laior 0.6.0__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.
- laior-0.6.0/LICENSE +21 -0
- laior-0.6.0/MANIFEST.in +8 -0
- laior-0.6.0/PKG-INFO +264 -0
- laior-0.6.0/README.md +200 -0
- laior-0.6.0/laior/__init__.py +7 -0
- laior-0.6.0/laior/agent.py +452 -0
- laior-0.6.0/laior/benchmarking.py +338 -0
- laior-0.6.0/laior/distributions/EuclideanNormal/__init__.py +15 -0
- laior-0.6.0/laior/distributions/EuclideanNormal/arguments.py +3 -0
- laior-0.6.0/laior/distributions/EuclideanNormal/distribution.py +26 -0
- laior-0.6.0/laior/distributions/EuclideanNormal/layers.py +35 -0
- laior-0.6.0/laior/distributions/EuclideanNormal/prior.py +18 -0
- laior-0.6.0/laior/distributions/HWNormal/__init__.py +15 -0
- laior-0.6.0/laior/distributions/HWNormal/arguments.py +3 -0
- laior-0.6.0/laior/distributions/HWNormal/distribution.py +56 -0
- laior-0.6.0/laior/distributions/HWNormal/layers.py +44 -0
- laior-0.6.0/laior/distributions/HWNormal/prior.py +24 -0
- laior-0.6.0/laior/distributions/LearnablePGMNormal/__init__.py +24 -0
- laior-0.6.0/laior/distributions/LearnablePGMNormal/arguments.py +4 -0
- laior-0.6.0/laior/distributions/LearnablePGMNormal/distribution.py +81 -0
- laior-0.6.0/laior/distributions/LearnablePGMNormal/layers.py +218 -0
- laior-0.6.0/laior/distributions/LearnablePGMNormal/prior.py +18 -0
- laior-0.6.0/laior/distributions/PGMNormal/__init__.py +24 -0
- laior-0.6.0/laior/distributions/PGMNormal/arguments.py +4 -0
- laior-0.6.0/laior/distributions/PGMNormal/distribution.py +71 -0
- laior-0.6.0/laior/distributions/PGMNormal/layers.py +82 -0
- laior-0.6.0/laior/distributions/PGMNormal/prior.py +18 -0
- laior-0.6.0/laior/distributions/PoincareNormal/__init__.py +16 -0
- laior-0.6.0/laior/distributions/PoincareNormal/arguments.py +4 -0
- laior-0.6.0/laior/distributions/PoincareNormal/ars.py +133 -0
- laior-0.6.0/laior/distributions/PoincareNormal/distribution.py +56 -0
- laior-0.6.0/laior/distributions/PoincareNormal/hyperbolic_radius.py +360 -0
- laior-0.6.0/laior/distributions/PoincareNormal/hyperbolic_uniform.py +43 -0
- laior-0.6.0/laior/distributions/PoincareNormal/layers.py +45 -0
- laior-0.6.0/laior/distributions/PoincareNormal/prior.py +18 -0
- laior-0.6.0/laior/distributions/__init__.py +0 -0
- laior-0.6.0/laior/distributions/utils.py +21 -0
- laior-0.6.0/laior/environment.py +548 -0
- laior-0.6.0/laior/examples/encoder_example.py +58 -0
- laior-0.6.0/laior/interpretation.py +1205 -0
- laior-0.6.0/laior/mixin.py +829 -0
- laior-0.6.0/laior/model.py +412 -0
- laior-0.6.0/laior/module.py +463 -0
- laior-0.6.0/laior/ode_functions.py +313 -0
- laior-0.6.0/laior/unified_models/__init__.py +53 -0
- laior-0.6.0/laior/unified_models/base_model.py +361 -0
- laior-0.6.0/laior/unified_models/cellblast_model.py +349 -0
- laior-0.6.0/laior/unified_models/clear_model.py +258 -0
- laior-0.6.0/laior/unified_models/gmvae_model.py +560 -0
- laior-0.6.0/laior/unified_models/scalex_model.py +352 -0
- laior-0.6.0/laior/unified_models/scdac_model.py +269 -0
- laior-0.6.0/laior/unified_models/scdeepcluster_model.py +300 -0
- laior-0.6.0/laior/unified_models/scdhmap_model.py +454 -0
- laior-0.6.0/laior/unified_models/scdiffusion_model.py +479 -0
- laior-0.6.0/laior/unified_models/scgcc_model.py +374 -0
- laior-0.6.0/laior/unified_models/scgnn_model.py +210 -0
- laior-0.6.0/laior/unified_models/scsmd_model.py +387 -0
- laior-0.6.0/laior/unified_models/sivae_model.py +460 -0
- laior-0.6.0/laior/utils.py +481 -0
- laior-0.6.0/laior.egg-info/PKG-INFO +264 -0
- laior-0.6.0/laior.egg-info/SOURCES.txt +65 -0
- laior-0.6.0/laior.egg-info/dependency_links.txt +1 -0
- laior-0.6.0/laior.egg-info/requires.txt +18 -0
- laior-0.6.0/laior.egg-info/top_level.txt +1 -0
- laior-0.6.0/pyproject.toml +92 -0
- laior-0.6.0/setup.cfg +4 -0
- laior-0.6.0/setup.py +11 -0
laior-0.6.0/LICENSE
ADDED
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2025 Zeyu Fu
|
|
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.
