Yusuf Roohani, PhD, Leading of the Ark Institute Machine Learning Group, is one of a team of researchers who train artificial intelligence (AI) models using transcriptome data to predict how cellular gene expression patterns will change in different cellular states. These so-called virtual cells help researchers discover new drugs that can shift cells from “disease” to “health.”
However, building virtual cells is not an easy feat.
“When you look at the cells, they live in a dynamic system,” Luhani said in an interview with Gen Edge. “Cells are constantly fluid, messy and experimentally dependent.”
Virtual cell models should explain biological complexities such as cell type, genetic background, and cell context. Furthermore, many existing single-cell datasets are affected by substantial technical noise, including limited reproducibility of perturbation effects across independent experiments that reduce model performance.
Without standardized benchmarks and dedicated datasets, this field struggles to assess whether the virtual cell model captures generalizable biological insights rather than dataset-specific patterns.
In a step towards benchmarking and accelerating virtual cells, the ARC Institute has announced its first “Virtual Cell Challenge,” a public competition sponsored by Nvidia, 10x Genomics and Ultima Genomics. The assignment is explained in a New commentary Published in cell Leuhani as the main author.
This initiative follows the footsteps of Critical evaluation of protein structure prediction (CASP) Semi-annual experiments evaluating the latest latest models in competition, structural biology. Dr. Patrick HSU, ARC co-founder and core researcher, emphasized that CASP competitions have changed protein structure predictions over 25 years, allowing breakthroughs such as Alphafold, the ultimate Nobel Prize-winning algorithm.
“We believe that using the same approach, ARC can fundamentally change the way biology research is done and accelerate progress towards a comprehensive virtual cell that can identify targets to better treat complex diseases,” HSU published.
Dr. Emma Lundberg, Associate Professor at Stanford University and co-director of Human Protein Atlas, based at KTH Royal Institute of Technology in Stockholmwe agree that establishing benchmarks is an important challenge for evaluating and comparing virtual cell models.
“I'm hoping for that [Arc’s] Challenges help coordinate communities and accelerate work towards performance and useful virtual cell models. Hopefully it will be the first of many standardized challenges in this field,” she said. Gen Edge.
Theofanis Karaletsos, senior director of AI at Chan Zuckerberg Initiative (CZI), is an active developer of virtual cells who has pushed CZI's recent models. scgenept For single cell perturbations, and Transcriptformer For heterogeneous predictions.
“CZI focuses on building cutting-edge models and providing a standardized evaluation framework for better understanding of the cells of the scientific community,” Karaletsos said. Gen Edge. “Community benchmarking is important and we believe open competition like ARC is a powerful mechanism for accelerating innovation and collective progress.”
A non-profit research institute based in Palo Alto, The ARC Institute was established in 2021 Dr. HSU and Silvana Konermann, Assistant Professor of Biochemistry at Stanford University, and current executive director of ARC. Since then, the lab has been known to place big bets on data-driven AI. In collaboration with Nvidia earlier this year, ARC released what it described as the biggest AI model of biology at the time. EVO 2.
New Context
An important challenge with AI models is making predictions outside of training data. ARC competition is appreciated Changes in gene activity can be predicted when competing virtual cells generalize to the context of new cells.
For the first competition, ARC generated a new single-cell transcriptome dataset of 300,000 H1 human embryonic stem cells (H1 HESCs) with 300 gene perturbations.
The model is evaluated on three metrics: 1) Performance in predicting differentially expressed genes. 2) Performance distinguishing between different perturbation effects. 3) General error in terms of deviation from the number of expression.
The interim performance of the competitive model will be shared on the live leaderboard during the midst of the competition. The three teams with top models will receive prizes worth $100,000, $50,000 and $25,000 worth of prizes, combining a cash award with NVIDIA DGX Cloud Credits.
Competition registration is Please open it now. Teams from individual contributors, academic institutions, biotechnology companies, and independent research institutions are eligible to participate. The final ranking is determined solely by the model performance of the final test set. This will be released in late October, one week before the final submission deadline. The winners will be announced in December.
Current status
As a baseline, competitors in the Virtual Cell Challenge go head-on with the first virtual cell model of ARC. It is designed to predict how a variety of stem, cancer, and immune cells respond to drugs, cytokines, or genetic perturbations. The state was released earlier this week for non-commercial purposes and explained in Pre-print It has been posted on ARC's website that has not yet been peer-reviewed.
According to the authors, the state improved discrimination of perturbation effects on multiple large datasets by more than 50%, identifying differentially expressed genes across genetic, signaling, and chemical perturbations with more than twice as accurate as existing models.
To promote flexibility and scalability, states consist of two interlock modules known as state transition models (ST) and state embedding models (SE).
ST learns perturbation effects using data from over 100 million perturbed cells in 70 contexts. In contrast to existing models that focus on making predictions for a single cell at once, ST leverages different bidirectional transarchitectures to predict whole cell collection. This approach brings about advancement by allowing for flexible capture of biological and technological heterogeneity without relying on explicit assumptions about distribution.
SE is trained on observed single-cell data from 167 million human cells and learns intercellular gene expression variations across diverse data sets. This module provides optimized representations for detecting biological perturbations and is robust to technical noise so that states can be effectively trained on multiple large datasets.
![The state is a transformer-based model for predicting perturbation effects across a set of cells [Arc Institute]](https://www.genengnews.com/wp-content/uploads/2025/06/PR_StateModel_Structure-002-1024x269.png)
Databound progress
Virtual Cell Challenge competitors are invited to train models for gene expression from public databases. Ark Virtual Cell AtlasIt consists of a large single-cell data set, scbasecount, and Tahoe-100m.
Fabian Theis, PhD, dIRECTOR of the AT Computational Biology Institute Helmholtz Munich is a well-known researcher working to predict genetic and chemical perturbations at the cellular level. He says data scale and quality improvements are key to moving the field forward.
“We are excited about the challenges for predicting perturbations from ARC going forward,” Theis said. Gen Edge. “Data scales have recently been expanded well enough to allow complex, generated AI models to outperform simpler linear models. It's exciting to see the behavior of true variance equations for the different model types evaluated with new data.”
Theis lab group is known as developers mobile phone, A flow matching-based framework, a generative modeling approach that can simulate single cell phenotypes induced by complex perturbations. Additionally, Theis is a scientific advisor to Open Exprice, the science group that hosts the associated C.Hallenges for benchmarking a variety of single-cell analysis methods.
Additional datasets are a fair game for virtual cell challenge model training x-atlas/orion, The largest published Preturb-seq dataset released last week by AI Drug Discovery Unicorn, Xaira Therapeutics. Dataset Provides the benefits of measurement Dose-dependent genetic effects for therapeutic applications, sUCH as a definition of the exact percentage inhibition that a drug target produces a desired effect.
Dr. CI CHU, Vice President of Early Discovery at Xaira, agrees that CASP has set a great precedent for protein structure prediction benchmarks.
“It's exciting to see ARC teams apply the same spirit to their virtual cell community,” Chu said. Gen Edge. “The progress of the field is ultimately tied to data. There is better quality public data that the community has to build, and better.–This is exactly why they also released X-Atlas/Orion. ”
Xaira is currently heading AI experts Bo Wang, PhD, SVP and Biomedical AI, who joined the team in April, and is building their own virtual cell model. The king from the University of Toronto The inventor of SCGPTa basic model of single-cell multiomics with downstream functions such as cell type annotation, perturbation response prediction, and gene network inference.
As researchers advance the next-generation AI model and mark the Virtual Cell Challenge leaderboard, the field monitors whether new treatment advancements follow. Let's start the challenge.
