scccvgben 0.1.0__tar.gz
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- scccvgben-0.1.0/LICENSE +21 -0
- scccvgben-0.1.0/PKG-INFO +126 -0
- scccvgben-0.1.0/README.md +71 -0
- scccvgben-0.1.0/pyproject.toml +73 -0
- scccvgben-0.1.0/scccvgben/__init__.py +6 -0
- scccvgben-0.1.0/scccvgben/baselines/__init__.py +5 -0
- scccvgben-0.1.0/scccvgben/baselines/deep_methods.py +250 -0
- scccvgben-0.1.0/scccvgben/baselines/runner.py +207 -0
- scccvgben-0.1.0/scccvgben/baselines/sklearn_methods.py +86 -0
- scccvgben-0.1.0/scccvgben/biovalidation/__init__.py +29 -0
- scccvgben-0.1.0/scccvgben/biovalidation/case_definition.py +116 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compose/__init__.py +11 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compose/case_figure.py +186 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compute/__init__.py +12 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compute/case_run.py +170 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compute/latent_gene_corr.py +100 -0
- scccvgben-0.1.0/scccvgben/biovalidation/compute/pathway.py +151 -0
- scccvgben-0.1.0/scccvgben/biovalidation/sidecar.py +270 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/__init__.py +19 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/case_style.py +34 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/dotplot.py +115 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/gene_grid.py +86 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/heatmap.py +84 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/panel.py +90 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/scatter.py +162 -0
- scccvgben-0.1.0/scccvgben/biovalidation/visualize/violin.py +79 -0
- scccvgben-0.1.0/scccvgben/config.py +22 -0
- scccvgben-0.1.0/scccvgben/data/__init__.py +5 -0
- scccvgben-0.1.0/scccvgben/data/datasets.csv +223 -0
- scccvgben-0.1.0/scccvgben/data/labels.py +58 -0
- scccvgben-0.1.0/scccvgben/data/loader.py +75 -0
- scccvgben-0.1.0/scccvgben/data/preprocessing.py +105 -0
- scccvgben-0.1.0/scccvgben/data/result_csv_normalizer.py +105 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/__init__.py +15 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CGVAE.py +180 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CGVAE_agent.py +178 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CGVAE_env.py +273 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CGVAE_module.py +305 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CGVAE_trainer.py +142 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/__init__.py +5 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/agent.py +338 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/environment.py +121 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/mixin.py +259 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/model.py +463 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/module.py +535 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/CODE/utils.py +140 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/__init__.py +5 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/mixin.py +356 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/cgvae/utils.py +332 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/dre.py +435 -0
- scccvgben-0.1.0/scccvgben/external/reference_core/lse.py +640 -0
- scccvgben-0.1.0/scccvgben/figures/__init__.py +75 -0
- scccvgben-0.1.0/scccvgben/figures/_long_form.py +150 -0
- scccvgben-0.1.0/scccvgben/figures/_naming.py +23 -0
- scccvgben-0.1.0/scccvgben/figures/_significance.py +248 -0
- scccvgben-0.1.0/scccvgben/figures/fonts.py +108 -0
- scccvgben-0.1.0/scccvgben/figures/metrics.py +318 -0
- scccvgben-0.1.0/scccvgben/figures/style.py +993 -0
- scccvgben-0.1.0/scccvgben/graphs/__init__.py +17 -0
- scccvgben-0.1.0/scccvgben/graphs/construction.py +205 -0
- scccvgben-0.1.0/scccvgben/models/__init__.py +4 -0
- scccvgben-0.1.0/scccvgben/models/decoders.py +82 -0
- scccvgben-0.1.0/scccvgben/models/encoder_registry.py +161 -0
- scccvgben-0.1.0/scccvgben/models/losses.py +65 -0
- scccvgben-0.1.0/scccvgben/models/scccvgben_model.py +173 -0
- scccvgben-0.1.0/scccvgben/stats/__init__.py +13 -0
- scccvgben-0.1.0/scccvgben/stats/tests.py +123 -0
- scccvgben-0.1.0/scccvgben/training/__init__.py +4 -0
- scccvgben-0.1.0/scccvgben/training/configs.py +47 -0
- scccvgben-0.1.0/scccvgben/training/dre.py +226 -0
- scccvgben-0.1.0/scccvgben/training/graph_sweep.py +158 -0
- scccvgben-0.1.0/scccvgben/training/lse.py +55 -0
- scccvgben-0.1.0/scccvgben/training/metrics.py +156 -0
- scccvgben-0.1.0/scccvgben/training/pair_sweep.py +174 -0
- scccvgben-0.1.0/scccvgben/training/prefetch.py +91 -0
- scccvgben-0.1.0/scccvgben/training/scccvgben_runner.py +239 -0
- scccvgben-0.1.0/scccvgben/training/trainer.py +91 -0
- scccvgben-0.1.0/scccvgben.egg-info/PKG-INFO +126 -0
