anticor-features

anticor_features

Anti-correlation based feature selection for single cell (and other) omics datasets.

Features

• Unsupervised feature selection based on gene-gene anti-correlations.
• Automatically filters out genes in mitochondrial, ribosomal, and other pathways (customizable).
• Scales to large datasets using HDF5-backed intermediate files.
• Integrated Python API and command-line interface.
• Passes null-dataset tests for robust selection.

Installation

Requires Python 3.6 or higher.

Install from PyPI:

pip install anticor_features

Or install from source:

git clone https://bitbucket.org/scottyler892/anticor_features.git
cd anticor_features
pip install .

Dependencies

• h5py
• numpy
• pandas
• scipy
• seaborn
• matplotlib
• numba
• ray
• gprofiler-official (>=0.3.5)
• psutil

Quickstart Python API

from anticor_features.anticor_features import get_anti_cor_genes

# exprs: array-like or HDF5 dataset with genes in rows and cells in columns
# feature_ids: list of gene IDs matching rows of exprs
# species: g:Profiler species code (e.g., "hsapiens" or "mmusculus")
anti_cor_table = get_anti_cor_genes(exprs, feature_ids, species="hsapiens")

# Filter selected genes
selected = anti_cor_table.loc[anti_cor_table["selected"], "gene"].tolist()
print(selected)

See the g:Profiler organism list for valid species codes: https://biit.cs.ut.ee/gprofiler/page/organism-list

Customization

• pre_remove_features: list of gene IDs to exclude before analysis.
• pre_remove_pathways: list of GO term codes whose genes will be removed.
• min_express_n: minimum number of cells a gene must be expressed in to be considered (set to -1 to disable filtering, e.g., for non-expression or non-single-cell data).
• scratch_dir: directory for temporary HDF5 files (default: system temp directory).
• bin_size: number of features per batch when computing correlation matrix.
• FPR and FDR: false positive rate and false discovery rate for negative correlations.
• num_pos_cor: minimum number of positive correlations to select a feature.

Using with Non-Expression or Other Omics Data

For datasets that are not single-cell or gene-expression matrices (e.g., bulk omics, proteomics, metabolomics, or other feature embeddings), you can skip the minimum-expression filter and run only the anti-correlation statistics by setting min_express_n=-1. For example:

anti_cor_df = get_anti_cor_genes(
    embed_df,
    feature_ids=embed_df.index.tolist(),
    pre_remove_features=[],
    pre_remove_pathways=[],
    min_express_n=-1
)

Setting min_express_n=-1 disables the minimum-expression requirement (only meaningful for count-based single-cell data), allowing all features to be included in the statistical analysis.

Scanpy Integration

When using Scanpy (AnnData), transpose the data matrix:

from anticor_features.anticor_features import get_anti_cor_genes

anti_cor_table = get_anti_cor_genes(
    adata.X.T,
    adata.var.index.tolist(),
    species="hsapiens"
)

import pandas as pd
adata.var = pd.concat([adata.var, anti_cor_table], axis=1)
selected = anti_cor_table.loc[anti_cor_table["selected"], "gene"].tolist()
adata.raw = adata
adata = adata[:, selected]

Command-Line Interface

python3 -m anticor_features.anticor_features \
  -i exprs.tsv \
  -species mmusculus \
  -out_file anti_cor_features.tsv \
  -scratch_dir /path/to/tmp \
  -use_default_pathway_removal

Options:

• -i, --infile: input expression matrix (TSV or HDF5).
• -species: g:Profiler species code (default: "hsapiens").
• -out_file: output file path for the results table.
• -hdf5: treat input as HDF5 with dataset key "infile".
• -ids: file with feature (gene) IDs (no header) for HDF5 input.
• -cols: file with sample (cell) IDs (with header) for HDF5 input.
• -scratch_dir: directory for temporary files.
• -use_default_pathway_removal: remove default mitochondrial, ribosomal, and related pathways.
• -h, --help: display full help message.

Performance

Computing time scales with number of features and batch size. Selecting anti-correlated features on ~10k genes and ~3k cells typically takes 1–2 minutes (network time for g:Profiler). Larger datasets may take longer.

License

This project is licensed under the GNU Affero General Public License v3.0 (AGPL-3.0).

Contact

Scott Tyler [e-mail]