Module 2: Population Structure & PCA

Visualizing genetic relationships

What is Population Structure?

Population structure means that individuals in your dataset are not randomly related — some are more similar to each other than to others.

This happens because of:

  • Geographic separation — Plants from different regions evolved separately
  • Breeding history — Different breeding programs created different genetic groups
  • Species/subspecies — Mixing samples from related species

Why Does It Matter for Breeding?

Population structure affects almost every genomic analysis:

Analysis How structure affects it
GWAS Can cause false positives — structure looks like trait association
Genomic selection Models may not work across populations
Diversity estimates Mixing populations inflates diversity
ImportantRule of thumb

Always check population structure before running GWAS or genomic prediction.

What is PCA?

PCA (Principal Component Analysis) is a way to visualize high-dimensional data in 2D or 3D.

The Intuition

Imagine you have 5,000 SNPs. Each sample can be thought of as a point in 5,000-dimensional space. We can’t visualize that!

PCA finds the main axes of variation and projects everything down to 2 or 3 dimensions while preserving as much of the structure as possible.

Diagram showing how PCA reduces dimensions while preserving structure

PCA concept diagram

Reading a PCA Plot

Each dot is one sample. Samples close together are genetically similar.

PCA plot showing two distinct clusters of samples

Example PCA plot

In this example: - Two clear clusters → likely two genetic groups - Samples in between → possible hybrids or admixed individuals - Outliers → check for mislabeling or contamination

Running PCA in BIGapp

Demo: PCA Analysis

Watch how to run PCA:

Animated GIF showing the PCA workflow in BIGapp

Running PCA in BIGapp

Steps:

  1. Go to “PCA” in the sidebar
  2. Upload your VCF file (atlantic_giant_pumpkin_diversity.vcf.gz)
  3. Click Run PCA
  4. Explore the interactive plot

Understanding the Output

BIGapp gives you:

  1. Interactive PCA plot — Hover over points to see sample IDs
  2. Variance explained — How much information each PC captures
  3. PC loadings — Which SNPs contribute most to each PC

Try It Yourself

  1. Run PCA on your filtered pumpkin data
  2. Look at the plot — do you see distinct clusters?
  3. Check the variance explained — how much does PC1 capture?

With the pumpkin dataset, you should see three overlapping clusters corresponding to the Eastern, Midwest, and Western US breeding regions. The clusters overlap because there’s been gene flow between regions — this is realistic for most breeding populations!

Interpreting Your Results

Scenario 1: Clear Clusters

PCA showing distinct population clusters

PCA with clear clusters

What it means: Your data contains distinct genetic groups.

What to do: - Consider whether to analyze groups separately - Account for structure in GWAS (BIGapp does this automatically) - For genomic selection, training within groups usually works better

Scenario 2: Continuous Variation

PCA showing continuous spread without distinct clusters

PCA with continuous variation

What it means: No strong substructure — samples vary along a gradient.

What to do: - This is often ideal for genomic prediction - GWAS should still include PCs as covariates

Scenario 3: Outliers

PCA showing main cluster with outlier samples

PCA with outliers

What it means: Some samples are very different from the rest.

What to do: - Check sample labels — are these mislabeled? - Check DNA quality — did something fail? - Are these a different species/variety? - Consider removing or analyzing separately

Coloring by Groups

If you have metadata (population, breeding program, location), you can color the PCA:

Animated GIF showing how to color PCA by metadata

PCA colored by population

This helps you see if genetic structure matches your expectations.

Steps:

  1. Upload the phenotype file (atlantic_giant_pumpkin_diversity_phenotypes.csv)
  2. In the PCA section, select Region as the grouping variable
  3. The plot updates with colors — you should see Eastern, Midwest, and Western clusters!
TipTry different groupings

Color by State to see a finer geographic gradient, or by Breeding_Program to see how programs cluster genetically.

How Many PCs to Use?

Look at the scree plot — it shows how much variance each PC explains.

Scree plot showing variance explained by each PC

Scree plot example

Rule of thumb: Keep PCs until the curve levels off (the “elbow”). Often this is 3-10 PCs.

For GWAS and genomic prediction, BIGapp uses PCs automatically to correct for structure.

DAPC: An Alternative to PCA

DAPC (Discriminant Analysis of Principal Components) is another method for visualizing structure. It’s better at separating known groups.

Use DAPC when: - You already know your group labels - You want to maximize separation between groups - You’re interested in which markers differentiate groups

BIGapp includes DAPC in the same section as PCA.

Exercise

Using your filtered pumpkin data:

  1. Run PCA
  2. How many distinct clusters do you see?
  3. What percentage of variance does PC1 explain? PC2?
  4. Are there any outliers?
  5. Color by Region — does genetic structure match geographic origin?

You should see three overlapping clusters that align with the Eastern, Midwest, and Western regions. PC1 typically captures 5-10% of variance in diverse breeding populations. If you color by State, you’ll see a geographic gradient within each regional cluster.

Key Takeaways

✅ Population structure means some samples are more related than others

✅ PCA visualizes this structure in 2D

✅ Clusters = genetic groups; outliers = check for problems

✅ Always check structure before GWAS or genomic prediction

✅ Use PCs as covariates to correct for structure

Next: Module 3: GWAS — Finding marker-trait associations