|
laior-0.6.0/MANIFEST.in
ADDED
laior-0.6.0/PKG-INFO
ADDED
|
@@ -0,0 +1,264 @@
|
|
|
1
|
+
Metadata-Version: 2.4
|
|
2
|
+
Name: laior
|
|
3
|
+
Version: 0.6.0
|
|
4
|
+
Summary: LAIOR: Lorentz Attentive Interpretable ODE Regularized VAE for single-cell omics analysis
|
|
5
|
+
Author-email: Zeyu Fu <fuzeyu99@126.com>
|
|
6
|
+
Maintainer-email: Zeyu Fu <fuzeyu99@126.com>
|
|
7
|
+
License: MIT License
|
|
8
|
+
|
|
9
|
+
Copyright (c) 2025 Zeyu Fu
|
|
10
|
+
|
|
11
|
+
Permission is hereby granted, free of charge, to any person obtaining a copy
|
|
12
|
+
of this software and associated documentation files (the "Software"), to deal
|
|
13
|
+
in the Software without restriction, including without limitation the rights
|
|
14
|
+
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
|
|
15
|
+
copies of the Software, and to permit persons to whom the Software is
|
|
16
|
+
furnished to do so, subject to the following conditions:
|
|
17
|
+
|
|
18
|
+
The above copyright notice and this permission notice shall be included in all
|
|
19
|
+
copies or substantial portions of the Software.
|
|
20
|
+
|
|
21
|
+
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
|
|
22
|
+
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
|
|
23
|
+
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
|
|
24
|
+
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
|
|
25
|
+
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
|
|
26
|
+
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
|
|
27
|
+
SOFTWARE.
|
|
28
|
+
|
|
29
|
+
Project-URL: Homepage, https://github.com/PeterPonyu/Liora
|
|
30
|
+
Project-URL: Repository, https://github.com/PeterPonyu/Liora
|
|
31
|
+
Project-URL: Issues, https://github.com/PeterPonyu/Liora/issues
|
|
32
|
+
Project-URL: Documentation, https://github.com/PeterPonyu/Liora#readme
|
|
33
|
+
Keywords: single-cell,rna-seq,scRNA-seq,vae,variational-autoencoder,neural-ode,ode,manifold,lorentz,hyperbolic,transformer,trajectory-inference,bioinformatics
|
|
34
|
+
Classifier: Development Status :: 3 - Alpha
|
|
35
|
+
Classifier: Intended Audience :: Science/Research
|
|
36
|
+
Classifier: Programming Language :: Python :: 3
|
|
37
|
+
Classifier: Programming Language :: Python :: 3.9
|
|
38
|
+
Classifier: Programming Language :: Python :: 3.10
|
|
39
|
+
Classifier: Programming Language :: Python :: 3.11
|
|
40
|
+
Classifier: Programming Language :: Python :: 3.12
|
|
41
|
+
Classifier: Topic :: Scientific/Engineering :: Bio-Informatics
|
|
42
|
+
Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
|
|
43
|
+
Classifier: Topic :: Scientific/Engineering :: Medical Science Apps.
