Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -21,7 +21,7 @@ Each round — a **loop** — proposes new culture protocols, runs them at the b
|
|
| 21 |
resulting cells. The measurements from each loop are published here as a separate file.
|
| 22 |
|
| 23 |
- **Code and full reproduction pipeline:** <https://github.com/theislab/LabCompass>
|
| 24 |
-
- **Contents:** per-loop measurements in `loops/`, trained models in `checkpoints/`, designed protocols in `solutions/`
|
| 25 |
- **Wet-lab experiments and measurements:** Göttgens Lab
|
| 26 |
- **License:** CC-BY-4.0
|
| 27 |
|
|
@@ -176,6 +176,25 @@ model = FlowMatching.load(path)
|
|
| 176 |
In the reproduction repository these are wired up through the `paths` config group, so pointing a run
|
| 177 |
at a downloaded loop is a matter of overriding the four checkpoint paths.
|
| 178 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 179 |
## Designed protocols (`solutions/`)
|
| 180 |
|
| 181 |
`solutions/` contains **every candidate protocol LabCompass generated**, across all sweeps behind the
|
|
|
|
| 21 |
resulting cells. The measurements from each loop are published here as a separate file.
|
| 22 |
|
| 23 |
- **Code and full reproduction pipeline:** <https://github.com/theislab/LabCompass>
|
| 24 |
+
- **Contents:** per-loop measurements in `loops/`, trained models in `checkpoints/`, derived datasets in `derived/`, designed protocols in `solutions/`
|
| 25 |
- **Wet-lab experiments and measurements:** Göttgens Lab
|
| 26 |
- **License:** CC-BY-4.0
|
| 27 |
|
|
|
|
| 176 |
In the reproduction repository these are wired up through the `paths` config group, so pointing a run
|
| 177 |
at a downloaded loop is a matter of overriding the four checkpoint paths.
|
| 178 |
|
| 179 |
+
## Derived datasets (`derived/`)
|
| 180 |
+
|
| 181 |
+
Processed artefacts the analysis depends on that are **not reconstructible** from `loops/` alone,
|
| 182 |
+
because the annotation and filtering steps that produced them are not part of the pipeline code.
|
| 183 |
+
|
| 184 |
+
| File | Contents | Size |
|
| 185 |
+
| --- | --- | --- |
|
| 186 |
+
| `derived/loop0_annotated_500k.h5ad` | Subsampled, annotated and filtered cells for the 116 designs of the initial screen. | 0.51 GB |
|
| 187 |
+
| `derived/loop1_annotated_500k.h5ad` | The same, extended to the 122 designs available at loop 1. | 0.54 GB |
|
| 188 |
+
| `derived/loop2_reannotated.h5ad` | Loop 2 cells after the cell-type reannotation; carries `cell_type_leiden` and `is_reannot_leiden`. This is the cell-type classifier's training input. | 0.75 GB |
|
| 189 |
+
| `derived/unique_protocols.h5ad` | The 103 unique protocols actually run, across 15 design axes. Used to snap generated designs onto realisable concentrations. | <1 MB |
|
| 190 |
+
|
| 191 |
+
The `*_annotated_500k` files carry cell-type labels (`cell_type`, `leiden`, `leiden_sub`) on a
|
| 192 |
+
subsample of the full measurements, and have had low-quality cells removed — which is why their cell
|
| 193 |
+
counts are lower than the corresponding `loops/` files and their `obs` has more columns.
|
| 194 |
+
|
| 195 |
+
In the reproduction repository these are expected at `project_folder/derived/`, mirroring the layout
|
| 196 |
+
here.
|
| 197 |
+
|
| 198 |
## Designed protocols (`solutions/`)
|
| 199 |
|
| 200 |
`solutions/` contains **every candidate protocol LabCompass generated**, across all sweeps behind the
|