Shantel A. Martinez

Logo

Wheat Molecular Genetics | Preharvest Sprouting

HOME
Return to R Protocols

Genomic Prediction:

      Five-Fold CV Two-step approach; rrbLUP, RKHS, and LASSO broken down and explained in an R markdown notebook (MISSING reliabilty adjustment)

D. Sweeney helped me realize I was doing a two-step GP process initially without adjusting for the reliability. So this meant I was calculating BLUPs for my phenotype data, and those BLUP values were what I was using as a phenotype to calculate the GEBVs. So I guess (meaning I still need to read more on this), when you do a two-step process, you need to divide your GEBVs by a reliability factor. And this last part, I had not done before. Which the two-step method is not wrong, however without an adjustment I could be inflating my prediction accuracies.
However what I’m working on now, is a one-step approach. The 1 -step approach is using my covariate variables while I am running the mixed.solve command in rrBLUP.

       PHS GS One-step Approach rrbLUP five-fold cross-validation broken down and explained in an R markdown notebook

Bandwidth Parameter for RKHS RKHS: Based on genetic distance and a kernel function with a smoothing parameter to regulate the distribution of QTL effects. Effective for detecting nonadditive gene effects.

When I was trying to understand modeling using Reproducing Kernel Hilbert Spaces (RKHS), I also needed to understand the parameters being used by the model in BGLR. This lead me down a rabbit hole of defining a bandwidth parameter h for my datasets.

NOTE1: I am still in the process of fully understanding h and RKHS, so my thoughts and process so far are shown in the R notebook (above link), however it will evolve as I better understand h and read more articles.
NOTE2: I had a meeting with the statistical consulting group and both the advisor and I went down another rabbit hole of determining the bandwidth parameter for RKHS. But, there is hope!

i) people are still working out how to best identify your optimal bandwidth parameter, so at least I am not alone.
ii) some papers that are useful are Campos et al., 2010, Pérez-Elizalde et al., 2015, and Peres and Campos, 2014
“The bandwidth parameter of the Gaussian kernel can be chosen using either cross-validation (CV) or Bayesian methods. From a Bayesian perspective, one possibility is to treat h as random; however, this is computationally demanding because the RK needs to be recomputed any time h is updated. To overcome this problem de los Campos et al. (2010) proposed using a multikernel approach (named kernel averaging, KA) consisting of: (a) defining a sequence of kernels based on a set of values of h, and (b) fitting a multikernel model with as many random effects as kernels in the sequence.” - Peres and Campos, 2014