|
|
44
|
+
Requires-Python: >=3.9
|
|
45
|
+
Description-Content-Type: text/markdown
|
|
46
|
+
License-File: LICENSE
|
|
47
|
+
Requires-Dist: torch>=2.1
|
|
48
|
+
Requires-Dist: torchdiffeq>=0.2.3
|
|
49
|
+
Requires-Dist: numpy>=1.23
|
|
50
|
+
Requires-Dist: scipy>=1.9
|
|
51
|
+
Requires-Dist: scikit-learn>=1.1
|
|
52
|
+
Requires-Dist: anndata>=0.9
|
|
53
|
+
Requires-Dist: tqdm>=4.65
|
|
54
|
+
Provides-Extra: test
|
|
55
|
+
Requires-Dist: pytest>=7.0; extra == "test"
|
|
56
|
+
Requires-Dist: pytest-cov>=4.0; extra == "test"
|
|
57
|
+
Provides-Extra: dev
|
|
58
|
+
Requires-Dist: pytest>=7.0; extra == "dev"
|
|
59
|
+
Requires-Dist: pytest-cov>=4.0; extra == "dev"
|
|
60
|
+
Requires-Dist: black>=23.0; extra == "dev"
|
|
61
|
+
Requires-Dist: isort>=5.12; extra == "dev"
|
|
62
|
+
Requires-Dist: flake8>=6.0; extra == "dev"
|
|
63
|
+
Dynamic: license-file
|
|
64
|
+
|
|
65
|
+
# LAIOR: Lorentz Attentive Interpretable ODE Regularized VAE
|
|
66
|
+
|
|
67
|
+
**A deep learning framework for single-cell omics analysis combining geometric manifold learning, latent dynamics, and comprehensive interpretability.**
|
|
68
|
+
|
|
69
|
+
[](https://opensource.org/licenses/MIT)
|
|
70
|
+
[](https://www.python.org/downloads/)
|
|
71
|
+
[](https://pypi.org/project/liora/)
|
|
72
|
+
|
|
73
|
+
LAIOR (Lorentz Attentive Interpretable ODE Regularized VAE) is a PyTorch-based unified framework for single-cell RNA-seq and ATAC-seq analysis (AnnData format) that learns low-dimensional embeddings from **raw count matrices** using count-likelihood objectives (NB/ZINB/Poisson/ZIP). It integrates (1) Lorentz geometric regularization for hierarchical structure, (2) dual-path information bottleneck architecture for coordinated biological programs, (3) Neural ODE regularization for temporal trajectory stability, and (4) transformer-based attention mechanisms for capturing long-range dependencies.
|
|
74
|
+
|
|
75
|
+
## Key Features
|
|
76
|
+
|
|
77
|
+
- **Advanced VAE Architecture**: Dimensionality reduction with count-based likelihoods (NB, ZINB, Poisson, ZIP)
|
|
78
|
+
- **Geometric Manifold Learning**: Lorentz (hyperbolic) or Euclidean regularization for hierarchical structure
|
|
79
|
+
- **Neural ODE Trajectories**: Latent dynamics solved with `torchdiffeq` (CPU by design). Supports ODE function types `legacy`, `time_mlp` (time-conditioned), and `gru`.
|
|
80
|
+
- **Comprehensive Interpretability**: Attribution analysis for Genes → Latents (discriminative) and Latents → Genes (reconstructive) pathways.
|
|
81
|
+
- **Flexible Encoders**: Standard MLP or Transformer-based with self-attention for capturing complex feature dependencies.
|
|
82
|
+
- **Information Bottleneck**: Hierarchical representation with controllable compression via the `i_dim` parameter.
|
|
83
|
+
|
|
84
|
+
## Data Requirements
|
|
85
|
+
|
|
86
|
+
- `adata.layers[layer]` must contain **raw, non-negative integer-like counts** (UMI counts).
|
|
87
|
+
LAIOR checks this heuristically and raises a `ValueError` if the layer looks normalized/log-transformed.
|
|
88
|
+
- LAIOR applies its own `log1p` + clipping / adaptive normalization internally for training.
|
|
89
|
+
|
|
90
|
+
## Installation
|
|
91
|
+
|
|
92
|
+
```bash
|
|
93
|
+
pip install liora
|
|
94
|
+
```
|
|
95
|
+
|
|
96
|
+
Or install from source:
|
|
97
|
+
|
|
98
|
+
```bash
|
|
99
|
+
git clone https://github.com/PeterPonyu/Liora.git
|
|
100
|
+
cd Liora
|
|
101
|
+
pip install -e .
