arcinstitute/PIE_splits

PIE: canonical splits

The train, validation and test splits of the PIE experiments: within one dataset (<dataset>_wdataset: Replogle-Nadig, Jiang, Tahoe-100M) and across datasets (<dataset>_xdataset, the named dataset held out: Replogle-Nadig, Jiang, X-Atlas/Orion, ARC VCC 25). The experiment configs pin this repo by commit, so pie train and pie eval download the split files they need automatically.

Datasets and knowledge sources: PIE collection

Contents

replogle_wdataset/<setting>/<cell_line>/{train,val,test}.json   setting: unseen_ctx, unseen_pert, unseen_ctx_pert
jiang_wdataset/<setting>/<cell_line>/{train,val,test}.json
tahoe_wdataset/<setting>/{train,val,test}.json
replogle_xdataset/{train,val,test_seen,test_unseen}.json
jiang_xdataset/{train,val,test_seen,test_unseen}.json
orion_xdataset/{train,val,test_seen,test_unseen}.json
arc_vcc_25_xdataset/{train,val,test_seen,test_unseen}.json

Each file is a JSON object that maps <dataset>.<context> to a list of perturbations, for example {"replogle.k562": ["AAMP", ...]}. A row is one (context, perturbation) pair of a PIE preprocessed dir. The files of one split dir are pairwise disjoint.

replogle_wdataset/

Replogle-Nadig only, four folds (hepg2, jurkat, k562, rpe1). A fold trains and validates on the other three cell lines; train.json and val.json are the same for all three settings of a fold.

Setting	Test cell lines	Test perturbations	Test rows (hepg2 / jurkat / k562 / rpe1)
unseen_ctx	held-out cell line	seen in training	867 / 978 / 884 / 983
unseen_pert	the three training lines	unseen in training	1,238 / 1,214 / 1,252 / 1,172
unseen_ctx_pert	held-out cell line	unseen in training	386 / 461 / 430 / 470

Train rows per fold: 2,670 / 2,526 / 2,603 / 2,550; validation rows: 298 / 280 / 290 / 284.

jiang_wdataset/

Jiang et al. only, 30 contexts (six cell lines under five stimuli), six folds (a549, bxpc3, hap1, ht29, k562, mcf7). A fold holds out the five contexts of one cell line and trains and validates on the other 25; train.json and val.json are the same for all three settings of a fold. A context with no row in a file is left out of that file.

Setting	Test contexts	Test rows (a549 / bxpc3 / hap1 / ht29 / k562 / mcf7)
unseen_ctx	held-out cell line	36 / 58 / 76 / 51 / 54 / 21
unseen_pert	the 25 training contexts	127 / 109 / 92 / 119 / 112 / 128
unseen_ctx_pert	held-out cell line	22 / 38 / 43 / 24 / 32 / 6

Train rows per fold: 248 / 231 / 227 / 240 / 238 / 274; validation rows: 28 / 25 / 23 / 27 / 25 / 32.

tahoe_wdataset/

Tahoe-100M only, one fold: five held-out cell lines and 45 training lines, with 265 training molecules at all three doses (795 perturbations); train.json and val.json are the same for all three settings.

Setting	Test cell lines	Test perturbations	Test rows
unseen_ctx	the five held-out lines	seen in training	3,377
unseen_pert	the 45 training lines	unseen in training	13,055
unseen_ctx_pert	the five held-out lines	unseen in training	1,451

Train rows: 32,173; validation rows: 3,588.

replogle_xdataset/

Trains on Tahoe, Jiang, ARC VCC 25 and Orion and tests zero-shot on the four Replogle-Nadig cell lines.

File	Contexts	Rows
train.json	tahoe 45, jiang 23, orion 2, arc_vcc_25 1	64,260
val.json	tahoe 4, jiang 2, orion 2, arc_vcc_25 1	957
test_seen.json	replogle 4 (perturbations in train.json)	2,286
test_unseen.json	replogle 4 (all other perturbations)	3,173

jiang_xdataset/, orion_xdataset/, arc_vcc_25_xdataset/

Each holds out one dataset and trains on the other four, with the same file layout as replogle_xdataset/. Contexts of a held-out cell line, and contexts derived from one, are removed from training and validation. The test files hold the held-out dataset's full panel.

Held out	Train: contexts, rows	Val rows	Test seen	Test unseen
jiang	tahoe 43, replogle 3, orion 2, arc_vcc_25 1; 65,597	1,172	227	234
orion	tahoe 45, jiang 28, replogle 4, arc_vcc_25 1; 56,921	613	841	12,443
arc_vcc_25	tahoe 45, jiang 28, replogle 4, orion 2; 69,403	1,225	121	69

Removed for jiang: Replogle-Nadig K562 and Tahoe-100M A-549 and HT-29. The ARC VCC 25 test rows span its two contexts, ARC_H1 and ARC_H1_VAL.

Usage

The experiment configs read these splits directly; pass a split file to pie eval or pie infer as an hf:// reference:

SPLITS=hf://datasets/arcinstitute/PIE_splits@396ab9563175ee887750c9eed7ccaea6f5fdbf50
uv run pie eval experiment_name=replogle_wdataset/k562 ckpt=best_auprc \
  split_path=$SPLITS/replogle_wdataset/unseen_ctx/k562/test.json row_set=unseen_ctx

To use your own splits, give a dir with train.json and val.json as data.split_dir=<dir> (local or hf://). To download the files:

hf download arcinstitute/PIE_splits --repo-type dataset --local-dir splits