arcinstitute/ST-HVG-Parse

STATE — Parse (X_hvg)

State Transition (ST) models using X_hvg inputs on Parse. This repository contains 4 fewshot and 4 zeroshot canonical models, with their saved configurations, loading files and available CellEval v1 outputs.

Paired representation · All 36 canonical models · STATE code and setup

Parse data notice: Certain uses of a model trained with Parse data may require a license from Parse Biosciences, Inc.

Model index

One row is one trained model. Additional saved checkpoints are snapshots of that run, not additional models. Fewshot uses the original best.ckpt; the replaced zeroshot bundles use final.ckpt.

Regime	Test context(s)	Canonical checkpoint	Saved evaluation
fewshot	B_Intermediate_Memory	fewshot/split_0/checkpoints/best.ckpt	CSV + H5AD bundle
fewshot	B_Naive	fewshot/split_1/checkpoints/best.ckpt	CSV + H5AD bundle
fewshot	CD4_Memory	fewshot/split_3/checkpoints/best.ckpt	CSV + H5AD bundle
fewshot	CD14_Mono	fewshot/split_4/checkpoints/best.ckpt	CSV + H5AD bundle
zeroshot	B_Intermediate_Memory	zeroshot/split_0/checkpoints/final.ckpt	CSV + H5AD bundle
zeroshot	B_Naive	zeroshot/split_1/checkpoints/final.ckpt	CSV + H5AD bundle
zeroshot	CD4_Memory	zeroshot/split_3/checkpoints/final.ckpt	CSV + H5AD bundle
zeroshot	CD14_Mono	zeroshot/split_4/checkpoints/final.ckpt	CSV + H5AD bundle

Directory mapping: split_0 = B Intermediate/Memory, split_1 = B Naive, split_3 = CD4 Memory, split_4 = CD14 Monocytes. These are zero-based Hub names; the corresponding rerun source splits were 1, 2, 4, 5. Plasmablast (split_2) is excluded from both regimes and both representations; remaining directories were not renumbered.

Known gap: X_state CD4 Memory fewshot has a checkpoint and loading files but no saved CellEval/H5AD outputs. No scores are imputed.

X_hvg versus X_state

Values are equal-weight means of context means, using paired perturbations with finite values for both models. DE-count Spearman is computed across perturbations within each context. The tables and figure use the same 3 contexts in both regimes.

CD4 Memory is omitted from both headline rows because of the fewshot evaluation gap. The all-available-context table also reports all four zeroshot contexts.

Fewshot

Metric	X_hvg	X_state
AUPRC	0.5264	0.5659
Pearson delta	0.7956	0.8044
DE-count Spearman	0.8479	0.8784
DE overlap at N	0.4261	0.4608
LFC Spearman	0.4474	0.4289
PDS L1	0.9306	0.9584

Zeroshot

Metric	X_hvg	X_state
AUPRC	0.3634	0.5257
Pearson delta	0.4074	0.7647
DE-count Spearman	0.6548	0.8376
DE overlap at N	0.2098	0.4290
LFC Spearman	0.3682	0.3373
PDS L1	0.8878	0.9722

Across the three dataset comparisons, the zeroshot reruns favor X_state on AUPRC, Pearson delta and DE-count Spearman. Fewshot results are closer for Tahoe and Parse; Replogle fewshot favors X_hvg. The advantage is metric-dependent: zeroshot Replogle overlap is lower with X_state, and zeroshot Tahoe/Parse LFC Spearman is lower.

These are descriptive comparisons of saved experiments, not a controlled estimate of representation effects across regimes. The original fewshot and rerun zeroshot recipes differ. All selected zeroshot runs trained for 120k steps; Tahoe uses the 768-width, sample-batch pair. For example, the original Tahoe fewshot horizons were 200k (X_hvg) and 80k (X_state). See each run’s unchanged config.yaml for architecture, batching and training settings.

Full precision: paired context scores, all available contexts, same-context summary. The reproduction script downloads only the pinned per-perturbation CSVs listed with SHA256 hashes in model_index.json. These are saved CellEval v1 scores, not CellEval2 rescoring.

Download and use

Download one complete run to retain its checkpoint, vocabulary maps, gene ordering and config. The command below excludes the large saved generations. Remove the exclusion to include adata_pred.h5ad and adata_real.h5ad.

hf download arcinstitute/ST-HVG-Parse \
  --revision e36fa08d27d22c807ab63506dae053b1641e5d3c \
  --include "zeroshot/split_0/*" \
  --exclude "*.h5ad" --local-dir ./state-model

Install the STATE package and prepare matching input features (X_hvg), gene order, perturbation labels and context metadata. Configs retain their original training-time filesystem paths; provide a TOML with paths to your local data and the intended split rather than relying on those absolute paths.

state tx predict \
  --output-dir ./state-model/zeroshot/split_0 \
  --checkpoint final.ckpt --toml /path/to/local-data-split.toml

For a retained fewshot run, use its directory and --checkpoint best.ckpt. Saved CSVs and generations can be inspected without rerunning prediction. The refresh preserves the saved evaluation outputs; it does not claim numerical parity from a different software version.

Collection and provenance

Dataset	X_hvg	X_state	Canonical models
Parse	ST-HVG-Parse	ST-SE-Parse	16
Replogle	ST-HVG-Replogle	ST-SE-Replogle	16
Tahoe	ST-HVG-Tahoe	ST-SE-Tahoe	4

Updated 2026-11-12: 18 zeroshot bundles replaced across the collection, 18 fewshot models retained, and Plasmablast removed from the default branches. Complete selected rerun bundles include checkpoints, configs, loading files, training logs, CellEval CSVs, PR arrays and predicted/real H5ADs. Previous artifacts remain available under the revision tag before-zeroshot-refresh-20261112; historical revisions are not part of the canonical 36-model index.

Use is governed by the model license and acceptable use policy.