- scccvgben-0.1.0/scccvgben.egg-info/SOURCES.txt +94 -0
- scccvgben-0.1.0/scccvgben.egg-info/dependency_links.txt +1 -0
- scccvgben-0.1.0/scccvgben.egg-info/requires.txt +38 -0
- scccvgben-0.1.0/scccvgben.egg-info/top_level.txt +1 -0
- scccvgben-0.1.0/setup.cfg +4 -0
- scccvgben-0.1.0/tests/test_audit_ai_tone.py +47 -0
- scccvgben-0.1.0/tests/test_audit_submission_ready.py +137 -0
- scccvgben-0.1.0/tests/test_dre_port.py +117 -0
- scccvgben-0.1.0/tests/test_encoder_registry.py +44 -0
- scccvgben-0.1.0/tests/test_fetch_geo_skeleton.py +269 -0
- scccvgben-0.1.0/tests/test_graph_constructions.py +52 -0
- scccvgben-0.1.0/tests/test_imports.py +12 -0
- scccvgben-0.1.0/tests/test_make_diff_vs_ccvgae.py +47 -0
- scccvgben-0.1.0/tests/test_public_surface_links.py +48 -0
- scccvgben-0.1.0/tests/test_scccvgben_model_forward.py +81 -0
- scccvgben-0.1.0/tests/test_smoke_pipelines.py +54 -0
- scccvgben-0.1.0/tests/test_stats.py +52 -0
- scccvgben-0.1.0/tests/test_symlinks.py +47 -0
scccvgben-0.1.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 scCCVGBen contributors
|
|
4
|
+
|
|
5
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
6
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
7
|
+
in the Software without restriction, including without limitation the rights
|
|
8
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
9
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
10
|
+
furnished to do so, subject to the following conditions:
|
|
11
|
+
|
|
12
|
+
The above copyright notice and this permission notice shall be included in all
|
|
13
|
+
copies or substantial portions of the Software.
|
|
14
|
+
|
|
15
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
16
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
17
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
18
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
19
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
20
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
21
|
+
SOFTWARE.
|
scccvgben-0.1.0/PKG-INFO
ADDED
|
@@ -0,0 +1,126 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: scccvgben
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: Single-cell scCCVGBen Benchmark — multi-encoder, multi-graph, baseline comparison on 200 scRNA+scATAC datasets
|
|
5
|
+
Author: Zeyu Fu
|
|
6
|
+
License: MIT
|
|
7
|
+
Project-URL: Homepage, https://github.com/PeterPonyu/scCCVGBen
|
|
8
|
+
Project-URL: Repository, https://github.com/PeterPonyu/scCCVGBen
|
|
9
|
+
Project-URL: Documentation, https://peterponyu.github.io/scCCVGBen/
|
|
10
|
+
Project-URL: Paper, https://doi.org/10.3389/fgene.2026.1822168
|
|
11
|
+
Keywords: single-cell,benchmark,graph neural network,variational autoencoder,scRNA-seq,scATAC-seq
|
|
12
|
+
Classifier: Development Status :: 4 - Beta
|
|
13
|
+
Classifier: Intended Audience :: Science/Research
|
|
14
|
+
Classifier: License :: OSI Approved :: MIT License
|
|
15
|
+
Classifier: Operating System :: OS Independent
|
|
16
|
+
Classifier: Programming Language :: Python :: 3
|
|
17
|
+
Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
|
|
18
|
+
Requires-Python: >=3.10
|
|
19
|
+
Description-Content-Type: text/markdown
|
|
20
|
+
License-File: LICENSE
|
|
21
|
+
Requires-Dist: torch>=2.0
|
|
22
|
+
Requires-Dist: torch-geometric
|
|
23
|
+
Requires-Dist: scanpy
|
|
24
|
+
Requires-Dist: anndata
|
|
25
|
+
Requires-Dist: numpy
|
|
26
|
+
Requires-Dist: pandas
|
|
27
|
+
Requires-Dist: scipy
|
|
28
|
+
Requires-Dist: scikit-learn
|
|
29
|
+
Requires-Dist: matplotlib
|
|
30
|
+
Requires-Dist: gseapy
|
|
31
|
+
Requires-Dist: networkx
|
|
32
|
+
Requires-Dist: pyyaml
|
|
33
|
+
Requires-Dist: umap-learn
|
|
34
|
+
Requires-Dist: seaborn
|
|
35
|
+
Requires-Dist: psutil
|
|
36
|
+
Requires-Dist: tqdm
|
|
37
|
+
Provides-Extra: atac
|
|
38
|
+
Requires-Dist: episcanpy; extra == "atac"
|
|
39
|
+
Provides-Extra: geo
|
|
40
|
+
Requires-Dist: GEOparse>=2.0.3; extra == "geo"
|
|
41
|
+
Provides-Extra: dev
|
|
42
|
+
Requires-Dist: pytest; extra == "dev"
|
|
43
|
+
Requires-Dist: ruff; extra == "dev"
|
|
44
|
+
Requires-Dist: mypy; extra == "dev"
|
|
45
|
+
Provides-Extra: sparse
|
|
46
|
+
Requires-Dist: torch-sparse; extra == "sparse"
|
|
47
|
+
Provides-Extra: full
|
|
48
|
+
Requires-Dist: episcanpy; extra == "full"
|
|
49
|
+
Requires-Dist: GEOparse>=2.0.3; extra == "full"
|
|
50
|
+
Requires-Dist: torch-sparse; extra == "full"
|
|
51
|
+
Requires-Dist: scib; extra == "full"
|
|
52
|
+
Requires-Dist: scvi-tools; extra == "full"
|
|
53
|
+
Requires-Dist: torchdiffeq; extra == "full"
|
|
54
|
+
Dynamic: license-file
|
|
55
|
+
|
|
56
|
+
# scCCVGBen
|
|
57
|
+
|
|
58
|
+
A 200-dataset single-cell representation-learning benchmark and a Python
|
|
59
|
+
package for the centroid-coupled variational graph attention autoencoder
|
|
60
|
+
(100 scRNA-seq + 100 scATAC-seq, balanced).
