NickDST/EMBER-data

Ember data release

Pretrained checkpoint, pretraining corpus, control cells and benchmark task data for Ember, a gene-centric foundation model that learns context-dependent gene representations from single-cell RNA-seq.

Fetch with the repository's scripts/download_data.sh (components checkpoint, pretrain, controls, tasks, embeddings); every file lands where the code expects it.

Path	Contents
checkpoints/ember_base/	released checkpoint (ember_base.pt), its config and gene vocabulary
data/cxg_250K_{train,val}.h5ad	pretraining corpus: 250K CELLxGENE cells, raw counts, 22,172 genes; val = held-out donors
data/controls/*.h5ad	raw-count control cells used to contextualise gene embeddings per cell line (K562/Jurkat/HepG2/RPE1 non-targeting Perturb-seq controls; HAP1; A549/MCF7/A375/HT-29; Tahoe-100M DMSO)
tasks/pert-response-*/	perturbation-response benchmarks (DE label per (pert, gene), 10 splits)
tasks/genetic-interaction/	HAP1 signed-qGI benchmark (10 splits)
tasks/{lincs,tahoe}-unseen-drugs/atlas.h5ad	per-(cell line, drug) log2FC atlases
gene_embeddings/ember/	Ember embeddings evaluated in the paper (base model per task, context-free, mismatched-context and pretraining-loss ablations)
gene_embeddings/baselines/	baseline gene embeddings evaluated in the paper

Gene embeddings

Each file is a pickle. Context-specific files are nested, embeds[cell_line][gene_symbol] -> float32 vector, built from that cell line's control cells; context-free files are flat, embeds[gene_symbol] -> vector. The suffix names the benchmark a file was built for and fixes its cell-line keys:

Suffix	Cell-line keys
_pert_response	HepG2, Jurkat, K562, RPE1
_genetic_interaction	HAP1
_lincs	DepMap ids: ACH-000219 (A375), ACH-000019 (MCF7), ACH-000552 (HT-29), ACH-000681 (A549)
_tahoe	Cellosaurus ids: CVCL_0131 (A-172), CVCL_0546 (SW480), CVCL_0459 (NCI-H460), CVCL_0179 (BT-474)
none	context-free (ember_static, esm2, gene2vec, genept, protrek, hyenadna, string_gnn)

ember/:

• ember_base_*: the released checkpoint, contextualised by the task's own cell lines
• ember_static: context-free Ember embeddings
• ember_mismatch_*: mismatched-context control; for each task cell line, embeddings built from CELLxGENE cells with moderately similar expression to that line
• ember_cscl_only_pert_response, ember_recon_only_pert_response: models pretrained with only the cell-subgraph contrastive loss or only the masked-expression reconstruction loss

baselines/: scGPT (whole-human, pan-cancer), Geneformer V2 (316M, 104M-CL), AIDO.Cell-3M, scPRINT-2, TranscriptFormer, PINNACLE, co-expression PCA, STRING-GNN (plain, + expression,

• scGPT), ESM2, Gene2vec, GenePT, ProTrek and HyenaDNA.

Any file can be passed to the Ember evaluation as embed_name=<file stem> after scripts/download_data.sh embeddings, which links them into embeddings/.

Schemas and provenance are documented in the repository (data/README.md, tasks/<task>/README.md).