satyasrinath1992/se600m-replogle-xstate

SE-600M Replogle X_state

Precomputed cell embeddings (obsm["X_state"]) for the Arc State-Replogle-Filtered corpus, produced with the published Hub SE-600M checkpoint (se600m_epoch16.ckpt).

This artifact is a ready-to-use reference chart for experiments that need SE-600M embeddings on Replogle CRISPRi cells without re-running state emb transform (~643k cells, ~21 GB).

Contents

File	Description
replogle_concat_xstate.h5ad	AnnData: native Replogle expression + obsm["X_state"]
replogle_concat_xstate.h5ad.sha256	SHA-256 of the h5ad
provenance.json	Pinned Hub revisions, shapes, and encode settings
MODEL_LICENSE.md	Arc State Model Non-Commercial License (redistribution)
MODEL_ACCEPTABLE_USE_POLICY.md	Arc acceptable-use policy
LICENSE.md	Arc accompanying license text

Dataset identity

Field	Value
Cells (n_obs)	643,413
Genes (n_vars)	6,546 (native Replogle gene axis)
Embedding	obsm["X_state"] shape (643413, 2058)
Cell lines	HepG2, Jurkat, K562, RPE1
Perturbation column	obs["gene"] (includes non-targeting)

Expression X is the native filtered Replogle count matrix after Nirvana42 ST-ready obs prep (no VCC gene-axis expansion). Embeddings were written under key X_state.

Upstream pins

Source	Hub id	Revision
Expression corpus	arcinstitute/State-Replogle-Filtered	d790193bb2c93726541a75ca3fa873a92ed44da5
SE checkpoint	arcinstitute/SE-600M se600m_epoch16.ckpt	5a9a80f44f7ce32ce57059933ef0d735d7c10ce5
Protein embeddings	SE-600M protein_embeddings.pt	same revision

How to load

import anndata as ad

adata = ad.read_h5ad("replogle_concat_xstate.h5ad")
X_state = adata.obsm["X_state"]  # (643413, 2058)

Or from the Hub:

from huggingface_hub import hf_hub_download
import anndata as ad

path = hf_hub_download(
    repo_id="satyasrinath1992/se600m-replogle-xstate",
    repo_type="dataset",
    filename="replogle_concat_xstate.h5ad",
)
adata = ad.read_h5ad(path)

How this was produced

1. Hydrate arcinstitute/State-Replogle-Filtered at the pinned revision.
2. Prepare native ST-ready expression (replogle_concat_st_ready.h5ad).
3. Encode with Hub SE-600M via Arc state emb transform (checkpoint-embedded config; batch size 128; two-GPU shard then merge).

Equivalent operator command in the Nirvana42 / Kalki tree:

python3 -m tree_of_life_kalki.evo2.scripts.preflight --fix --gate stage_e
state-exec python3 -m tree_of_life_common.state.scripts.prepare_replogle_native_st
state-exec python3 -m tree_of_life_kalki.state_embedding.scripts.materialize_se600m_xstate \
  --gpus auto

Intended use

• SE-600M baseline / reference X_state for Replogle perturbation modeling
• Warm-start comparison against from-scratch SE charts
• Avoiding repeated multi-GPU emb transform cost on the filtered corpus

This is not a Kalki multimodal SE campaign arm checkpoint. It is the published Arc SE-600M encoder applied to the published filtered Replogle corpus.

License and citation

This repository redistributes Output of the Arc State Model. Use is subject to the Arc Research Institute State Model Non-Commercial License (MODEL_LICENSE.md). Commercial use requires a separate license from Arc (legal@arcinstitute.org).

When using this dataset, cite:

Adduri, A. et al. (2025). Predicting cellular responses to perturbation across diverse contexts with State. https://huggingface.co/arcinstitute/SE-600M

Also acknowledge the Replogle / Nadig Perturb-seq data underlying arcinstitute/State-Replogle-Filtered.

Provenance

See provenance.json for machine-readable pins (Hub revisions, embedding shape, batch size, encode mode).