|
|
61
|
+
|
|
62
|
+
The repository carries:
|
|
63
|
+
|
|
64
|
+
- the dataset catalogue with GEO-verified metadata,
|
|
65
|
+
- the method and metric registries,
|
|
66
|
+
- the Python package used to regenerate the benchmark,
|
|
67
|
+
- the Hugo source for the public site,
|
|
68
|
+
- the Next.js companion explorer (`webapp/`).
|
|
69
|
+
|
|
70
|
+
## Install
|
|
71
|
+
|
|
72
|
+
```bash
|
|
73
|
+
pip install scccvgben
|
|
74
|
+
```
|
|
75
|
+
|
|
76
|
+
The reference graph model also needs `torch-sparse` (`pip install "scccvgben[sparse]"`, with a wheel that matches your PyTorch build). To regenerate the benchmark, clone the repository and run `pip install -e ".[dev,full]"`.
|
|
77
|
+
|
|
78
|
+
## Reproduce
|
|
79
|
+
|
|
80
|
+
See [`REPRODUCE.md`](REPRODUCE.md). Sweep scripts are idempotent and resume
|
|
81
|
+
per `(dataset, method)` row.
|
|
82
|
+
|
|
83
|
+
## Layout
|
|
84
|
+
|
|
85
|
+
- `scccvgben/` — Python package
|
|
86
|
+
- `scripts/` — data curation and sweep runners
|
|
87
|
+
- `data/` — dataset metadata CSVs
|
|
88
|
+
- `figures/` — generated figures (raw screenshots in `figures/site_shots/` stay local)
|
|
89
|
+
- `site/` — Hugo source for the public site
|
|
90
|
+
- `webapp/` — Next.js benchmark explorer
|
|
91
|
+
- `tests/` — smoke tests
|
|
92
|
+
|
|
93
|
+
## Figures
|
|
94
|
+
|
|
95
|
+
- **Figure 1 — site and benchmark overview.** Render with `scripts/make_figure1_site.py`.
|
|
96
|
+
- **Figure 2 — model architecture and benchmark axes.** Render with `scripts/make_figure2_model_architecture.py`.
|
|
97
|
+
- Generated figure binaries are not committed by the figure-pipeline PR.
|
|
98
|
+
|
|
99
|
+
## Sites
|
|
100
|
+
|
|
101
|
+
- Hugo public site: <https://peterponyu.github.io/scCCVGBen/>
|
|
102
|
+
- Next.js companion explorer: <https://peterponyu.github.io/scccvgben-next/>
|
|
103
|
+
|
|
104
|
+
## Citation
|
|
105
|
+
|
|
106
|
+
If you use scCCVGBen, please cite the article:
|
|
107
|
+
|
|
108
|
+
> Fu Z, Fu J, Chen C, Zhang K, Wang J, Ran T, Wang S. scCCVGBen for benchmarking of single-cell representation learning anchored on a centroid-coupled variational graph attention autoencoder across scRNA-seq and scATAC-seq. *Frontiers in Genetics*. 2026;17:1822168. https://doi.org/10.3389/fgene.2026.1822168
|
|
109
|
+
|
|
110
|
+
```bibtex
|
|
111
|
+
@article{fu2026scccvgben,
|
|
112
|
+
author = {Fu, Zeyu and Fu, Jiawei and Chen, Chunlin and Zhang, Keyang and Wang, Junping and Ran, Tianfei and Wang, Song},
|
|
113
|
+
title = {scCCVGBen for benchmarking of single-cell representation learning anchored on a centroid-coupled variational graph attention autoencoder across scRNA-seq and scATAC-seq},
|
|
114
|
+
journal = {Frontiers in Genetics},
|
|
115
|
+
volume = {17},
|
|
116
|
+
pages = {1822168},
|
|
117
|
+
year = {2026},
|
|
118
|
+
doi = {10.3389/fgene.2026.1822168}
|
|
119
|
+
}
|
|
120
|
+
```
|
|
121
|
+
|
|
122
|
+
GitHub's "Cite this repository" button reads the same record from `CITATION.cff`.
