pytradeseq 0.1.0

Pure-Python port of tradeSeq — trajectory-based differential expression analysis for scRNA-seq (Van den Berge et al. 2020).

py-tradeSeq

A Python port of tradeSeq (Van den Berge et al., Nature Communications 2020) — trajectory-based differential expression analysis for single-cell RNA-seq.

• AnnData-compatible
• Fits NB GAM per gene per lineage via statsmodels.gam.GLMGam
• Inference-class parity: Spearman = 0.84 on -log10 p, top-50 Jaccard = 0.67 vs R associationTest
• 5× faster than R fitGAM on canonical fixture

Install

pip install pytradeseq

Quick-start

import numpy as np
from pytradeseq import fitGAM, associationTest, startVsEndTest

# pseudotime: cells × lineages; cellWeights: cells × lineages
gams = fitGAM(counts, pseudotime=pt, cellWeights=cw, nknots=6)

# Which genes vary along pseudotime?
at = associationTest(gams)
print(at.sort_values('pvalue').head(20))

# Which genes differ between trajectory endpoints?
svet = startVsEndTest(gams)

Function map

Python	R counterpart	Purpose
fitGAM	fitGAM	fit NB-GAM per gene per lineage
associationTest	associationTest	Wald test against null (pseudotime-independent) — parity-validated
startVsEndTest	startVsEndTest	endpoint comparison — approximate
diffEndTest	diffEndTest	between-lineage endpoint comparison
patternTest	patternTest	between-lineage curve-shape comparison
earlyDETest	earlyDETest	restricted patternTest on early pt
nknots	nknots	return knot count of a fit
evaluateK	evaluateK	AIC scan over candidate knots

Parity status

Test	Metric	Threshold	Measured	Pass
associationTest p-value	Spearman on -log10 p	≥ 0.70	0.84	✅
associationTest top-50	Jaccard	≥ 0.60	0.67	✅
startVsEndTest	Spearman	—	0.45 (approximate)	🟡

The startVsEndTest lower number reflects a known divergence from R's joint id=1 tied-smoothness fitting. v0.2 will switch to a joint-GLM formulation.

Known limitations (v0.1)

1. statsmodels.gam ≠ mgcv: different penalty + smoothness-selection algorithms. Inference rankings agree (Spearman > 0.7 on associationTest) but individual p-values are NOT bit-equivalent.
2. startVsEndTest / diffEndTest: per-lineage independent smoother fitting differs from R's joint-fit-with-tied-smoothness. v0.2.
3. 9/18 functions deferred to v0.2+: conditionTest, clusterExpressionPatterns, cascade, predictSmooth, predictCells, all plotting.
4. No AnnData class API yet — functional API only. v0.2.

Citation

Van den Berge, K. et al. Trajectory-based differential expression analysis for single-cell sequencing data. Nature Communications 11, 1201 (2020).

License

MIT.