DrugDis data

Processed inputs for the code at github.com/Boom5426/DrugDis, which accompanies DrugDis: Disentangling general and context-specific effects in drug-response prediction (project page). The manuscript and Supplementary Information are in the code repository's manuscript/. To try the decomposition first, the code repository ships a 10-cell-line example in examples/data/ that needs no download.

From the root of the code repository, download the whole dataset (7.3 GB) and point the code at it:

python -m pip install huggingface_hub

from huggingface_hub import snapshot_download
snapshot_download("Boom5426/DrugDis", repo_type="dataset", local_dir="data_external/DrugDis")

export DRUGDIS_DATA=$PWD/data_external/DrugDis

One table can be read without the download:

import pandas as pd
response = pd.read_parquet("hf://datasets/Boom5426/DrugDis/drug_response.parquet")  # 4,146,204 rows

The dataset viewer shows the three response and sample tables. The expression matrices and representations have up to 15,962 columns, beyond what the viewer displays.

Files

file	size	contents	built by (in the code repository)
droma.sqlite	2.3 GB	the DROMA database of harmonized preclinical drug-response and omics data	DROMA (doi:10.5281/zenodo.18503188)
annotations/drug_anno_with_struc_info.csv, annotations/drug_anno_nci60_structure.csv	1.3 MB, 13.0 MB	compound name to SMILES tables, as used to build drug_response.parquet	provided as used
master_table.parquet	0.1 MB	one row per DROMA sample (13,823): project, model type, tumour type	drugdis/data/build_processed_tables.py
drug_response.parquet	55.9 MB	one response per (compound name, sample), 4,146,204 rows: canonical SMILES, Sensitivity (the DROMA response, in [0, 1]), resource	drugdis/data/build_processed_tables.py
gdsc_sensitivity_data.parquet	4.7 MB	GDSC1 and GDSC2 responses kept apart, 343,719 rows; rows before 229,420 are GDSC1	drugdis/data/build_gdsc_sensitivity.py
omics_mrna_raw/	570 MB	expression per cohort as stored in DROMA, samples × genes (.parquet, with .h5ad copies); scales differ between cohorts, and CCLE values are log2(TPM+1)	drugdis/data/build_processed_tables.py
omics_baseline/	203 MB	the 15,961 genes shared by CCLE and GDSC, and all cohorts on those genes	drugdis/data/build_processed_tables.py
omics_baseline_frozen/	73 MB	the CCLE-derived transcriptomic input matrix of the benchmark dataset (1,406 samples × 15,961 genes, log2(TPM+1)) and its provenance record	drugdis/data/build_ccle_substrate.py
Molecule_Embeddings/	3.7 GB	ECFP4 (radius 2, 2,048 bits) and eleven pretrained molecular representations, as {canonical SMILES: vector} pickles; the file name gives the dimension	ECFP4: drugdis/data/encode_ecfp4.py; others: the published models cited in the manuscript
Gene_Embeddings/	441 MB	gene-level expression (Baseline) and eleven pretrained transcriptomic embeddings, one row per sample (2,923)	the published models cited in the manuscript

Key columns

file	columns
drug_response.parquet	Sample_ID, SMILES (canonical), Sensitivity, Project (the response resource), Original_DrugName
master_table.parquet	Sample_ID, Project, Model_Type, Tumor_Type, Molecular_Subtype, HarmonizedIdentifier, Patient_ID, AlternateName
gdsc_sensitivity_data.parquet	Drug_Name, Sample_ID, Sensitivity, SMILES
omics_baseline_frozen/baseline_mrna_ccle_anchored.parquet	index (sample ID), then one column per gene
Gene_Embeddings/*_embeddings.parquet	feature_id (sample ID), then one column per dimension (embed_0 onwards; 0 onwards for BulkFormer and Tahoe_x1); Baseline has one column per gene and the sample ID in the index

drug_response.parquet and gdsc_sensitivity_data.parquet also store a pandas row index, the row positions of the source table before filtering. pandas reads it as the index. The viewer and datasets show it as the column __index_level_0__.

Where each file enters the manuscript

Figure and table numbers refer to the manuscript and Supplementary Information in the code repository; the result-table index maps every canonical table to its figures.

file	used for
droma.sqlite, annotations/	source of every processed table; decomposition within each response resource (Supplementary Table 3)
drug_response.parquet, master_table.parquet, omics_baseline_frozen/	the benchmark dataset (Fig. 1b, Supplementary Table 1) and every model trained on it
gdsc_sensitivity_data.parquet	cross-assay reproducibility (Fig. 2e, 4e; Supplementary Table 6), the error geometry of Fig. 4, and the GDSC1 selection with an independent GDSC2 test (Fig. 5e–h)
omics_mrna_raw/, omics_baseline/	construction of the transcriptomic inputs and the organoid inputs of the zero-shot evaluation
Molecule_Embeddings/, Gene_Embeddings/	the representation benchmark (Fig. 3; Supplementary Tables 8, 9); ECFP4 is the compound input of M0, M1 and M2

The benchmark dataset of the manuscript (3,141,680 drug–sample pairs, 986 cell lines, 54,180 compounds) is defined from these files by configs/substrate_config.frozen.json in the code repository. Its manifests/ holds the train, validation and test splits for held-out cell lines (LCLO_seed*.json) and held-out compounds (LSO_seed*.json) at seeds 3407 to 3411.

