anndata-io

AnnData I/O: 10x readers and backed metadata

For every task, apply Caveman. Before broad repository navigation, use CocoIndex Code; for web-page research, use Crawl4AI. Read and follow the repository-root docs/agent-tools.md and its linked skills, and retain the required execution evidence.

This maintained, narrowly scoped overlay corrects selected I/O examples in the pinned AnnData skill. It does not replace or validate the rest of that skill. The original entrypoint and I/O reference are retained in provenance/.

Read 10x data

Install Scanpy in the user's chosen environment when these readers are needed. For a 10x HDF5 file:

import scanpy as sc

counts_h5 = sc.read_10x_h5(h5_path)

For a 10x Matrix Market directory:

import scanpy as sc

counts_mtx = sc.read_10x_mtx(mtx_dir)

Inspect observation and feature names, feature types and matrix dimensions before downstream analysis. The readers' defaults filter to gene expression; pass the documented arguments if other feature types are needed. These are Scanpy readers, not AnnData readers.

Save changes to observation metadata

AnnData backed mode does not automatically persist changes to obs. Select a subset that fits in memory, materialize it, close the backed file, and write to a new output path. The selection and label below are supplied by the caller:

import anndata as ad

backed = ad.io.read_h5ad(input_path, backed="r")
try:
    subset = backed[obs_selection, :].to_memory()
finally:
    backed.file.close()
subset.obs["sample"] = sample_label
subset.write_h5ad(output_path)

Choose a distinct output file and verify the result after reopening it. The selection determines which observations are written; this example does not write the complete original dataset unless the selection includes all rows. Materializing the selection requires enough memory for its data and annotations. For larger-than-memory changes, choose and validate an appropriate workflow for the specific data rather than assuming this example will fit.

See the AnnData backed reader documentation, Scanpy HDF5 reader and Scanpy Matrix Market reader. The recorded component control uses synthetic data; it is not an agent benchmark.