_IO

Core I/O helpers for saving AnnData objects and converting them into analysis-ready DataFrame tables.

This module provides:

• save_dataset

• make_df_obs_adataX

These helpers are used by higher-level workflows such as model fitting and expectation correction. In particular, _model_fit.md depends on make_df_obs_adataX(...), and _expectation_based_covar_correction.md uses save_dataset(...) in its wrapper workflow.

save_dataset

save_dataset(...) writes one AnnData object to a bundle of files with a shared basename.

Full signature

def save_dataset(
    _adata: ad.AnnData,
    output_path: str | Path,
    logger: logging.Logger | None = None,
    save_obsm: bool = True,
    obsm_keys: Sequence[str] | None = None,
) -> None:

import adata_science_tools as adtl

adtl.save_dataset(
    adata,
    "results/corrected_dataset.h5ad",
)

To skip .obsm CSV exports or save only selected .obsm keys:

adtl.save_dataset(
    adata,
    "results/corrected_dataset.h5ad",
    save_obsm=False,
)

adtl.save_dataset(
    adata,
    "results/corrected_dataset.h5ad",
    obsm_keys=["pre_values", "post_values"],
)

Output files

If the output path is results/corrected_dataset.h5ad, the helper writes:

• results/corrected_dataset.h5ad

• results/corrected_dataset.obs.csv

• results/corrected_dataset.var.csv

• results/corrected_dataset.X.csv

• one CSV per layer, named like results/corrected_dataset.layer.<layer_name>.csv

• one CSV per saved .obsm table, named like results/corrected_dataset.obsm.<obsm_key>.csv

If the path does not end in .h5ad, the helper treats it as the basename and still writes the same set of files with .h5ad and CSV suffixes.

Important behavior

• Parent directories are created automatically.

• adata.X is converted to dense before writing the .X.csv file if needed.

• Each layer is also written to CSV.

• Layer names containing / are normalized to _ in the layer CSV filenames.

• By default, each 2D .obsm value aligned to observations is written to CSV.

• .obsm names containing / are normalized to _ in CSV filenames.

• Requested missing .obsm keys raise KeyError.

• Sanitized .obsm filename collisions raise ValueError before files are written.

• Unsupported .obsm values, such as arrays that are not 2D, are logged and skipped.

• The helper always writes the .h5ad plus CSV bundle together. There is no flag to save only one of those outputs.

.obsm table exports

For .obsm entries that are already pandas.DataFrame objects, save_dataset(...) preserves the DataFrame index and columns.

For array-like or sparse 2D .obsm entries, the helper uses adata.obs_names as the row index. If the number of .obsm columns equals adata.n_vars, it uses adata.var_names as column labels; otherwise it writes deterministic columns named dim_0, dim_1, and so on.

This is useful for source-value tables created by ref_vs_target_adata(save_source_values_obsm=True), which stores paired reference and target source matrices in .obsm["pre_values"] and .obsm["post_values"] by default:

post_minus_pre = adtl.ref_vs_target_adata(
    adata,
    pair_by_key="Subject_ID",
    save_source_values_obsm=True,
)

adtl.save_dataset(post_minus_pre, "results/post_minus_pre.h5ad")

This writes:

• results/post_minus_pre.obsm.pre_values.csv

• results/post_minus_pre.obsm.post_values.csv

make_df_obs_adataX

make_df_obs_adataX(...) builds a pandas.DataFrame from AnnData expression data and, optionally, prepends adata.obs.

Full signature

def make_df_obs_adataX(
    adata,
    layer: str | None = None,
    index: str | None = None,
    varcolumns: list[str] | str | None = None,
    include_obs: bool = True,
    use_raw: bool = False
):

df = adtl.make_df_obs_adataX(
    adata,
    layer="pgml",
    include_obs=True,
)

Typical uses:

• build a combined obs_X_df for OLS or MixedLM fitting,

• inspect an expression matrix together with observation metadata,

• choose alternate feature labels or a different observation index.

Data source selection

Current precedence is:

• adata.raw.X when use_raw=True and layer is None,

• adata.raw.layers[layer] when use_raw=True and layer is provided,

• adata.layers[layer] when layer is present on the main object,

• otherwise adata.X.

Implementation note:

• The function prints which source it used rather than logging through a Logger.

Practical caution:

• In standard AnnData, adata.raw typically exposes X and var, but not arbitrary layers. The current implementation still attempts adata.raw.layers[layer] when both use_raw=True and layer are supplied, so the safest raw-data usage is use_raw=True with layer=None.

Column labeling

By default, expression columns are labeled with adata.var_names.

You can override that with varcolumns:

• None: use adata.var_names

• str: use one column from adata.var

• list of length 1: same as a single-string column selection

• list of length 2 or more: build a pandas.MultiIndex from multiple adata.var columns

Example:

df = adtl.make_df_obs_adataX(
    adata,
    varcolumns=["feature_class", "gene_name"],
    include_obs=False,
)

Row indexing

By default, the DataFrame index is adata.obs_names.

If index is provided, it is interpreted as a column name in adata.obs and that column becomes the DataFrame index.

Example:

df = adtl.make_df_obs_adataX(
    adata,
    index="SubjectID",
    include_obs=True,
)

include_obs

When include_obs=True, the function concatenates adata.obs in front of the expression matrix columns.

When include_obs=False, the result contains only the expression values.

This distinction matters in _model_fit.md, where the fit wrappers build a combined obs_X_df for downstream formula-based model fitting.

Dense conversion and scale

If the selected matrix is sparse, the helper converts it to dense with toarray() before building the DataFrame.

That is convenient for downstream formula-based modeling, but it can increase memory use substantially on large matrices.

Current limitations and caveats

The docs for this module are based on the current implementation rather than direct _io regression tests in this repo.

Important current caveats:

• The default varcolumns path is based on adata.var_names even when use_raw=True.

• If adata.raw.var_names differ from adata.var_names, column labels may not match the raw matrix automatically.

• The helper imports AnnData locally but does not enforce the input type beyond expected attribute access.