|
|
123
|
+
|
|
124
|
+
## License
|
|
125
|
+
|
|
126
|
+
MIT — see [`LICENSE`](LICENSE).
|
|
@@ -0,0 +1,71 @@
|
|
|
1
|
+
# scCCVGBen
|
|
2
|
+
|
|
3
|
+
A 200-dataset single-cell representation-learning benchmark and a Python
|
|
4
|
+
package for the centroid-coupled variational graph attention autoencoder
|
|
5
|
+
(100 scRNA-seq + 100 scATAC-seq, balanced).
|
|
6
|
+
|
|
7
|
+
The repository carries:
|
|
8
|
+
|
|
9
|
+
- the dataset catalogue with GEO-verified metadata,
|
|
10
|
+
- the method and metric registries,
|
|
11
|
+
- the Python package used to regenerate the benchmark,
|
|
12
|
+
- the Hugo source for the public site,
|
|
13
|
+
- the Next.js companion explorer (`webapp/`).
|
|
14
|
+
|
|
15
|
+
## Install
|
|
16
|
+
|
|
17
|
+
```bash
|
|
18
|
+
pip install scccvgben
|
|
19
|
+
```
|
|
20
|
+
|
|
21
|
+
The reference graph model also needs `torch-sparse` (`pip install "scccvgben[sparse]"`, with a wheel that matches your PyTorch build). To regenerate the benchmark, clone the repository and run `pip install -e ".[dev,full]"`.
|
|
22
|
+
|
|
23
|
+
## Reproduce
|
|
24
|
+
|
|
25
|
+
See [`REPRODUCE.md`](REPRODUCE.md). Sweep scripts are idempotent and resume
|
|
26
|
+
per `(dataset, method)` row.
|
|
27
|
+
|
|
28
|
+
## Layout
|
|
29
|
+
|
|
30
|
+
- `scccvgben/` — Python package
|
|
31
|
+
- `scripts/` — data curation and sweep runners
|
|
32
|
+
- `data/` — dataset metadata CSVs
|
|
33
|
+
- `figures/` — generated figures (raw screenshots in `figures/site_shots/` stay local)
|
|
34
|
+
- `site/` — Hugo source for the public site
|
|
35
|
+
- `webapp/` — Next.js benchmark explorer
|
|
36
|
+
- `tests/` — smoke tests
|
|
37
|
+
|
|
38
|
+
## Figures
|
|
39
|
+
|
|
40
|
+
- **Figure 1 — site and benchmark overview.** Render with `scripts/make_figure1_site.py`.
|
|
41
|
+
- **Figure 2 — model architecture and benchmark axes.** Render with `scripts/make_figure2_model_architecture.py`.
|
|
42
|
+
- Generated figure binaries are not committed by the figure-pipeline PR.
|
|
43
|
+
|
|
44
|
+
## Sites
|
|
45
|
+
|
|
46
|
+
- Hugo public site: <https://peterponyu.github.io/scCCVGBen/>
|
|
47
|
+
- Next.js companion explorer: <https://peterponyu.github.io/scccvgben-next/>
|
|
48
|
+
|
|
49
|
+
## Citation
|
|
50
|
+
|
|
51
|
+
If you use scCCVGBen, please cite the article:
|
|
52
|
+
|
|
53
|
+
> Fu Z, Fu J, Chen C, Zhang K, Wang J, Ran T, Wang S. scCCVGBen for benchmarking of single-cell representation learning anchored on a centroid-coupled variational graph attention autoencoder across scRNA-seq and scATAC-seq. *Frontiers in Genetics*. 2026;17:1822168. https://doi.org/10.3389/fgene.2026.1822168
|
|
54
|
+
|
|
55
|
+
```bibtex
|
|
56
|
+
@article{fu2026scccvgben,
|
|
57
|
+
author = {Fu, Zeyu and Fu, Jiawei and Chen, Chunlin and Zhang, Keyang and Wang, Junping and Ran, Tianfei and Wang, Song},
|
|
58
|
+
title = {scCCVGBen for benchmarking of single-cell representation learning anchored on a centroid-coupled variational graph attention autoencoder across scRNA-seq and scATAC-seq},
|
|
59
|
+
journal = {Frontiers in Genetics},
|
|
60
|
+
volume = {17},
|
|
61
|
+
pages = {1822168},
|
|
62
|
+
year = {2026},
|
|
63
|
+
doi = {10.3389/fgene.2026.1822168}
|
|
64
|
+
}
|
|
65
|
+
```
|
|
66
|
+
|
|
67
|
+
GitHub's "Cite this repository" button reads the same record from `CITATION.cff`.
