Arc Institute Is Building a Virtual Cell: The AI Model That Predicts What Happens Inside a Cell Before Anyone Runs the Experiment

#ai #machinelearning #biotech #datascience

The most important AI model you have never heard of is not trying to write code or pass the bar exam. It is trying to predict what happens inside a single human cell when you turn a gene off.

In October, Yusuf Roohani, who leads a machine learning group at the Arc Institute in Palo Alto, gave a seminar at Columbia University titled "Building a Virtual Cell: An AI Platform for Engineering Cell State." The talk is on YouTube, and it is a rare look at a project that people in AI keep calling the next AlphaFold. But instead of predicting how a protein folds, this one predicts how an entire cell reacts when you perturb it. Silencing a gene, dosing a drug, applying a signaling molecule. The model outputs what the cell's gene expression profile will look like afterwards, before anyone runs the experiment in a wet lab.

If that works at scale, drug discovery changes shape. Today about 90% of drugs fail in clinical trials, mostly because of weak efficacy or side effects. If a model can filter thousands of candidate interventions in silico and hand biologists only the promising ones, the expensive part of the pipeline gets smaller.

The talk covered four projects. Here is what each one does, and what the results actually say.

1. State: the foundation model for cell response

State is Arc's first virtual cell model, released in June 2025 alongside a paper in Cell. It is a transformer that takes two inputs: a set of cells in their current state and a perturbation, such as a CRISPR gene knockdown or a drug. The output is the predicted gene expression distribution after the perturbation.

The scale is the headline. State was trained on perturbation data from more than 100 million cells across roughly 70 cell contexts, which the team says is more than any other model at the time of release. On the Tahoe-100M cancer cell line dataset, Arc reported that State improved discrimination of perturbation effects by over 50% and roughly doubled the accuracy of identifying true differentially expressed genes compared to prior deep learning baselines like GEARS and scGPT. On some metrics it was the first model to consistently beat simple linear baselines, a bar that sounds low and is not.

The more interesting part of the talk was what Roohani said the model cannot do. State is best at a task the team calls context generalization: given T cell data from 50 donors, predict how donor 51's T cells respond to the same perturbations. When a cell type has little or no training data, performance drops hard. Someone in the audience asked when this reaches AlphaFold-level reliability. His answer: the bottleneck is data quality, not architecture, and we are likely more than a decade away from an AlphaFold moment for cells.

2. The Virtual Cell Challenge: the field's report card

Arc also runs an open competition, the Virtual Cell Challenge, framed as a Turing test for cell models: can a model stand in for an actual Perturb-seq experiment? The first challenge drew over 5,000 registrants from 114 countries and more than 1,200 teams.

The dataset was deliberately hard. Around 300,000 single-cell profiles from H1 human embryonic stem cells, a cell type poorly represented in training data, with 300 CRISPRi perturbations sequenced at over 50,000 UMIs per cell and roughly 1,000 cells per perturbation. Most public datasets carry 30 to 100 cells per perturbation, so this is close to an order of magnitude deeper.

The sober finding came in the wrap-up: on mean absolute error, nearly every submission did worse than a naive baseline that just predicts the average cell. Teams did make real progress on biologically meaningful metrics like perturbation discrimination and identifying differentially expressed genes, and the winning approaches mixed deep learning with classical statistics. Pure end-to-end learning has not solved this problem yet. The challenge is annual, and the 2026 edition with new cell contexts is already live.

3. scBaseCount: the biggest biological dataset ever curated by AI agents

The third project is the one with the most direct lesson for machine learning engineers. Public sequencing data lives in the SRA archive, and the raw reads are a mess: inconsistent metadata, inconsistent processing, samples described in free text. The largest existing single-cell repository, CZ CELLxGene, held around 107 million cells.

Arc built a hierarchical agent system to fix this. One agent constantly scans SRA for 10x Genomics datasets. Other agents extract and normalize metadata, map free-text descriptions onto controlled vocabularies and ontologies, and feed clean records into a standardized recounting pipeline. Supervisors check the outputs of the agents below them, because, as Roohani put it, agents that check each other produce much more validated results.

The result is scBaseCount: over 500 million cells, more than four times the size of CELLxGene, spanning 27 species and 75 tissues, all processed with the exact same alignment and gene annotation. Batch effects are measurably lower than in manually assembled repositories, and the thing keeps growing on its own because the agent just re-runs. In the Q&A, the team noted the agent workflow was built with LangGraph.

The honest caveat: this is data cleaning at scale, not a magic system. Roohani shared a tweet that captures the deployment experience perfectly: LLMs feel like geniuses on personal side projects, then feel unbelievably stupid the moment you put them in production. The multi-level agent design is Arc's answer to that gap.

4. The bio-discovery agent: AI that designs the experiment

The last project closes the loop. Instead of cleaning data passively, the agent proposes which genes to knock down in a T cell activation screen, gets the lab results back, updates its beliefs, and proposes the next round. Over successive iterations it selects more true hits than both random selection and Bayesian optimization. Giving the agent both the biological context and the experimental readouts produced the strongest results; the agent learned to reason over both.

This is where the talk connects to the current agent engineering conversation. The pattern is the same one being worked out in software: propose, execute, observe, update. The difference is that the execution step is a wet lab, and the feedback loop is measured in days instead of seconds.

Why this matters outside biology

Two takeaways transfer directly to anyone building AI systems right now.

First, data beats architecture. Arc's own evaluation of the first challenge showed models losing to naive baselines on raw accuracy, while a purpose-built, deeply sequenced, uniformly processed dataset showed stronger signal than almost everything public. The team is now collaborating with academic labs to standardize even the data submission process to SRA. That is a decade-scale bet on data infrastructure, and it rhymes with what every serious ML team eventually concludes.

Second, agents earn trust through verification loops. The scBaseCount hierarchy works because outputs are checked, and the discovery agent works because reality scores every prediction. Neither system asks you to believe an LLM's output. Both make the LLM's output falsifiable.

The full seminar is worth watching, especially the Q&A, where Roohani is unusually candid about how far the field is from a working virtual cell. AlphaFold took decades of accumulated protein structures before its moment arrived. The cell version is waiting for its data foundation, and Arc is building it at half a billion cells at a time.

DEV Community