Skill: Nonlinear Dimensionality Reduction

Use When

• User wants to visualize cell populations or sample clusters in 2D after scRNA-seq, flow cytometry, or omics analysis
• User wants a t-SNE or UMAP plot colored by cell type, cluster, or metadata
• User needs a nonlinear embedding that preserves local neighborhood structure
• Use stat-pca for linear embedding first, then this skill for refined visualization

Inputs

• Required:
  • Data matrix (CSV or TSV): rows = samples/cells, columns = features; OR a precomputed PCA coordinate table
• Optional:
  • Embedding method: umap (default) or tsne
  • Input type: raw (default) or pca (precomputed PC coordinates)
  • For t-SNE: perplexity (default: 30), max iterations (default: 1000), learning rate (default: 200)
  • For UMAP: n_neighbors (default: 15), min_dist (default: 0.1), metric (default: euclidean)
  • Metadata table (CSV or TSV) for coloring the plot
  • Color-by column from metadata
  • Random seed (default: 42)
  • Output directory (default: ./embedding_output)

Workflow

1. Read data matrix or precomputed PCA coordinates.
2. If raw data: run PCA first (top 50 PCs) and use PCs as input to embedding (standard practice for large omics datasets).
3. Fit t-SNE or UMAP using specified parameters and random seed.
4. Extract 2D embedding coordinates for each sample/cell.
5. If metadata provided: join embedding coordinates with metadata; generate scatter plot colored by specified column.
6. If no metadata: generate plain scatter plot.
7. Report: number of samples embedded, method and parameters used, runtime.
8. Write embedding coordinates (TSV) and scatter plot (PDF) to output directory.
9. 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

• Embedding coordinates (TSV): sample_id, dim1, dim2
• Scatter plot (PDF): 2D embedding colored by metadata or uniform color
• Parameters used (printed to stdout): method, all hyperparameters, random seed
• Reproducibility artifacts: analysis_code.py or analysis_code.R, plus session_info.txt

Limits

• t-SNE and UMAP are non-deterministic (set seed for reproducibility); results vary with hyperparameters.
• t-SNE distances between clusters are not interpretable — use only for local structure visualization.
• UMAP better preserves global structure than t-SNE and is generally faster.
• Both methods are sensitive to the scale of input data; ensure PCA pre-processing for high-dimensional input.
• Common failure: memory error for very large inputs (> 100k cells) — subsample or use approximate UMAP.