|
|
68
|
+
|
|
69
|
+
## License
|
|
70
|
+
|
|
71
|
+
MIT — see [`LICENSE`](LICENSE).
|
|
@@ -0,0 +1,73 @@
|
|
|
1
|
+
[build-system]
|
|
2
|
+
requires = ["setuptools>=68", "wheel"]
|
|
3
|
+
build-backend = "setuptools.build_meta"
|
|
4
|
+
|
|
5
|
+
[project]
|
|
6
|
+
name = "scccvgben"
|
|
7
|
+
version = "0.1.0"
|
|
8
|
+
description = "Single-cell scCCVGBen Benchmark — multi-encoder, multi-graph, baseline comparison on 200 scRNA+scATAC datasets"
|
|
9
|
+
readme = "README.md"
|
|
10
|
+
requires-python = ">=3.10"
|
|
11
|
+
license = {text = "MIT"}
|
|
12
|
+
authors = [{name = "Zeyu Fu"}]
|
|
13
|
+
keywords = ["single-cell", "benchmark", "graph neural network", "variational autoencoder", "scRNA-seq", "scATAC-seq"]
|
|
14
|
+
classifiers = [
|
|
15
|
+
"Development Status :: 4 - Beta",
|
|
16
|
+
"Intended Audience :: Science/Research",
|
|
17
|
+
"License :: OSI Approved :: MIT License",
|
|
18
|
+
"Operating System :: OS Independent",
|
|
19
|
+
"Programming Language :: Python :: 3",
|
|
20
|
+
"Topic :: Scientific/Engineering :: Bio-Informatics",
|
|
21
|
+
]
|
|
22
|
+
dependencies = [
|
|
23
|
+
"torch>=2.0",
|
|
24
|
+
"torch-geometric",
|
|
25
|
+
"scanpy",
|
|
26
|
+
"anndata",
|
|
27
|
+
"numpy",
|
|
28
|
+
"pandas",
|
|
29
|
+
"scipy",
|
|
30
|
+
"scikit-learn",
|
|
31
|
+
"matplotlib",
|
|
32
|
+
"gseapy",
|
|
33
|
+
"networkx",
|
|
34
|
+
"pyyaml",
|
|
35
|
+
"umap-learn",
|
|
36
|
+
"seaborn",
|
|
37
|
+
"psutil",
|
|
38
|
+
"tqdm",
|
|
39
|
+
]
|
|
40
|
+
|
|
41
|
+
[project.optional-dependencies]
|
|
42
|
+
atac = ["episcanpy"]
|
|
43
|
+
geo = ["GEOparse>=2.0.3"]
|
|
44
|
+
dev = ["pytest", "ruff", "mypy"]
|
|
45
|
+
sparse = ["torch-sparse"]
|
|
46
|
+
full = ["episcanpy", "GEOparse>=2.0.3", "torch-sparse", "scib", "scvi-tools", "torchdiffeq"]
|
|
47
|
+
|
|
48
|
+
[project.urls]
|
|
49
|
+
Homepage = "https://github.com/PeterPonyu/scCCVGBen"
|
|
50
|
+
Repository = "https://github.com/PeterPonyu/scCCVGBen"
|
|
51
|
+
Documentation = "https://peterponyu.github.io/scCCVGBen/"
|
|
52
|
+
Paper = "https://doi.org/10.3389/fgene.2026.1822168"
|
|
53
|
+
|
|
54
|
+
[tool.setuptools.packages.find]
|
|
55
|
+
where = ["."]
|
|
56
|
+
include = ["scccvgben*"]
|
|
57
|
+
exclude = ["tests*", "scripts*", "site*", "workspace*"]
|
|
58
|
+
|
|
59
|
+
[tool.setuptools.package-data]
|
|
60
|
+
scccvgben = ["data/*.csv"]
|
|
61
|
+
|
|
62
|
+
[tool.ruff]
|
|
63
|
+
line-length = 100
|
|
64
|
+
target-version = "py310"
|
|
65
|
+
|
|
66
|
+
[tool.pytest.ini_options]
|
|
67
|
+
markers = [
|
|
68
|
+
"integration: end-to-end pipeline smoke tests (skipped by default, use -m integration to run)",
|
|
69
|
+
]
|
|
70
|
+
|
|
71
|
+
[tool.mypy]
|
|
72
|
+
python_version = "3.10"
|
|
73
|
+
ignore_missing_imports = true
|
|
@@ -0,0 +1,250 @@
|
|
|
1
|
+
"""deep_methods.py — Deep generative baselines with ELBO variants.
|
|
2
|
+
|
|
3
|
+
Each method follows the convention:
|
|
4
|
+
run_<method>(X: np.ndarray, n_components: int = 10) -> np.ndarray
|
|
5
|
+
|
|
6
|
+
Returns z_mean (latent mean) of shape (n_cells, n_components).
