> ## Content Index
> Fetch the complete content index at: https://nexi.fund/llms.txt
> Use this file to discover other available public pages before exploring further.

# The Virtual Cell Race: AI Biology Moves From DNA to Whole-Cell Simulation
- URL: https://nexi.fund/virtual-cell-ai-models-2026/
- Published: 2026-09-17T10:00:01.000Z
- Updated: 2026-09-17T10:00:01.000Z
- Description: In August 2026, GenBio AI and Arc Institute pushed biological AI from reading DNA to simulating a whole cell. The models are impressive; the data, the benchmarks and the GPU bill will decide who wins.
- Author: Nexi.fund Labs
- Tags: AI & Infrastructure, #mode-5, #hook-paradox, #track-F

For three years, biology has been promised its ChatGPT moment. The sales pitch barely changes: a model that reads the language of life, writes sequences nature never got around to, and compresses a decade at the bench into an afternoon of inference.

What landed in August 2026 is stranger and smaller than that pitch. On August 18, GenBio AI — a Palo Alto company co-founded by Nobel laureate David Baker and Eric Xing, president of the Mohamed bin Zayed University of Artificial Intelligence — introduced AIDO Cell. GenBio calls it the first world model of a human cell. The system walks the full biological hierarchy, from DNA and RNA through protein to the whole cell, and predicts how a drug or a switched-off gene ripples upward. Two days later, Arc Institute opened the 2026 Virtual Cell Challenge on the opposite principle: no training set, six cell lines no entrant has seen perturbed, and $100,000 for whoever predicts how each responds to a gene knockdown — switching one gene off.

Both programs chase the same artifact. Both keep hitting the same wall. The binding constraint on simulated biology turned out to be compute, not biology.

## Three things the August milestones settled

🎯

**A virtual cell is now a scored benchmark.** Arc's challenge grades models zero-shot on cell lines they have never seen, so the leaderboard measures generalization rather than memorization.  
  
**The winning architecture is a stack.** A DNA foundation model, a perturbational dataset, and an agentic build harness sit underneath every credible cell simulator.  
  
**The compute vendors are inside the race.** NVIDIA's BioNeMo is the substrate GenBio builds on and a sponsor of Arc's challenge, which turns GPU capacity into a strategic input for biology labs. 

The two announcements matter less for what they shipped than for what they measured. A cell simulator became something a third party can grade. The build is assembling into layers. And the supplier of those layers now has a stake in the result.

## The stack underneath a cell

Start with DNA as a language. Arc's Evo 2 is a genomic foundation model trained on more than 9.3 trillion nucleotides drawn from over 128,000 genomes, built as a 40-billion-parameter network with a one-million-token context window. It reads and writes DNA, RNA and protein sequences at single-nucleotide resolution — the level at which evolution actually operates. The work was published in Nature in March 2026, and the weights are open. The 40B model has passed six million API calls; the repository has cleared 88,000 downloads.

Data is the second layer. Arc's Virtual Cell Atlas gathers more than 600 million cells, including Tahoe-100M — 100 million cells from roughly 60,000 drug-perturbation experiments, mapping 50 cancer models against more than 1,100 treatments. Tahoe built it on its Mosaic platform with Parse Biosciences' GigaLab and Ultima Genomics sequencing. Scale matters here because a cell simulator fails on the experiments it never saw.

## From sequence to simulation

The third layer is the simulator. Arc shipped State, its first virtual cell model, in June 2025\. GenBio's AIDO Cell, released in August 2026, tries the full hierarchy in one continuous system: perturb the cell at any level, from a single gene to a protein, then read out the whole response. Release 1.0 covers two human cell lines, K-562 and Hep-G2, and arrives with a Virtual Cell Benchmark.

This is the direction we flagged in July, when [AI protein-design platforms were already compressing the discovery loop](https://nexithon.com/ai-frontier-models-protein-design-platform?ref=nexi.fund). The virtual cell pushes that work up a level, from single molecules to an entire cell.

```

TIMELINE: from DNA language to a simulated cell
─────────────────────────────────────────────────────────────
  2024        2025         2025         2026          2026
  🧬          🧬           🧪           🧪            ◉ NOW
  Evo 1       Evo 2        State        AIDO Cell     Virtual Cell
  7B params   preprint     first        whole-cell    Challenge
              + BioNeMo    virtual      world model   zero-shot
                           cell model
─────────────────────────────────────────────────────────────
Sources: Arc Institute, GenBio AI, Nature (2026)

```

Evo 1 (Science, 2024), Evo 2 (Nature, March 2026), State (June 2025), AIDO Cell (August 2026), Virtual Cell Challenge 2026.

