pyDoubletFinder

Faithful Python port of the R DoubletFinder algorithm for scRNA-seq doublet detection

---

pyDoubletFinder is a line-by-line Python port of the R DoubletFinder algorithm, designed as a drop-in replacement for projects using scanpy / AnnData without requiring an R environment. Replicates the exact Seurat preprocessing pipeline including LogNormalize, VST, ScaleData, full Euclidean distance matrix, and pANN scoring.

Features

• Line-by-line port - replicates the exact R DoubletFinder algorithm
• Native VST - reimplementation of Seurat v3's FindVariableFeatures(method="vst") on raw counts
• R-matching loess - uses scikit-misc (degree=2) to match R's stats::loess exactly
• Full preprocessing pipeline - LogNormalize, VST, ScaleData, PCA, distance matrix, pANN
• 94.3% classification agreement with R on matched data (4926 cells)
• 99.5% HVG overlap confirms faithful VST reproduction
• Parameter sweep - param_sweep_and_summarize() for automatic pK selection via bimodality coefficient
• SCTransform approximation - experimental support via Pearson residuals
• scanpy / AnnData native - no R dependencies required

Installation

pip install doubletfinder-py

For exact R-matching loess (recommended):

pip install "doubletfinder-py[loess]"

This installs scikit-misc which provides skmisc.loess — a degree-2 loess matching R's stats::loess. Without it, the library falls back to statsmodels.lowess (degree-1, local linear), which is a close but not identical approximation.

Quick Start

import scanpy as sc
from pydoubletfinder import doublet_finder, model_homotypic

adata = sc.read_10x_h5("sample.h5")
adata.var_names_make_unique()
sc.pp.filter_cells(adata, min_genes=200)
sc.pp.filter_genes(adata, min_cells=3)
adata.layers["counts"] = adata.X.copy()

pK   = 0.09
nExp = int(0.075 * adata.n_obs)

# Optional: adjust for homotypic doublets
homo_prop = model_homotypic(adata.obs["cell_type"].values)
nExp = int(nExp * (1 - homo_prop))

adata = doublet_finder(adata, PCs=10, pK=pK, nExp=nExp)

col_class = f"DF.classifications_0.25_{pK}_{nExp}"
print(adata.obs[col_class].value_counts())

For pK tuning, annotations, reuse and sparse data see docs/usage.md.

Gallery

pANN Distribution	PC Selection	Multi-Sample Batch

Examples

10 runnable scripts covering all features — see docs/examples.md for the full list with previews.

cd examples && python generate_all.py

Automatic pK selection

from pydoubletfinder import param_sweep_and_summarize

sweep_df = param_sweep_and_summarize(adata, PCs=10)
best_pK  = float(sweep_df.loc[sweep_df["BCreal"].idxmax(), "pK"])

Note: the parameter sweep is computationally expensive. For most datasets, a fixed pK=0.09 is a reasonable starting point.

Benchmark vs R

Tested on snRNA-seq mouse EAM data (sample42, D0, 4926 cells) using identical doublet pairs (same random seed exported from R):

Metric	Value
Classification agreement	94.32%
pANN Pearson r	0.8236
pANN Spearman r	0.8477
HVG overlap (VST)	1990 / 2000 (99.5%)
Cohen's κ	0.5899

Where the ~6% discrepancy comes from

Source	Impact	Details
PCA solver	~5.5%	R uses irlba (Seurat), Python uses ARPACK (scanpy.tl.pca)
HVG selection (VST)	~0.5%	10 different genes out of 2000 — negligible

The ~6% discrepancy is a fundamental property of the port — R's irlba and Python's SVD solvers use different numerical paths. All 280 cells classified differently (140 in each direction of the confusion matrix) have pANN values within ~0.01 of the decision threshold. No solver swap can reliably fix this without reimplementing irlba line-for-line in Python.

API

doublet_finder(adata, PCs, pK, nExp, pN=0.25, ...)

