scDrugPerturb-Bench The world's largest open drug perturbation prediction dataset with key-gene annotations
AI for Biology

Build the Virtual Cell

Our vision is to build an intelligent virtual cell that can reason over biological mechanisms, simulate cellular responses and help solve the most complex challenges in biopharmaceutical discovery and development.

Explore our roadmap scDrugPerturb-Bench Projects
A computable virtual cell with molecular signals and perturbation-response trajectories
Roadmap

Building the foundations for virtual cells.

We are building from open, reproducible perturbation resources toward models that predict intervention responses, connect molecular mechanisms to phenotypes and support therapeutic discovery.

01

scDrugPerturb-Bench preprint

The scDrugPerturb-Bench manuscript defines mechanism-aware evaluation tasks and reporting standards for single-cell drug response prediction.

Completed
02

Drug perturbation dataset release

Version 1 of our curated perturbation dataset is now available, with ongoing maintenance and updates as the resource continues to grow.

v1 released
03

Gene perturbation datasets

Versioned genetic-intervention datasets with biological context, provenance, evidence and reproducible splits.

In progress
04

Single-cell foundation models

General-purpose models that learn transferable representations of cell state across datasets, conditions and biological contexts.

In progress
05

Perturbation prediction models

Mechanism-aware models for predicting cellular response across drug and gene perturbations, spanning compounds, doses, cell states and experimental systems.

In progress
06

Applications

Virtual-cell applications for drug discovery tasks such as compound retrieval, target discovery, response prediction and therapeutic prioritization.

Long-term
Publication
Preprint Active development

A mechanism-annotated benchmark reveals limited fidelity to drug-response signatures in single-cell perturbation models

Unlike benchmarks built only from expression matrices, scDrugPerturb-Bench links each evaluation case to a biological conclusion extracted from its source paper. It tests whether a model recovers the direction, magnitude and specificity of drug response beyond expression similarity.

2.5Msingle cells
181datasets
101drugs
4,500model evaluations
Literature-grounded test case

In HUES8 pancreatic progenitors, 0.5 µM latrunculin A for 5 days upregulates NEUROG3 versus untreated cells.

Benchmark 32094658_01GSE1376591,728 cells (686 control + 1,042 treated)
Research page Preprint Code
Overview of scDrugPerturb-Bench data curation, evaluation framework and dataset composition
Figure 1. Benchmark design, mechanism-aware evaluation and dataset coverage.
Projects
Agent GitHub

Paper2Perturb

Turn scientific papers and public single-cell datasets into evidence-linked perturbation metadata and standardized h5ad files.

View on GitHub
6agent skills
5data sources
2validation skills
Supported data sources GEO · ArrayExpress · Zenodo · CELLxGENE · GitHub
End-to-end workflow Paper evidence -> Structural single-cell data -> Reproduce reported results -> Validate paper conclusions
Open sourceApache 2.0
Agent GitHub

ScAgent

One processing standard for heterogeneous public single-cell data, from source files to analysis-ready AnnData.

View on GitHub
2input workflows
1processing standard
.h5adanalysis-ready output
The problem Public expression matrices are produced with different processing software and quality-control criteria, introducing avoidable technical batch effects.
The ScAgent approach Reprocess every dataset with one software stack and the same QC, normalization, cell annotation and analysis workflow.
Public repositoryPython

Open research, released with evidence.

Follow Simucella as datasets, models and evaluation resources become available.

View on GitHub Roadmap