pyproda

A pure-Python port of Bioconductor proDA (Ahlmann-Eltze & Anders 2020) — a probabilistic dropout model for differential abundance analysis of mass-spec proteomics data with MNAR missing values.

• 100% R proDA function coverage — every exported R function is ported (see table below)
• Per-protein censored maximum-likelihood fit (pd_lm) reimplemented in NumPy + scipy.optimize.L-BFGS-B
• Probit dropout curve P(observed | y) = invprobit((y - rho)/zeta) matched bit-exactly to R proDA::invprobit
• Per-sample dropout-curve hyperparameter MLE (rho_s, zeta_s) by Nelder-Mead optim
• Empirical-Bayes moderate_variance (Smyth-style σ² shrinkage) and moderate_location (Cauchy β prior)
• n_subsample chunked outer EM for large matrices
• Wald test for any user-specified contrast (test_diff); convenience pd_row_t_test / pd_row_f_test
• Probabilistic distance dist_approx, predict, median_normalization, and the full S4 accessor set
• No rpy2, no R install — all algorithmic code is reimplemented directly
• Wald t_statistic and pval agree with R proDA at Pearson r > 0.99 on the synthetic 300x12 benchmark
• 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. All algorithmic work is developed upstream in omicverse and synced here.

Install

pip install pyproda

Quick-start

import numpy as np
import pandas as pd
from pyproda import proda, test_diff

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

# Per-sample group labels (or a (n_samples, p) design matrix).
groups = ["control"] * 3 + ["treatment"] * 3

fit = proda(M, design=groups)

# Wald test: "is the treatment vs control coefficient != 0?"
res = test_diff(fit, np.array([0.0, 1.0]))
res.head()
#    name        pval      adj_pval    diff   t_statistic   se      df    avg_abundance  n_obs
#    prot_0042   3.2e-7    2.1e-4      2.13   8.4           0.25    10    18.91          6
#    ...

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

from pyproda import (
    invprobit,                # proDA::invprobit
    fit_one_protein,           # proDA::pd_lm.fit
    fit_dropout_curves,        # proDA::dropout_curves
    proda,                     # proDA::proDA
    test_diff,                 # proDA::test_diff
    pd_row_t_test,             # proDA::pd_row_t_test
    pd_row_f_test,             # proDA::pd_row_f_test
    dist_approx,               # proDA::dist_approx
    median_normalization,      # proDA::median_normalization
    predict,                   # proDA::predict
    # S4 accessors:
    abundances, coefficients, coefficient_variance_matrices,
    convergence, design, feature_parameters, hyper_parameters,
    reference_level, result_names,
)

p_obs = invprobit(y, rho=18.0, zeta=0.5)                  # probit sigmoid
hp    = fit_dropout_curves(M, Pred)                        # rho_s, zeta_s
fit_i = fit_one_protein(M[0], X, rho=hp.rho, zeta=hp.zeta) # censored MLE

# R-style functional accessors on a fitted object:
fit = proda(M, design=groups)
coefficients(fit)           # (n_proteins x p) DataFrame
feature_parameters(fit)     # n_approx / df / s2 / n_obs DataFrame
hyper_parameters(fit)       # rho / zeta + location & variance priors
predict(fit)                # per-protein x sample linear predictor
dist_approx(fit)            # probabilistic protein-protein distances

# Convenience two-/multi-group tests on raw matrices:
pd_row_t_test(X, Y)                       # row-wise censored t
pd_row_f_test(X, Y, Z)                    # multi-group F
pd_row_f_test(M, groups=["a","a","b","b"])

The ProDAFit container exposes every field the R proDAFit S4 object does — coefficients, coef_variance_matrices, s2, df, n_obs, hyper_parameters (with rho, zeta, location/variance priors), design, convergence. The same names are also available as module-level functions (coefficients(fit) etc.) to mirror the R generic-function API one-to-one.

R-parity status

The remaining gap (relative to bit-exact) is dominated by the outer EM for the dropout-curve hyperparameters — R proDA runs up to 20 iterations of (rho, zeta) <-> coefs with both Pred and Pred_var updates; v0.1 runs 3-5 fixed-point iterations and only feeds Pred_var from the censored fit. The per-protein coefficient at fixed (rho, zeta) agrees with R to ~7 decimal places.

Benchmark

300 proteins x 12 samples, 12% MNAR-censored, 15% planted DE, 5 outer iterations:

scipy.optimize.L-BFGS-B is well-matched to R's C-compiled nlminb per protein — neither has a clear edge on raw inner-loop speed. The end-to-end win is the absence of Rscript boot time (~5-6 s on the CMAP env).

Run it yourself:

python examples/benchmark.py --runs 2 --n-proteins 300 --n-per-group 6

Reproducing R results exactly

tests/r_reference_driver.R runs the R proDA pipeline (proDA(...) -> test_diff(...)) on the same TSV input the Python side dumps. tests/test_r_parity.py (skipped if no CMAP / proDA 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 function exported by R proDA is now ported (pyproda 0.1.1):

Notes on approximations: the nested-model pd_row_f_test uses an observed-sample SSR ratio rather than a full re-fit of the censored likelihood under both models — this agrees with R to a few percent on well-determined rows. moderate_variance applies post-hoc Smyth shrinkage (s2_post = (df0 s0² + df s²)/(df0+df)) rather than embedding the inverse-chi-square prior inside the per-protein objective, which is why the σ² parity is r ≈ 0.996 rather than bit-exact.

Relationship to omicverse

Developed upstream in omicverse:

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

Citation

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

Ahlmann-Eltze C., Anders S. proDA: Probabilistic Dropout Analysis for Identifying Differentially Abundant Proteins in Label-Free Quantitative Mass Spectrometry. bioRxiv (2020). DOI: 10.1101/661496

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

License

GPL-3 — matches the upstream Bioconductor package.