Introduction to scOmics data structures

1 Introduction

起始于 raw counts matrix。

2 Seurat

Click to see a larger version of the image

The Seurat object contains the following important data structures:

assays：可以包含很多 assay（Assay in Seurat 3, and Assay5 in Seurat 5），如 RNA 和 SCT。每个 assay 以 layers 形式存储 gene by cell matrix（行为 genes/features，列为 cells/barcodes），如 counts、data 和 scale.data。
meta.data：存储与细胞相关的元信息。基因相关的元信息存储在对应 assay 的 meta.data 中。
graphs：it contains a list of graphs (usually a nearest-neighbor graph) that are used for clustering and other analyses.
reductions：存储细胞或基因在低维空间中的信息，如 PCA、UMAP 和 t-SNE。对于细胞：it is called the matrix of cell embeddings, which has the coordinates of each cell in a low-dimensional space that summarize the most important variation in the data. 对于基因：it is called the matrix of gene loadings, which describe how much each gene contributes to each PC. They tell you which genes drive the variation captured by a given PC.

其它的还有如 clusters、commands、misc 等。

Seurat object 一般存储为 RDS 文件即可。

3 AnnData

AnnData 一般存储为 H5AD 文件。

import anndata as ad

The active data matrix, a scipy sparse matrix, is in ad.X, which is the matrix of normalized and log(1+x)-transformed counts. Its rows are cells/observations and its columns are genes/variables.

Alternative versions of the active data in is layers, a Python dictionary, which stores each layer under a separate key, such as ad.layers["raw"].

对于 Seurat 而言，上述 matrix 均存储在某一个 assay 的 layers 中，如 RNA assay 中存储的 counts、data 和 scale.data。

Simple annotations (metadata) for cells and genes are in obs and var, Pandas Data Frame objects.

对于 Seurat 而言，就是 meta.data，或给定 assay 的 meta.data（对于基因）。

Index cells and genes with obs_names and var_names which directly mirror the indices of obs and var.

Subset AnnData with either cells, genes or both by name index, numerical index, and/or boolean index, such as ad[:4, ["LYZ", "FOS"]], which returns a view of the AnnData.

Multidimensional annotations for cells and genes are in obsm and varm, a Python dictionary with separate keys, such as ad.obsm["X_pca"].

对于 Seurat 而言，主要存储在 reductions 中。

Annotations for cell-cell and gene-gene pairs are in obsp (all matrices in obsp must have shape n_cells x n_cells and use obs_names as an index to both dimensions) and varp (all matrices in varp must have shape n_genes x n_genes and use var_names as an index to both dimensions), a Python dictionary.

对于 Seurat 而言，如 graphs。

Unstructured annotations or general metadata is in uns, a Python dictionary.

对于 Seurat 而言，诸如 misc。

Note: subsetting AnnData like ad[:4, ["LYZ", "FOS"]] returns a view. There are two cases where a copy will be returned.

Call .copy() like ad[:4, ["LYZ", "FOS"]].copy().
Modify any of elements of the view - the view object will be turned into copy.

4 MuData

mudata is the extension of anndata to multimodal datasets. It builds intuitively on anndata by treating the data from each modality as a separate AnnData object, but combining them to a joint, multimodal way in the MuData object.