|
|
7
|
+
All tiny VAEs: encoder MLP → (mean, logvar) → decoder MLP.
|
|
8
|
+
Training: 50 epochs, Adam lr=1e-4, batch size = min(256, N).
|
|
9
|
+
"""
|
|
10
|
+
|
|
11
|
+
from __future__ import annotations
|
|
12
|
+
|
|
13
|
+
import logging
|
|
14
|
+
|
|
15
|
+
import numpy as np
|
|
16
|
+
import torch
|
|
17
|
+
import torch.nn as nn
|
|
18
|
+
import torch.nn.functional as F
|
|
19
|
+
from torch.utils.data import DataLoader, TensorDataset
|
|
20
|
+
|
|
21
|
+
log = logging.getLogger(__name__)
|
|
22
|
+
|
|
23
|
+
_EPOCHS = 200
|
|
24
|
+
_LR = 1e-4
|
|
25
|
+
_LATENT_DIM = 10
|
|
26
|
+
_HIDDEN = 128
|
|
27
|
+
|
|
28
|
+
|
|
29
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
30
|
+
# Shared building blocks
|
|
31
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
32
|
+
|
|
33
|
+
class _BaseVAE(nn.Module):
|
|
34
|
+
"""Minimal VAE backbone shared by all deep baselines."""
|
|
35
|
+
|
|
36
|
+
def __init__(self, input_dim: int, hidden: int, latent_dim: int) -> None:
|
|
37
|
+
super().__init__()
|
|
38
|
+
self.encoder_fc = nn.Sequential(
|
|
39
|
+
nn.Linear(input_dim, hidden), nn.ReLU(),
|
|
40
|
+
nn.Linear(hidden, hidden), nn.ReLU(),
|
|
41
|
+
)
|
|
42
|
+
self.fc_mean = nn.Linear(hidden, latent_dim)
|
|
43
|
+
self.fc_logvar = nn.Linear(hidden, latent_dim)
|
|
44
|
+
self.decoder = nn.Sequential(
|
|
45
|
+
nn.Linear(latent_dim, hidden), nn.ReLU(),
|
|
46
|
+
nn.Linear(hidden, input_dim),
|
|
47
|
+
)
|
|
48
|
+
|
|
49
|
+
def encode(self, x: torch.Tensor):
|
|
50
|
+
h = self.encoder_fc(x)
|
|
51
|
+
return self.fc_mean(h), self.fc_logvar(h)
|
|
52
|
+
|
|
53
|
+
def reparameterise(self, mean: torch.Tensor, logvar: torch.Tensor) -> torch.Tensor:
|
|
54
|
+
std = torch.exp(0.5 * logvar)
|
|
55
|
+
return mean + std * torch.randn_like(std)
|
|
56
|
+
|
|
57
|
+
def decode(self, z: torch.Tensor) -> torch.Tensor:
|
|
58
|
+
return self.decoder(z)
|
|
59
|
+
|
|
60
|
+
def forward(self, x):
|
|
61
|
+
mean, logvar = self.encode(x)
|
|
62
|
+
z = self.reparameterise(mean, logvar)
|
|
63
|
+
x_hat = self.decode(z)
|
|
64
|
+
return x_hat, mean, logvar
|
|
65
|
+
|
|
66
|
+
|
|
67
|
+
def _standard_elbo(x, x_hat, mean, logvar):
|
|
68
|
+
recon = F.mse_loss(x_hat, x, reduction="sum")
|
|
69
|
+
kl = -0.5 * torch.sum(1 + logvar - mean.pow(2) - logvar.exp())
|
|
70
|
+
return recon + kl
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
def _train_vae(
|
|
74
|
+
model: _BaseVAE,
|
|
75
|
+
X: np.ndarray,
|
|
76
|
+
loss_fn,
|
|
77
|
+
epochs: int = _EPOCHS,
|
|
78
|
+
lr: float = _LR,
|
|
79
|
+
) -> np.ndarray:
|
|
80
|
+
"""Generic train loop; returns z_mean for all cells."""
