Hugging Face's logo

Datasets:
Tino3141
/
aml12-drug-response-eval

Tasks:
Tabular Regression
Tabular Classification
Languages:
English
Size:
n<1K
Tags:
biology
single-cell
drug-response
aml
scrna-seq
benchmark
License:
Dataset card Data Studio Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This dataset contains derived data (differentially expressed genes and metadata) from primary AML patient single-cell RNA-seq samples originally released by Ianevski et al. (2024). Although the data is anonymised and aggregated, you must agree to use it only for non-commercial academic research and to cite both the original scTherapy paper and the upstream Zenodo deposition.

Log in or Sign Up to review the conditions and access this dataset content.

AML-12 Drug Response Evaluation Set

Derived single-cell-derived gene expression profiles and ground-truth ex-vivo drug sensitivity scores (DSS) for 12 acute myeloid leukemia (AML) patient samples. Built as the held-out evaluation set for benchmarking single-cell drug-response models against the scTherapy paper.

This is not the raw scRNA-seq data. The raw .RDS files are available from the upstream Zenodo deposition linked below. This dataset contains only derived per-patient differential expression tables, cell metadata, drug name → SMILES mappings, and the ground-truth DSS table extracted from the scTherapy supplementary Excel.

Contents 

aml12-drug-response-eval/
├── README.md
├── degs/
│   ├── patient{1..12}_degs.csv            # full DEG table (malignant vs healthy)
│   ├── patient{1..12}_degs_filtered.csv   # DEGs intersected with LINCS L1000 landmark genes
│   └── patient{1..12}_meta.csv            # per-cell metadata (cluster, %mito, malignant flag, …)
├── lincs_landmark_genes.csv               # 978 LINCS L1000 landmark genes (gene vocab)
├── model_inputs/
│   ├── patient_deg_vectors.parquet        # 12 × 978 dense DEG matrix (patients × landmark genes)
│   ├── drug_fingerprints.parquet          # 239 × 1024 ECFP4 Morgan fingerprints
│   ├── full_inputs.parquet                # 20 076 × 2003 — full (patient × drug × dose) cross-product
│   └── example_full_input.parquet         # one example concatenated input row (deg|fp|dose) → 2003-d
├── drug_name_to_smiles.json               # 242 paper drugs → canonical SMILES (PubChem-resolved)
├── ground_truth_dss.parquet               # tidy ground-truth DSS table (1 052 rows)
├── ground_truth_dss.csv                   # same, as csv
└── ground_truth_supplementary.xlsx        # original paper supplementary Excel (Fig5 sheet)

Files in detail

degs/patient{N}_degs.csv

Per-patient Seurat FindMarkers output: malignant vs. healthy cells. Columns: p_val, avg_log2FC, pct.1, pct.2, p_val_adj, gene_symbol.

degs/patient{N}_degs_filtered.csv

Same DEG table, intersected with the LINCS L1000 landmark gene space and annotated. Additional columns: gene_id, ensembl_id, gene_title, gene_type, src, feature_space.

degs/patient{N}_meta.csv

Per-cell metadata exported from the Seurat object. Columns: orig.ident, nCount_RNA, nFeature_RNA, percent.mt, RNA_snn_res.0.5, malignant, barcode.

ground_truth_dss.parquet / .csv

Tidy table extracted from the Fig5 sheet of the scTherapy supplementary Excel. Each row is one (drug, patient, baseline-model) triple with its experimentally measured DSS:

column type description
Drug string drug name (as in paper)
DSS float drug sensitivity score (higher = more potent)
Label string best or worst per the predicting model
Patient int AML patient id (1–12)
Model string scTherapy, SCDRUG, or Beyond

1 052 rows total. Note: DSS values are ground truth (ex-vivo measurement); they do not depend on the model. Only the best/worst labels differ per model.

drug_name_to_smiles.json

242 unique drug names from the supplementary Excel resolved to canonical SMILES via PubChem. 239 / 242 successfully resolved.

ground_truth_supplementary.xlsx

Verbatim copy of the scTherapy paper supplementary Excel (41467_2024_52980 _MOESM4_ESM.xlsx) for reference. The Fig5 sheet is the source of ground_truth_dss.parquet.

lincs_landmark_genes.csv

The 978 LINCS L1000 landmark genes that define the gene vocabulary used to build the model input. Sorted alphabetically by gene_symbol. This is the fixed gene order used in model_inputs/patient_deg_vectors.parquet and the deg_* columns of model_inputs/example_full_input.parquet. Columns: gene_id, gene_symbol, ensembl_id, gene_title, gene_type, src, feature_space.

model_inputs/patient_deg_vectors.parquet

Pre-built per-patient 978-dim DEG vectors (wide format), ready to feed to the model. Shape (12, 979):

  • Patient (int) — patient id
  • <gene_symbol> × 978 (float32) — avg_log2FC for that landmark gene, with 0.0 for genes that didn't pass the DEG filter for this patient.

