Source: https://github.com/aipoch/medical-research-skills

When to Use

Use AnnData when you need to:

• Load, inspect, or export annotated single-cell datasets stored as .h5ad (or zarr) for downstream tools.
• Keep a matrix (cells × features) tightly coupled with observation/feature metadata (e.g., cell types, batches, gene annotations).
• Work efficiently with large, sparse count matrices (e.g., scRNA-seq) and avoid loading everything into memory (backed mode).
• Combine multiple experiments/batches/modalities into a unified object while tracking provenance.
• Subset/filter/transform data while preserving alignment between the matrix and metadata.

Key Features

• Unified container: X (data matrix) plus aligned annotations: obs, var, uns, and multi-dimensional slots (obsm, varm, obsp, varp), plus layers and optional raw.
• Interoperable I/O: Native .h5ad and zarr, plus common genomics formats (e.g., 10x, loom, mtx, csv).
• Scalable workflows: Sparse matrices and backed mode (backed="r") for large datasets.
• Safe subsetting: Slicing preserves alignment across matrix and annotations; supports views vs copies.
• Concatenation utilities: ad.concat(...) with join/merge strategies and batch labeling; experimental lazy collections.

Reference notes: the original material mentions additional guides under references/ (e.g., references/data_structure.md, references/io_operations.md, references/concatenation.md, references/manipulation.md, references/best_practices.md) for deeper explanations of each topic.

Dependencies

• anndata (latest compatible with your environment; install via pip/uv)
• numpy
• pandas
• scipy (recommended for sparse matrices)
• Optional ecosystem tools (only if needed):
  • scanpy
  • muon
  • torch (for deep learning) and anndata experimental loader utilities

Example Usage

A complete runnable example that creates an AnnData object, writes/reads .h5ad, subsets, concatenates batches, and demonstrates backed mode.

import numpy as np
import pandas as pd
import anndata as ad
from scipy.sparse import csr_matrix

# ----------------------------
# 1) Create an AnnData object
# ----------------------------
rng = np.random.default_rng(0)

n_cells, n_genes = 100, 500
X = rng.poisson(1.0, size=(n_cells, n_genes)).astype(np.float32)

obs = pd.DataFrame(
    {
        "cell_type": (["T cell", "B cell"] * (n_cells // 2)),
        "sample": (["A", "B"] * (n_cells // 2)),
        "quality_score": rng.random(n_cells),
    },
    index=[f"cell_{i}" for i in range(n_cells)],
)

var = pd.DataFrame(
    {"gene_name": [f"Gene_{j}" for j in range(n_genes)]},
    index=[f"ENSG{j:05d}" for j in range(n_genes)],
)

adata = ad.AnnData(X=X, obs=obs, var=var)

# Use sparse storage for typical count-like matrices
adata.X = csr_matrix(adata.X)

# Convert string columns to categoricals to reduce memory and speed up ops
adata.strings_to_categoricals()

print(f"Created: {adata.n_obs} obs × {adata.n_vars} vars")

# ----------------------------
# 2) Write and read .h5ad
# ----------------------------
adata.write_h5ad("example.h5ad", compression="gzip")

adata2 = ad.read_h5ad("example.h5ad")
print(f"Reloaded: {adata2.n_obs} obs × {adata2.n_vars} vars")

# ----------------------------
# 3) Subset (keeps alignment)
# ----------------------------
t_cells = adata2[adata2.obs["cell_type"] == "T cell", :]
high_quality = adata2[adata2.obs["quality_score"] > 0.8, :]

print(f"T cells: {t_cells.n_obs}")
print(f"High quality: {high_quality.n_obs}")

# ----------------------------
# 4) Concatenate batches
# ----------------------------
adata_a = adata2[adata2.obs["sample"] == "A", :].copy()
adata_b = adata2[adata2.obs["sample"] == "B", :].copy()

combined = ad.concat(
    [adata_a, adata_b],
    axis=0,                 # concatenate observations (cells)
    join="inner",           # keep shared variables
    label="batch",          # add a column in .obs
    keys=["A", "B"],        # batch labels
)

print(combined.obs["batch"].value_counts().to_dict())

# ----------------------------
# 5) Backed mode for large files
# ----------------------------
adata_backed = ad.read_h5ad("example.h5ad", backed="r")
# Slicing in backed mode is metadata-friendly; load to memory when needed:
subset_mem = adata_backed[:10, :50].to_memory()
print(f"Backed subset loaded: {subset_mem.shape}")

Implementation Details

Data model (core slots)

• X: primary data matrix (dense numpy.ndarray or sparse scipy.sparse), shape (n_obs, n_vars).
• obs: per-observation metadata (pandas.DataFrame), indexed by obs_names (e.g., cell IDs).
• var: per-variable metadata (pandas.DataFrame), indexed by var_names (e.g., gene IDs).
• layers: named alternative matrices aligned to X (e.g., "counts", "log1p").
• obsm / varm: multi-dimensional embeddings aligned to obs/var (e.g., PCA, UMAP coordinates).
• obsp / varp: pairwise graphs/matrices (e.g., kNN graph in obsp["connectivities"]).
• uns: unstructured metadata (dict-like), often used for parameters and plotting configs.
• raw (optional): snapshot of unfiltered/untransformed data for reproducibility.

Views vs copies

• Slicing like adata_subset = adata[mask, :] typically returns a view (lightweight reference).
• Use .copy() when you need an independent object (e.g., before in-place modifications).

Backed mode (large datasets)

• ad.read_h5ad(path, backed="r") keeps the matrix on disk and loads data lazily.
• Convert a slice to memory with .to_memory() when you need in-memory computation.
• Backed mode is best for filtering by metadata, chunked processing, and avoiding OOM.

Concatenation behavior

• ad.concat([...], axis=0) stacks observations; axis=1 stacks variables.
• join="inner" keeps intersection of variables; join="outer" unions variables (may introduce missing values).
• label + keys records dataset/batch provenance in .obs[label].
• Merge strategies control how conflicting .uns and annotation columns are handled (choose based on your data governance needs).

Practical performance parameters

• Prefer sparse matrices (csr_matrix) for count-like data.
• Convert repeated strings to categoricals (adata.strings_to_categoricals()).
• Use compression when writing .h5ad (e.g., compression="gzip") to reduce storage; consider zarr for chunked/cloud-friendly access.