pyedger 0.1.2__tar.gz → 0.2.0__tar.gz

{pyedger-0.1.2 → pyedger-0.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pyedger
-Version: 0.1.2
+Version: 0.2.0
 Summary: Pure-Python port of Bioconductor edgeR — negative-binomial models for differential expression of count data (estimateDisp, glmFit, glmLRT, glmQLFTest, exactTest).
 Author-email: Zehua Zeng <starlitnightly@163.com>
 License: LGPL-3.0-or-later
@@ -33,7 +33,7 @@ Dynamic: license-file
 A **pure-Python port of [Bioconductor edgeR](https://bioconductor.org/packages/release/bioc/html/edgeR.html)** (Robinson, McCarthy & Smyth, *Bioinformatics* 2010) — negative-binomial models for differential expression of count data.
 
 - **No `rpy2`**, no R install — the edgeR negative-binomial GLM workflow reimplemented in NumPy / SciPy, vectorised over genes
-- Matches Bioconductor edgeR 4.0 numerically (see [Numerical parity](#numerical-parity-with-edger))
+- Matches Bioconductor edgeR 4.10 numerically (see [Numerical parity](#numerical-parity-with-edger))
 - The canonical pipelines: `DGEList → calcNormFactors → estimateDisp → glmQLFit → glmQLFTest` and the classic `exactTest`
 - TMM normalization, common / trended / tagwise dispersion, GLM and quasi-likelihood F-tests, `filterByExpr`, `cpm` / `aveLogCPM`
 - Both Python-style (`glm_fit`, `estimate_disp`, `top_tags`) and R-style (`glmFit`, `estimateDisp`, `topTags`) names exported
@@ -93,27 +93,31 @@ pyedger.topTags(et)
 
 ## Numerical parity with edgeR
 
-`tests/test_r_parity.py` checks pyedger against edgeR 4.0.16 (R 4.3.3, limma 3.58.1,
-locfit 1.5-9) on a paired design with zero-count library pairs; the reference
-values are regenerated by `tests/data/make_reference.R`. Since 0.1.1 the GLM
-kernels follow edgeR's compiled code step by step and run vectorised over all
-genes:
+`tests/test_r_parity.py` checks pyedger against edgeR 4.10.5 (R 4.5.3, limma 3.68.5,
+locfit 1.5-9.12) on a paired design with zero-count library pairs, a Milo-style
+design and limma hyperparameter cases; the reference values are regenerated by
+`tests/data/make_reference.R`. The GLM kernels follow edgeR's compiled code step
+by step and run vectorised over all genes:
 
 | Step | Ported from | Agreement with R |
 |---|---|---|
-| `glmFit` (Levenberg, one-way shortcut, start values) | `glm_levenberg.cpp`, `glm_one_group.cpp` | ~1e-9 |
-| `estimateDisp` with a design (Cox-Reid APL, zero-fit groups) and classic mode | `R_compute_apl.cpp`, `adj_coxreid.cpp`, `q2qnbinom` | ~1e-10 |
+| `glmFit` (Levenberg, one-way shortcut, start values) | `glm.c` | ~1e-9 |
+| `estimateDisp` with a design (Cox-Reid APL, zero-fit groups) and classic mode | `compute_apl.c`, `q2qnbinom` | ~1e-10 |
 | Trend: `locfitByCol` | locfit tree evaluation + interpolation | bit-identical |
-| `maximizeInterpolant`, `aveLogCPM` | `interpolator.cpp`, `R_ave_log_cpm.cpp` | ~1e-13 |
-| `squeezeVar` / `fitFDist` with a covariate | limma (natural-spline trend) | ~1e-14 |
-| `glmQLFit(legacy=False)` bias-adjusted deviance/df | `ql_glm.c`, `ql_weights.c`, `.computePriorS2` | ~1e-10 |
-
-`glmQLFit(legacy=False)` and `estimateGLMCommonDisp` depend on edgeR's
-`optimize(tol=1e-5)` over dispersion^(1/4), so R's own value is only defined to
-~4e-5 relative; agreement there is at that level.
-
-Defaults follow edgeR 4.0: `glmQLFit(legacy=True, abundance_trend=True)` and
-`estimateDisp(prior_df=None)` (prior df estimated).
+| `squeezeVar` (`fitFDist`, `fitFDistUnequalDF1`, `fitFDistRobustly`) | limma 3.68 | ~1e-11 |
+| `glmQLFit(legacy=False)` (new QL) given a dispersion | `ql_glm.c` | ~1e-9 |
+| `glmQLFit(legacy=True)`, `exactTest`, `glmTreat` | | ~1e-10 |
+
+`glmQLFit(legacy=False)` without a dispersion and `estimateGLMCommonDisp` estimate
+the NB dispersion with edgeR's `optimize(tol=1e-5)` over dispersion^(1/4); R's own
+value is only defined to ~4e-5 relative, and agreement there is at that level
+(p-values ~1e-4 relative on the test data). On real data the agreement is usually
+much closer: on four GSE274113 pseudobulk ATAC comparisons (~170k peaks x 26
+samples, `~ library + condition`, default new QL) p-values agreed to < 1e-7
+relative with identical FDR < 0.1 calls.
+
+Defaults follow edgeR 4.10: `glmQLFit(legacy=False, abundance_trend=True)`,
+`estimateDisp(prior_df=None)` with the `chooseLowessSpan` span.
 
