anndata_skill

Overview

This skill provides a compact, practical interface to the AnnData annotated-matrix data structure used across single-cell and multimodal genomics. It helps create, read, write, subset, and manipulate .h5ad / zarr datasets and integrates smoothly with the scverse ecosystem (Scanpy, scvi-tools, muon). Use it to manage large, metadata-rich experiments with memory- and disk-efficient patterns.

How this skill works

The skill exposes common AnnData operations: constructing AnnData objects with X, obs, var and multi-dimensional annotations (layers, obsm, varm, obsp, varp, uns), reading and writing h5ad/zarr/loom/10x formats, and performing efficient subsetting and concatenation. It supports backed mode for out-of-core workflows, sparse matrix storage, and helpers for converting strings to categoricals and preserving raw data before filtering.

When to use it

• Loading or saving single-cell datasets (.h5ad, .zarr, .loom, 10x)
• Creating AnnData objects with cell and gene metadata
• Subsetting, filtering, or transforming annotated matrices
• Concatenating multiple batches or modalities with tracking labels
• Working with very large datasets using backed mode or sparse storage
• Preparing data for Scanpy, scvi-tools, muon, or deep learning pipelines

Best practices

• Use sparse matrices for inherently sparse counts (scipy.sparse.csr_matrix) to save memory
• Convert repeated strings to categoricals with strings_to_categoricals()
• Store raw before aggressive filtering: adata.raw = adata.copy()
• Use backed='r' for read-only access to very large h5ad files and load subsets to memory for processing
• Track sample/batch provenance via ad.concat(label='batch', keys=[...]) to simplify batch correction

Example use cases

• Preprocess a 10x single-cell experiment, compute highly variable genes, and save a processed .h5ad for downstream analysis
• Concatenate three experiment batches with batch labels and run Combat or other batch-correction workflows
• Open a 100+ GB h5ad in backed mode, filter cells by metadata, and load only the filtered subset to memory
• Create a MuData container combining RNA and protein AnnData objects for multimodal analysis
• Build an AnnLoader-backed DataLoader to stream batches into a PyTorch training loop