Tino3141/lincs-pharmacodb-training

LINCS-PharmacoDB Training Data

Pre-built training data for the scTherapy LightGBM drug-response model.

Files

File	Description
training_data.npz	NumPy archive with X (N, 2003) and Y (N,) arrays
feature_names.json	2003 feature labels: 978 gene symbols + 1024 ECFP4 bits + dose
matched_samples.parquet	Metadata join between LINCS signatures and PharmacoDB

Feature vector

Each row of X is [978 LINCS landmark gene z-scores | 1024 ECFP4 fingerprint bits | 1 dose].

Target Y is the inhibition score (100 - viability) from PharmacoDB.

Data sources

• Gene expression: LINCS L1000 Level 5 (z-scores), 978 landmark genes
• Drug fingerprints: ECFP4 (Morgan radius=2, 1024 bits) from canonical SMILES
• Inhibition scores: PharmacoDB dose-response data

Usage

from huggingface_hub import hf_hub_download
import numpy as np, json

npz_path = hf_hub_download("Tino3141/lincs-pharmacodb-training", "training_data.npz", repo_type="dataset")
data = np.load(npz_path)
X, Y = data["X"], data["Y"]

fn_path = hf_hub_download("Tino3141/lincs-pharmacodb-training", "feature_names.json", repo_type="dataset")
with open(fn_path) as f:
    feature_names = json.load(f)