|
|
102
|
+
```
|
|
103
|
+
|
|
104
|
+
## Quick Start
|
|
105
|
+
|
|
106
|
+
### Basic Usage
|
|
107
|
+
|
|
108
|
+
```python
|
|
109
|
+
import scanpy as sc
|
|
110
|
+
from liora import LAIOR
|
|
111
|
+
|
|
112
|
+
# Load your data
|
|
113
|
+
adata = sc.read_h5ad('data.h5ad')
|
|
114
|
+
|
|
115
|
+
# Train with default settings
|
|
116
|
+
model = LAIOR(
|
|
117
|
+
adata,
|
|
118
|
+
layer='counts',
|
|
119
|
+
hidden_dim=128,
|
|
120
|
+
latent_dim=10,
|
|
121
|
+
i_dim=2,
|
|
122
|
+
)
|
|
123
|
+
model.fit(epochs=100)
|
|
124
|
+
|
|
125
|
+
# Extract embeddings
|
|
126
|
+
latent = model.get_latent()
|
|
127
|
+
```
|
|
128
|
+
|
|
129
|
+
### Advanced Configuration
|
|
130
|
+
|
|
131
|
+
```python
|
|
132
|
+
# Transformer encoder + Neural ODE trajectory inference
|
|
133
|
+
model = LAIOR(
|
|
134
|
+
adata,
|
|
135
|
+
layer='counts',
|
|
136
|
+
hidden_dim=128,
|
|
137
|
+
latent_dim=10,
|
|
138
|
+
i_dim=2,
|
|
139
|
+
# Encoder configuration
|
|
140
|
+
encoder_type='transformer',
|
|
141
|
+
attn_embed_dim=64,
|
|
142
|
+
attn_num_heads=4,
|
|
143
|
+
attn_num_layers=2,
|
|
144
|
+
attn_seq_len=32,
|
|
145
|
+
# ODE configuration
|
|
146
|
+
use_ode=True,
|
|
147
|
+
ode_type='time_mlp',
|
|
148
|
+
ode_time_cond='concat',
|
|
149
|
+
ode_hidden_dim=64,
|
|
150
|
+
ode_solver_method='dopri5',
|
|
151
|
+
ode_rtol=1e-5,
|
|
152
|
+
ode_atol=1e-7,
|
|
153
|
+
# Loss weights
|
|
154
|
+
lorentz=5.0,
|
|
155
|
+
beta=1.0,
|
|
156
|
+
)
|
|
157
|
+
model.fit(epochs=200, patience=20)
|
|
158
|
+
|
|
159
|
+
# Extract results
|
|
160
|
+
latent = model.get_latent() # Latent embeddings
|
|
161
|
+
bottleneck = model.get_bottleneck() # Information bottleneck
|
|
162
|
+
pseudotime = model.get_time() # Predicted pseudotime
|
|
163
|
+
transitions = model.get_transition() # Transition matrix
|
|
164
|
+
```
|
|
165
|
+
|
|
166
|
+
> Note: `get_time()` and `get_transition()` require `use_ode=True`.
|
|
167
|
+
|
|
168
|
+
## Configuration Guide
|
|
169
|
+
|
|
170
|
+
### Architecture Parameters
|
|
171
|
+
|
|
172
|
+
| Parameter | Type | Default | Description |
|
|
173
|
+
|-----------|------|---------|-------------|
|
|
174
|
+
| `hidden_dim` | int | 128 | Hidden layer dimension |
|
|
175
|
+
| `latent_dim` | int | 10 | Primary latent space size |
|
|
176
|
+
| `i_dim` | int | 2 | Information bottleneck dimension |
|
|
177
|
+
|
|
178
|
+
### Encoder Options
|
|
179
|
+
|
|
180
|
+
| Parameter | Type | Default | Description |
|
|
181
|
+
|-----------|------|---------|-------------|
|
|
182
|
+
| `encoder_type` | str | `'mlp'` | `'mlp'` or `'transformer'` |
|
|
183
|
+
| `attn_embed_dim` | int | 64 | Transformer embedding dimension |
|
|
184
|
+
| `attn_num_heads` | int | 4 | Number of attention heads |
|
|
185
|
+
| `attn_num_layers` | int | 2 | Transformer encoder layers |
|
|
186
|
+
| `attn_seq_len` | int | 32 | Token sequence length |
|
|
187
|
+
|
|
188
|
+
### ODE Configuration
|
|
189
|
+
|
|
190
|
+
| Parameter | Type | Default | Description |
|
|
191
|
+