 ## Citation
 

{pyedger-0.1.2 → pyedger-0.2.0}/README.md

@@ -3,7 +3,7 @@
 A **pure-Python port of [Bioconductor edgeR](https://bioconductor.org/packages/release/bioc/html/edgeR.html)** (Robinson, McCarthy & Smyth, *Bioinformatics* 2010) — negative-binomial models for differential expression of count data.
 
 - **No `rpy2`**, no R install — the edgeR negative-binomial GLM workflow reimplemented in NumPy / SciPy, vectorised over genes
-- Matches Bioconductor edgeR 4.0 numerically (see [Numerical parity](#numerical-parity-with-edger))
+- Matches Bioconductor edgeR 4.10 numerically (see [Numerical parity](#numerical-parity-with-edger))
 - The canonical pipelines: `DGEList → calcNormFactors → estimateDisp → glmQLFit → glmQLFTest` and the classic `exactTest`
 - TMM normalization, common / trended / tagwise dispersion, GLM and quasi-likelihood F-tests, `filterByExpr`, `cpm` / `aveLogCPM`
 - Both Python-style (`glm_fit`, `estimate_disp`, `top_tags`) and R-style (`glmFit`, `estimateDisp`, `topTags`) names exported
@@ -63,27 +63,31 @@ pyedger.topTags(et)
 
 ## Numerical parity with edgeR
 
-`tests/test_r_parity.py` checks pyedger against edgeR 4.0.16 (R 4.3.3, limma 3.58.1,
-locfit 1.5-9) on a paired design with zero-count library pairs; the reference
-values are regenerated by `tests/data/make_reference.R`. Since 0.1.1 the GLM
-kernels follow edgeR's compiled code step by step and run vectorised over all
-genes:
+`tests/test_r_parity.py` checks pyedger against edgeR 4.10.5 (R 4.5.3, limma 3.68.5,
+locfit 1.5-9.12) on a paired design with zero-count library pairs, a Milo-style
+design and limma hyperparameter cases; the reference values are regenerated by
+`tests/data/make_reference.R`. The GLM kernels follow edgeR's compiled code step
+by step and run vectorised over all genes:
 
