AI summary by TwinMind
StatQuest: Hierarchical Clustering
StatQuest with Josh Starmer
11m 19s
Introduction to Hierarchical Clustering and Heat Maps
- Hierarchical clustering often associated with heat maps
- Heat map structure:
- Columns represent different samples
- Rows represent measurements from different genes
- Red signifies high gene expression
- Blue/purple means lower gene expression
- Hierarchical clustering orders rows/columns based on similarity
- Makes correlation patterns visible
- Comparison shown: heat map without vs with clustering
- Heat maps often accompanied by dendrograms
Basic Clustering Process - Step-by-Step Example
- Simple example: 3 samples, 4 genes
- Step 1: Determine which gene most similar to gene #1
- Gene 1 vs Gene 2: very different (opposite expression patterns)
- Gene 1 vs Gene 3: similar (both red in sample 1, both blue in sample 3)
- Gene 1 vs Gene 4: also similar
- Gene 1 most similar to Gene 3
- Step 2: Find gene most similar to Gene 2
- Gene 2 most similar to Gene 4
- Step 3: Compare all combinations, find most similar pair
- Genes 1 and 3 are most similar overall
- Merge into Cluster #1
- Step 4: Repeat process treating cluster as single unit
- Cluster 1 most similar to Gene 4
- Gene 2 most similar to Gene 4
- Genes 2 and 4 most similar combination
- Merge into Cluster #2
- Final step: Merge remaining two clusters
Dendrograms
- Visual representation showing:
- Similarity levels (branch heights)
- Order of cluster formation
- Cluster 1 (genes 1,3): formed first, shortest branch (most similar)
- Cluster 2 (genes 2,4): formed second, second shortest branch
- Final cluster (all genes): formed last, longest branch
Distance Metrics for Similarity
- Euclidean distance most commonly used
- Formula: √[(difference in sample 1)² + (difference in sample 2)²]
- Example calculation: genes with values 1.6 vs 0.5 and 0.5 vs -1.9
- Result: √[1.1² + 2.4²] = 3.2
- Geometric interpretation: hypotenuse of triangle using Pythagorean theorem
- Extends to more samples by adding more squared differences
- Manhattan distance alternative
- Sum of absolute values of differences
- No squaring/square root operations
- Geometric interpretation: total length of line segments placed head-to-tail
- Distance metric choice is arbitrary
- No biological/physical reason to prefer one
- Choose based on data insights
- Visual comparison shows similar but slightly different heat maps
Cluster Comparison Methods
- Multiple approaches for comparing clusters to other points/clusters:
- Centroid method: compare to average of cluster measurements
- Single linkage: compare to closest point in cluster
- Complete linkage: compare to furthest point in cluster (R hclust default)
- Different methods produce similar but distinct heat maps
- Visual examples show data points on XY plane with different linkage approaches
Summary and Best Practices
- Cluster formation based on similarity notion (user decides)
- Most programs have reasonable defaults for:
- Distance metrics
- Cluster comparison methods
- Dendrogram branch heights indicate similarity levels
- Shorter branches = more similar clusters
