arcinstitute/PIE_sources

PIE: knowledge sources

The knowledge-source embeddings that PIE reads for perturbations, genes and cell contexts, built with pie sources. Together with the PIE dataset repos they are everything the replogle_wdataset and replogle_xdataset experiments need.

Datasets and knowledge sources: PIE collection

Contents

One dir per source, in the pie sources format (format version 1): meta.json (keys, layout, dimension, provenance) and embeddings.npy, plus offsets.npy for the token layout and descriptions.json (the embedded texts) for the text sources. esm2, ncbi_text and depmap_gene_effect also hold aliases.yaml: reviewed key renames (for example TAZ -> TAFAZZIN), used only when a key is missing.

Source	Index	Keys	Layout	Dim, dtype	Built from	Experiments
esm2	perturbation	19,203	dense	1280, float32	ESM-2 650M (facebook/esm2_t33_650M_UR50D) on reviewed human UniProt sequences	w, x
ncbi_text	perturbation	20,599	token	2048, float16	Qwen3.5-35B-A3B-Base token states over NCBI Gene records (GO, MedGen, gene groups)	w, x
string_space	perturbation	19,699	dense	512, float16	STRING v12.0 human protein network embeddings	w, x
depmap_gene_effect	perturbation	18,435	dense	1186, float32	DepMap CRISPR gene effect across 1,186 cell lines	w, x
perturbation_text	perturbation	12,441	dense	3072, float32	text-embedding-3-large over NCBI Gene and PubChem descriptions	w, x
gene_text	gene	31,684	dense	3072, float32	text-embedding-3-large over NCBI Gene descriptions	w, x
context_text	context	88	dense	3072, float32	text-embedding-3-large over Cellosaurus 56.0 cell line records	w, x
smiles	perturbation	1,131	dense	384, float32	ChemBERTa-77M-MTR over canonical SMILES	x
l1000_tas	perturbation	663	dense	83, float32	LINCS L1000 transcriptional activity scores (GSE92742, GSE70138)	x
prism_secondary	perturbation	348	dense	480, float32	PRISM secondary screen log2 viability across 480 cell lines	x
jump_morphology	perturbation	807	dense	737, float32	JUMP Cell Painting compound profiles (cpg0016)	x

Experiments: w = replogle_wdataset, x = replogle_xdataset. Chemical perturbation keys are drug-dose strings (<drug>_<dose>uM). Each meta.json records the input URLs and sha256s, model revisions and build parameters.

Usage

The experiment configs pin this repo by commit, so pie train and pie eval download only the source dirs they use. To download a copy:

hf download arcinstitute/PIE_sources --repo-type dataset \
  --local-dir "$PIE_DATA_ROOT/sources"

replogle_wdataset needs only the first seven sources:

hf download arcinstitute/PIE_sources --repo-type dataset \
  --local-dir "$PIE_DATA_ROOT/sources" \
  --include "esm2/*" --include "ncbi_text/*" --include "string_space/*" \
  --include "depmap_gene_effect/*" --include "perturbation_text/*" \
  --include "gene_text/*" --include "context_text/*"