|
|
81
|
+
torch.manual_seed(42)
|
|
82
|
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
83
|
+
model = model.to(device)
|
|
84
|
+
|
|
85
|
+
Xt = torch.tensor(X, dtype=torch.float32)
|
|
86
|
+
bs = min(256, len(Xt))
|
|
87
|
+
loader = DataLoader(TensorDataset(Xt), batch_size=bs, shuffle=True)
|
|
88
|
+
|
|
89
|
+
opt = torch.optim.Adam(model.parameters(), lr=lr)
|
|
90
|
+
model.train()
|
|
91
|
+
for _ in range(epochs):
|
|
92
|
+
for (batch,) in loader:
|
|
93
|
+
batch = batch.to(device)
|
|
94
|
+
x_hat, mean, logvar = model(batch)
|
|
95
|
+
loss = loss_fn(batch, x_hat, mean, logvar)
|
|
96
|
+
opt.zero_grad()
|
|
97
|
+
loss.backward()
|
|
98
|
+
opt.step()
|
|
99
|
+
|
|
100
|
+
model.eval()
|
|
101
|
+
with torch.no_grad():
|
|
102
|
+
mean_all, _ = model.encode(Xt.to(device))
|
|
103
|
+
return mean_all.cpu().numpy()
|
|
104
|
+
|
|
105
|
+
|
|
106
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
107
|
+
# DIPVAE (Disentangled Inferred Prior)
|
|
108
|
+
# ELBO + lambda * (off-diagonal covariance terms of E[z] penalised)
|
|
109
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
110
|
+
|
|
111
|
+
def run_DIPVAE(X: np.ndarray, n_components: int = _LATENT_DIM) -> np.ndarray:
|
|
112
|
+
"""DIPVAE: standard ELBO + DIP-II covariance penalty (lambda=10)."""
|
|
113
|
+
lam = 10.0
|
|
114
|
+
|
|
115
|
+
def loss_fn(x, x_hat, mean, logvar):
|
|
116
|
+
base = _standard_elbo(x, x_hat, mean, logvar)
|
|
117
|
+
# DIP-II: penalise off-diagonal entries of Cov[E[z]]
|
|
118
|
+
cov = torch.mm(mean.T, mean) / mean.shape[0]
|
|
119
|
+
off_diag = cov - torch.diag(torch.diag(cov))
|
|
120
|
+
dip = lam * off_diag.pow(2).sum()
|
|
121
|
+
return base + dip
|
|
122
|
+
|
|
123
|
+
model = _BaseVAE(X.shape[1], _HIDDEN, n_components)
|
|
124
|
+
return _train_vae(model, X, loss_fn)
|
|
125
|
+
|
|
126
|
+
|
|
127
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
128
|
+
# InfoVAE (Mutual Information VAE / MMD-VAE)
|
|
129
|
+
# ELBO + alpha * MMD(q(z) || p(z))
|
|
130
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
131
|
+
|
|
132
|
+
def _rbf_kernel(x: torch.Tensor, y: torch.Tensor, sigma: float = 1.0) -> torch.Tensor:
|
|
133
|
+
dist = torch.cdist(x, y).pow(2)
|
|
134
|
+
return torch.exp(-dist / (2 * sigma ** 2))
|
|
135
|
+
|
|
136
|
+
|
|
137
|
+
def _mmd(z: torch.Tensor) -> torch.Tensor:
|
|
138
|
+
n = z.shape[0]
|
|
139
|
+
z_prior = torch.randn_like(z)
|
|
140
|
+
kzz = _rbf_kernel(z, z).sum() / (n * n)
|
|
141
|
+
kpp = _rbf_kernel(z_prior, z_prior).sum() / (n * n)
|
|
142
|
+
kzp = _rbf_kernel(z, z_prior).sum() / (n * n)
|
|
143
|
+
return kzz + kpp - 2 * kzp
|
|
144
|
+
|
|
145
|
+
|
|
146
|
+
def run_InfoVAE(X: np.ndarray, n_components: int = _LATENT_DIM) -> np.ndarray:
|
|
147
|
+
"""InfoVAE (MMD-VAE): ELBO + 100 * MMD(q(z)||N(0,I))."""
|
|
148
|
+
alpha = 100.0
|
|
149
|
+
|
|
150
|
+
def loss_fn(x, x_hat, mean, logvar):
|
|
151
|
+
z = mean + torch.exp(0.5 * logvar) * torch.randn_like(mean)
|
|
152
|
+
return _standard_elbo(x, x_hat, mean, logvar) + alpha * _mmd(z)
|
|
153
|
+
|
|
154
|
+
model = _BaseVAE(X.shape[1], _HIDDEN, n_components)
|
|
155
|
+
return _train_vae(model, X, loss_fn)
|
|
156
|
+
|
|
157
|
+
|
|
158
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
159
|
+
# TCVAE (Total Correlation VAE)
|
|
160
|
+
# ELBO - beta * TC(q(z)) using minibatch-weighted estimator
|
|
161
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
162
|
+
|
|
163
|
+
def run_TCVAE(X: np.ndarray, n_components: int = _LATENT_DIM) -> np.ndarray:
|
|
164
|
+
"""TCVAE: decomposes ELBO; penalises total-correlation term (beta=6)."""
