linear regression for removing the population effects

Question

[LJ]

"*To adjust for *population stratification, a linear regression of protein level on population label was performed and the residuals were normalized by transforming the quantiles of the residual values to their respective quantiles of a N(0,1) distribution**."what does this sentence mean? Is it a way to remove the population effect? So if it is ,how can i do this in R?

Answer 1

[dariober]

That's what I understand and how I would do it. First let's create some simulated data of protein level for 4 populations, each population with different mean (i.e. there is stratification that we want to remove):

N<- 100
a<- rnorm(n= N, mean= 1, sd= 2)
b<- rnorm(n= N, mean= 2, sd= 2)
c<- rnorm(n= N, mean= 3, sd= 2)
d<- rnorm(n= N, mean= 4, sd= 2)
dat<- data.frame(prot_lev= c(a, b, c, d), pop= rep(c('a', 'b', 'c', 'd'), each= N))
boxplot(prot_lev ~ pop, data= dat)

To adjust for *population stratification, a linear regression of protein level on population label was performed

So let's model protein level as a function of the population and get the residuals from the model:

lmreg<- lm(prot_lev ~ pop, data= dat)
dat$prot_lev_res<- lmreg$residuals

boxplot(prot_lev_res ~ pop, data= dat)

mean(dat$prot_lev_res) ## ~ zero as it should be 
sd(dat$prot_lev_res)   ## ~ 2 as per simulation

the residuals were normalized by transforming the quantiles of the residual values to their respective quantiles of a N(0,1) distribution

Data has already mean equal to zero, we need to make the stdev equal 1:

dat$prot_lev_qnorm<- lmreg$residuals/sd(lmreg$residuals)

mean(dat$prot_lev_qnorm) ## ~ 0
sd(dat$prot_lev_qnorm)  ## ~ 1

boxplot(prot_lev_qnorm ~ pop, data= dat)

[LJ]

Thanks for you reply.Two questions for you : (1)So you mean the author used the residuals from the linear model to represent the protein levels and the residuals were protein levels after removing the population effects ? (2)As you can see in the following code:

a<-rnorm(100)
b<-rnorm(50)
c<-rnorm(30)
dat<-data.frame(prot_lev= c(a, b, c), pop=c(rep(0,100),rep(1,50),rep(2,30)))
lmreg<- lm(prot_lev ~ pop, data= dat)
summary(lmreg)

And it goes like this:

Call: lm(formula = prot_lev ~ pop, data = dat)

Residuals: Min 1Q Median 3Q Max -2.24687 -0.66452 0.02218 0.76330 2.24357

Coefficients: Estimate Std. Error t value Pr(>|t|)
(Intercept) 0.07544 0.09237 0.817 0.4152
pop -0.16256 0.09505 -1.710 0.0889 .

--- Signif. codes: 0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 0.9636 on 178 degrees of freedom

Multiple R-squared: 0.01617, Adjusted R-squared: 0.01064

F-statistic: 2.925 on 1 and 178 DF, p-value: 0.08894

you can see the (Intercept) , pop and the model significance are greater than 0.05, so is it reasonable that i use the residuals in this model?

Answer 2

[Shab86]

Population stratification means any differences you might find in the allele frequencies within a population due to different ancestry. This could be due to non-random mating or admixture of populations in the past. This can be a problem for GWAS, where association is due to the underlying population structure and not a disease-associated locus.

Some papers to help you out:

1. http://genepath.med.harvard.edu/~reich/Reich%20and%20Goldstein.pdf (One of the earlier works)
2. http://www.hindawi.com/journals/ijg/2015/501617/ (example of how pop strat affects GWAS)
3. http://www.nature.com/nrg/journal/v11/n7/full/nrg2813.html

[LJ]

Thanks for your reply. But i'm not doing GWAS,what i need is: i have a 3000 genes expression data in 100 samples,and i konw the samples population.So how do i remove the population effects to normalize the expression data using the linear regression in R? What i need finally is the 3000 expression data in these samples after population effects removal,and i plan to use the data to do genes co-varying.So how do i normalize the data in R just like the sentence said?

[Sam]

Maybe you can try and input the population as a categorical variables in the analysis. Not sure if you are working on RNA Seq but DESeq2 got a very nice tutorial on how to perform such analysis. You can find it here

Answer 3

[LJ]

nobody has answer???

[Benn]

You got one answer, but it was not what you asked. Maybe that means that the question was not clearly phrased?