Notes on construction

• Overlap rule. When the same compound name is measured on the same sample by more than one response resource, drug_response.parquet keeps the measurement of the first resource in a fixed order (CCLE, CTRP1, CTRP2, FIMM, GDSC1, GDSC2, GRAY, HKUPDO, LICOB, NCI60, PDTXBreast, Prism, Tavor, UHNBreast, UMPDO1, UMPDO2, UMPDO3, Xeva, gCSI).
• Shared SMILES. Different compound names can share one canonical SMILES, for example stereoisomers whose canonical SMILES carries no stereochemistry, or one compound named differently by two resources (AVL-292 in PRISM, spebrutinib in NCI60). 551 SMILES are shared, so 38,622 (SMILES, sample) pairs have 2 to 7 rows (79,562 rows in all), and 15,493 of these pairs combine rows from more than one resource. The benchmark dataset averages Sensitivity over the rows of each pair (drugdis/splits/make_manifests.py), which leaves 4,105,264 pairs before its filters.
• Tavor. DROMA types the 53 Tavor samples as PDC. master_table.parquet labels them "Cell Line", and the benchmark dataset excludes the Tavor project, so Tavor enters neither the cell-line benchmark nor the organoid cohorts.
• Organotin compounds. Two NCI60 organotin compounds carry a five-valent [Sn-] in their SMILES. Their fingerprints are stored under the neutralised [Sn] form, which matches no SMILES in drug_response.parquet, so neither compound is in the benchmark dataset. RDKit 2023.09.4 rejects the [Sn-] form, so a rebuild with it lacks their 109 NCI60 rows and their two fingerprints; every other row and fingerprint is reproduced exactly.
• The upstream resources are NCI60, PRISM, CTRP1, CTRP2, GDSC1, GDSC2, CCLE, GRAY, gCSI, FIMM and UHNBreast for cell lines, and UMPDO1, UMPDO2, UMPDO3, HKUPDO and LICOB for organoids. The input-data provenance of DROMA is doi:10.5281/zenodo.17498421.

License

The processed files in this dataset are released under CC BY 4.0. droma.sqlite is redistributed from DROMA, which is released under CC BY 4.0 (doi:10.5281/zenodo.18503188). The response and omics measurements come from the resources listed above, which keep their original terms; Supplementary Data 1 of the manuscript gives the DROMA table and primary publication of each, and work that uses them should cite those publications. The code in the GitHub repository is released separately under the MIT License.

Pretrained-model embeddings

The CC BY 4.0 grant above does not cover Molecule_Embeddings/ and Gene_Embeddings/ beyond ECFP4 and Baseline. Each other file holds the output of the published model it is named after, computed by us with that model, and remains subject to that model's terms, which this dataset does not change. The table lists the license each model's authors publish for its code and, where stated separately, its weights; "not stated" means we found no license statement. Check the terms of the model you use before reuse.

License published with each model

embeddings	license published with the model
ChemBERTa2	code MIT (DeepChem); weights not stated
Chemprop	MIT
GeminiMol	AFL-3.0
InfoAlign	weights MIT
KPGT	Apache-2.0
MolCLR	MIT
Mole-BERT	not stated
MolT5	code BSD-3-Clause; weights Apache-2.0
Ouroboros	code MIT; weights not stated
UniMol, UniMolV2	MIT
BulkFormer	code MIT; weights not stated
Cell2Sentence	code Apache-2.0; weights CC BY 4.0
CellPLM	code BSD-2-Clause; weights not stated
Geneformer	Apache-2.0
scFoundation	code Apache-2.0; weights under the scFoundation model license, for non-commercial research only
scGPT	code MIT; weights not stated
SCimilarity	code Apache-2.0; weights CC BY-SA 4.0
STATE	weights under the Arc Institute State Model Non-Commercial License
Tahoe-x1	Apache-2.0
tGPT	code GPL-3.0; weights not stated
UCE	code MIT; weights CC BY 4.0

Gene_Embeddings/STATE_embeddings.parquet is Output of the State model, redistributed under its license, which requires this citation: Adduri, A. et al. (2025) Predicting cellular responses to perturbation across diverse contexts with State. Users of the scFoundation, SCimilarity and STATE embeddings in particular should read those models' terms.

Citation

Please cite the manuscript when using this dataset, together with DROMA and the primary publications of the response resources you use.

@article{Li2026DrugDis,
  title  = {DrugDis: Disentangling general and context-specific effects in drug-response prediction},
  author = {Li, Bo and Liu, Chengliang and Peng, Yuzhong and Zhang, Bob and Wang, Qing and
            Zeng, Pinxian and Li, Mengran and Huang, Shenghui and Deng, Chuxia and Zhang, Yang},
  year   = {2026}
}