|
|
165
|
+
beta = 6.0
|
|
166
|
+
N = float(len(X))
|
|
167
|
+
|
|
168
|
+
def loss_fn(x, x_hat, mean, logvar):
|
|
169
|
+
recon = F.mse_loss(x_hat, x, reduction="sum")
|
|
170
|
+
# log q(z|x) per sample
|
|
171
|
+
log_q_zx = -0.5 * (logvar + (mean - mean).pow(2) / logvar.exp() + np.log(2 * np.pi)).sum(-1)
|
|
172
|
+
# log p(z)
|
|
173
|
+
log_pz = -0.5 * (mean.pow(2) + logvar.exp() + np.log(2 * np.pi)).sum(-1)
|
|
174
|
+
# minibatch TC approximation (subtract aggregate)
|
|
175
|
+
_z = mean + torch.exp(0.5 * logvar) * torch.randn_like(mean)
|
|
176
|
+
log_prod_qz = -0.5 * (
|
|
177
|
+
_z.unsqueeze(0) - mean.unsqueeze(1)
|
|
178
|
+
).pow(2).div(logvar.exp().unsqueeze(1)).sum(-1).logsumexp(0) + np.log(N)
|
|
179
|
+
tc = (log_q_zx - log_prod_qz - log_pz).mean()
|
|
180
|
+
kl = -0.5 * (1 + logvar - mean.pow(2) - logvar.exp()).sum()
|
|
181
|
+
return recon + kl + (beta - 1) * tc
|
|
182
|
+
|
|
183
|
+
model = _BaseVAE(X.shape[1], _HIDDEN, n_components)
|
|
184
|
+
return _train_vae(model, X, loss_fn)
|
|
185
|
+
|
|
186
|
+
|
|
187
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
188
|
+
# HighBetaVAE (beta-VAE with large beta)
|
|
189
|
+
# ELBO with beta=10 KL weight
|
|
190
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
191
|
+
|
|
192
|
+
def run_HighBetaVAE(X: np.ndarray, n_components: int = _LATENT_DIM) -> np.ndarray:
|
|
193
|
+
"""beta-VAE with beta=10: over-penalises KL to encourage disentanglement."""
|
|
194
|
+
beta = 10.0
|
|
195
|
+
|
|
196
|
+
def loss_fn(x, x_hat, mean, logvar):
|
|
197
|
+
recon = F.mse_loss(x_hat, x, reduction="sum")
|
|
198
|
+
kl = -0.5 * torch.sum(1 + logvar - mean.pow(2) - logvar.exp())
|
|
199
|
+
return recon + beta * kl
|
|
200
|
+
|
|
201
|
+
model = _BaseVAE(X.shape[1], _HIDDEN, n_components)
|
|
202
|
+
return _train_vae(model, X, loss_fn)
|
|
203
|
+
|
|
204
|
+
|
|
205
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
206
|
+
# scVI wrapper
|
|
207
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
208
|
+
|
|
209
|
+
def run_scVI(X: np.ndarray, n_components: int = _LATENT_DIM) -> np.ndarray:
|
|
210
|
+
"""scVI latent via scvi-tools (requires scvi-tools installed).
|
|
211
|
+
|
|
212
|
+
X should be raw integer count matrix (n_cells × n_genes).
|
|
213
|
+
Returns z_mean of shape (n_cells, n_components).
|
|
214
|
+
"""
|
|
215
|
+
try:
|
|
216
|
+
import scvi
|
|
217
|
+
import anndata as ad
|
|
218
|
+
except ImportError as exc:
|
|
219
|
+
raise NotImplementedError(
|
|
220
|
+
"scVI requires scvi-tools and anndata. "
|
|
221
|
+
"Install with: pip install scvi-tools anndata. "
|
|
222
|
+
f"Original error: {exc}"
|
|
223
|
+
) from exc
|
|
224
|
+
|
|
225
|
+
import pandas as pd
|
|
226
|
+
|
|
227
|
+
n_cells, n_genes = X.shape
|
|
228
|
+
adata = ad.AnnData(
|
|
229
|
+
X=X.astype(np.float32),
|
|
230
|
+
obs=pd.DataFrame(index=[str(i) for i in range(n_cells)]),
|
|
231
|
+
var=pd.DataFrame(index=[str(j) for j in range(n_genes)]),
|
|
232
|
+
)
|
|
233
|
+
scvi.model.SCVI.setup_anndata(adata)
|
|
234
|
+
vae = scvi.model.SCVI(adata, n_latent=n_components)
|
|
235
|
+
vae.train(max_epochs=_EPOCHS, plan_kwargs={"lr": _LR})
|
|
236
|
+
latent = vae.get_latent_representation()
|
|
237
|
+
return np.array(latent, dtype=np.float32)
|
|
238
|
+
|
|
239
|
+
|
|
240
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
241
|
+
# Dispatch registry
|
|
242
|
+
# ──────────────────────────────────────────────────────────────────────────────
|
|
243
|
+
|
|
244
|
+
DEEP_REGISTRY: dict[str, callable] = {
|
|
245
|
+
"scVI": run_scVI,
|
|
246
|
+
"DIP": run_DIPVAE,
|
|
247
|
+
"INFO": run_InfoVAE,
|
|
248
|
+
"TC": run_TCVAE,
|
|
249
|
+
"highBeta": run_HighBetaVAE,
|
|
250
|
+
}
|