pydeqms

A pure-Python port of Bioconductor DEqMS (Zhu et al., MCP 2020) — peptide-count-aware moderated t-tests for differential mass-spec proteomics.

• Three-step pipeline mirroring R: lmFit → eBayes → spectraCounteBayes → outputResult
• No rpy2, no R install — limma lmFit, Smyth 2004 eBayes, and the DEqMS count-prior reimplemented directly in NumPy / SciPy
• Bit-for-bit reproduction of limma lmFit (coefs / sigma) and eBayes (df.prior / s2.prior / t / p); DEqMS sca.t / sca.P.Value match R within numerical loess slack — see tests/test_r_parity.py
• 4.8× faster than R limma + DEqMS on the canonical 5000 × 12 benchmark (Pearson r = 1.0 on sca.t / sca.P; 100% top-500 DE gene agreement)
• AnnData-friendly: accepts np.ndarray or pd.DataFrame (rows = proteins, columns = samples)

This is a standalone mirror of the canonical implementation that lives in omicverse (omicverse.protein.tl.de(method='deqms')). All algorithmic work is developed upstream in omicverse and synced here.

Install

pip install pydeqms

pydeqms depends on scikit-misc for the R-faithful loess; this is a pip-installable C/Cython package.

Quick-start (high-level pipeline)

import numpy as np
import pandas as pd
from pydeqms import deqms

# M: log2 protein intensity matrix (proteins × samples, NaN = missing)
M = pd.read_csv("proteinGroups_log2.tsv", sep="\t", index_col=0)

# design: samples × groups (e.g. two-group treatment)
design = pd.DataFrame({
    "control":   [1, 1, 1, 0, 0, 0],
    "treatment": [0, 0, 0, 1, 1, 1],
}, index=M.columns)

# count: per-protein peptide / PSM count (one column per protein)
count = pd.read_csv("peptide_counts.tsv", sep="\t", index_col=0)["count"]

result = deqms(
    M, design, count=count,
    contrast=np.array([-1.0, 1.0]),       # treatment - control
    fit_method="loess",                    # or 'nls' or 'spline'
)
result.head()
# gene        logFC  AveExpr     t  P.Value  adj.P.Val  count  sca.t  sca.P.Value  sca.adj.pval
# prot_0042   2.13    18.91    ...    ...      ...      14     5.86   3.2e-7       2.1e-4
# ...

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

from pydeqms import lm_fit, ebayes, spectra_count_ebayes, output_result

fit = lm_fit(M, design)                                   # limma::lmFit
ebayes(fit)                                                # limma::eBayes
spectra_count_ebayes(fit, count=count, fit_method="loess") # DEqMS::spectraCounteBayes
result = output_result(fit, coef_col=0)                    # DEqMS::outputResult

The LinearModelFit container exposes every field the R MArrayLM object does — coefficients, sigma, df_residual, stdev_unscaled, Amean, t, p_value, s2_post, df_prior, s2_prior, sca_t, sca_p, sca_postvar, sca_priorvar, sca_dfprior, fit_method, loess_model.

R-parity status

The 5% slack on sca.t is due to small differences between R stats::loess and skmisc.loess — both wrap the same Cleveland 1979 C backend but apply slightly different edge weights. The ranking and downstream calls are unaffected (Pearson r = 1.0, top-500 gene-set 100% identical).

Benchmark

5000 proteins × 12 samples, 10% planted DE, loess fit method (R default):

Run it yourself:

python examples/benchmark.py --runs 3 --n-proteins 5000 --n-per-group 6

Reproducing R results exactly

tests/r_reference_driver.R runs the original limma + DEqMS pipeline on the same TSV input the Python side dumps. tests/test_r_parity.py (skipped if no CMAP / DEqMS env) checks every numerical field — see the R-parity table above.

# Run the R-parity tests
pytest tests/test_r_parity.py -v

R function coverage

Every public function in DEqMS 1.24.0 has a Python counterpart.

Future work

Relationship to omicverse

Developed upstream in omicverse:

• Canonical implementation: omicverse.protein.tl.de(adata, method='deqms')
• Standalone mirror (this repo): same code, same API, minus the omicverse packaging

Citation

If you use this package, please cite both the DEqMS paper and Smyth 2004 (for the underlying empirical-Bayes framework):

Zhu Y., Orre L.M., Tran Y.Z., Mermelekas G., Johansson H.J., Malyutina A., Anders S., Lehtiö J. DEqMS: A method for accurate variance estimation in differential protein expression analysis. Mol Cell Proteomics 19, 1047–1057 (2020). DOI: 10.1074/mcp.TIR119.001646

Smyth, G. K. Linear models and empirical Bayes methods for assessing differential expression in microarray experiments. Statistical Applications in Genetics and Molecular Biology 3, Article 3 (2004).

…and acknowledge omicverse / this repo for the Python port.

License

LGPL-3 — matches the upstream Bioconductor package.