Above all three layers sits the build harness. GenBio runs VCHarness, an agentic system that generates, tests and refines candidate cell components, wired into NVIDIA's BioNeMo, Megatron and NIM microservices. Biology's model factory is now an AI infrastructure project.

40B parameters 

#### Evo 2 genomic foundation model

Trained on 9.3T nucleotides with a one-million-token context window. · *Nature, 2026*

100M single cells 

#### Tahoe-100M perturbational dataset

About 60,000 drug experiments across 50 cancer models; the Atlas holds 600M+ cells. · *Arc Virtual Cell Atlas, 2025*

> We are far from solving cell biology, but we have built a system that can be interrogated as though it were a cell.— Eric Xing, co-founder and chief scientist, GenBio AI

## Who is actually racing

Arc Institute is the nonprofit anchor. Headquartered in Palo Alto, it builds the open stack — Evo, State, the Atlas — and runs the competition that sets the benchmark. It named Usman Muzaffar as chief technology officer in July 2026 to push the Virtual Cell Initiative. Its challenge is backed by NVIDIA, 10x Genomics and Ultima Genomics, and the final test set lands on October 22, 2026.

GenBio AI is the commercial challenger. It assembles pre-trained models for DNA, RNA, protein and cell into one programmable system, and it is already selling the promise of downloadable cells for open-ended experiments. Its offices in Palo Alto, Abu Dhabi and Paris signal where it expects the buyers to be.

Around them sit the incumbents of computational drug design. Isomorphic Labs, spun out of Alphabet's DeepMind, applies the AlphaFold lineage to molecular design. Insitro pairs lab-generated data with machine learning across metabolic and neurological disease. Recursion runs one of the largest automated wet-lab fleets in the industry. None of them owns a full-hierarchy cell simulator yet. That is the gap both August announcements claim to narrow.

## Where the promise breaks

The Evo 2 team put a number on the distance still to travel. Fine-tuning the model on bacteriophage genomes produced 285 candidate designs; 16 of them propagated and inhibited the target bacteria without touching unrelated strains. That is a working rate near 6% — progress, and a long way from engineering.

Cell biology adds its own friction. AIDO Cell 1.0 simulates two cell lines. Generalization across cell contexts — the exact capability Arc's 2026 challenge measures — is unsolved, and that is why the competition withholds a training set. Roughly one in 10,000 compounds that enter development reaches the clinic, the ratio the virtual cell is meant to improve. About a fifth of human proteins still have no known function, which limits what any simulator can be asked to predict.

Biosafety is the quiet constraint. Arc excluded eukaryotic viruses from Evo 2's training on purpose, and its red-team reviews found that generated sequences for pathogenic viral proteins read as effectively random. The team has said plainly that future, more capable models will need stronger alignment work.

## What to watch next

📊

**Key signals to track**  
  
The Virtual Cell Challenge final test set releases on October 22, 2026 — the zero-shot scores will show whether generalization is real or leaderboard noise.  
  
Whether AIDO Cell expands past K-562 and Hep-G2 in its next release, or stays a two-cell demonstration.  
  
Whether a pharmaceutical company commits to replacing any part of a trial with an in-silico readout.  
  
Whether the Arc Atlas crosses one billion cells, and whether NVIDIA's biology stack becomes a paid line item for labs. 

The virtual cell is no longer a thought experiment. It has a benchmark, a dataset, a vendor stack and a scoreboard. What it does not yet have is proof that a simulation predicts a patient. The next twelve months of leaderboard results will say more about that than any launch.

Compute decides how fast biology gets there.

## Sources

[ Genome modelling and design across all domains of life with Evo 2 The peer-reviewed Evo 2 paper: 40B parameters, a one-million-token context and 9.3 trillion nucleotides of training data. Nature ](https://doi.org/10.1038/s41586-026-10176-5?ref=nexi.fund) 

The primary source for the model layer of the stack — and for the 16-of-285 bacteriophage result.

[ The 2026 Virtual Cell Challenge: zero-shot prediction across unseen cell lines Arc's benchmark sets the task: predict gene-knockdown responses in six cell lines the model has never been trained on, for a $100,000 prize. Arc Institute ](https://arcinstitute.org/news/virtual-cell-challenge-2026?ref=nexi.fund) 

The benchmark design is the story: it measures generalization, which is the capability the field has not yet shown.

[ AIDO Cell: A General-Purpose Simulator for Cell Biology GenBio AI's technical description of its whole-cell world model, covering the K-562 and Hep-G2 prototype and Virtual Cell Benchmark 1.0. GenBio AI ](https://genbio.ai/aido-cell-simulator?ref=nexi.fund) 

The commercial counterweight to Arc's open stack — and the clearest statement yet of what a programmable cell is meant to sell.