A cell atlas embeds each cell's gene-expression profile (thousands of genes) into a low-dimensional map via dimensionality reduction (PCA → UMAP/t-SNE), so cells of the same type cluster together spatially.
d_ij = ||x_i - x_j||
P(j|i) ∝ exp(-d_ij² / 2σ_i²) (neighbor prob., t-SNE/UMAP-style)
batch-corrected: x_i' = x_i - f(batch_i)
- Cluster count — number of distinct cell-type populations shown as separated point clouds.
- Batch integration — pulls the same cell type from different experimental batches together, simulating batch-effect correction (e.g. Harmony/scVI).
- Point size — visual density control for the point cloud (no biological meaning).
- Cells rendered — total single cells drawn; larger atlases (Human Cell Atlas) contain 10s of millions.
- Color toggle — switch between coloring by inferred cell type (biology) or by batch/donor (technical variation) to see why integration matters.
Real-world use: projects like the Human Cell Atlas and Tabula Sapiens integrate millions of single-cell profiles across labs into unified reference maps used to annotate new samples and study disease states.