 | Step | Ported from | Agreement with R |
 |---|---|---|
-| `glmFit` (Levenberg, one-way shortcut, start values) | `glm_levenberg.cpp`, `glm_one_group.cpp` | ~1e-9 |
-| `estimateDisp` with a design (Cox-Reid APL, zero-fit groups) and classic mode | `R_compute_apl.cpp`, `adj_coxreid.cpp`, `q2qnbinom` | ~1e-10 |
+| `glmFit` (Levenberg, one-way shortcut, start values) | `glm.c` | ~1e-9 |
+| `estimateDisp` with a design (Cox-Reid APL, zero-fit groups) and classic mode | `compute_apl.c`, `q2qnbinom` | ~1e-10 |
 | Trend: `locfitByCol` | locfit tree evaluation + interpolation | bit-identical |
-| `maximizeInterpolant`, `aveLogCPM` | `interpolator.cpp`, `R_ave_log_cpm.cpp` | ~1e-13 |
-| `squeezeVar` / `fitFDist` with a covariate | limma (natural-spline trend) | ~1e-14 |
-| `glmQLFit(legacy=False)` bias-adjusted deviance/df | `ql_glm.c`, `ql_weights.c`, `.computePriorS2` | ~1e-10 |
-
-`glmQLFit(legacy=False)` and `estimateGLMCommonDisp` depend on edgeR's
-`optimize(tol=1e-5)` over dispersion^(1/4), so R's own value is only defined to
-~4e-5 relative; agreement there is at that level.
-
-Defaults follow edgeR 4.0: `glmQLFit(legacy=True, abundance_trend=True)` and
-`estimateDisp(prior_df=None)` (prior df estimated).
+| `squeezeVar` (`fitFDist`, `fitFDistUnequalDF1`, `fitFDistRobustly`) | limma 3.68 | ~1e-11 |
+| `glmQLFit(legacy=False)` (new QL) given a dispersion | `ql_glm.c` | ~1e-9 |
+| `glmQLFit(legacy=True)`, `exactTest`, `glmTreat` | | ~1e-10 |
+
+`glmQLFit(legacy=False)` without a dispersion and `estimateGLMCommonDisp` estimate
+the NB dispersion with edgeR's `optimize(tol=1e-5)` over dispersion^(1/4); R's own
+value is only defined to ~4e-5 relative, and agreement there is at that level
+(p-values ~1e-4 relative on the test data). On real data the agreement is usually
+much closer: on four GSE274113 pseudobulk ATAC comparisons (~170k peaks x 26
+samples, `~ library + condition`, default new QL) p-values agreed to < 1e-7
+relative with identical FDR < 0.1 calls.
+
+Defaults follow edgeR 4.10: `glmQLFit(legacy=False, abundance_trend=True)`,
+`estimateDisp(prior_df=None)` with the `chooseLowessSpan` span.
 
 ## Citation
 

{pyedger-0.1.2 → pyedger-0.2.0}/pyedger/__init__.py

@@ -31,6 +31,7 @@ from .core import (
 )
 
 calcNormFactors = calc_norm_factors
+normLibSizes = calc_norm_factors          # edgeR >= 4.0 name for calcNormFactors
 effectiveLibSizes = effective_lib_sizes
 equalizeLibSizes = equalize_lib_sizes
 estimateCommonDisp = estimate_common_disp
@@ -65,6 +66,7 @@ __all__ = [
     "aveLogCPM",
     "calc_norm_factors",
     "calcNormFactors",
+    "normLibSizes",
     "effective_lib_sizes",
     "effectiveLibSizes",
     "equalize_lib_sizes",

{pyedger-0.1.2 → pyedger-0.2.0}/pyedger/_glm_batch.py

@@ -20,7 +20,7 @@ LOG_LOW_VALUE = np.log(LOW_VALUE)
 
 # --------------------------------------------------------------------------- deviance
 def unit_nb_deviance(y, mu, phi):
-    """nbdev.c compute_unit_nb_deviance, elementwise with broadcasting."""
+    """compute_nbdev.c compute_unit_nb_deviance (edgeR 4.10), elementwise with broadcasting."""
     y = np.asarray(y, dtype=float) + 1e-8
     mu = np.asarray(mu, dtype=float) + 1e-8
     y, mu, phi = np.broadcast_arrays(y, mu, np.asarray(phi, dtype=float))
@@ -38,7 +38,7 @@ def unit_nb_deviance(y, mu, phi):
         nb = rest & ~(prod > 1e6)
         inv = 1.0 / phi[nb]
         out[nb] = 2 * (y[nb] * np.log(y[nb] / mu[nb]) + (y[nb] + inv) * np.log((mu[nb] + inv) / (y[nb] + inv)))
-    return out
+    return np.maximum(out, 0.0)                 # edgeR >= 4.4 floors unit deviances at 0
 
 
 def nb_deviance(Y, Mu, Phi, W):