omicverse.metabol.serrf

Contents

omicverse.metabol.serrf#

omicverse.metabol.serrf(adata, *, qc_col, qc_label='QC', batch_col=None, top_k=10, n_estimators=100, min_qc_samples=5, layer=None, seed=0)[source]#

SERRF — QC-based Random Forest drift correction (Fan 2019).

For each feature and batch:

  1. Rank other features by absolute correlation with the target feature across QC samples only.

  2. Fit RandomForestRegressor with the target as response and the top-top_k correlated features as predictors, trained on QC injections only.

  3. Predict the target abundance for every sample in the batch from its co-feature vector.

  4. Corrected value = raw / predicted * mean(QC_raw).

The ratio scales each sample onto the QC baseline; features whose QC mean is zero or negative, or whose predictions collapse, are left uncorrected.

Parameters:

  • qc_col (str) – Column in adata.obs that tags QC vs real samples.

  • qc_label (str (default: 'QC')) – Value in qc_col marking QC rows. Default "QC".

  • batch_col (Optional[str] (default: None)) – Optional column in adata.obs for per-batch correction. Without a batch column the full run is treated as one batch.

  • top_k (int (default: 10)) – Number of co-features to use as RF predictors. Fan 2019 uses 10; larger values add runtime and marginal accuracy.

  • n_estimators (int (default: 100)) – RF tree count. Default 100 (the Fan R package uses 500 but 100 is within 1% AUC on benchmarks at 5× the speed).

  • min_qc_samples (int (default: 5)) – If a batch has fewer QC samples than this, correction is skipped for that batch (features left as raw). Default 5.

  • layer (Optional[str] (default: None)) – Which adata.layers entry to correct. None → use adata.X.

  • seed (int (default: 0)) – RandomForest seed.

Returns:

New AnnData with corrected .X. The original matrix is preserved in out.layers['raw']. Per-feature CV% before / after is stored in out.var['cv_qc_raw'] / cv_qc_serrf''] to help users audit improvement.

Return type:

AnnData

Contents