Upload README.md with huggingface_hub

README.md

@@ -31,23 +31,50 @@ Stim8hr 132,000 / Stim48hr 132,000 = **396,000**, across 4 donors (D1–D4). Gen
 
 ## Embedding recipe (frozen; produced upstream, not re-run by the workbench)
 
-- Normalize: `sc.pp.normalize_total(target_sum=1e4)` → `sc.pp.log1p` (log-normalized "CP10k")
+Verified against the authors' notebook `src/2_embedding/ntc_embedding.ipynb`
+(`emdann/GWT_perturbseq_analysis_2025`), which ran **scvi-tools 1.3.0**:
+
 - HVG: `highly_variable_genes(n_top_genes=5000, flavor='seurat_v3')` on raw counts
-- scVI: `n_latent=30, n_layers=2, n_hidden=128, dropout_rate=0.2`, NB likelihood,
-  `batch_key=donor`, `max_epochs=300` (early stopping); `scvi.settings.seed=0`
+  — matches the authors' code exactly.
+- scVI: `SCVI(n_latent=30, n_layers=2, dropout_rate=0.2, gene_likelihood="nb",
+  use_layer_norm="both", use_batch_norm="none", encode_covariates=True)`,
+  `setup_anndata(batch_key="donor_id")`; `n_hidden=128` and `dispersion="gene"`
+  are the scVI defaults (confirmed in the notebook's printed model repr, not passed
+  explicitly). Trained: `train(max_epochs=300, early_stopping=True,
+  early_stopping_patience=45, train_size=0.9, batch_size=1024,
+  limit_train_batches=20)`. The authors set **no scVI seed** — their embedding is
+  not seed-reproducible.
 - Graph/UMAP: `neighbors(n_neighbors=100, use_rep='X_scVI', random_state=42)` → `umap()`
-- Clustering: Leiden `resolution=0.8` → 13 clusters, stored in `obs['L0.8']`
-
-> Parameters are transcribed from the authors' analysis notebook
-> (`emdann/GWT_perturbseq_analysis_2025`) as reported by the depositor; they have
-> **not** been independently re-verified against that notebook. "Paper-exact" is a
-> depositor claim, not a validated reproduction.
+  — matches exactly.
+- Normalization: the authors apply plain `normalize_total()` (**default median
+  target_sum, not 1e4**) then `log1p`, and do so **after** the scVI/UMAP step
+  (scVI is trained on raw counts); the log-normalized matrix is what lands in
+  stored `.X`.
+
+**Divergences in spot's own reproduction / description** (not in the authors' code):
+
+- `scvi.settings.seed=0` is set by spot's `run_scvi_embedding.py` for
+  determinism; the authors' notebook has no seed.
+- `normalize_total(target_sum=1e4)` was used by spot; the authors use the default
+  (median) target_sum. If byte-parity with the authors' `.X` matters, drop the 1e4.
+- **Leiden `resolution=0.8` → `obs['L0.8']` is spot's own clustering, not the
+  authors'.** `ntc_embedding.ipynb` performs **no Leiden/Louvain clustering** of
+  the NTC cells at all (the only Leiden in the repo is in
+  `6_functional_interaction`, clustering *perturbations*, not cells). The 13-cluster
+  NTC partition is introduced by spot.
+
+> Bottom line: the scVI architecture/likelihood/HVG/neighbor params are
+> **paper-exact**, but the recipe is **not** a verbatim reproduction — the
+> normalization target_sum differs, spot adds a seed the authors did not set, and
+> the Leiden clustering is spot's addition. Treat the embedding as
+> "paper-parameter scVI, spot-clustered," not "paper-exact reproduction."
 
 ## Contents
 
 - `ntc_clustered.h5ad` — **396,000 × 18,130** (gene-symbol `var_names`).
-  - `X` = **log1p of total-count-normalized counts** (target_sum=1e4). **Raw counts
-    are NOT retained** in the slimmed object.
+  - `X` = **log1p of total-count-normalized counts**. **Raw counts are NOT
+    retained** in the slimmed object. (See Embedding recipe re: target_sum — the
+    authors normalize to the default median; spot's variant used 1e4.)
   - `obs`: `perturbed_gene_name` (all `NTC`), `condition` {Rest, Stim8hr, Stim48hr},
     `donor` {D1–D4}, `donor_id`, `L0.8` (Leiden cluster id).
   - Slimmed to `X + obs + var` only; gzip ~14 GB → ~3.86 GB. **`obsm` is dropped —
@@ -57,8 +84,10 @@ Stim8hr 132,000 / Stim48hr 132,000 = **396,000**, across 4 donors (D1–D4). Gen
 
