Arc's new challenge uses 220k cell profiles to build an AI that simulates CRISPR

Arc has thrown down a gauntlet that could fundamentally change how we do biology. The Virtual Cell Challenge asks a deceptively simple question: can you train a model to predict what happens to a cell when you silence a specific gene with CRISPR? The goal is a “virtual cell,” a neural network simulator that lets researchers test thousands of drug candidates without ever touching a petri dish. It’s a play to tighten the feedback loop in biological research, and if it works, the impact would be enormous.

The challenge provides a hefty dataset of roughly 300,000 single-cell RNA sequencing profiles, with 220,000 cells forming the core training set. For each cell, you get a sparse transcriptome—a vector of raw RNA molecule counts for different genes. Crucially, about 38,000 of these cells are unperturbed controls. They’re the baseline, and they’re absolutely essential because measuring a cell’s transcriptome destroys it. You can’t see a cell’s state before and after a perturbation; you have to infer the change by comparing a perturbed cell to a population of control cells that are inherently heterogeneous. The core technical problem is separating the true signal of the gene edit from the noise of natural biological variation.

Arc isn’t leaving participants out in the cold. They’ve released STATE, a strong baseline model built on two transformer-based components. The State Transition Model (ST) is the actual simulator, using a Llama backbone. It ingests a set of covariate-matched control cells and a one-hot vector indicating the targeted gene, then outputs a predicted, perturbed transcriptome. It’s trained using Maximum Mean Discrepancy to align the distributions of real and predicted cell states. The second piece, the State Embedding Model (SE), is a BERT-like autoencoder. It’s designed to learn rich semantic embeddings of a cell’s state, with the goal of helping the ST model generalize better across different cell types.

The mechanics of gene expression are key to understanding the data. DNA is transcribed into pre-mRNA, which then undergoes alternative splicing—a process where non-protein-coding introns are cut out and protein-coding exons are stitched together in different combinations. The dataset reflects this complexity. A clear example is the gene TMSB4X, the most frequently silenced gene in the data. When you compare its RNA count in a control cell versus a CRISPR-edited cell, the silenced cell shows a dramatically reduced number of transcripts. That stark difference is the kind of signal the model needs to learn, but it has to do so while accounting for the messy reality that no two cells are ever truly identical to begin with.