pydecontx

A pure-Python re-implementation of DecontX (Yang et al., Genome Biology 2020) for decontamination of ambient / cross-contaminating RNA in droplet single-cell RNA-seq data.

• AnnData-native — drop-in for the scanpy ecosystem
• No rpy2, no R install — the variational-EM, the native/contamination multinomial mixture, and the Minka fixed-point Dirichlet MLE are all implemented directly in NumPy/SciPy
• Same function surface as the R workflow (decontX → per-cell contamination fraction + decontaminated count matrix)
• Numerical reproducibility against the Bioconductor reference — contamination fraction, decontaminated matrix and per-cell theta agree at Pearson r > 0.99 (see tests/test_r_parity.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 DecontX without the full omicverse stack.

Install

pip install pydecontx

Dependencies: numpy, scipy, pandas, anndata.

Quick-start (function API)

import pydecontx as dx

# 1) a synthetic contaminated dataset (genes x cells) with ground truth
sim = dx.simulate_contamination(C=300, G=100, K=3, delta=(1, 10))

# 2) run DecontX with cluster labels
res = dx.decontx(sim["observed_counts"], z=sim["z"])

res.contamination            # per-cell contamination fraction (0-1)
res.decontx_counts           # decontaminated sparse count matrix
res.decontaminated_counts()  # ... integer-rounded
res.to_dataframe()           # per-cell summary table

Quick-start (AnnData)

import scanpy as sc, pydecontx as dx

adata = sc.read_10x_h5("filtered.h5")          # cells × genes, raw counts in .X
sc.pp.pca(adata); sc.pp.neighbors(adata); sc.tl.leiden(adata)

# write results back into the AnnData
adata = dx.decontx(adata, z="leiden", copy=True)
adata.obs["decontX_contamination"]      # per-cell contamination
adata.layers["decontX_counts"]          # decontaminated counts

x may be an AnnData (cells × genes), a pandas.DataFrame, a NumPy array or a SciPy sparse matrix (genes × cells). z is required — DecontX needs a broad clustering of cell types; when x is an AnnData it may name a column of .obs. Empty droplets are not required, but an optional background matrix of raw empty droplets can be supplied to anchor the contamination distribution.

The model

DecontX models each cell's observed counts as a Bayesian two-component multinomial mixture:

• a native distribution phi_k — the gene probabilities of the cell's own population k;
• a contamination distribution eta_k — a weighted blend of every other population's native distribution;
• a per-cell latent theta_j ~ Beta giving the proportion of native counts, with a Bernoulli native/contaminant label per transcript.

Inference is variational EM: variational distributions over theta and the transcript labels maximise the ELBO; phi, eta and the Dirichlet hyper-parameter delta are re-estimated each iteration (the latter by a Minka fixed-point Dirichlet MLE — a port of MCMCprecision::fit_dirichlet). The output is a per-cell contamination fraction and a decontaminated (native) count matrix.

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

from pydecontx import (
    decontx, DecontXResult,
    decontx_initialize, decontx_em, decontx_loglik,
    calculate_native_matrix, fit_dirichlet,
)

# variational-EM building blocks
init = decontx_initialize(counts, theta, z)          # phi / eta initialisation
step = decontx_em(counts, colsums, theta, eta, phi, z)
ll   = decontx_loglik(counts, theta, eta, phi, z)

# decontaminated-matrix computation
native = calculate_native_matrix(counts, theta, eta, phi, z)

# Minka fixed-point Dirichlet MLE (= MCMCprecision::fit_dirichlet)
alpha = fit_dirichlet(proportions)["alpha"]

What's included

Reproducing R results

examples/compare_R_vs_Python.ipynb runs Bioconductor decontX (via Rscript) and pydecontx on the same real PBMC 3k dataset with identical cluster labels, and shows the per-cell contamination fraction, the decontaminated count matrix and the per-cell theta all agree at Pearson r > 0.99.

tests/test_r_parity.py runs the same synthetic two/three-population contaminated count matrix — with identical cluster labels — through Bioconductor decontX (R) and pydecontx, and asserts agreement on the contamination fraction, the decontaminated matrix, the final theta, and the background-anchored mode.

Unavoidable difference. The variational EM is deterministic given its initial theta, but theta is seeded by a Beta draw: R uses its Mersenne-Twister RNG (stats::rbeta under withr::with_seed), NumPy uses PCG64. The two initialisations differ, so the converged estimates agree to high correlation rather than bit-exactly. R also runs the EM inner loops in C++; pydecontx reproduces them in vectorised NumPy.

Relationship to omicverse

Developed upstream in omicverse:

• Canonical implementation: synced to omicverse
• Standalone mirror (this repo): same code, same API, minus the omicverse packaging

Citation

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

Yang, S. et al. Decontamination of ambient RNA in single-cell RNA-seq with DecontX. Genome Biology 21, 57 (2020).

and acknowledge omicverse / this repo for the Python port.

License

Apache-2.0. The upstream Bioconductor decontX / celda packages are MIT-licensed; pydecontx is an independent re-implementation from the published algorithm and the decontX source.