---
license: other
pretty_name: Haruka v2 preprocessed and fitted spatial datasets
tags: [spatial-transcriptomics, single-cell, anndata, contrastive-vae]
---
# Haruka v2: preprocessed inputs and fitted outputs for the reproduction notebooks

Ready-to-fit AnnData files for every figure of Haruka v2 (code: https://github.com/C0nc/haruka_v2, branch `v2`, folder `notebooks/`).
Each folder matches one notebook; with these files the notebook can start at the **fit** stage.
`fitted/` holds the same AnnData objects with the paper's default Haruka fit added (z, s, predicted neighbourhoods) for
further analysis; the notebooks unpack them and reuse the fit instead of refitting (`HARUKA_REFIT=1` refits).

```bash
hf download C1vy/haruka_v2_data --repo-type dataset --local-dir haruka_v2_data --exclude "fitted/*"   # inputs (~3.7 GB)
hf download C1vy/haruka_v2_data --repo-type dataset --local-dir haruka_v2_data --include "fitted/*"   # fitted (~5.8 GB)
hf download C1vy/haruka_v2_data 06_fibrosis/gse250346_unaffected_vs_all_fibrosis.h5ad --repo-type dataset --local-dir .
```

| file | size | notebook | source / preprocessing |
|---|---|---|---|
| `02_fig2_benchmark/mouse_brain_A8_paired.h5ad` | 31 MB | `02_fig2_benchmark` | aging coronal mouse-brain MERFISH section paired with a copy carrying 3 planted composition-shift domains (`build_paired_sections.py`, seed 20260815) |
| `02_fig2_benchmark/mouse_brain_A9_paired.h5ad` | 32 MB | `02_fig2_benchmark` | aging coronal mouse-brain MERFISH section paired with a copy carrying 3 planted composition-shift domains (`build_paired_sections.py`, seed 20260815) |
| `02_fig2_benchmark/mouse_brain_B17_paired.h5ad` | 38 MB | `02_fig2_benchmark` | aging coronal mouse-brain MERFISH section paired with a copy carrying 3 planted composition-shift domains (`build_paired_sections.py`, seed 20260815) |
| `02_fig2_benchmark/mouse_brain_B3_paired.h5ad` | 31 MB | `02_fig2_benchmark` | aging coronal mouse-brain MERFISH section paired with a copy carrying 3 planted composition-shift domains (`build_paired_sections.py`, seed 20260815) |
| `02_fig2_benchmark/mouse_brain_B5_paired.h5ad` | 36 MB | `02_fig2_benchmark` | aging coronal mouse-brain MERFISH section paired with a copy carrying 3 planted composition-shift domains (`build_paired_sections.py`, seed 20260815) |
| `03_fig3_ad_plaque/ad_starmap.h5ad` | 56 MB | `03_fig3_ad_plaque` | STARmap PLUS AD mouse, 13 months, control vs disease (`prepare_studies.py ad`) |
| `03_fig3_ad_plaque/codex_wholeslice_all7_no_oligo.h5ad` | 58 MB | `03_fig3_ad_plaque` | human AD CODEX whole slices, 4 AD / 3 control donors, 34 proteins, normalised (`build_wholeslice.py`) |
| `03_fig3_ad_plaque/codex_wholeslice_ctl3VsCtl14.h5ad` | 26 MB | `03_fig3_ad_plaque` | arm-swap control: Ctl-3 relabelled disease vs Ctl-1/4 (`build_ctl3_vs_ctl14.py`) |
| `03_fig3_ad_plaque/codex_abeta_objects.tar.gz` | 2 MB | `03_fig3_ad_plaque` | CODEX-CNS amyloid-beta plaque objects (position, class, intensities) for the 4 AD donors and Ctl-3, from the original release; extracted into the repository |
| `04_exercise/exercise_old_full.h5ad` | 1,064 MB | `04_exercise` | Sun et al. aging/exercise brain MERFISH, old control vs old exercise, 594,896 cells x 300 genes (`prepare_exercise.py`) |
| `05_petracer/petracer_M2-1_lineage.h5ad` | 293 MB | `05_petracer` | PEtracer M2-1: non-tumour control vs malignant cells in tumours, lineage columns for evaluation only (`build_m2_tumour_vs_normal.py` + `build_m2_lineage.py`) |
| `06_fibrosis/gse250346_unaffected_vs_all_fibrosis.h5ad` | 291 MB | `06_fibrosis` | GSE250346 lung Xenium, unaffected vs fibrosis, 1.63 M cells x 343 genes (`export_gse250346_unaffected_vs_fibrosis.R` + `build_h5ad.py`) |
| `06_fibrosis/he_annot_matched.csv.gz` | 22 MB | `06_fibrosis` | GSE250346 H&E region annotations matched to cells (`match_he_annotations.py`) |
| `07_kras/mouselung_kras_xenium_clean_noC3.h5ad` | 1,731 MB | `07_kras` | KRAS inhibitor (MRTX1133) mouse lung Xenium 5K, 920,681 cells; X log-normalised, raw counts in layers["counts"] |

