5.4.6.Clustering for single-cell genomics

About clustering

• Identify groups of similar cells (measurement objects)
• What clustering methods did we learn about?
• Could you explain the main ideas please?

Hierarchical and K-means clustering

→ See the clustering slides from the course materials / lectures of the EDA course.

Clustering in Seurat is a bit more complicated

Seurat clusters cells:

• Instead of genes, based on principal components:
  • each PC represents a ‘metagene’,
    • which combines information from multiple correlated genes.
• Seurat clustering is based on a community detection approach similar to SNN-Cliq
  • Algorithm: graph-based clustering

Graph-based clustering is based on K-nearest neighbours (KNN) classification

KNN classification

• Supervised approach (we have labels: red, blue)
• Is X red or blue?

Wikipedia

Graph-based clustering

Think of finding groups of friends among the people in UG!

Steps of graph-based clustering

• Embed cells in a graph structure
  • This graph is derived from KNN graph
    • calculated on the euclidean distance of cells in PCA space
    • refine the edge weights between any two cells based on the overlap in their local neighborhoods
• Find “cliques” or “communities” link in the graph