Code for "Are Current AI Virtual Cell Models Useful for Scientific Discovery?"

Code to reproduce experiments from "Are Current AI Virtual Cell Models Useful for Scientific Discovery?"

The python package to compute PerturbHD hit discovery metrics is located in packages/perturb-hd and can be installed with pip install perturb-hd.

Installation

Dependencies are managed with uv. Run uv sync to install packages, and run python files as uv run [filename.py]

Overview

• analysis/scripts/: code to download and preprocess data, train models, and compute evaluation metrics
• analysis/notebooks/: code to generate figures
• packages/perturb-hd/: python package to compute PerturbHD hit discovery metrics

Download precomputed data for figures

Precomputed data to reproduce figures (model predictions, labels, metrics, LLM predictions + reasoning traces) can be downloaded from figshare by run uv download_precomputed_data.py [--skip-gene]

Skip gene skips downloading the large predicted and estimated gene-level perturbation effect files (~10GB).

Citation

@article{Bereket2026PerturbHD,
  author  = {Bereket, Michael D and Leskovec, Jure},
  title   = {Are Current AI Virtual Cell Models Useful for Scientific Discovery?},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.04.23.719015}
}

Paper Link

Code to reproduce experiments