## Fitted AnnData (`fitted/`)

The input AnnData plus the default Haruka v2 fit used in the figures, matched cell by cell:

| key | content |
|---|---|
| `obsm['X_haruka_z']` | background latent z (32-d), shared by control and condition cells |
| `obsm['X_haruka_s']` | salient latent s (32-d). The decoder sets s = 0 for control cells, so analyse s on condition cells only |
| `obsm['haruka_D_z0']`, `obsm['haruka_D_zs']` | predicted composition of the k = 20 spatial neighbours from [z, 0] and [z, s] |
| `obsm['haruka_salient_shift']` | D([z,s]) − D([z,0]): what the condition changed in the cell's neighbourhood |
| `obsm['haruka_null_shift_0..2']` | the same with s shuffled across cells (exercise, KRAS) |
| `obsm['haruka_M_z0']`, `obsm['haruka_M_shift']` | Fig. 2/3 only: predicted per-type neighbour mean expression (16-d SVD per type) and its salient shift |
| `obs['haruka_condition_arm']`, `obs['haruka_in_fit']` | condition-arm cells; cells included in the fit |
| `uns['haruka']` | composition column names (cell types), condition key/value, all fit arguments (JSON) |

```python
import anndata as ad
a = ad.read_h5ad('fitted/07_kras/mouselung_kras_xenium_clean_noC3_haruka_fitted.h5ad')
cond = a.obs['haruka_condition_arm'].to_numpy()
z, s = a.obsm['X_haruka_z'][cond], a.obsm['X_haruka_s'][cond]
shift = a.obsm['haruka_salient_shift'][cond]        # columns: a.uns['haruka']['composition_columns']
```

| file | size | fit |
|---|---|---|
| `fitted/02_fig2_benchmark/mouse_brain_A8_paired_haruka_fitted.h5ad` | 238 MB | Fig. 2 simulation, λ = 10 default fit (Poisson likelihood) |
| `fitted/02_fig2_benchmark/mouse_brain_A9_paired_haruka_fitted.h5ad` | 259 MB | Fig. 2 simulation, λ = 10 default fit (Poisson likelihood) |
| `fitted/02_fig2_benchmark/mouse_brain_B17_paired_haruka_fitted.h5ad` | 302 MB | Fig. 2 simulation, λ = 10 default fit (Poisson likelihood) |
| `fitted/02_fig2_benchmark/mouse_brain_B3_paired_haruka_fitted.h5ad` | 275 MB | Fig. 2 simulation, λ = 10 default fit (Poisson likelihood) |
| `fitted/02_fig2_benchmark/mouse_brain_B5_paired_haruka_fitted.h5ad` | 297 MB | Fig. 2 simulation, λ = 10 default fit (Poisson likelihood) |
| `fitted/03_fig3_ad_plaque/ad_haruka_fitted.h5ad` | 104 MB | Fig. 3 default fit |
| `fitted/03_fig3_ad_plaque/codex_wholeslice_all7_no_oligo_haruka_fitted.h5ad` | 175 MB | Fig. 3 default fit |
| `fitted/03_fig3_ad_plaque/codex_wholeslice_ctl3VsCtl14_haruka_fitted.h5ad` | 70 MB | Fig. 3 default fit |
| `fitted/04_exercise/exercise_old_full_haruka_fitted.h5ad` | 570 MB | exercise default fit (+ 3 shuffled-s nulls) |
| `fitted/05_petracer/petracer_M2-1_lineage_haruka_fitted.h5ad` | 336 MB | PEtracer default fit |
| `fitted/06_fibrosis/gse250346_unaffected_vs_all_fibrosis_haruka_fitted.h5ad` | 1,465 MB | lung fibrosis default fit |
| `fitted/07_kras/mouselung_kras_xenium_clean_noC3_haruka_fitted.h5ad` | 2,434 MB | KRAS default fit (+ 3 shuffled-s nulls) |