|-----------|------|---------|-------------|
|
|
192
|
+
| `use_ode` | bool | False | Enable Neural ODE |
|
|
193
|
+
| `ode_type` | str | `'time_mlp'` | `'legacy'`, `'time_mlp'`, or `'gru'` |
|
|
194
|
+
| `ode_time_cond` | str | `'concat'` | `'concat'`, `'film'`, or `'add'` |
|
|
195
|
+
| `ode_hidden_dim` | int | None | ODE network hidden size |
|
|
196
|
+
| `ode_solver_method` | str | `'rk4'` | Solver: `'rk4'`, `'dopri5'`, `'adams'`, etc. |
|
|
197
|
+
| `ode_step_size` | float | None | Fixed-step size or `'auto'` |
|
|
198
|
+
| `ode_rtol` | float | None | Relative tolerance (adaptive solvers) |
|
|
199
|
+
| `ode_atol` | float | None | Absolute tolerance (adaptive solvers) |
|
|
200
|
+
|
|
201
|
+
### Loss Configuration
|
|
202
|
+
|
|
203
|
+
| Parameter | Type | Default | Description |
|
|
204
|
+
|-----------|------|---------|-------------|
|
|
205
|
+
| `recon` | float | 1.0 | Reconstruction loss weight |
|
|
206
|
+
| `irecon` | float | 0.0 | Bottleneck reconstruction weight |
|
|
207
|
+
| `lorentz` | float | 0.0 | Manifold regularization weight |
|
|
208
|
+
| `beta` | float | 1.0 | KL divergence weight (β-VAE) |
|
|
209
|
+
| `dip` | float | 0.0 | DIP-VAE loss weight |
|
|
210
|
+
| `tc` | float | 0.0 | Total Correlation loss weight |
|
|
211
|
+
| `info` | float | 0.0 | MMD loss weight |
|
|
212
|
+
| `loss_type` | str | `'nb'` | `'nb'`, `'zinb'`, `'poisson'`, or `'zip'` |
|
|
213
|
+
|
|
214
|
+
### Training Parameters
|
|
215
|
+
|
|
216
|
+
| Parameter | Type | Default | Description |
|
|
217
|
+
|-----------|------|---------|-------------|
|
|
218
|
+
| `lr` | float | 1e-4 | Learning rate |
|
|
219
|
+
| `batch_size` | int | 128 | Mini-batch size |
|
|
220
|
+
| `train_size` | float | 0.7 | Training set proportion |
|
|
221
|
+
| `val_size` | float | 0.15 | Validation set proportion |
|
|
222
|
+
| `test_size` | float | 0.15 | Test set proportion |
|
|
223
|
+
|
|
224
|
+
## Methods
|
|
225
|
+
|
|
226
|
+
### Training
|
|
227
|
+
```python
|
|
228
|
+
model.fit(epochs=400, patience=25, val_every=5, early_stop=True)
|
|
229
|
+
```
|
|
230
|
+
|
|
231
|
+
### Extraction
|
|
232
|
+
```python
|
|
233
|
+
latent = model.get_latent() # Latent representations
|
|
234
|
+
bottleneck = model.get_bottleneck() # Information bottleneck embeddings
|
|
235
|
+
pseudotime = model.get_time() # Pseudotime (ODE mode)
|
|
236
|
+
transitions = model.get_transition() # Transition probabilities (ODE mode)
|
|
237
|
+
```
|
|
238
|
+
|
|
239
|
+
## Citation
|
|
240
|
+
|
|
241
|
+
If you use Liora in your research, please cite:
|
|
242
|
+
|
|
243
|
+
```bibtex
|
|
244
|
+
@software{laior2025,
|
|
245
|
+
title = {LAIOR: Lorentz Attentive Interpretable ODE Regularized VAE},
|
|
246
|
+
author = {Zeyu Fu and Jiawei Fu and Chunlin Chen and Keyang Zhang and Junping Wang and Song Wang},
|
|
247
|
+
year = {2025},
|
|
248
|
+
url = {https://github.com/PeterPonyu/Liora}
|
|
249
|
+
}
|
|
250
|
+
```
|
|
251
|
+
|
|
252
|
+
## License
|
|
253
|
+
|
|
254
|
+
MIT License - see [LICENSE](LICENSE) file for details.
