pymfuzz 0.1.2

Pure-Python port of Bioconductor Mfuzz — soft clustering of time-series gene-expression data by fuzzy c-means.

py-mfuzz

Pure-Python port of the Bioconductor package Mfuzz — soft clustering of time-series gene-expression data by fuzzy c-means (Futschik & Carlisle, J. Bioinform. Comput. Biol. 2005; Kumar & Futschik, Bioinformation 2007).

pymfuzz reproduces the full computational and visualisation API of Mfuzz with no R dependency — only numpy, scipy, pandas, matplotlib and anndata. The fuzzy c-means core is a faithful numpy port of e1071's cmeans C routine, the same algorithm R's mfuzz() wraps.

Why

Mfuzz operates on a Bioconductor ExpressionSet. pymfuzz instead accepts a plain genes × timepoints numpy.ndarray, pandas.DataFrame or anndata.AnnData, and returns numpy / pandas / dataclasses — drop-in for Python single-cell / bulk pipelines.

Install

pip install pymfuzz

From source:

pip install -e .

Quick start

import pymfuzz as mf

# 1. load a genes x timepoints time-course (Mfuzz's data(yeast))
data = mf.load_yeast()

# 2. preprocessing
data = mf.filter_NA(data, thres=0.25)   # drop genes with many NAs
data = mf.fill_NA(data, mode="knn")     # impute remaining NAs
data = mf.standardise(data)             # per-gene z-score

# 3. estimate the fuzzifier and cluster
m  = mf.mestimate(data)                 # Schwammle & Jensen (2010)
cl = mf.mfuzz(data, c=16, m=m, random_state=0)

# 4. extract core genes and plot
cores = mf.acore(data, cl, min_acore=0.5)
fig   = mf.mfuzz_plot(data, cl, mfrow=(4, 4))

API

Group	Functions
Data structures	ExpressionMatrix, as_expression_matrix, FClust, KMeansResult, AcoreCluster, PartcoefResult
Preprocessing	standardise, standardise2, filter_NA, fill_NA, filter_std
Clustering	mestimate, mfuzz, cmeans
Diagnostics	acore, Dmin, cselection, partcoef, overlap
Hard clustering	kmeans2
Plotting	mfuzz_plot, mfuzz_plot2, kmeans2_plot, overlap_plot
Datasets	load_yeast, make_synthetic_timecourse

Mapping to the R package

Mfuzz (R)	pymfuzz (Python)
standardise / standardise2	standardise / standardise2
mestimate	mestimate
mfuzz (e1071::cmeans)	mfuzz / cmeans
acore	acore
Dmin, cselection, partcoef	Dmin, cselection, partcoef
filter.NA, fill.NA, filter.std	filter_NA, fill_NA, filter_std
overlap, overlap.plot	overlap, overlap_plot
mfuzz.plot, mfuzz.plot2	mfuzz_plot, mfuzz_plot2
kmeans2, kmeans2.plot	kmeans2, kmeans2_plot

R parity

Validated against Mfuzz 2.66.0 / e1071 1.7.17 on the bundled yeast cell-cycle time-course (data(yeast)):

Routine	Agreement vs R
standardise	bit-exact (rel-diff ≈ 1e-15)
mestimate	bit-exact (rel-diff ≈ 1e-15)
fill_NA(knn)	bit-exact (max abs diff ≈ 1e-15)
mfuzz	membership Pearson r = 1.0, centres r = 1.0, hard-assignment ARI ≈ 0.99
Dmin	curve Pearson r = 1.0

standardise, mestimate and fill_NA are deterministic and match R to machine precision. Fuzzy c-means uses random initialisation, so a bit-exact match across RNGs is not expected; instead clustering agreement is asserted (Hungarian-matched membership correlation and Adjusted Rand Index). Because the yeast fuzzifier (m ≈ 1.15) is close to 1, fuzzy c-means is sharp and slow to converge — both sides take the best of several converged restarts so they reach the same optimum.

Run the parity tests (needs the CMAP R environment):

python -m pytest tests/ -q

License

GPL-2 — the same license as the original Bioconductor Mfuzz package. See LICENSE.

Citation

If you use pymfuzz, please cite the original Mfuzz papers:

• L. Kumar, M. Futschik (2007). Mfuzz: a software package for soft clustering of microarray data. Bioinformation 2(1):5–7.
• M. Futschik, B. Carlisle (2005). Noise-robust soft clustering of gene expression time-course data. J. Bioinform. Comput. Biol. 3(4):965–988.