Fit time per dataset (one H200; neighbourhood-target construction + 150 epochs): Fig. 2 sections ~3 min each,
STARmap 15 min, CODEX 1.5-3.5 min, PEtracer 10 min, exercise 48 min, KRAS 1.6 h, lung fibrosis 6.8 h.

## Saved baseline results (`baselines/`)

The notebooks reproduce the Haruka results only; the comparison methods are not refitted by default. Their saved
results are extracted into the repository instead (`RUN_BASELINE=1` refits them):

| file | size | contents |
|---|---|---|
| `baselines/02_fig2_benchmark_baselines.tar.gz` | 2 MB | Fig. 2: scVI / ContrastiveVI / expression + CellCharter scores per section, composition controls, CINEMA-OT rows, B5 map cache |
| `baselines/03_fig3_ad_plaque_baselines.tar.gz` | 3 MB | Fig. 3: scVI latent (32-d) for STARmap AD |
| `baselines/05_petracer_baselines.tar.gz` | 10 MB | PEtracer: scVI latent (32-d) |

## SHA-256

```
3cf8c39cf4f02d835017a08a18941ac694b29e3986b9ef72706dbcf0878c197d  02_fig2_benchmark/mouse_brain_A8_paired.h5ad
b79e7e0dfe2aa7fb90d4decd22ce8b44e98fb06423520af9ba56adb6bded4c08  02_fig2_benchmark/mouse_brain_A9_paired.h5ad
da857a27c0ccd9560e3062b4f06ff3fc539ab9defd0f83edcbeea630534109b7  02_fig2_benchmark/mouse_brain_B17_paired.h5ad
6987890c11e2f49e8fd4f91de12249e6664138cdb3b6b969f38e897104b8ec45  02_fig2_benchmark/mouse_brain_B3_paired.h5ad
69e406c53e073cacc58c4c3ebebfaccad8b2455e2af6ce7088f0324737b05a9e  02_fig2_benchmark/mouse_brain_B5_paired.h5ad
2e6bdc50bd99fbf2959bba37caa5ca87685c1855b0c23e2b3fe95f2267fda094  03_fig3_ad_plaque/ad_starmap.h5ad
5a06c6ea693c331858febc1645974a457d77fb1d6e7bc97782b2adb75926e68f  03_fig3_ad_plaque/codex_wholeslice_all7_no_oligo.h5ad
9b9e5bbd203c46e2abd59747b7956c7aefde5b0397d25dec2f6abde78f631fb0  03_fig3_ad_plaque/codex_wholeslice_ctl3VsCtl14.h5ad
a9780e19bfb1fde24d88218d2c9779c6b494fa9b0d00252341fe9a4aa5193b53  04_exercise/exercise_old_full.h5ad
8e9ead69840327e6d2013839557bfda68fa5466d8a11ff761bc392b8147d04c8  05_petracer/petracer_M2-1_lineage.h5ad
d9b0f1f8c67d0d06004efe2772b412a4ce93590cbea75aa2ec21f04485c3b440  06_fibrosis/gse250346_unaffected_vs_all_fibrosis.h5ad
97fb92aaf0beabf41e50bee7aaaec1fc4312c7d7af40e639e3c1d8d8b042b5aa  06_fibrosis/he_annot_matched.csv.gz
bd191fd8d259aebc43fd068cf26759db11341a11eb891915fc4c3b6930166b9f  07_kras/mouselung_kras_xenium_clean_noC3.h5ad
