Arc Institute opens 2026 Virtual Cell Challenge with zero-shot cell predictions

Why it matters

Arc is turning virtual-cell performance into a recurring, experimentally grounded benchmark. The zero-shot, multi-cell-line task targets the generalization needed before these models can guide experiments across unfamiliar biology.

Arc Institute opened pre-registration for its 2026 Virtual Cell Challenge in a three-post thread on X, setting a harder task that will measure whether AI models can predict gene-knockdown responses across multiple cell lines without task-specific examples.

The Palo Alto research organization was founded in 2021 by Silvana Konermann (@SKonermann), Patrick Hsu (@pdhsu), and Patrick Collison after the three worked together on Fast Grants. Konermann, a Stanford biochemistry professor who earned her neuroscience doctorate at MIT, studies functional genomics and Alzheimer's disease. Hsu, a Stanford pathology professor and early CRISPR researcher, runs a lab combining genome engineering with machine learning.

The full competition will launch on August 20th. Arc is encouraging prospective competitors to assemble teams during the pre-registration period.

A harder generalization test

The 2026 challenge asks entrants to predict, on a zero-shot basis, how specified gene knockdowns will change the behavior of multiple cell lines. Arc will score those predictions against new experimental perturbation data generated from the same cell lines.

Gene knockdown experiments reduce the activity of selected genes, allowing researchers to measure the resulting changes in RNA expression and cellular state. A virtual-cell model attempts to predict those changes computationally. Accurate predictions could eventually help researchers choose which genetic interventions or drug candidates to test in the lab, narrowing a vast experimental search space.

The multi-cell-line, zero-shot format raises the difficulty from Arc's inaugural 2025 competition. That contest centered on a single H1 human embryonic stem-cell context and used a few-shot design: Arc released a training subset from the target cell type before evaluating models on unseen perturbations. The 2025 dataset covered about 300,000 cells and 300 genetic perturbations, according to Arc's account of its experimental design.

Arc said the first challenge drew over 5,000 registrants from 114 countries. More than 1,200 teams submitted results, while over 300 made final submissions. BioMap Research's BM_xTVC team won first place and $100,000 with a hybrid system combining deep learning, protein embeddings, public perturbation datasets, and statistical features.

The 2025 results also exposed the field's limitations. Arc found that perturbation-prediction models did not consistently beat naive baselines across every evaluation metric. Several of the strongest entries combined neural models with classical statistical methods, rather than relying exclusively on end-to-end deep learning. Moving to unfamiliar cell lines directly tests whether those approaches learned transferable biology or patterns tied to one experimental context.

Arc is building the test and a model to take it

The competition sits inside Arc's broader Virtual Cell Initiative, which combines large-scale biological data generation, machine-learning research, and experimental validation. Arc describes the annual contest as carrying a $100,000 grand prize and as an effort to establish common benchmarks for models that predict changes in gene expression after genetic or chemical perturbations.

Arc also develops virtual-cell systems itself. In June 2025, Arc released STATE, a model trained on observational data from nearly 170 million cells and perturbational data from over 100 million cells across 70 cellular contexts. Arc followed in January 2026 with Stack, an open-source model designed to predict cellular responses in new contexts through in-context learning.

That dual role makes the benchmark design consequential. Arc is producing experimental data, setting evaluation criteria, administering the contest, and building models in the same category. An open competition gives outside teams a shared dataset and scoring framework, while giving Arc evidence about which architectures and training strategies generalize beyond its own systems.

The challenge draws its format from CASP, the long-running protein-structure prediction competition that helped establish comparable tests for computational biology. Arc's 2026 version applies that model to cellular response prediction, where evaluation remains fragmented across datasets, cell types, perturbations, and scoring methods.

NVIDIA, 10x Genomics, and Ultima Genomics are returning as sponsors. Their involvement also reflects the infrastructure behind the contest: GPU computation for training models, single-cell analysis tools for measuring gene expression, and sequencing systems for producing the experimental results used as ground truth.

Pre-registration opened on August 4th. The August 20th launch will begin the next attempt to determine whether virtual-cell models can carry what they learn from existing experiments into cellular contexts they have never seen.