File size: 4,294 Bytes
8aea526
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
---
license: cc-by-4.0
language:
- en
pretty_name: SpatialAgent Human Expert Reference Data
tags:
- biology
- spatial-transcriptomics
- single-cell
- gene-panel-design
- cell-type-annotation
- benchmark
size_categories:
- 100K<n<1M
configs:
- config_name: panel_workflows
  data_files: panel_design/workflows.csv
- config_name: annotation_workflows
  data_files: annotation/workflows.csv
---

# SpatialAgent — Human Expert Reference Data

Anonymized reference data produced by human scientists for two spatial-transcriptomics
tasks used to benchmark **SpatialAgent**:

1. **Gene panel design** — expert-designed targeted gene panels for the human
   **dorsolateral prefrontal cortex (DLPFC / PFC)**.
2. **Cell-type & tissue-niche annotation** — expert annotations of a **developing human
   heart** MERFISH dataset (228,633 cells × 238 genes).

All scientist identities are removed. Each task uses its **own independent numbering**, so
the same person generally has a *different* id in the two tasks (this is intentional — the
two studies were anonymized separately). No real names appear anywhere in this repository.
Each expert's methodology is documented (by anonymized id) in the `workflows.csv` files.

## Repository layout

```
panel_design/
  workflows.csv              # id (1–10) -> free-text description of the panel-design approach
  {1..10}.csv                # one full panel per expert (ranked gene lists)
  split/{id}_top{50,100,150}.csv   # top-N subsets of each panel
annotation/
  workflows.csv              # id (1–8) -> cell-type & niche annotation approach
  combined_annotations_anonymized.h5ad   # all experts (anonymized) + model/baseline predictions
  human_annotations_anonymized.h5ad       # human experts only (anonymized), no model columns
  per_expert_raw/            # the original per-expert annotation files, anonymized
    expert{1,2,5,6,7}.h5ad
    expert{3,4}.csv
    expert7_niche.h5ad
```

See `panel_design/README.md` and `annotation/README.md` for the column-level details of
each subset.

## Panel design (DLPFC)

10 experts each submitted a ranked panel (typically top 50 / 100 / 150 genes) with a short
rationale per gene. Formats are heterogeneous (experts used different tools), so columns
differ between files; the common fields are a gene symbol, a ranking/priority, and a
free-text reasoning column. `split/` holds the top-50/100/150 truncations used for
size-matched evaluation. Workflows range from purely algorithmic (Persist, greedy kNN
reconstruction) to literature-driven marker curation — see `panel_design/workflows.csv`.

## Annotation (developing human heart, MERFISH)

8 experts annotated the same 228,633 cells. The two combined `.h5ad` objects share an
identical cell index and embeddings:

- `X` — log1p-normalized expression (238 genes); `layers['raw_count']` — raw counts.
- `obsm``X_pca`, `X_umap`, `spatial` (tissue coordinates).
- Per-expert columns: `cell_type_tier{1,2,3}_expert{N}`, `tissue_niche_tier{1,2}_expert{N}`,
  and consolidated `cell_type_expert{N}` / `tissue_niche_expert{N}`.
- Consensus reference labels: `cell_type`, `tissue_niche`.

`combined_annotations_anonymized.h5ad` additionally contains model / baseline predictions
(`cell_type_agent`, `tissue_niche_agent`, `cell_type_gpt`, `cell_type_sctab`,
`cell_type_popv`, `cell_type_biomni_run_{1,2,3}`, `cell_type_spatialagent_run_4`) for direct
benchmarking; `human_annotations_anonymized.h5ad` is the human-only subset (those columns
dropped). `per_expert_raw/` preserves each expert's original file (with their native,
heterogeneous column schema) for full transparency.

### Caveats
- **annotation expert 1** did not produce tissue-niche labels (niche fields are empty/NA).
- **annotation expert 3**'s labels are of uncertain origin and are likely mis-ordered — use with care.
- **annotation expert 8** has no standalone raw file; their annotations exist only inside the combined objects.
- **panel expert 3** submitted a previously designed panel for the wrong tissue.

## License & citation

Released under **CC-BY-4.0** (adjust if your venue requires otherwise). If you use this
data, please cite the SpatialAgent paper. The two `workflows.csv` files correspond to the
Extended Data tables describing human-scientist workflows.