R/phewas.R

Defines functions phewas

Documented in phewas

phewas <-
function(phenotypes,genotypes,data,covariates=c(NA),adjustments=list(NA), outcomes, predictors, cores=1, additive.genotypes=T, 
         significance.threshold, alpha=0.05, unadjusted=F, return.models=F, min.records=20, MASS.confint.level=NA,quick.confint.level) {
  if(missing(phenotypes)) {
    if(!missing(outcomes)) phenotypes=outcomes
    else stop("Either phenotypes or outcomes must be passed in.")
  }
  if(missing(genotypes)) {
    if(!missing(predictors)) genotypes=predictors
    else stop("Either genotypes or predictors must be passed in.")
  }
  association_method=phe_as
  if(unadjusted) {
    association_method=phe_as_unadjusted
    if(!is.na(covariates) | !is.na(adjustments)) warning("Covariates and adjustments are ignored in unadjusted mode.")
  }
  #If input is formatted as a set of data frames, reshape it
  if(missing(data)) {
    phe=phenotypes
    gen=genotypes
    cov=covariates
    adjustment=adjustments
    id=intersect(names(phenotypes),names(genotypes))
    if(length(id)==0) {stop("There is no shared column to merge phenotypes and genotypes!")}
    message(paste("Merging data using these shared columns: ",id))
    phenotypes=names(phenotypes)
    phenotypes=phenotypes[!(phenotypes %in% id)]
    genotypes=names(genotypes)
    genotypes=genotypes[!(genotypes %in% id)]
    if(length(phenotypes)<1 || length(genotypes)<1) 
      {stop("Either phenotypes or genotypes contained no non-shared columns, yielding no variables for analysis after the merge.")}
    data=merge(phe,gen,by=id)
    if(!is.null(names(covariates)[-1])) {
      covariates=names(covariates)
      if(sum(id %in% covariates)!=length(id)){stop(paste("The shared ID column(s) do not all exist in covariates: ",id))}
      covariates=covariates[!(covariates %in% id)]
      data=merge(data,cov,by=id)
    }
    if(!is.null(names(adjustments)[-1])) {
      adjustments=names(adjustments)
      if(sum(id %in% adjustments)!=length(id)){stop(paste("The shared ID column(s) do not all exist in adjustments: ",id))}
      adjustments=as.list(c(NA,adjustments[!(adjustments %in% id)]))
      data=merge(data,adjustment,by=id)
    }
  }
  para=(cores>1)
  
  #Check to make sure that there were >=1 phenotypes and genotypes
  if(length(phenotypes)<1 || length(genotypes)<1) {stop("You must provide at least one genotype/predictor and one phenotype/outcome for analysis.")}
  
  #Create the list of combinations to iterate over
  full_list=data.frame(t(expand.grid(phenotypes,genotypes,adjustments,stringsAsFactors=F)),stringsAsFactors=F)

  #If parallel, run the parallel version.
  if(para) {
    #Check to make sure there is no existing phewas cluster.
    if(exists("phewas.cluster.handle")) {
      #If there is, kill it and remove it
      message("Old cluster detected (phewas.cluster.handle), removing...")
      try(stopCluster(phewas.cluster.handle), silent=T)
      rm(phewas.cluster.handle, envir=.GlobalEnv)
    }
    message("Starting cluster...")
    assign("phewas.cluster.handle", makeCluster(cores), envir = .GlobalEnv)
    message("Cluster created, finding associations...")
    clusterExport(phewas.cluster.handle,c("data", "covariates"), envir=environment())
    #Loop across every phenotype- iterate in parallel
    result <-parLapplyLB(phewas.cluster.handle, full_list, association_method, additive.genotypes, confint.level=MASS.confint.level, min.records,return.models)
    #Once we have succeeded, stop the cluster and remove it.
    stopCluster(phewas.cluster.handle)
    rm(phewas.cluster.handle, envir=.GlobalEnv)
  } else {
    #Otherwise, just use lapply.
    message("Finding associations...")
    result=lapply(full_list,FUN=association_method, additive.genotypes, min.records,return.models, confint.level=MASS.confint.level, data, covariates)
  }

  if(return.models) {
    message("Collecting models...")
    models=lapply(result,function(x){attributes(x)$model})
    names(models)=sapply(models,function(x){paste0(as.character(terms(x))[c(2,1,3)],collapse=" ")})
  }
  
  message("Compiling results...")
  successful.phenotypes=na.omit(sapply(result,function(x){attributes(x)$successful.phenotype}))
  n.tests=length(successful.phenotypes)
  successful.phenotypes=unique(successful.phenotypes)
  successful.genotypes=unique(na.omit(sapply(result,function(x){attributes(x)$successful.genotype})))
  sig=bind_rows(result)
  
  #Report warning if any convergence errors
  if(max(grepl(pattern = "[Error: The model did not converge]", sig$note, fixed=TRUE))){
    warning("Not all models converged, check the notes column for details.")
  }
  
  message("Cleaning up...")
  