a0ef884f984432f17a654daf3c773d073c4f06e77f64f2195311aba16009d32b  03_fig3_ad_plaque/codex_abeta_objects.tar.gz
96714eb2b38b3b1903e2973a2df784fad6de882dedf8302ecae76c5d013b7a50  fitted/02_fig2_benchmark/mouse_brain_A8_paired_haruka_fitted.h5ad
79851b39b96b064a2823986d92c1cec9e04c557eaa6961876379b657f0e769c5  fitted/02_fig2_benchmark/mouse_brain_A9_paired_haruka_fitted.h5ad
cdb1fc5353578ae858d205102848f5a54af56aade55ad2104af6aef4719a4518  fitted/02_fig2_benchmark/mouse_brain_B17_paired_haruka_fitted.h5ad
f66094a83d363796721758e5791b8b65d0a908ae5b72c354c2cb6f3739a45376  fitted/02_fig2_benchmark/mouse_brain_B3_paired_haruka_fitted.h5ad
92485410ac07fb2d2749b0dc248c1a80d9692be1166648e396a1fbb528b14c17  fitted/02_fig2_benchmark/mouse_brain_B5_paired_haruka_fitted.h5ad
37a1e0ce3801fe9238dcc394ea2f1c3e055296d689c400f8e19f5133c5770bae  fitted/03_fig3_ad_plaque/ad_haruka_fitted.h5ad
8b66da9d44eb57a55eae8c63cb6bad1971d755b3b60102075010454942697d43  fitted/03_fig3_ad_plaque/codex_wholeslice_all7_no_oligo_haruka_fitted.h5ad
cbe90ee1897067760d5189b276a916d35ad3355f4b172decc3a3000ed9e7c429  fitted/03_fig3_ad_plaque/codex_wholeslice_ctl3VsCtl14_haruka_fitted.h5ad
f6914465455282f0ca2d50621c2891d1e5044b2d5636152f43f7e4ed710d662e  fitted/04_exercise/exercise_old_full_haruka_fitted.h5ad
de2b6fef2365ff4b3392d3cb038f7231ee9a7d375df801c10f9d1dd535654af8  fitted/05_petracer/petracer_M2-1_lineage_haruka_fitted.h5ad
bc54a42a3fdb7f383436e3e5a8037262617b0b7d02cacce0ead641fdd982debc  fitted/06_fibrosis/gse250346_unaffected_vs_all_fibrosis_haruka_fitted.h5ad
4912d39272b3ada1281a0b26b93668d7554c814118c5da6a4282c20736b9a2ee  fitted/07_kras/mouselung_kras_xenium_clean_noC3_haruka_fitted.h5ad
d2b58ed6b6454ad467dc64a6aed41cba3693c5af01290be4e6ccccd51955c241  baselines/02_fig2_benchmark_baselines.tar.gz
0a1a2ea2e75f4e5d2d24bdc8cfc195070fcaff1007d07d462030950325080d7b  baselines/03_fig3_ad_plaque_baselines.tar.gz
c686adeab993ad4b9576c1fbde57111a18e3cd0c6a1f62053f47ffe377a9f270  baselines/05_petracer_baselines.tar.gz
```

Licences follow the original data releases (GEO GSE250346, STARmap PLUS, CODEX-CNS, PEtracer, Sun et al.); the KRAS Xenium data are shared by the data owners for reproduction of this work.