← back to domains
§ 05 · domain

Cellular & Systems Modeling

Cellular & Systems Modeling ("Virtual Cells")

Benchmarks & competitions

1 tools
Virtual Cell Challenge
Arc Institute, launched 2025

annual benchmark for context generalization (predict perturbation effects in held-out cell types).

Datasets enabling this

3 tools
Arc Virtual Cell Atlas

consolidated perturbation data resource.

Perturb-seq atlases

genome-wide CRISPR screens (Replogle, Norman, etc.) at K562 and other lines.

Tahoe-100M
Vevo Therapeutics

100M cells, 50 cancer lines, 1100 drugs, the largest public perturbation atlas.

Industry players

5 tools
Arc Institute
State, Stack models; Virtual Cell Challenge

Twitter/X feeds for real-time field developments.

Cellarity

companies pursuing virtual-cell-style approaches with image and transcriptomic phenotypes.

Insitro

companies pursuing virtual-cell-style approaches with image and transcriptomic phenotypes.

Recursion

companies pursuing virtual-cell-style approaches with image and transcriptomic phenotypes.

Xaira

companies pursuing virtual-cell-style approaches with image and transcriptomic phenotypes.

Perturbation prediction

9 tools
AlphaCell

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

CPA

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

CellFlow

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

GEARS

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

Lingshu-Cell

generative cellular world model, leading on Virtual Cell Challenge benchmarks.

ProteinTalks

proteomics-based foundation model using ODE network architecture for drug response prediction.

STATE

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

scDFM

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

scLAMBDA

various architectures predicting how a cell's transcriptome shifts under genetic or chemical perturbation.

Single-cell foundation models

6 tools
Critical evaluations
Kedzierska et al. 2024, others

have shown these models often don't dramatically outperform simpler baselines on real perturbation tasks, there's a recognized "foundation model honeymoon" reckoning happening.

Geneformer
Universal Cell Embedding

transformer-based, trained on 30-100M cells.

Nicheformer

transformer-based, trained on 30-100M cells.

UCE
Universal Cell Embedding

transformer-based, trained on 30-100M cells.

scFoundation
Universal Cell Embedding

transformer-based, trained on 30-100M cells.

scGPT
Universal Cell Embedding

transformer-based, trained on 30-100M cells.