|
|
255
|
+
|
|
256
|
+
## Contributing
|
|
257
|
+
|
|
258
|
+
Contributions are welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
|
|
259
|
+
|
|
260
|
+
## Support
|
|
261
|
+
|
|
262
|
+
- **Issues**: [GitHub Issues](https://github.com/PeterPonyu/Liora/issues)
|
|
263
|
+
- **Email**: fuzeyu99@126.com
|
|
264
|
+
- **Documentation**: [GitHub Repository](https://github.com/PeterPonyu/Liora)
|
laior-0.6.0/README.md
ADDED
|
@@ -0,0 +1,200 @@
|
|
|
1
|
+
# LAIOR: Lorentz Attentive Interpretable ODE Regularized VAE
|
|
2
|
+
|
|
3
|
+
**A deep learning framework for single-cell omics analysis combining geometric manifold learning, latent dynamics, and comprehensive interpretability.**
|
|
4
|
+
|
|
5
|
+
[](https://opensource.org/licenses/MIT)
|
|
6
|
+
[](https://www.python.org/downloads/)
|
|
7
|
+
[](https://pypi.org/project/liora/)
|
|
8
|
+
|
|
9
|
+
LAIOR (Lorentz Attentive Interpretable ODE Regularized VAE) is a PyTorch-based unified framework for single-cell RNA-seq and ATAC-seq analysis (AnnData format) that learns low-dimensional embeddings from **raw count matrices** using count-likelihood objectives (NB/ZINB/Poisson/ZIP). It integrates (1) Lorentz geometric regularization for hierarchical structure, (2) dual-path information bottleneck architecture for coordinated biological programs, (3) Neural ODE regularization for temporal trajectory stability, and (4) transformer-based attention mechanisms for capturing long-range dependencies.
|
|
10
|
+
|
|
11
|
+
## Key Features
|
|
12
|
+
|
|
13
|
+
- **Advanced VAE Architecture**: Dimensionality reduction with count-based likelihoods (NB, ZINB, Poisson, ZIP)
|
|
14
|
+
- **Geometric Manifold Learning**: Lorentz (hyperbolic) or Euclidean regularization for hierarchical structure
|
|
15
|
+
- **Neural ODE Trajectories**: Latent dynamics solved with `torchdiffeq` (CPU by design). Supports ODE function types `legacy`, `time_mlp` (time-conditioned), and `gru`.
|
|
16
|
+
- **Comprehensive Interpretability**: Attribution analysis for Genes → Latents (discriminative) and Latents → Genes (reconstructive) pathways.
|
|
17
|
+
- **Flexible Encoders**: Standard MLP or Transformer-based with self-attention for capturing complex feature dependencies.
|
|
18
|
+
- **Information Bottleneck**: Hierarchical representation with controllable compression via the `i_dim` parameter.
|
|
19
|
+
|
|
20
|
+
## Data Requirements
|
|
21
|
+
|
|
22
|
+
- `adata.layers[layer]` must contain **raw, non-negative integer-like counts** (UMI counts).
|
|
23
|
+
LAIOR checks this heuristically and raises a `ValueError` if the layer looks normalized/log-transformed.
|
|
24
|
+
- LAIOR applies its own `log1p` + clipping / adaptive normalization internally for training.
|
|
25
|
+
|
|
26
|
+
## Installation
|
|
27
|
+
|
|
28
|
+
```bash
|
|
29
|
+
pip install liora
|
|
30
|
+
```
|
|
31
|
+
|
|
32
|
+
Or install from source:
|
|
33
|
+
|
|
34
|
+
```bash
|
|
35
|
+
git clone https://github.com/PeterPonyu/Liora.git
|
|
36
|
+
cd Liora
|
|
37
|
+
pip install -e .
