pyfastcar

A pure-Python re-implementation of FastCAR (Gocht / Berg et al., BMC Genomics 2023) for fast, deterministic per-gene correction of ambient RNA in droplet-based single-cell RNA-seq data.

• AnnData-native — drop-in for the scanpy ecosystem
• No rpy2, no R install — the empty-droplet detection, per-gene gMax/frC profiling, and count subtraction are all implemented directly in NumPy/SciPy
• Same function surface as the R workflow (determine.background.to.remove → remove.background, plus describe.ambient.RNA.sequence / recommend.empty.cutoff)
• Bit-for-bit reproducibility against the R reference — FastCAR is fully deterministic (no RNG anywhere), so pyfastcar matches R FastCAR exactly (see tests/test_exact_match.py)

This is a standalone mirror of the canonical implementation that lives in omicverse. All algorithmic work is developed upstream in omicverse and synced here for users who want FastCAR without the full omicverse stack.

Install

pip install pyfastcar

Dependencies: numpy, scipy, pandas, anndata.

What it does

Droplet-based scRNA-seq libraries are contaminated by ambient RNA — transcripts released by lysed cells that end up in every droplet. FastCAR estimates this contamination from the empty droplets of the raw, unfiltered count matrix and subtracts it, gene by gene, from the real cells.

The algorithm is fully deterministic — a handful of vectorised operations on the raw count matrix:

1. Empty droplets are the libraries whose total UMI count is below an empty_droplet_cutoff (thE, default 100). All-zero "unused barcodes" are excluded from the empty-droplet population.
2. For each gene, compute gMax — the highest count of that gene in any single empty droplet — and frC — the fraction of (non-zero) empty droplets that contain the gene.
3. A gene is corrected only when its frC clears the allowable contamination fraction contamination_chance_cutoff (frAA, default 0.005). For those genes, gMax is subtracted from every cell's count, with negative results floored at zero.

Quick-start (AnnData API)

import anndata as ad
from pyfastcar import correct_anndata

# raw = unfiltered AnnData (cells × genes): real cells + empty droplets
raw   = ad.read_h5ad("raw_feature_bc_matrix.h5ad")
cells = ad.read_h5ad("filtered_feature_bc_matrix.h5ad")   # the real cells

corrected = correct_anndata(
    raw, cells,
    empty_droplet_cutoff=100, contamination_chance_cutoff=0.005,
)

corrected.uns["fastcar"]["ambient_rna_profile"]   # per-gene subtract amounts
corrected.uns["fastcar"]["diagnostics"]           # gMax / frC table
corrected.uns["fastcar"]["n_genes_corrected"]

The wrapper transposes internally, so you always work in the natural cells × genes AnnData orientation.

Low-level functional API (mirrors R one-to-one)

from pyfastcar import (
    determine_background_to_remove, remove_background,
    describe_ambient_rna_sequence, recommend_empty_cutoff,
)

# full = unfiltered genes × droplets matrix (real cells + empty droplets)
# cells = filtered genes × cells matrix of real cells
profile = determine_background_to_remove(
    full, cells, empty_droplet_cutoff=100, contamination_chance_cutoff=0.005)

corrected = remove_background(cells, profile)        # ambient-corrected matrix

# pass return_table=True for a per-gene diagnostic DataFrame:
profile, table = determine_background_to_remove(full, cells, return_table=True)

Choosing the empty-droplet cutoff

desc = describe_ambient_rna_sequence(
    full, start=50, stop=500, by=25, contamination_chance_cutoff=0.005)
cutoff = recommend_empty_cutoff(desc)

What's included

Reproducing R results exactly

FastCAR has no stochastic component — given the same raw matrix and the same empty_droplet_cutoff / contamination_chance_cutoff, the output is fully determined. pyfastcar therefore reproduces R FastCAR bit-for-bit: the per-gene subtraction amounts, the corrected integer matrix and the threshold-profiling table are all identical.

tests/test_exact_match.py runs the R reference (FastCAR::determine.background.to.remove

• remove.background + describe.ambient.RNA.sequence) inside the CMAP environment on a deterministic synthetic raw matrix, saves every intermediate, and checks that the Python port reproduces them element-for-element.

pip install -e ".[dev]"
pytest tests/ -q              # smoke tests + R-parity tests
pytest tests/test_smoke.py -q # smoke tests only (no R required)

examples/compare_R_vs_Python.ipynb runs both implementations on a real raw 10x dataset (data/pbmc1k_raw.h5ad, a subset of 10x Genomics' pbmc_1k_v3 raw feature-barcode matrix that still contains empty droplets) and visualizes the bit-exact agreement with omicverse.

Relationship to omicverse

Developed upstream in omicverse:

• Canonical implementation: lives in the omicverse single-cell preprocessing stack
• Standalone mirror (this repo): same code, same API, minus the omicverse packaging

Citation

If you use this package, please cite the original FastCAR paper:

Gocht A.M., Berg M., et al. FastCAR: fast correction for ambient RNA to facilitate differential gene expression analysis in single-cell RNA-sequencing datasets. BMC Genomics 24, 2023.

and acknowledge omicverse / this repo for the Python port.

License

Apache-2.0. The upstream R package FastCAR is GPL-3; this is an independent re-implementation of its published algorithm.