Skill: Sample Clustering

Use When

• User wants to discover groups of similar samples or features in high-dimensional data
• User wants to create a heatmap with hierarchical clustering of genes and samples
• User wants to determine the optimal number of clusters for k-means
• Use cases: gene expression pattern discovery, cell type grouping, patient stratification

Inputs

• Required:
  • Data matrix (CSV or TSV): rows = samples (or features), columns = features (or samples)
• Optional:
  • Matrix orientation: samples_as_rows (default) or samples_as_columns
  • Clustering method: hierarchical (default) or kmeans
  • For hierarchical: linkage method (ward default, complete, average, single)
  • Distance metric: euclidean (default), correlation, or cosine
  • For k-means: number of clusters k (default: auto-select using elbow method, k = 2–10)
  • Scale features before clustering: true (default) or false
  • Metadata table for heatmap annotation (optional)
  • Output directory (default: ./cluster_output)

Workflow

1. Read data matrix; scale features if requested.
2. If hierarchical clustering: a. Compute distance matrix using specified metric. b. Perform hierarchical clustering on both rows (samples) and columns (features). c. Generate a clustered heatmap with dendrogram.
3. If k-means: a. If k not specified: compute within-cluster sum of squares for k = 2–10; plot elbow curve; select k at elbow. b. Fit k-means with selected k; report cluster assignments. c. Generate PCA-based cluster scatter plot.
4. Compute silhouette score for final cluster solution; report mean and per-sample scores.
5. Assign and report cluster labels for each sample.
6. Write cluster assignments (TSV), silhouette scores (TSV), heatmap or cluster plot (PDF) to output directory.
7. Save the executable analysis script (analysis_code.py or analysis_code.R) and software version manifest (session_info.txt) used to produce results.

Output Contract

• Cluster assignments table (TSV): sample_id, cluster_label
• Silhouette scores table (TSV): sample_id, silhouette_score, cluster_label
• Clustered heatmap (PDF, hierarchical) or cluster scatter plot (PDF, k-means)
• Elbow curve (PDF, k-means only when k is auto-selected)
• Mean silhouette score (printed to stdout)
• Reproducibility artifacts: analysis_code.py or analysis_code.R, plus session_info.txt

Limits

• K-means assumes spherical, similarly sized clusters; performs poorly on elongated or irregular shapes.
• Hierarchical clustering with Ward linkage and Euclidean distance is sensitive to outliers; remove outliers before clustering.
• Very large matrices (> 10k samples or > 50k features) may be slow; consider subsampling or running PCA first.
• Common failure: data with many zero entries (e.g., single-cell) — apply dimensionality reduction first.