Patch catalog

Every patch that ships with AutoZyme, grouped by ecosystem. Per-tier breakdowns and historical runs live on the benchmarks page.

Python (pip install "git+https://github.com/ElliotXie/autozyme.git#subdirectory=autozyme_py")

Patch Upstream Targets Speedup (median) Concordance gate Tested against
scanpy scanpy normalize_total, pca, scale, leiden, highly_variable_genes, rank_genes_groups 1.6×–130× pearson ≥ 0.999 / ARI ≥ 0.95 scanpy 1.11.5
scvelo scvelo recover_dynamics 1.9×–3× pearson ≥ 0.999 scvelo 0.3.5.dev1+gf63c0e705
sccoda sccoda + tensorflow_probability run_nuts 1.4×–1.7× per-sample mean diff ≤ 0.05 sccoda 0.1.9
cell2location cell2location + pyro train_full + train_amortised 2.1×–2.3× mean cell-type abundance ≤ 1% drift cell2location 0.1.5
squidpy_cooccurrence squidpy co_occurrence 5.8×–31× mean cooccurrence delta ≤ 0.01 squidpy 1.6.5
cellphonedb cellphonedb statistical_analysis 11×–19× LR-pair recall ≥ 0.99 cellphonedb 5.0.1
mdanalysis_rmsd MDAnalysis RMSD.run 3.3×–4.1× max diff ≤ 1e-3 Å MDAnalysis 2.10.0
prody prody calcCovariance + calcModes 34×–112× eigenvalue ratio ≤ 1.001 prody 2.6.1
statsmodels statsmodels GLM.fit 15×–22× coef relative diff ≤ 1e-6 statsmodels 0.14.6
lifelines lifelines CoxPHFitter.fit 1.4×–1.7× survival-curve max diff ≤ 1e-6 lifelines 0.30.3
dipy dipy TensorModel.fit 5×–9.6× per-voxel diff ≤ 1e-4 dipy 1.12.1
fipy fipy TransientTerm.solve 31×–63× per-cell diff ≤ 1e-4 fipy 4.0.2+4.gb847e2c
obspy obspy Trace.filter (bandpass) 3.7×–5.1× sample-wise diff ≤ 1e-6 obspy 1.5.0
xclim xclim indices.atmos.* (subset) 350×–871× per-grid relative diff ≤ 1e-4 xclim 0.60.0
sarsen sarsen + xarray_sentinel terrain_correction 5×–21× pixel-wise relative diff ≤ 0.01 sarsen 0.9.6.dev5+g6c5e37d1d
astropy_boxleastsquares astropy BoxLeastSquares.autopower 2.8×–3.9× peak-period diff ≤ 1e-4 astropy 6.1.0

R (remotes::install_github("ElliotXie/autozyme", subdir = "autozyme_r"))

Patch Upstream Targets Speedup (median) Concordance gate Tested against
seurat Seurat FindAllMarkers, FindNeighbors, FindVariableFeatures, IntegrateLayers (CCA), NormalizeData, RunPCA, ScaleData, SCTransform 2.3×–156× identical (deterministic) or pearson ≥ 0.999 Seurat 5.4.0
cellchat CellChat computeCommunProb 13×–234× LR-pair recall ≥ 0.99 CellChat 2.2.0.9001
nichenetr nichenetr predict_ligand_activities 575×–1027× pearson ≥ 0.999 nichenetr 2.2.1.1
scriabin scriabin GenerateCCIM 1.9×–2.9× LR-pair recall ≥ 0.99 scriabin 0.0.0.9000
slingshot slingshot getCurves 2.1×–3.6× pseudotime pearson ≥ 0.999 slingshot 2.16.0
tradeseq tradeSeq fitGAM 1.4×–4× coef pearson ≥ 0.999 tradeSeq 1.22.0
mast MAST zlm + lrTest 2.3×–3.7× p-value rank pearson ≥ 0.99 MAST 1.35.2
bayesspace BayesSpace spatialEnhance 7.7×–9.4× label ARI ≥ 0.95 BayesSpace 1.21.2
infercnv infercnv run 34×–62× per-gene CNV state agreement ≥ 0.99 infercnv 1.24.0
rctd spacexr run.RCTD 7.4×–19× cell-type fraction pearson ≥ 0.999 spacexr 2.2.1
clusterprofiler clusterProfiler compareCluster (enrichGO ORA) 12×–71× pathway recall ≥ 0.99 clusterProfiler 4.16.0
decontx celda decontX 1.2×–1.9× contamination-fraction pearson ≥ 0.999 celda 1.24.0
fgsea fgsea fgseaMultilevel 2.5×–3.2× per-pathway NES diff ≤ 0.01 fgsea 1.34.2
maftools maftools read.maf + summarizeMaf 11×–27× gene-count exact match maftools 2.24.0
vegan vegan adonis2 82×–182× F-statistic relative diff ≤ 1e-4 vegan 2.8.0
wgcna WGCNA blockwiseModules 13×–59× module-assignment ARI ≥ 0.95 WGCNA 1.74

Reading the table

  • Targets: the upstream functions/methods this patch replaces. Other functions in the upstream package are unaffected.
  • Speedup (median): median across the tier sweep. Tiny tier ≈ best case; large/OOD tier ≈ worst case. See benchmarks for per-tier rows.
  • Concordance gate: the metric and threshold the patched output must satisfy vs vanilla. Patches that drift fail CI and don’t ship — every published patch in this table has passed its gate on every tier in its test matrix.
  • Tested against: the upstream version pinned in CI. Patches generally work on a range of nearby versions; the tested_upstream_versions field in each patch declares which versions have been explicitly validated (see Python API reference → inspect()).

Don’t see what you want accelerated?

Vote for it on Suggest & Vote — community demand is the primary signal for what enters the queue next.