# 单因素分析
tra.cox <- t(apply(tra.data[,3:c(ncol(tra.data))],2,function(x){
  vl=as.numeric(x)
  tm=tra.data$OS.time
  ev=tra.data$OS
  #ev=ifelse(ev=='Alive',0,1)
  dat=data.frame(tm,ev,vl)[which(tm > 0 & !is.na(vl)),]
  return(coxFun(dat))
}))
colnames(tra.cox)=c('p.value','HR','Low 95%CI','High 95%CI')
length(which(tra.cox[,1]<0.05))

tra.cox <- na.omit(tra.cox)
filter_genes <- rownames(tra.cox[tra.cox[,1]<0.05, ])
write.csv(tra.cox[tra.cox[,1]<0.05, ],
          file = 'results/S6.csv',
          quote = F)
library(glmnet)
set.seed(num)
fit1=glmnet(as.matrix(tra.data[,filter_genes])
            #,factor(samps)
            ,cbind(time=tra.data$OS.time,
                   status=tra.data$OS)
            ,family="cox"
            #,family="binomial"
            #,type.measure="deviance"
            ,nlambda=100
            , alpha=1) 

cv.fit<-cv.glmnet(as.matrix(tra.data[,filter_genes])
                  #,factor(samps)
                  ,cbind(time=tra.data$OS.time,
                         status=tra.data$OS)
                  ,family="cox"
                  #,family="binomial"
                  #,type.measure="deviance"
                  ,nlambda=100
                  , alpha=1)
sig.coef <- coefficients(cv.fit,s=cv.fit$lambda.min)[which(coefficients(cv.fit,s=cv.fit$lambda.min)[,1]!=0),1]
cv.fit$lambda.min

mg_plot_lasso <- function(fit,cv_fit,lambda=NULL,show_text=T,figLabels=c('A','B')){
  if(is.null(lambda)){
    lmda=cv_fit$lambda.min
  }else{
    lmda=lambda
  }
  fit.coef=fit$beta[(apply(fit$beta,1,function(x){
    return(sum(x!=0))
  })>0),]
  
  fit.coef=as.matrix(fit.coef)
  colnames(fit.coef)=fit$lambda
  #fit$lambda==cv_fit$lambda
  library(ggplot2)
  dat=data.table::melt(t(as.matrix(fit.coef)))
  dat_z=dat[which(dat$value==0),]
  dat=dat[which(dat$value!=0),]
  dat.sv=rbind()
  for (u in unique(dat_z[,2])) {
    t.z=dat_z[which(dat_z[,2]==u),1]
    t.zx=max(t.z)
    dat.sv=rbind(dat.sv,c(t.zx,u,0))
    t.zn=min(t.z)
    if(t.zx!=t.zn){
      dat.sv=rbind(dat.sv,c(t.zn,u,0))
    }
  }
  colnames(dat.sv)=colnames(dat_z)
  #dat_z=dat_z[dat_z[,2]%in%names(which(fit.coef[,which(fit$lambda==lmda)]!=0)),]
  dat=crbind2DataFrame(rbind(dat,dat.sv))
  mn=min(-log(dat$Var1))
  mx=max(-log(dat$Var1))
  if(show_text){
    mx=(mx-mn)*0.1+mx
  }
  p=ggplot(dat, aes(x=-log(Var1), y=value,colour=Var2))+geom_line()+theme_bw()+theme(legend.position = "none")
  p=p+coord_cartesian(xlim=c(mn, mx))+xlab('-ln(lambda)')+ylab('Coefficients')
  if(show_text){
    fl=fit.coef[which(fit.coef[,which(fit$lambda==lmda)]!=0),ncol(fit.coef)]
    for_label=data.frame(Var1=rep(min(dat$Var1),length(fl)),Var2=names(fl),value=fl)
    p=p+ggrepel::geom_label_repel(
      aes(label = Var2,color=Var2),
      data = for_label,hjust = 0
    )
  }
  p=p+geom_vline(aes(xintercept=-log(lmda)), colour="#BB0000", linetype="dashed")
  p=p+annotate('text',x=-log(lmda),y=min(dat[,3]),label=paste0('lambda=',round(lmda,4)))
  tgc=data.frame(lambda=cv_fit$lambda,cvm=cv_fit$cvm,cvup=cv_fit$cvup,cvlo=cv_fit$cvlo,cvsd=cv_fit$cvsd
                 ,col=ifelse(cv_fit$lambda>=cv_fit$lambda.min&cv_fit$lambda<=cv_fit$lambda.1se,ifelse(cv_fit$lambda==lmda,'A','C'),'B'))
  p1=ggplot(tgc, aes(x=log(lambda), y=cvm)) + xlab('ln(lambda)')+ ylab('Parial Likelihood Deviance')+
    geom_errorbar(aes(ymin=cvm-cvsd, ymax=cvm+cvsd)) +
    geom_point(aes(colour=col))
  p1=p1+theme_bw()+theme(legend.position = "none")
  gal=ggpubr::ggarrange(p,p1, ncol = 2, nrow = 1
                        #,align = "hv"
                        ,labels = figLabels)
  return(gal)
}
lasso.pdf <- mg_plot_lasso(fit1,
                           cv.fit,
                           # lambda = cv.fit$lambda.min,
                           show_text=T,
                           figLabels=c('A','B'))
lasso.pdf
ggsave(plot = lasso.pdf,
       filename = 'PDFs/lasso.pdf',
       width = 10, height = 5)

pdf('PDFs/lasso.pdf',width = 10,height = 5)
par(mfrow=c(1,2))
plot(fit1, xvar="lambda")
plot(cv.fit)
dev.off()


# 基因的多因素
tcga_dat1 <- cbind(time=tra.data$OS.time,
                   status=tra.data$OS,
                   tra.data[,names(sig.coef)])

fmla <- as.formula(paste0("Surv(time, status) ~"
                          ,paste0(names(sig.coef),collapse = '+')))


cox <- coxph(fmla, data =as.data.frame(tcga_dat1))
lan <- coef(cox)
round(lan, 3)
genes <- names(cox$coefficients)
tra.cox[genes,]