API Reference

API Reference

Complete reference for every public function and class in HarmonizePy.

fromharmonizepyimportharmonize, combat, remove_batch_effect, HarmonizeConfigfromharmonizepyimportadjust_combat, adjust_limma

---

harmonize()

The main pipeline entry point. Handles structural missingness, sorts and blocks batches, groups features by affiliation, corrects each group, and reassembles the output.

harmonize(
data,
description,
*,
config=None,
algorithm="ComBat",
combat_mode=1,
needed_values=None,
sort=None,
block=None,
unique_removal=True,
output_file=None,
) ->pd.DataFrame

Parameters

Parameter	Type	Default	Description
data	DataFrame, str, or Path	required	Features x samples matrix. If a path, read from that TSV/CSV/Parquet file. Feature names must be in df.index, sample names in df.columns.
description	DataFrame, str, or Path	required	Batch description. Must have columns ID, sample, batch. If a path, read from that CSV file.
config	HarmonizeConfig or None	None	If provided, all algorithm settings come from this object. Individual keyword arguments below are ignored.
algorithm	"ComBat" or "limma"	"ComBat"	Batch correction algorithm.
combat_mode	int (1-4)	1	ComBat variant. Ignored when algorithm="limma".
needed_values	int or None	None	Minimum non-missing observations per batch for a feature to qualify. None auto-selects: 2 for modes 1, 3 and limma; 1 for modes 2, 4.
sort	"sparsity", "jaccard", "seriation", or None	None	Batch sorting strategy applied before blocking.
block	int >= 2 or None	None	Group this many consecutive sorted batches into one block.
unique_removal	bool	True	Rescue singleton-affiliation features by cropping to the nearest shared pattern.
output_file	str, Path, or None	None	Write the corrected matrix to this path. Format inferred from extension: .tsv, .csv, .parquet, .pq.

Returns

pd.DataFrame with the same shape, index, and column order as the input. NaN positions are preserved where data was insufficient for correction.

Raises

ValueError on invalid arguments, mismatched sample IDs, or duplicate feature names.

Examples

fromharmonizepyimportharmonize# From file pathsresult=harmonize("data.tsv", "batch.csv")
# From DataFramesresult=harmonize(data_df, desc_df, algorithm="limma")
# Sort and block for many-batch datasetsresult=harmonize(data_df, desc_df, sort="sparsity", block=2)
# Non-parametric ComBat for small batchesresult=harmonize(data_df, desc_df, combat_mode=3, needed_values=1)
# Write output directlyresult=harmonize("data.tsv", "batch.csv", output_file="corrected.parquet")

---

HarmonizeConfig

Frozen dataclass for storing a reproducible run configuration. Pass as config= to harmonize() to override all individual keyword arguments.

fromharmonizepyimportHarmonizeConfigcfg=HarmonizeConfig(
algorithm="ComBat",
combat_mode=1,
sort_strategy="sparsity",
block_size=2,
unique_removal=True,
)

Fields

Field	Type	Default	Description
algorithm	str	"ComBat"	"ComBat" or "limma"
combat_mode	int	1	ComBat mode 1-4
needed_values	int or None	None	Minimum observations per batch (auto-selected if None)
sort_strategy	str or None	None	"sparsity", "jaccard", "seriation", or None
block_size	int or None	None	Block size >= 2, or None
unique_removal	bool	True	Rescue singleton features

HarmonizeConfig is frozen and slots-based. Once created, fields cannot be changed. Create a new instance to use different settings.

---

combat()

Low-level ComBat engine. Operates directly on NumPy arrays. Does not handle structural missingness: for that, use harmonize().

combat(
data,
batch,
*,
par_prior=True,
mean_only=False,
ref_batch=None,
) ->ndarray

Parameters

Parameter	Type	Default	Description
data	ndarray, shape (n_features, n_samples)	required	Expression matrix. Per-cell NaN is allowed.
batch	array-like, shape (n_samples,)	required	Integer batch label per sample.
par_prior	bool	True	Use parametric prior (True = modes 1/2, False = modes 3/4).
mean_only	bool	False	Correct location only, skip scale adjustment (True = modes 2/4).
ref_batch	int or None	None	Reference batch to leave unadjusted. Other batches are adjusted toward this reference.

Returns

ndarray with the same shape as data. NaN positions preserved.

Example

importnumpyasnpfromharmonizepyimportcombatdata=np.random.randn(1000, 30) # 1000 features, 30 samplesbatch=np.array([1]*10+ [2]*10+ [3]*10)
corrected=combat(data, batch)
corrected_np=combat(data, batch, par_prior=False) # non-parametriccorrected_loc=combat(data, batch, mean_only=True) # location only

---

remove_batch_effect()

Low-level limma engine. Operates directly on NumPy arrays. Does not handle structural missingness: for that, use harmonize().

remove_batch_effect(data, batch) ->ndarray

data is an ndarray of shape (n_features, n_samples). Per-cell NaN is allowed. batch is a 1-D array-like of integer batch labels, length n_samples.

Returns

ndarray with the same shape as data. NaN positions preserved.

Example

importnumpyasnpfromharmonizepyimportremove_batch_effectcorrected=remove_batch_effect(data, batch)

---

adjust_combat()

Wrapper that maps HarmonizR-style integer modes (1-4) to the combat() parameters. Operates on DataFrames rather than raw ndarrays.

adjust_combat(sub_df, batch_labels, mode=1, ref_batch=None) ->pd.DataFrame

sub_df is a features x samples DataFrame. Returns a corrected DataFrame with the same shape, index, and columns. Intended for custom pipeline use when you are managing sub-matrices yourself.

---

adjust_limma()

Wrapper applying limma correction to a DataFrame.

adjust_limma(sub_df, batch_labels) ->pd.DataFrame

sub_df is a features x samples DataFrame. Returns a corrected DataFrame with the same shape, index, and columns.

---

Semi-public submodule API

These functions are importable from their submodules for building custom pipelines. They are not re-exported from the top-level harmonizepy namespace.

fromharmonizepy.sortingimportsort_batchesfromharmonizepy.blockingimportbuild_block_listfromharmonizepy.affiliationimportremove_unique_combinations

Function	Module	Description
sort_batches(data, batch_labels, strategy)	harmonizepy.sorting	Returns a reordered copy of batch_labels according to the chosen strategy.
build_block_list(batch_labels, block_size)	harmonizepy.blocking	Returns a list of batch-group assignments for the given block size.
remove_unique_combinations(affiliation_list)	harmonizepy.affiliation	Rescues singleton-affiliation features by cropping to the nearest shared pattern.

These interfaces are stable for the current version but are not covered by the same backward-compatibility guarantee as the top-level public API.