scValue

scValue is a Python package for value-based subsampling (“sketching”) of large scRNA-seq datasets for downstream machine / deep learning tasks.

scValue assigns each cell a data value (DV) using out-of-bag (OOB) estimates from a random forest classifier. Intuitively, a cell gets a higher DV if it is more helpful for correctly distinguishing its cell type in OOB predictions. scValue then builds a subsample by:

1. Computing DV for each cell (OOB-based).
2. Allocating sketch size across cell types (DV-weighted or proportional).
3. Selecting cells within each cell type using a DV-guided strategy (binning or top-pick).

Implemented using: Python 3.10, scikit-learn 1.5.2, SciPy 1.14.0, NumPy 1.26.4, pandas 1.5.3, joblib 1.4.2.
Experiments in the paper used: Scanpy 1.10.2 and AnnData 0.10.8.

---

Installation

Install from PyPI:

pip install scvalue

Quick start

Load modules

import scanpy as sc
from scvalue import SCValue

Read and preprocess an input scRNA-seq dataset

The example mTC dataset (mouse T cells; 89,429 cells) is available in .h5ad format. Download it (example with wget):

wget -O mTC.h5ad "https://datasets.cellxgene.cziscience.com/6c253303-8255-4004-8464-5e90dd5846cb.h5ad"

It is already log-normalised. Here we compute PCA on the top 3,000 HVGs. By default, scValue expects features from adata.obsm["X_pca"] (use_rep="X_pca").

adata = sc.read_h5ad('mTC.h5ad')

sc.pp.highly_variable_genes(adata, n_top_genes=3000)
sc.pp.pca(adata, n_comps=50) # stores PCs in adata.obsm["X_pca"]

Subsample by scValue

Subsample 10% of cells and return an AnnData sketch:

scv = SCValue(
    adata=adata,
    sketch_size=0.1,          # float in (0, 1] or an integer cell count
    use_rep="X_pca",          # default: "X_pca"; use "X" to use adata.X
    cell_type_key="cell_type",
    n_trees=100,
    type_weighting="inv_cv",  # "inv_cv" (default), "sqrt_n_sd", or "proportional"
    strategy="FB",            # "FB" (default), "MTB", or "TP"
    write_dv=False,           # if True, writes dv_values.csv and dv_summary.csv
    seed=42,
)

adata_sub = scv.value()       # also available as scv.adata_sub

After running:

• DV is stored in adata.obs["dv"] (and included in adata_sub.obs["dv"]).
• scv.dv stores a DataFrame of cell type + DV for the full dataset.
• scv.adata_sub is the subsampled AnnData.

The sketch (adata_sub) can be visualised with the usual Scanpy workflow (neighbours/UMAP/etc.). The function plot_umap in reproducibility/eval_time_hd_gini.py can be used to reproduce the mTC demonstration in our paper.

All datasets used in our study are publicly available, as detailed in the data availability section.

Parameters

Core inputs

• sketch_size (required):
  • float in (0, 1] = fraction of cells to keep, or
  • int = exact number of cells to keep.
• use_rep (default "X_pca"): feature matrix used to fit the Random Forest and compute DV.
  • "X_pca" expects adata.obsm["X_pca"] (e.g., from sc.pp.pca).
  • "X" uses adata.X (note: sparse matrices may be densified internally).
• cell_type_key (default "cell_type"): column name in adata.obs containing cell type labels.

Step 1: data value computation

• n_trees (default 100): number of trees in random forest
• seed (default 42): random seed for reproducibility.

Step 2: sketch size determination

• type_weighting for each cell type
  • "inv_cv" (default): weights ∝ mean(DV) * sqrt(N) / std(DV)
  • "sqrt_n_sd": weights ∝ sqrt(N) * std(DV)
  • "proportional": proportional to the original cell-type frequencies

Step 3: value-guided cell selection

• strategy
  • "FB" (full binning, default): DV is binned into 0.1 intervals from 0.0 to 1.0; selects top-DV cells within each bin.
  • "MTB" (mean-threshold binning): same bins as FB, but only bins above the type mean DV are considered.
  • "TP" (top-pick): no binning; simply selects the highest-DV cells per type.

Saving DV

• write_dv (default False): if True, writes:
  • dv_values.csv (per-cell DV + type)
  • dv_summary.csv (per-type DV summary statistics)

Reference

Li Huang, Weikang Gong, Dongsheng Chen, scValue: value-based subsampling of large-scale single-cell transcriptomic data for machine and deep learning tasks, Briefings in Bioinformatics, Volume 26, Issue 3, May 2025, bbaf279, https://doi.org/10.1093/bib/bbaf279