recursieve 1.0.3

RecurSieve is an iterative gene selection algorithm that identifies discriminative genes between two biological groups in single-cell datasets. It uses an ensemble learning approach with a blend of random forest classifiers and gaussian mixture models to progressively select genes that best separate the groups, collapsing selected genes into a lower dimensional representational axis at each iteration.

Installation

Install from PyPI:

pip install recursieve

RecurSieve requires Python 3.12 or newer.

To install the development version from GitHub:

pip install git+https://github.com/ArpiarSaundersLab/RecurSieve.git

Usage

import anndata as ad
from recursieve import recursieve

adata = ad.read_h5ad("data/alz.h5ad")
model = recursieve(
    adata=adata,
    group1="control",
    group2="disease",
    field_name="condition",
    max_iterations=100,
)

# all genes found
panel = model.genes

# genes unique to RecurSieve
non_de_genes = model.unique_gene_panel 

Example

examples/recursieve_example.ipynb is a short notebook that runs RecurSieve on simulated data, with no download needed. It builds a two-group dataset of 2,000 cells and 2,000 genes. In 10 of the genes, the control group is unimodal and the disease group is bimodal around the same center, and those genes move together. Because the group means match, differential expression has little to find. The notebook runs RecurSieve with default parameters and checks which of the 10 genes appear in unique_gene_panel.

Parameters

Robustness checks

seed_stability checks whether the same genes are selected with different random seeds. coexpression_null checks whether the selected genes are found more often than by chance, by rerunning on randomized data.

Each seed and each shuffle is a full RecurSieve run, so these checks take much longer than a single run. On a 10-core laptop, 25 shuffles took about 1.5 hours for 25,000 cells and about 3.5 hours for 50,000 cells.

Both methods return a dictionary of pandas DataFrames and do not write files. The example below runs both checks and saves the main tables as CSV files.

seeds = model.seed_stability(n_seeds=10)
null = model.coexpression_null(n_shuffles=100)

seeds["gene_frequency"].to_csv("seed_gene_frequency.csv", index=False)
null["pvalues"].to_csv("null_pvalues.csv", index=False)
null["gene_frequency"].to_csv("null_gene_frequency.csv", index=False)

License

The RecurSieve software is released under the MIT License. The data, figures, and tables in manuscript/ are dedicated to the public domain under CC0 1.0.

Citation

RecurSieve: A Python Package Using Iterative Machine Learning to Detect Subtle Gene Expression Patterns in Single-Cell Data
bioRxiv link coming soon