def apply_qc_filters(unidata: UnimodalData):
""" Apply QC filters to filter out low quality cells """
if "passed_qc" in unidata.obs:
prior_n = unidata.shape[0]
unidata._inplace_subset_obs(unidata.obs["passed_qc"])
cols = ["passed_qc"]
if unidata.uns.get("__del_demux_type", False):
cols.append("demux_type")
if "assignment" in unidata.obs:
# remove categories that contain no elements
series = unidata.obs["assignment"].value_counts(sort=False)
unidata.obs["assignment"] = pd.Categorical(
unidata.obs["assignment"],
categories=series[series > 0].index.astype(str))
# del unidata.uns["__del_demux_type"]
unidata.obs.drop(columns=cols, inplace=True)
if len(unidata.obsm) > 0:
unidata.obsm.clear()
if len(unidata.varm) > 0:
unidata.varm.clear()
for key in list(unidata.uns):
if key not in {'genome', 'modality', 'norm_count', 'df_qcplot'}:
del unidata.uns[key]
logger.info(
f"After filtration, {unidata.shape[0]} out of {prior_n} cell barcodes are kept in UnimodalData object {unidata.get_uid()}."
)
def estimate_background_probs(hashing_data: UnimodalData,
random_state: int = 0) -> None:
"""For cell-hashing data, estimate antibody background probability using KMeans algorithm.
Parameters
----------
hashing_data: ``UnimodalData``
Annotated data matrix for antibody.
random_state: ``int``, optional, default: ``0``
Random seed set for reproducing results.
Returns
-------
``None``
Update ``hashing_data``: Filtered cell barcodes with 0 counts
Update ``hashing_data.uns``:
* ``hashing_data.uns["background_probs"]``: estimated antibody background probability.
Example
-------
>>> estimate_background_probs(hashing_data)
"""
hashing_data.obs["counts"] = hashing_data.X.sum(axis=1).A1
# Remove barcodes with 0 total counts
idx = hashing_data.obs["counts"] == 0
ncell_zero = idx.sum()
if ncell_zero > 0:
logger.warning(
f"Detected {ncell_zero} cell barcodes with 0 hashtag counts, which are removed from the hashing data object."
)
hashing_data._inplace_subset_obs(~idx)
counts_log10 = np.log10(hashing_data.obs["counts"].values.reshape(-1, 1))
kmeans = KMeans(n_clusters=2, random_state=random_state).fit(counts_log10)
signal = 0 if kmeans.cluster_centers_[0] > kmeans.cluster_centers_[1] else 1
hashing_data.obs["hto_type"] = "background"
hashing_data.obs.loc[kmeans.labels_ == signal, "hto_type"] = "signal"
idx = np.isin(hashing_data.obs["hto_type"], "background")
pvec = hashing_data.X[idx, ].sum(axis=0).A1
back_probs = pvec / pvec.sum()
idx = back_probs <= 0.0
if idx.sum() > 0:
logger.warning(
f"Detected {idx.sum()} antibody barcodes {','.join(hashing_data.var_names[idx])} with 0 counts in the background! These barcodes are likely not in the experiment and thus removed."
)
hashing_data._inplace_subset_var(~idx)
back_probs = back_probs[~idx]
hashing_data.uns["background_probs"] = back_probs
logger.info("Background probability distribution is estimated.")
def apply_qc_filters(unidata: UnimodalData):
""" Apply QC filters to filter out low quality cells """
if "passed_qc" in unidata.obs:
prior_n = unidata.shape[0]
unidata._inplace_subset_obs(unidata.obs["passed_qc"])
cols = ["passed_qc"]
if unidata.uns.get("__del_demux_type", False):
cols.append("demux_type")
del unidata.uns["__del_demux_type"]
unidata.obs.drop(columns=cols, inplace=True)
logger.info(
f"After filtration, {unidata.shape[0]} out of {prior_n} cell barcodes are kept in UnimodalData object {unidata.get_uid()}."
)
