---
license: mit
tags:
- biology
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
dataset_info:
  features:
  - name: counts
    sequence: int32
  - name: counts_norm
    sequence: float32
  - name: counts_log
    sequence: float32
  - name: counts_log_norm
    sequence: float32
  - name: gene_names
    sequence: string
  - name: control_counts
    sequence: float32
  - name: control_counts_norm
    sequence: float32
  - name: control_counts_log
    sequence: float32
  - name: control_counts_log_norm
    sequence: float32
  - name: delta_counts
    sequence: float32
  - name: delta_counts_norm
    sequence: float32
  - name: delta_counts_log
    sequence: float32
  - name: delta_counts_log_norm
    sequence: float32
  - name: cell_line
    dtype: string
  - name: perturbation
    dtype: string
  - name: compound_concentration
    dtype: float64
  - name: compound_unit
    dtype: string
  - name: compound_smiles
    dtype: string
  - name: mechanism
    dtype: string
  - name: moa
    dtype: string
  - name: biological_effect
    dtype: string
  - name: experimental_id
    dtype: string
  - name: timepoint
    dtype: string
  - name: text
    dtype: string
  - name: text_embeddings
    sequence: float32
  - name: chembert_embeddings
    sequence: float32
  splits:
  - name: train
    num_bytes: 176083286910
    num_examples: 49392
  download_size: 65016664023
  dataset_size: 176083286910
---

**I AM NOT AFFILIATED WITH NOVARTIS IN ANY WAY; THIS IS SIMPLY AN UPLOAD OF THEIR DATASET, "[NOVARTIS/DRUG-SEQ U2OS MOABOX DATASET](https://zenodo.org/records/14291446)."**

# Novartis DRUG-seq U2OS MoABox Dataset

This dataset profiles transcriptomic responses of the **U-2 OS** human osteosarcoma cell line to a broad collection of small molecule perturbations. It contains **49,392 observations** spanning **3,742 unique compounds** tested at **4 distinct dosages + `0.0`**, each annotated with their respective **mechanisms of action (MoA)**.

Each observation records gene expression data for **59,594 genes**.

The dataset was generated using the DRUG-seq platform, enabling high-throughput, unbiased transcriptomic readouts suited for drug discovery applications.

## Additional Information

- Perturbation dosages range across 4 unique concentration values + `0.0`.
- Observations include multiple experimental replicates and plate layouts.
- **Normalized counts** were scaled so that the total expression per cell sums to `1e4`.
- **Control counts** represent the average expression of each gene across all control cells.
- **Delta values** are computed as the difference between each sample's expression and the corresponding control mean.
- **SMILES** strings and **mechanism of action (MoA)** annotations curated by Novartis, retrieved from the [ChEMBL](https://www.ebi.ac.uk/chembl/) database and enhanced with additional sources.

## Citations

> Hadjikyriacou, A., Yang, C., Henault, M., *et al.*  
> **Novartis DRUG-seq U2OS MoABox Dataset**  
> [Novartis DRUG-seq GitHub Repository](https://github.com/Novartis/DRUG-seq)

> Hadjikyriacou, A., Yang, C., Henault, M., Ge, R., Mansur, L., Lindeman, A., Russ, C., Renner, S., Hild, M., Jenkins, J., Gubser-Keller, C., Li, J., Ho, D. J., Neri, M., Sigoillot, F. D., & Ihry, R. (2025).  
> **Novartis/DRUG-seq U2OS MoABox Dataset (1.0.0) [Data set].** Zenodo.  
> https://doi.org/10.5281/zenodo.14291446

> Li, J., Ho, D. J., Henault, M., Yang, C., Neri, M., Ge, R., Renner, S., Mansur, L., Lindeman, A., Tumkaya, T., Russ, C., Hild, M., Gubser Keller, C., Jenkins, J. L., Worringer, K. A., Sigoillot, F. D., & Ihry, R. J. (2021).  
> **DRUG-seq Provides Unbiased Biological Activity Readouts for Drug Discovery.** bioRxiv.  
> https://doi.org/10.1101/2021.06.07.447456  
> [Full text PDF](https://www.biorxiv.org/content/early/2021/06/08/2021.06.07.447456.full.pdf)