 **Determinism:** `SEED=12345` in `stage1_pipeline.py` fixes the **downstream
 nomenclature scoring** (`score_genes` panels + permutation-null FDR), i.e. the
-workbench's *calls made from* this object. It does **not** regenerate the h5ad;
-the embedding used different seeds (scVI seed 0, neighbors random_state 42).
+workbench's *calls made from* this object. It does **not** regenerate the h5ad.
+Note the embedding itself is only partially seeded: `neighbors` used
+`random_state=42`, but the authors set no scVI training seed, so the scVI latent
+space (and hence UMAP/clustering) is not byte-reproducible from their code.
 
 ## State labels are the workbench's own
 
@@ -66,16 +95,19 @@ The 5 state programs the workbench assigns (Naive / Activated / Cycling / Memory
 Treg) are **defined by spot**, not by the Marson paper (which defines no
 cell-state taxonomy — its only clustering is of ~3,341 perturbations, not cells).
 Functional calls (Th1/Th2/Th17/Tfh/Treg/CD4-CTL) are scored against the framework
-in: Masopust D. et al. "Guidelines for T cell nomenclature." *Nat Rev Immunol*
-2026;26(4):298–313. doi:10.1038/s41577-025-01238-2.
+in: Masopust D, Awasthi A, Bosselut R, et al. "Guidelines for T cell
+nomenclature." *Nat Rev Immunol* 2026;26(4):298–313.
+doi:10.1038/s41577-025-01238-2 (verified via Crossref).
 
 ## License & attribution
 
-Released under the **MIT License**, mirroring the upstream CZI dataset
-(`license: https://mit-license.org/` in the CZI Croissant metadata for all 12
-AnnData splits, and on the portal dataset page). MIT permits redistribution and
-derivatives (incl. commercial) provided the copyright and permission notice are
-retained. Attribution:
+Released under the **MIT License**, mirroring the upstream CZI dataset. Evidence:
+the CZI Croissant metadata (`metadata/*.jsonld`) for **all 12** AnnData splits
+declares `license: https://mit-license.org/`, `version: 1.0.0`, `datePublished:
+2025-12-22`, and `citeAs: unpublished`, with `url:
+https://virtualcellmodels.cziscience.com/dataset/genome-scale-tcell-perturb-seq`.
+MIT permits redistribution and derivatives (incl. commercial) provided the
+copyright and permission notice are retained. Attribution:
 
 > Data derived from "Primary Human CD4+ T Cell Perturb-seq" (Zhu, Dann, … Marson,
 > 2025), CZI Virtual Cells Platform, MIT License. Cite bioRxiv
@@ -86,14 +118,16 @@ A verbatim `LICENSE` (MIT text + the copyright line above) is included in the re
 
 ## Ethics
 
-Source material is from primary human CD4⁺ T cells (4 consented adult donors,
-leukapheresis). The upstream dataset page states it contains **no personal
-identifying information (PII)**. This NTC subset carries only coded donor IDs
-(D1–D4) and condition — no demographic fields (age/sex/ethnicity/etc. live only in
-the upstream `sample_metadata.suppl_table.csv`, which is not redistributed here).
+Source material is from primary human CD4⁺ T cells (4 adult donors, leukapheresis;
+consent per the upstream study). This NTC subset carries only coded donor IDs
+(D1–D4) and condition — **no demographic fields**. The upstream release does hold
+donor demographics (age/sex/ethnicity/weight/height/smoker/blood-type) in
+`sample_metadata.suppl_table.csv`; that table is **not** redistributed here, and
+should not be, as it is the identifying layer. Confirm the upstream consent/PII
+statement on the CZI dataset page before public release.
 
 ## Versioning & contact
 
 - Card version: v1.1 (corrected 2026-07-10) · Object: spot Stage-1, SEED=12345.
 - Upstream: CZI VCP v1.0.0 (2025-12-22).
-- Contact / issues: Kirit Singh (Hugging Face [@KiritSingh](https://huggingface.co/KiritSingh); issues via [github.com/Kirit-Singh/spot](https://github.com/Kirit-Singh/spot/issues)).
+- Contact / issues: <add maintainer name + email or HF handle>.