Introduction to GWAS & Genomic Selection

1 · What is GWAS?

Genome-Wide Association Study (GWAS) is a statistical approach that scans the entire genome of many individuals — plants, animals, or humans — to find regions of DNA (SNP markers) that are statistically associated with a trait of interest.

In plant breeding, that trait might be grain yield in sorghum, flowering time in rice, or drought tolerance in maize. Instead of looking at one gene at a time, GWAS tests hundreds of thousands of SNPs simultaneously, asking: "Is this marker position significantly more different between high-yielding and low-yielding plants than we'd expect by chance?"

The core logic of GWAS

Imagine you have 300 sorghum lines. Each line has been genotyped at 50,000 SNP positions and phenotyped for grain yield. For every SNP, you run a linear regression:

# For each SNP marker j (j = 1 … 50,000):
y  =  μ  +  β_j · x_j  +  ε

# where:
#   y     = phenotype (e.g. grain yield, kg/ha)
#   μ     = overall mean
#   β_j   = allele substitution effect at SNP j
#   x_j   = marker genotype (0, 1, or 2 copies of alt allele)
#   ε     = residual error

The test asks: is β_j significantly different from zero? If yes, that SNP is "associated" with yield. A Manhattan plot then visualises the −log₁₀(p-value) for every SNP across every chromosome, and peaks show candidate regions.

2 · What is Genomic Selection?

Genomic Selection (GS), introduced by Meuwissen, Hayes and Goddard in 2001, is a method that uses marker data from across the entire genome to predict the breeding value of an individual — without needing to phenotype it in a field trial.

The workflow has two phases:

• Training phase: A reference population (genotyped + phenotyped) is used to fit a statistical model that links marker genotypes to observed phenotypes.
• Prediction phase: The fitted model predicts the Genomic Estimated Breeding Value (GEBV) for new candidates that have only been genotyped.

The GBLUP model

The most widely used GS model is GBLUP (Genomic Best Linear Unbiased Prediction). It assumes all SNP effects are normally distributed with equal variance — a ridge regression in marker space:

# GBLUP model:
y  =  Xβ  +  Zu  +  ε

# where:
#   y     = phenotype vector (n individuals)
#   β     = fixed effects (e.g. environment, block)
#   u     = random genomic effects ~ N(0, Gσ²_g)
#   G     = genomic relationship matrix (GRM)
#   ε     = residual ~ N(0, Iσ²_e)
#
# u is the GEBV vector — the breeding value we want to predict

In R, the rrBLUP package implements GBLUP via the mixed.solve() function. The BGLR package extends this to Bayesian models (BayesA, BayesB, BayesC, BayesR) that allow locus-specific effect variances — more powerful when a few large-effect QTL exist.

3 · Understanding Marker Data

Both GWAS and GS require genotype data: a matrix where rows are individuals and columns are SNP markers. Each cell holds the count of the alternate allele (0, 1, or 2), called additive dosage coding.

Typical SNP panel sizes

• GBS (Genotyping-by-Sequencing): 5,000–200,000 SNPs depending on genome size and coverage.
• SNP arrays: 9K, 50K, 90K chips exist for wheat, maize, sorghum.
• Whole-genome sequencing (WGS): Millions of SNPs; usually thinned by LD pruning for GS.
• Sorghum context: The sorghum genome is ~730 Mb. GBS panels of 10,000–100,000 SNPs are commonly used in GS research.

4 · GWAS vs Genomic Selection — Key Differences

5 · The Overall Workflow (This Module)

Across the 10 lessons of this module, you will build the complete pipeline from raw genotype data to a publishable GS analysis:

6 · Setting Up Your R Environment

All analyses in this module use R. Install the required packages now so you are ready from Lesson 2 onward.

# ── Install core GS / GWAS packages ──────────────────────

# rrBLUP: GBLUP, G-matrix, ridge regression
install.packages("rrBLUP")

# BGLR: Bayesian shrinkage models (BayesA/B/C/R, RKHS)
install.packages("BGLR")

# GAPIT3: GWAS with mixed model (MLM, BLINK, FarmCPU)
if (!requireNamespace("devtools", quietly = TRUE))
  install.packages("devtools")
devtools::install_github("jiabowang/GAPIT3")

# Visualisation
install.packages("qqman")         # Manhattan + QQ plots
install.packages("CMplot")        # Circular / rectangular Manhattan
install.packages("ggplot2")       # General plotting
install.packages("tidyverse")     # Data wrangling

# Data handling
install.packages("data.table")    # Fast reading of large genotype files
install.packages("Matrix")        # Sparse matrix support for large GRMs

# ── Verify installation ───────────────────────────────────
library(rrBLUP)
library(BGLR)
library(ggplot2)
library(data.table)

cat("rrBLUP version:", packageVersion("rrBLUP"), "\n")
cat("BGLR version:",   packageVersion("BGLR"),   "\n")

# Expected output (versions may differ slightly):
# rrBLUP version: 4.6.1
# BGLR version: 1.1.1

Create your project folder

# Create the folder structure for this module
mkdir -p ~/genomic-selection-bpv/{data/raw,data/processed,results,scripts}

# Move into the project folder
cd ~/genomic-selection-bpv

# Confirm the structure
tree .

# Expected output:
.
├── data
│   ├── processed
│   └── raw
├── results
└── scripts

7 · Exercises

# Simulate a marker matrix: 10 individuals × 20 SNPs
set.seed(42)
n_ind  <- 10   # number of individuals
n_snp  <- 20   # number of SNP markers

# Sample genotypes: 0, 1, or 2 (additive dosage)
M <- matrix(
  sample(0:2, n_ind * n_snp, replace = TRUE,
         prob = c(0.5, 0.3, 0.2)),   # minor allele freq ~0.25
  nrow = n_ind, ncol = n_snp
)

# Name rows (individuals) and columns (SNPs)
rownames(M) <- paste0("IND", 1:n_ind)
colnames(M) <- paste0("SNP", 1:n_snp)

# Inspect
dim(M)       # should be 10 x 20
M[1:5, 1:6]   # preview first 5 rows, 6 columns
table(M)     # counts of 0, 1, 2 across all cells