SParseSampler (SPS) is a Python package for efficient subsampling of large-scale single-cell RNA-seq and flow cytometry datasets while preserving rare cell populations. The method employs an unsupervised approach that maintains dataset structure and rare cell types without requiring explicit labels.

• PCA-based dimensionality reduction with automatic parameter selection via an EVR-based heuristic
• Variance-weighted binning in the reduced dimensional space
• Iterative cell selection prioritizing sparsest bins
• Preserves rare populations without requiring cell type labels

• Computational efficiency comparable to random sampling
• Superior rare cell retention compared to existing methods (Hopper, Atomic Sketch)
• Performance comparable to scSampler with improved speed
• Successfully tested on datasets up to 34 million cells
• Validated at multiple rarity levels (1%, 0.5%, 0.1%)

pip install sparsesampler

SParseSampler uses an EVR-based heuristic for automatic parameter selection. Both the number of principal components (p) and the Bin Resolution Factor (K) are derived from the explained variance ratio (EVR) of the principal components using the EVR index (feature_index).

• EVR Index (feature_index)

  • Default: 12
  • Controls both the number of principal components and the bin resolution
  • The number of principal components p is set to feature_index + 1
  • The Bin Resolution Factor K is computed as K = 2 / EVR_i, where EVR_i is the explained variance ratio of the i-th principal component
  • Recommended ranges:
    • Flow cytometry data: EVR indices 7–20
    • scRNA-seq data: EVR indices 12–25
  • The default value of 12 lies within both optimal ranges

• Sample Size (size)

  • Number of cells to subsample from the dataset

• Seed (seed)

  • Random seed for reproducibility

• Single-cell RNA sequencing data
• Flow cytometry data

• Comprehensive comparison against state-of-the-art methods
• Validated on large-scale datasets:
  • MCC dataset (scRNA-seq): 3.2M cells, 3,065 genes
  • LCMV dataset (flow cytometry): 34M cells, 31 features
• Consistent performance across varying dataset sizes and rarity levels (1%, 0.5%, 0.1%)
• Downstream validation: Random Forest classifiers trained on SPS-subsampled data achieve substantially higher F1 scores for rare cell types compared to random subsampling

• Flow cytometry (LCMV, 34M cells, 31 features): ~11 seconds
• scRNA-seq (MCC, 3M cells, 3,065 genes): ~5 minutes

import sparsesampler.sampling as sps
import numpy as np

# Load your data (n_samples × n_features)
# Example 1: From NumPy array
X = np.load('your_data.npy')

# Example 2: From CSV file
import pandas as pd
X = pd.read_csv('your_data.csv').values

# Example 3: scRNA-seq data (AnnData format)
import scanpy as sc
adata = sc.read_h5ad('your_data.h5ad')
X = adata.X  # Use .toarray() if sparse matrix

# Run SParseSampler with default parameters (EVR index = 12)
indices, elapsed_time = sps.sample(X=X, size=100000)

# Run with custom EVR index (e.g., for flow cytometry data)
indices, elapsed_time = sps.sample(X=X, size=50000, feature_index=8)

# Get subsampled data
X_sampled = X[indices]

We recommend applying standard quality control filtering prior to SPS, including:

• Removal of cells with abnormally high/low UMI counts
• Filtering cells with high mitochondrial gene percentages
• Doublet detection and removal (e.g., using Scrublet or DoubletFinder)

The following animation shows how points are selected from a 2D toy dataset using PCA binning. Points are selected category by category (cells with 1 point, 2 points, etc.), and the process is visualized step by step:

• All points start as skyblue.
• When a category is considered, the cells are highlighted in yellow and the points in those cells are shown in gray for visibility.
• Selected points turn red and remain red in all subsequent frames.
• The process continues until the target number of points is reached.

To generate the animation yourself, run:

python docs/generate_visualization.py

If you use SParseSampler in your research, please cite:

# Add citation when available

This project is licensed under the MIT License - see the LICENSE file for details.

Contributions are welcome! Please feel free to submit a Pull Request.