Core doublet prediction function. Adds two columns to adata.obs:

• pANN_{pN}_{pK}_{nExp} — doublet score (proportion of artificial nearest neighbours)
• DF.classifications_{pN}_{pK}_{nExp} — "Singlet" or "Doublet"

Parameters:

Parameter	Type	Default	Description
adata	AnnData	—	Input object. Raw counts in adata.layers["counts"], adata.raw.X, or adata.X.
PCs	int or list[int]	—	Number of PCs or list of 1-based PC indices.
pK	float	—	Neighbourhood proportion for pANN computation.
nExp	int	—	Expected number of doublets (classification threshold).
pN	float	0.25	Proportion of artificial doublets to generate.
reuse_pANN	str or None	None	Existing adata.obs column with precomputed pANN — skips heavy computation.
sct	bool	False	Use SCTransform-like normalisation (experimental).
annotations	array or None	None	Cell-type labels. Adds DF.doublet.contributors_* columns.
scale_factor	float	1e4	Target sum for normalisation.
n_top_genes	int	2000	Number of HVGs for VST.
loess_span	float	0.3	Span for loess in VST.
scale_max	float	10	Clip value for ScaleData.
random_state	int	0	PCA seed.

model_homotypic(annotations)

Estimates the proportion of homotypic doublets from cell type annotations. Returns sum(p_i^2) where p_i is the proportion of cell type i. Replicates R's modelHomotypic.

param_sweep_and_summarize(adata, PCs, ...)

Runs a pN–pK parameter sweep and returns a DataFrame with columns pN, pK, BCreal (bimodality coefficient). Select the pK that maximises BCreal.

Differences from R DoubletFinder

Aspect	R	Python
Normalisation	NormalizeData (Seurat)	sc.pp.normalize_total + sc.pp.log1p
HVG selection	FindVariableFeatures(method="vst")	Native reimplementation (_seurat_vst)
Scaling	ScaleData (Seurat)	sc.pp.scale
PCA	irlba via RunPCA	ARPACK via sc.tl.pca
Distance matrix	fields::rdist	scipy.spatial.distance.cdist
Loess (VST)	stats::loess (degree=2)	skmisc.loess (degree=2) or statsmodels.lowess fallback

Benchmarks

To reproduce the benchmark comparing this implementation against R DoubletFinder:

SAMPLE_H5=/path/to/sample.h5 bash benchmarks/benchmark.sh

Requires Docker. On first run, builds an image with R 4.4 + Seurat + Python (~10 min). Subsequent runs reuse the cached image.

Results are written to benchmarks/results/:

• comparison_report.txt — full metrics summary
• plots/pann_scatter.png — pANN correlation scatter
• plots/pann_hist.png — pANN distribution overlay
• plots/confusion.png — classification confusion matrix
• plots/hvg_overlap.png — HVG overlap bar chart

Citation

If you use pyDoubletFinder in a publication, please cite both this package and the original DoubletFinder paper:

APA:

dam2452. (2026). pyDoubletFinder: Python port of the R DoubletFinder algorithm (Version 1.0.0). https://github.com/dam2452/pydoubletfinder

McGinnis, C.S., Murrow, L.M. & Gartner, Z.J. (2019). DoubletFinder: Doublet Detection in Single-Cell RNA Sequencing Data Using Artificial Nearest Neighbors. Cell Systems, 8, 329–337.e4. https://doi.org/10.1016/j.cels.2019.03.003

BibTeX:

@software{pydoubletfinder2026,
  title   = {pyDoubletFinder: Python port of the R DoubletFinder algorithm},
  author  = {dam2452},
  year    = {2026},
  version = {1.0.0},
  url     = {https://github.com/dam2452/pydoubletfinder}
}

@article{mcginnis2019doubletfinder,
  title     = {{DoubletFinder}: Doublet Detection in Single-Cell {RNA} Sequencing Data Using Artificial Nearest Neighbors},
  author    = {McGinnis, Christopher S. and Murrow, Lydia M. and Gartner, Zev J.},
  journal   = {Cell Systems},
  volume    = {8},
  number    = {4},
  pages     = {329--337.e4},
  year      = {2019},
  doi       = {10.1016/j.cels.2019.03.003}
}

Contributing

Contributions are welcome! Here's how you can help:

1. Bug reports - Open an issue with a minimal reproducible example
2. Feature requests - Open an issue describing the use case
3. Code contributions - Fork, create a feature branch, and open a pull request

Development setup

git clone https://github.com/dam2452/pydoubletfinder.git
cd pydoubletfinder
pip install -e ".[dev]"
pytest tests/

License

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

Reference

McGinnis, C.S., Murrow, L.M. & Gartner, Z.J. DoubletFinder: Doublet Detection in Single-Cell RNA Sequencing Data Using Artificial Nearest Neighbors. Cell Systems 8, 329–337.e4 (2019). https://doi.org/10.1016/j.cels.2019.03.003