model_inputs/drug_fingerprints.parquet

ECFP4 Morgan fingerprints for 239 drugs with valid SMILES. Shape (239, 1025):

  • Drug (str) — drug name (lowercased to match ground_truth_dss)
  • fp_<i> × 1024 (uint8) — bit i of the Morgan fingerprint (AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=1024))

Cast to float32 before feeding to the model.

model_inputs/full_inputs.parquet

The full cross-product of model inputs — every (patient × drug × dose) triple expanded into the full 2003-dim concatenated vector. Shape (20076, 2006):

  • Patient (int) × Drug (str) × dose_uM (float) — the 3 metadata columns
  • deg_<gene_symbol> × 978 (float32) — patient DEG vector
  • fp_<i> × 1024 (float32) — drug fingerprint
  • log10_dose (float32) — log10(dose_uM)

20 076 rows = 12 patients × 239 drugs × 7 doses (0.016, 0.13, 0.52, 1.0, 4.2, 17.0, 33.0 µM). Use this for bit-perfect reproducibility — feed each row's last 2003 columns directly to the LightGBM model with no preprocessing.

model_inputs/example_full_input.parquet

A single-row excerpt of full_inputs.parquet for (Patient=1, Drug=vincristine, dose=1.0 µM). Shape (1, 2006). Useful as a tiny sanity check that your input pipeline reproduces ours bit-for-bit.

Columns:

  • Patient (int=1), Drug (str="vincristine"), dose_uM (float=1.0)
  • deg_<gene_symbol> × 978 — DEG values (same order as lincs_landmark_genes.csv sorted by gene_symbol)
  • fp_<i> × 1024 — ECFP4 Morgan fingerprint (AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=1024))
  • log10_dose — log10(dose_uM)

The model expects: [deg(978) | fingerprint(1024) | log10_dose(1)] → 2003-d float32 vector.

How to use 

from huggingface_hub import snapshot_download
import pandas as pd

path = snapshot_download(
    repo_id="Tino3141/aml12-drug-response-eval",
    repo_type="dataset",
)

dss = pd.read_parquet(f"{path}/ground_truth_dss.parquet")
degs_p1 = pd.read_csv(f"{path}/degs/patient1_degs_filtered.csv")

# Pre-built 978-dim DEG vectors, ready to concat with a fingerprint + dose
deg_matrix = pd.read_parquet(f"{path}/model_inputs/patient_deg_vectors.parquet")
landmarks = pd.read_csv(f"{path}/lincs_landmark_genes.csv")

Loading the full cross-product (zero preprocessing) 

full = pd.read_parquet(f"{path}/model_inputs/full_inputs.parquet")
X = full.iloc[:, 3:].values.astype(np.float32)   # (20076, 2003) — feed straight to model
ids = full[["Patient", "Drug", "dose_uM"]]       # (20076, 3) — row identifiers

Building the full 2003-dim model input from scratch 

import numpy as np, pandas as pd, json
from rdkit import Chem, DataStructs
from rdkit.Chem import AllChem

deg_matrix = pd.read_parquet(f"{path}/model_inputs/patient_deg_vectors.parquet")
smiles = json.load(open(f"{path}/drug_name_to_smiles.json"))

deg_vec = deg_matrix[deg_matrix["Patient"] == 1].iloc[0, 1:].values.astype(np.float32)  # (978,)
mol = Chem.MolFromSmiles(smiles["vincristine"])
fp_arr = np.zeros(1024, dtype=np.float32)
DataStructs.ConvertToNumpyArray(
    AllChem.GetMorganFingerprintAsBitVect(mol, radius=2, nBits=1024), fp_arr
)
log_dose = np.float32(np.log10(1.0))  # dose in µM

x = np.concatenate([deg_vec, fp_arr, [log_dose]])  # (2003,)

Provenance

  • Raw scRNA-seq (not included here): Zenodo 10.5281/zenodo.13340927
  • Ground truth DSS + best/worst labels: scTherapy paper supplementary Excel (41467_2024_52980_MOESM4_ESM.xlsx, sheet Fig5)
  • DEG tables: derived in this project via Seurat FindMarkers (malignant vs. healthy cells per patient)
  • Drug SMILES: PubChem PUG-REST lookups

Citation

If you use this dataset, please cite the original scTherapy paper:

@article{ianevski2024scTherapy,
  title   = {Single-cell transcriptomes identify patient-tailored therapies for selective co-targeting of cancer clones},
  author  = {Ianevski, Aleksandr and others},
  journal = {Nature Communications},
  year    = {2024},
  doi     = {10.1038/s41467-024-52980-5}
}

…and the Zenodo deposition for the raw single-cell data: 10.5281/zenodo.13340927.

License

The derived files in this repository are released under CC-BY-4.0, matching the upstream Zenodo deposition. The supplementary Excel is reproduced here under the original Nature Communications open-access licensing.

Access

This repository is gated — request access via the Hugging Face UI. Acceptance is automatic once you fill in the form, but you commit to using the data only for non-commercial academic research.
Downloads last month
12
Total file size:
16.6 MB