← All projects

Training a 'virtual cell' model for gene expression inference

Train a virtual-cell model from public scRNA-seq perturbation data and benchmark it against entries from the November 2025 Virtual Cell Challenge.

  • Theme: Single-cell, transcriptomics, machine learning
  • Repo:

Team

Why this project

The Virtual Cell Challenge is a fresh, well-defined benchmark with harmonized public data — exactly the kind of target a hackathon team can ship something measurable against in three days.

What a team could build in one day

  • Stand up the STATE base model from Arc Institute’s harmonized datasets (a video tutorial walks through the setup).
  • Run inference on a held-out perturbation set.
  • Compare predictions to challenge baselines and the published entries.

Minimum viable demo: trained model + a small evaluation report against one or two challenge tasks.

Stretch directions

  • Novel architectures or multi-task heads.
  • Data augmentations.
  • Mixed real + simulated perturbation training.

Resources

Compute

High — STATE training is ~2 hours on an A100. Inference is much cheaper.