obsm

obsm is a field of the AnnData data structure used in single-cell omics analysis (notably with Scanpy). It stores per-observation multidimensional representations, such as coordinate matrices for cells in low-dimensional embeddings or other per-observation projections.

Technically, obsm is a dictionary-like container that maps string keys to numpy arrays with shape (n_obs, n_dims). Here, n_obs is the number of observations (for example, cells) and n_dims is the dimensionality of the embedding (commonly 2 or 3). Each key identifies a distinct embedding or projection. Typical keys include 'X_pca', 'X_umap', 'X_tsne', and 'X_diffmap', though the names can vary.

Access and usage are straightforward. For example, coordinates = adata.obsm['X_umap'] yields the UMAP coordinates for all observations. The rows of obsm arrays align with the rows of adata.obs, meaning the i-th row in an obsm array corresponds to the i-th observation described in adata.obs.

obsm complements other AnnData fields. While adata.X stores the primary expression data and adata.varm can hold per-variable embeddings, obsm stores per-observation embeddings. This organization facilitates downstream visualization and analysis, such as plotting 2D or 3D embeddings, clustering, or trajectory inference results that are tied to individual observations.

Memory considerations apply, as large embeddings increase the in-memory footprint. Users typically populate obsm with embeddings produced from analyses like PCA or manifold learning and then reuse or share those coordinates for plotting and further analysis.