  • phenotypes: matrix or data frame of covariates.
  • covariates: list of covariates, each a matrix or data frame of covariates.
  • covariate_map: Default = NULL. If NULL, function assumes by all covariates are to be used for each phenotype. If this is not desired behavior, user can supply a data frame with two columns one called variable and a second called phenotype. In the variable column is the name of covariates, with phenotypes being specified as 'all' (to use the variable for all phenotypes) or the name of a phenotype. Variables can be entered multiple times if they go to multiple phenotypes (but not all). A phenotype specific data set of covariates with use 'all' covariate + phenotype specific covariates.
  • types: optional vector specifying data types ('continuous', 'binary', 'count'). Default = NULL (phenotype data types will be inferred). Note that 'count' will never be inferred, only 'binary' or 'continuous'.

• run_hypothesis_test: run hypothesis tests from intermediate mixWAS components

  • score: score vector from mixWAS intermediate output
  • V_inv: Inverse Varariance Matrix from mixWAS intermediate output
  • z: Standardized Z-scores from mixWAS intermediate output
  • q: # of phenotypes}

• combine_site_results: Combine results from running mixWAS on each site individually

  • mixWAS_components: list of mixWAS_single_site output of length = # of sites
  • q: # of phenotypes
  • phenotypes: Optional list of phenotypes, each a matrix of phenotypes (one per column), with names of phenotypes specified as column names. If phenotype_index is NULL, will be used to infer phenotypes. One of phenotypes and phenotype_index must be specified
  • phenotype_index: list of vectors giving the index (numeric) of which phenotypes are in each site's matrix. If NULL (default), will be inferred from phenotypes.
  • return_p: logical, if TRUE return P-values, else return components like score/variance. Default = TRUE.

Simulations

Simulations from the paper are available in the simulations folder.