|
|
38
|
+
```
|
|
39
|
+
|
|
40
|
+
## Quick Start
|
|
41
|
+
|
|
42
|
+
### Basic Usage
|
|
43
|
+
|
|
44
|
+
```python
|
|
45
|
+
import scanpy as sc
|
|
46
|
+
from liora import LAIOR
|
|
47
|
+
|
|
48
|
+
# Load your data
|
|
49
|
+
adata = sc.read_h5ad('data.h5ad')
|
|
50
|
+
|
|
51
|
+
# Train with default settings
|
|
52
|
+
model = LAIOR(
|
|
53
|
+
adata,
|
|
54
|
+
layer='counts',
|
|
55
|
+
hidden_dim=128,
|
|
56
|
+
latent_dim=10,
|
|
57
|
+
i_dim=2,
|
|
58
|
+
)
|
|
59
|
+
model.fit(epochs=100)
|
|
60
|
+
|
|
61
|
+
# Extract embeddings
|
|
62
|
+
latent = model.get_latent()
|
|
63
|
+
```
|
|
64
|
+
|
|
65
|
+
### Advanced Configuration
|
|
66
|
+
|
|
67
|
+
```python
|
|
68
|
+
# Transformer encoder + Neural ODE trajectory inference
|
|
69
|
+
model = LAIOR(
|
|
70
|
+
adata,
|
|
71
|
+
layer='counts',
|
|
72
|
+
hidden_dim=128,
|
|
73
|
+
latent_dim=10,
|
|
74
|
+
i_dim=2,
|
|
75
|
+
# Encoder configuration
|
|
76
|
+
encoder_type='transformer',
|
|
77
|
+
attn_embed_dim=64,
|
|
78
|
+
attn_num_heads=4,
|
|
79
|
+
attn_num_layers=2,
|
|
80
|
+
attn_seq_len=32,
|
|
81
|
+
# ODE configuration
|
|
82
|
+
use_ode=True,
|
|
83
|
+
ode_type='time_mlp',
|
|
84
|
+
ode_time_cond='concat',
|
|
85
|
+
ode_hidden_dim=64,
|
|
86
|
+
ode_solver_method='dopri5',
|
|
87
|
+
ode_rtol=1e-5,
|
|
88
|
+
ode_atol=1e-7,
|
|
89
|
+
# Loss weights
|
|
90
|
+
lorentz=5.0,
|
|
91
|
+
beta=1.0,
|
|
92
|
+
)
|
|
93
|
+
model.fit(epochs=200, patience=20)
|
|
94
|
+
|
|
95
|
+
# Extract results
|
|
96
|
+
latent = model.get_latent() # Latent embeddings
|
|
97
|
+
bottleneck = model.get_bottleneck() # Information bottleneck
|
|
98
|
+
pseudotime = model.get_time() # Predicted pseudotime
|
|
99
|
+
transitions = model.get_transition() # Transition matrix
|
|
100
|
+
```
|
|
101
|
+
|
|
102
|
+
> Note: `get_time()` and `get_transition()` require `use_ode=True`.
|
|
103
|
+
|
|
104
|
+
## Configuration Guide
|
|
105
|
+
|
|
106
|
+
### Architecture Parameters
|
|
107
|
+
|
|
108
|
+
| Parameter | Type | Default | Description |
|
|
109
|
+
|-----------|------|---------|-------------|
|
|
110
|
+
| `hidden_dim` | int | 128 | Hidden layer dimension |
|
|
111
|
+
| `latent_dim` | int | 10 | Primary latent space size |
|
|
112
|
+
| `i_dim` | int | 2 | Information bottleneck dimension |
|
|
113
|
+
|
|
114
|
+
### Encoder Options
|
|
115
|
+
|
|
116
|
+
| Parameter | Type | Default | Description |
|
|
117
|
+
|-----------|------|---------|-------------|
|
|
118
|
+
| `encoder_type` | str | `'mlp'` | `'mlp'` or `'transformer'` |
|
|
119
|
+
| `attn_embed_dim` | int | 64 | Transformer embedding dimension |
|
|
120
|
+
| `attn_num_heads` | int | 4 | Number of attention heads |
|
|
121
|
+
| `attn_num_layers` | int | 2 | Transformer encoder layers |
|
|
122
|
+
| `attn_seq_len` | int | 32 | Token sequence length |
|
|
123
|
+
|
|
124
|
+
### ODE Configuration
|
|
125
|
+
|
|
126
|
+
| Parameter | Type | Default | Description |
|
|
127
|
+
|-----------|------|---------|-------------|
|
|
128
|
+
| `use_ode` | bool | False | Enable Neural ODE |
|
|
129
|
+
| `ode_type` | str | `'time_mlp'` | `'legacy'`, `'time_mlp'`, or `'gru'` |
|
|
130
|
+
| `ode_time_cond` | str | `'concat'` | `'concat'`, `'film'`, or `'add'` |
