Patterns and Clustering

Patterns and clustering

Moderated F test

Genome-wide moderated F (any group differs) tests, for each gene, whether any group mean differs from the others, using all samples and empirical-Bayes variance moderation. It is the appropriate first test in designs with more than two groups.

Pattern clustering

1. The most variable genes (500, 1,000 or 2,000) are selected.
2. Each gene's group-mean profile is z-scored.
3. Profiles are clustered by k-means with seeded k-means++ initialization, so repeated runs give identical results.

Choosing k

With auto (silhouette), k from 2 to 10 is evaluated and the k with the highest mean silhouette width is chosen. The per-k silhouette scores are reported.

Best mean silhouette	Interpretation
Above 0.5	Well-separated clusters
0.3 to 0.5	Moderate structure
Below 0.3	Profiles form a continuum; treat clusters as a descriptive partition

The silhouette criterion replaced an earlier elbow criterion, which selected four or five clusters regardless of the true structure in benchmarks with known cluster numbers.

Trajectory heatmap

For numeric designs, the Trajectory heatmap shows every selected gene's profile over the ordered groups.

Gene-set module score

Paste a list of genes into Gene-set module score to plot their mean z-score per sample.

Reference

Rousseeuw PJ (1987) J Comput Appl Math 20:53.