That's a complicated task: teach you how to perform GWAS ASAP from my computer...

As you even mentioned, there are many different software for this. You chose to use GAPIT in R. You can use SAS, GenABEL (in R as well), qxkap, ASReml, PLINK, and more..

All of these methods are based on a single-SNP type of analysis, testing one SNP at a time, treating them as fixed effects. Some of these software allows for random effects (SAS/ASReml/qkpak) and some also allow the use of relationship matrices (ASReml and qxpak).

Other methods, as those implemented in GenSel, fit all (of several) SNPs simultaneously in the model, as random effects. These are known as Bayesian methods (BayesA, BayesB, BayesC, etc..), and are usually used for genomic prediction, but have been shown to be very good for GWAS as well.

I've found these slides online showing how to use the R package you chose: http://pbgworks.org/sites/pbgworks.org/files/GAPIT_with_SYslides.pdf

Performing GWAS is not complicated, but you need to pay attention to your model.. What are ALL possible sources of variation that you have? Do you have 300 different genotypes, or 300 individuals within genotype? What's your sample size?

Some of these software also allows for interactions between SNPs and effects in the model (e.g. SNP x Location).

Nick

Thank you very much Nick for your informative answer. 

Nick VL Serão Thank you for response . Do you know a method in which we analysis all SNPs together? I mean the modle similar to :

Y= mu + snp1+ snp2 + ....+ SNPn.

So the model select the most important SNPs. Such as LOGISTIC regression with maximum R-square, for example.

Sincerely,

Seyed Abbas Rafat

Thanks for the questions.

Generally, in models where all SNPs are fitted simultaneously, SNP effects are considered random effects, such as in ridge regression (we call it regression GBLUP), BayesA, BayesB, BayesC, etc. In some of these methods, such as in BayesB and BayesC, there is also model selection going on.

If you are using SNPs as fixed effects (which is seen in most cases), you must consider the degrees-of-freedom of your model to see how many SNPs could be fitted in the model.

Using "n" as the sample size (ie, number of individuals) , "p" as the number of other effects in the model, and "m" as the number of SNPs you have, you can fit a value "m*" of SNPs such that m* + p + 2 = n

Therefore, it is required that m < n ... so, is this your case?

Otherwise, you really need to perform some sort of model selection (e.g., stepwise) or some sort of method that reduces dimensionality (e.g., partial least squares).

All of these have nothing to do with your response variable, which seems to be binary, and thus, the use of Logistic Regression would be appropriate (although Logistic is not the only method).

As for your last comment about R2, please understand that if m > n and you fit a model where m* + p + 2 = n, then, by definition, your R2 will be 1. That does not mean AT ALL that your model is the best one, even with R2 = 1.

Using R2 that is obtained by using the same data to develop the model is highly not indicated. AIC, BIC, etc, are better model fit statistics to select the model.

Please let me know if these make more sense to you now.

Thanks, Nick

Talk to someone who knows how to do it and learn from him/her.

@Nick VL Serão Thank you so much for your response. All Best.

Nick VL Serão In calculation of Degree of Freedom, the level of each effect is important. For example we consider a case where some SNPs have 2 levels and some 3. So, based on the formula (m* + p + 2 = n) , how we can consider the level for each SNPs?

I mean, the level of effect must be considered. Whenever there are more levels for each effect, we need more sample size. Please see the following example in SAS. Output is in attachment file.Thank you in advance. All the Best.

data new;

input obs snp1 snp2 snp3 snp4 snp5 snp6 y;

cards;

1 0 0 0 1 1 1 50

2 0 0 0 1 1 1 52

3 0 0 0 1 1 1 50

4 1 1 1 0 0 0 60

5 1 1 1 0 0 0 70

6 1 1 1 0 0 0 45

;

run;

title1 " Degree of freedom Error ";

title2 " m* + p + 2 = n 4 + 0 +2 =6 ";

proc anova;

class snp1 snp2 snp3 snp4 ;

model y= snp1 snp2 snp3 snp4 ;

run;

data new2;

input obs snp1 snp2 snp3 snp4 snp5 snp6 y;

cards;

1 0 0 0 0 0 0 50

3 1 1 1 1 1 1 50

5 2 2 2 2 2 2 70

4 0 0 0 0 0 0 50

5 1 1 1 1 1 1 50

6 2 2 2 2 2 2 70

;

run;

title1 " Degree of freedom Error ";

title2 " m* + p + 2 = n 4+0+2=6 ";

proc anova;

class snp1 snp2 snp3 ;

model y= snp1 snp2 snp3 ;

run;

Seyed Abbas Rafat , indeed, given that you are analyzing the SNP data as genotypes, those with 3 levels will have 2 degrees-of-freedom.

Due to my bias as a breeder, I think of SNPs as covariates (regression based on the number of the reference allele), in which we fit the allele substitution effect, and not the 3 classes of phenotypes, as I am (mostly) interested in the additive effects, since that’s the only heritable part.

Hope this makes sense!

Let me know if you have any other questions. Thanks, Nick

Gapit has a user manual which I think is very comprehensive and has several models you can test (both single-locus and multi-locus).

For multi-locus GWAS, you could as well explore R packages such as rMVP and mrMLM (or mrMLM.GUI which has a graphical user interface). For multi-trait GWAS, you may try GEMMA, mtmlSEM, MTMM.... (check literature).

But make sure you read enough literature to familiarize yourself with GWAS steps: from experimentation to analysis, and understand what the results you get mean/imply (pay attention to how your QQ and Manhattan plots look like, understand the requirements for you particular mapping panel, how to analyze the phenotypic data and what to use to run the GWAS, genotyping types and QC steps, Extent of LD decay, kinship and population structure, etc).

All the best...

https://zzlab.net/GAPIT/gapit_help_document.pdf