|
|
131
|
+
| `ode_hidden_dim` | int | None | ODE network hidden size |
|
|
132
|
+
| `ode_solver_method` | str | `'rk4'` | Solver: `'rk4'`, `'dopri5'`, `'adams'`, etc. |
|
|
133
|
+
| `ode_step_size` | float | None | Fixed-step size or `'auto'` |
|
|
134
|
+
| `ode_rtol` | float | None | Relative tolerance (adaptive solvers) |
|
|
135
|
+
| `ode_atol` | float | None | Absolute tolerance (adaptive solvers) |
|
|
136
|
+
|
|
137
|
+
### Loss Configuration
|
|
138
|
+
|
|
139
|
+
| Parameter | Type | Default | Description |
|
|
140
|
+
|-----------|------|---------|-------------|
|
|
141
|
+
| `recon` | float | 1.0 | Reconstruction loss weight |
|
|
142
|
+
| `irecon` | float | 0.0 | Bottleneck reconstruction weight |
|
|
143
|
+
| `lorentz` | float | 0.0 | Manifold regularization weight |
|
|
144
|
+
| `beta` | float | 1.0 | KL divergence weight (β-VAE) |
|
|
145
|
+
| `dip` | float | 0.0 | DIP-VAE loss weight |
|
|
146
|
+
| `tc` | float | 0.0 | Total Correlation loss weight |
|
|
147
|
+
| `info` | float | 0.0 | MMD loss weight |
|
|
148
|
+
| `loss_type` | str | `'nb'` | `'nb'`, `'zinb'`, `'poisson'`, or `'zip'` |
|
|
149
|
+
|
|
150
|
+
### Training Parameters
|
|
151
|
+
|
|
152
|
+
| Parameter | Type | Default | Description |
|
|
153
|
+
|-----------|------|---------|-------------|
|
|
154
|
+
| `lr` | float | 1e-4 | Learning rate |
|
|
155
|
+
| `batch_size` | int | 128 | Mini-batch size |
|
|
156
|
+
| `train_size` | float | 0.7 | Training set proportion |
|
|
157
|
+
| `val_size` | float | 0.15 | Validation set proportion |
|
|
158
|
+
| `test_size` | float | 0.15 | Test set proportion |
|
|
159
|
+
|
|
160
|
+
## Methods
|
|
161
|
+
|
|
162
|
+
### Training
|
|
163
|
+
```python
|
|
164
|
+
model.fit(epochs=400, patience=25, val_every=5, early_stop=True)
|
|
165
|
+
```
|
|
166
|
+
|
|
167
|
+
### Extraction
|
|
168
|
+
```python
|
|
169
|
+
latent = model.get_latent() # Latent representations
|
|
170
|
+
bottleneck = model.get_bottleneck() # Information bottleneck embeddings
|
|
171
|
+
pseudotime = model.get_time() # Pseudotime (ODE mode)
|
|
172
|
+
transitions = model.get_transition() # Transition probabilities (ODE mode)
|
|
173
|
+
```
|
|
174
|
+
|
|
175
|
+
## Citation
|
|
176
|
+
|
|
177
|
+
If you use Liora in your research, please cite:
|
|
178
|
+
|
|
179
|
+
```bibtex
|
|
180
|
+
@software{laior2025,
|
|
181
|
+
title = {LAIOR: Lorentz Attentive Interpretable ODE Regularized VAE},
|
|
182
|
+
author = {Zeyu Fu and Jiawei Fu and Chunlin Chen and Keyang Zhang and Junping Wang and Song Wang},
|
|
183
|
+
year = {2025},
|
|
184
|
+
url = {https://github.com/PeterPonyu/Liora}
|
|
185
|
+
}
|
|
186
|
+
```
|
|
187
|
+
|
|
188
|
+
## License
|
|
189
|
+
|
|
190
|
+
MIT License - see [LICENSE](LICENSE) file for details.
|
|
191
|
+
|
|
192
|
+
## Contributing
|
|
193
|
+
|
|
194
|
+
Contributions are welcome! See [CONTRIBUTING.md](CONTRIBUTING.md) for guidelines.
|
|
195
|
+
|
|
196
|
+
## Support
|
|
197
|
+
|
|
198
|
+
- **Issues**: [GitHub Issues](https://github.com/PeterPonyu/Liora/issues)
|
|
199
|
+
- **Email**: fuzeyu99@126.com
|
|
200
|
+
- **Documentation**: [GitHub Repository](https://github.com/PeterPonyu/Liora)
|