I have been using Tassel for GWAS in wheat for several years. I tried GAPIT too before, but for some reason, GAPIT is quite slow when I have over 20 K markers and 260 lines. So most of the time, I still use Tassel. In recent years, new GWAS methods have been reported, but I have not had a chance to try them yet. Miao et al (2008) compared different GWAS software/methods on the trait-variant association in crops under varying genetic architectures and found that FarmCPU provides the greatest statistical power for moderately complex traits. A few people also recommended FarmCPU to me. So I went to the FarmCPU webiste, but I found Zhiwu’s group has developed a new software called “BLINK”, which supposed to outperform FarmCPU in both speed and accuracy. I tested it with my wheat 90K SNP data and kernel shape data, but I did not get any significant SNPs with the message “The signal from LM is too weak!”. I read the preprint paper of BLINK and found that BLINK only allows examination of SNPs with P values of 0.01 after a multiple test correction. I think that is why BLINK is not working in wheat. I used the R version of BLINK, in which I could change the primary criteria, but I am not confident about this modification. I will contact the authors to confirm.
Wheat is a highly inbred species with very long LD blocks, but most of these software developed for GWAS are for maize, an outcrossing crop with a very short LD decay distance. And a lot of GWAS are also reported in outcrossing species. When I first touched GWAS in wheat, using the Bonferroni criteria or the FDR criteria, I could not find any significant markers. Later, after discussing with my wheat colleagues, I think the problem is that wheat has a long distance of LD decay, which is good for such a huge genome species because I can use a small number of markers to find MTAs, but the bad thing is that the associations are usually not that significant. What I do now is to see consistency rather than just highly significant P values. A few steps for GWAS in my research:
Estimate the overall LD decay distance and critical
$r^2$. In wheat, the average LD is about 10 cM depending on the populations. In our latest GWAS paper with a collection of 262 elite spring wheat materials from North America, the LDs ranged from 0.55 cM to 12.8 cM. In wheat, 1 cM is about 1 Mb on average.Estimate the number of highly informative markers based on the critical
$r^2$value using Haploview. In our latest GWAS paper, we used 22,226 SNPs, but there are only 1090 highly informative markers. So if we use Bonferroni correction at$\alpha = 0.1$, the significant raw P value should be$0.1/1090$, about$1e-4$, which is the P value we used for “highly significant SNPs”.Significant SNPs were identified using raw P values based on consistency across environments, since we have at least 3 trials for field experiments. In our latest GWAS paper, we defined “significant SNPs” as those that have raw P values < 0.05 in at least 3 environments and raw P values < 0.01 in at least 1 environment (we have 4 to 10 trials for different traits).
To better reduce the false positives, we also validated the significant SNPs using 8 nested association mapping populations (NAMs), which also allows us to have donor materials for gene clonings and breeding.