  #Add significance thresholds
  attributes(sig)$alpha=alpha
  attributes(sig)$n.tests=n.tests
  if(!missing(significance.threshold)) {
    message("Finding significance thresholds...")
    thresh=match(c("p-value","bonferroni","fdr","simplem-genotype","simplem-phenotype","simplem-product"),significance.threshold)
    sm.g=1
    sm.p=1
    #p.value
    if(!is.na(thresh[1])) {
      sig$p.value=sig$p<=alpha
    }
    #bonferroni
    if(!is.na(thresh[2])) {
      sig$bonferroni=sig$p<=alpha/n.tests
      attributes(sig)$bonferroni=alpha/n.tests
    }
    #fdr
    if(!is.na(thresh[3])) {
      sig$fdr=p.adjust(sig$p,method="fdr")<=alpha
    }
    #simplem-genotype
    if(!is.na(thresh[4])|!is.na(thresh[6])) {
      if(length(successful.genotypes)>1){
        eigs=eigen(cor(data[,genotypes],use="pairwise.complete.obs",method="spearman"))[[1]]
        max.eig=sum(eigs)
        sm.g=which.max(cumsum(eigs)>.995*max.eig)
      } else {
        sm.g=1
      }
      sig$simplem.genotype=sig$p<=alpha/sm.g
      attributes(sig)$simplem.genotype=alpha/sm.g
      attributes(sig)$simplem.genotype.meff=sm.g
    }
    #simplem-phenotype
    if(!is.na(thresh[5])|!is.na(thresh[6])) {
      if(length(successful.phenotypes>1)) {
        eigs=try(cor(data[,successful.phenotypes],use="pairwise.complete.obs",method="spearman"),silent=T)
        if(class(eigs)!="try-error"){
          eigs[is.na(eigs)]=0
          eigs=eigen(eigs)[[1]]
          max.eig=sum(eigs)
          sm.p=which.max(cumsum(eigs)>.995*max.eig)
        } else {
          warning("Phentoype correlation generation failed; this is typically due to sparse phenotypes.")
          sm.p=NA
        }
      } else {
        sm.p=1
      }
      sig$simplem.phenotype=sig$p<=alpha/sm.p
      attributes(sig)$simplem.phenotype=alpha/sm.p    
      attributes(sig)$simplem.phenotype.meff=sm.p
    }
    #simplem-product
    if(!is.na(thresh[6])) {
      sm=sm.g*sm.p
      sig$simplem.product=sig$p<=alpha/sm
      attributes(sig)$simplem.product=alpha/sm
      attributes(sig)$simplem.product.meff=sm
    }
  }
  if(!missing(outcomes)) names(sig)[names(sig)=="phenotype"]="outcome"
  if(!missing(predictors)) names(sig)[names(sig)=="snp"]="predictor"
  if(return.models){sig=list(results=sig,models=models)}
  if(!missing(quick.confint.level)) {
    if(quick.confint.level>=1|quick.confint.level<=0) {warning("Quick confidence interval requested, but a value in the range (0,1) was not supplied")}
    else {
      sig.names=names(sig)
      two.sided=(1-quick.confint.level)/2
      sig=sig %>% mutate(lower.q=beta+se*qnorm(two.sided),upper.q=beta+se*qnorm(two.sided,lower.tail=F))
      sig=sig  %>% mutate(lower.q=ifelse(sig$type=="logistic",exp(lower.q),lower.q),
                      upper.q=ifelse(sig$type=="logistic",exp(upper.q),upper.q))
      sig=sig[,c(sig.names[1:5],"lower.q","upper.q",sig.names[6:length(sig.names)])]
    }
  }
  return(sig)
}
PheWAS/PheWAS documentation built on Nov. 16, 2018, 6:28 p.m.