omicverse.metabol.impute

Contents

omicverse.metabol.impute#

omicverse.metabol.impute(adata, *, method='qrilc', missing_threshold=0.5, n_neighbors=5, q=0.01, seed=0)[source]#

Impute missing values (NaN / 0) in adata.X.

Parameters:

  • method (Literal['knn', 'half_min', 'qrilc', 'zero'] (default: 'qrilc')) –

    • "knn" — per-feature kNN on the sample × sample Euclidean distance restricted to non-missing positions

    • "half_min" — replace missing with half the minimum non-missing value of that feature (classic MNAR-friendly default)

    • "qrilc" — Quantile Regression Imputation of Left-Censored: draws from TruncatedNormal(mean=mu_below_q, sd=sigma) where mu and sigma are estimated from values below the q-quantile

    • "zero" — replace missing with 0 (for downstream methods that treat 0 as a valid observation)

  • missing_threshold (float (default: 0.5)) – Before imputation, drop features whose missingness exceeds this fraction. Default 0.5 (drop a feature missing in >50% of samples).

  • n_neighbors (int (default: 5)) – kNN neighborhood size (only used when method='knn').

  • q (float (default: 0.01)) – Quantile defining the “below-detection-limit” band for QRILC.

  • seed (int (default: 0)) – RNG seed for the QRILC truncated-normal draws. Default 0 — pass a different integer to bootstrap or change the imputation realization across pipeline runs.

Returns:

New object with imputed .X and a var['missing_frac'] column recording the pre-imputation missingness of each feature.

Return type:

AnnData

Contents