bio-data-visualization-dimensionality-reduction-plots

What it does

Produces and interprets PCA, t-SNE, UMAP, and PHATE plots for high-dimensional omics data, with emphasis on what each method preserves (variance, local structure, manifold, transitions), hyperparameter sensitivity, and the limits of 2D embeddings. Includes PCA biplot/scree/loadings, t-SNE PCA initialization (Kobak-Berens 2019), UMAP n_neighbors/min_dist trade-offs, and the Chari-Pachter 2023 critique. Applies to bulk PCA, single-cell embeddings, and multi-omics integration projections. Uses the primary tool: scanpy.

How it works

Guides choosing a projection method based on the plot's reveal: variance explained (PCA), local neighborhood structure (t-SNE), manifold approximation with some global structure (UMAP), or continuous transitions (PHATE). Encourages setting hyperparameters deliberately and communicating the projection's limits. Describes algorithmic taxonomy and provides examples in Python and R for PCA, t-SNE, UMAP, and PHATE, including relevant parameters (e.g., PCA/UMAP/t-SNE parameters) and warnings about embedding distortions. Emphasizes modern t-SNE improvements via PCA initialization and high learning rate, and UMAP defaults with seed control. Includes explicit failure modes and fixes (e.g., misinterpreting UMAP distances, seed reproducibility, perplexity handling, PCA scaling) and practical fixes for plotting, save paths, and DPI considerations. Reporting all hyperparameters used is recommended, along with explicit interpretation limits.

When to use it

Use when visualizing high-dimensional data such as bulk PCA, single-cell embeddings, or multi-omics projections to understand variance, local structure, or transitions, while avoiding over-interpretation of 2D distances and preserving awareness of embedding limitations. Triggered when selecting a projection method and tuning hyperparameters for accurate interpretation and reproducibility.

What it can touch

The skill references and prescribes usage patterns for Python tools (e.g., sklearn.decomposition.PCA, openTSNE, umap-learn, phate, scanpy.tl.umap / scanpy.tl.tsne / scanpy.tl.pca) and R counterparts (e.g., prcomp, PCAtools::pca, Seurat::RunPCA / RunUMAP / RunTSNE, phateR). It requires/uses the tool claude-code as declared. It instructs to report hyperparameters (perplexity, n_neighbors, min_dist, random_state) and to verify versions if compatibility concerns arise.

Caveats

Notes include the well-documented limits of 2D embeddings (Chari-Pachter 2023; Kobak-Berens 2019), dependency on proper scaling and normalization for PCA, seed requirements for reproducibility, and potential misinterpretation of distances in UMAP/t-SNE. It emphasizes that distances between clusters in 2D plots may be meaningless and that loadings are PCA-specific while UMAP/t-SNE coordinates do not provide linear interpretability for gene loadings. The license is NOASSERTION; the skill ships executable scripts and uses the primary tool scanpy.