| --- |
| license: cc-by-nc-4.0 |
| tags: |
| - biology |
| - single-cell |
| - perturbation |
| - cell-embeddings |
| extra_gated_prompt: >- |
| These embeddings are released for non-commercial academic research. Please |
| tell us who you are and what you plan to use them for. |
| extra_gated_fields: |
| Name: text |
| Affiliation: text |
| Intended use: text |
| I will use these embeddings for non-commercial research only: checkbox |
| --- |
| |
| # BioReasonCell — cell embeddings |
|
|
| Mean-pooled cell-state embeddings for the perturbation datasets used by |
| [BioReasonCell](https://huggingface.co/wanglab). One vector per biological context, consumed by the |
| model as the `<|CELL_START|> … <|CELL_END|>` block. |
|
|
| ## `mean_pool/` |
| |
| | folder | vectors | contexts | |
| |---|---:|---| |
| | `mean_pool/genetic/` | 47 | CRISPRi / CRISPRa screens — Kaufman2025 (6 lines × 5 stimuli), Replogle (K562 gwps, RPE1), Norman, Papalexi, Marson (D1–D4), Xaira (HCT116, HEK293T), Xu2026, HepG2 / Jurkat / iPSC | |
| | `mean_pool/chemical/` | 47 | Tahoe drug screen — one per cell line | |
|
|
| Every file is a **1-D `torch.float32` tensor of shape `(2058,)`**, the per-context mean of STATE |
| (`st-se-replogle-full`) cell embeddings over that context's control cells. 936 KB total. |
|
|
| These are **cell-state** vectors, not perturbation vectors: the same file is used for every |
| perturbation measured in that context, and the perturbation itself reaches the model through the |
| prompt text. The genetic and chemical sets are disjoint in filename, so the two folders can be |
| merged into one directory if you prefer a single source. |
|
|
| ## Filenames are lookup keys — do not rename |
|
|
| A file's name (minus `.pt`) is exactly the key the loader resolves. Given a dataset row's |
| `cell_file`, the key is built by stemming **twice**: |
|
|
| ```python |
| from pathlib import Path |
| key = Path(Path(row["cell_file"]).stem).stem # collate.py stems an already-stemmed column |
| emb = torch.load(f"{cell_file_dir}/{key}.pt") # -> (2058,) float32 |
| ``` |
|
|
| The double stem is deliberate and load-bearing. `tahoe_Panc 03.27_merged.h5ad` resolves to |
| `tahoe_Panc 03.pt`, because `.27_merged` is read as a suffix — so that file is correctly named and |
| renaming it to `tahoe_Panc 03.27_merged.pt` would break the lookup. Names also contain spaces |
| (`tahoe_SW 1088_merged.pt`, `tahoe_AN3 CA_merged.pt`); keep them. |
|
|
| ## Use |
|
|
| ```python |
| from huggingface_hub import snapshot_download |
| |
| local = snapshot_download("wanglab/BioReasonCell-embeddings", repo_type="dataset") |
| # then point the training / eval config at the folder you want: |
| # cell_file_dir: {local}/mean_pool/chemical |
| ``` |
|
|
| The loader (`bioreason_cell/dataset/cell_embedding_store.py`) takes a directory of per-cell `.pt` |
| files and casts each to bfloat16, unsqueezing 1-D tensors to `(1, 2058)`. |
|
|
| ## Coverage |
|
|
| Verified against the sampled batches under the double-stem transform: **47/47** keys present for |
| chemical (`chemical_tahoe_v1.2_50k`, `chemical_tahoe_10k`) and **47/47** for genetic |
| (`genetic_v7.5_50k`, `genetic_v7.6_50k`, `genetic_v7.6_relabel`, `genetic_batch1_5k`), with no |
| missing keys in either. Train and test splits draw from the same context pool, so one folder covers |
| both. |
|
|
| ## Related |
|
|
| - `wanglab/BioReasonCell-ReasoningData` — perturbation rows, gene/pathway/cell annotations |
| - `wanglab/BioReasonCell-ExperimentData` — per-experiment train/validation/test splits |
|
|
