VCHarness builds virtual‑cell models
A project called VCHarness described an autonomous AI 'harness' that builds perturbation‑response models for virtual cells using Monte Carlo Tree Search and foundation models, and reported improved performance on CRISPR differential‑expression tasks across cell lines. The authors released the approach as a benchmark for computational cell models. ((x.com))
A virtual cell is software that tries to predict how a real cell will react, and a new preprint says an automated system called VCHarness can build those models faster than human teams. (biorxiv.org) The paper, posted on bioRxiv on April 11, 2026, describes VCHarness as a closed-loop system that combines biological foundation models, an artificial-intelligence coding agent, Monte Carlo tree search, structured memory, and distributed execution. (biorxiv.org) In plain terms, the system writes candidate model code, runs experiments, scores the results, and then searches for a better design, repeating that cycle with little human intervention. The authors said it reduced model-development time from months to days on perturbation-response tasks. (biorxiv.org) The task here is narrower than a full digital copy of a cell. VCHarness focuses on perturbation response, meaning it predicts how gene expression changes after a genetic or chemical intervention such as Clustered Regularly Interspaced Short Palindromic Repeats, or CRISPR, knockdown. (biorxiv.org) That prediction problem sits near the center of drug discovery and disease biology, because researchers want to know which genes or compounds will change a cell’s behavior before running every experiment in the lab. A 2024 Cell perspective called virtual cells a priority for the field and argued that recent artificial-intelligence and omics advances put the goal within reach. (cell.com) VCHarness comes from GenBio AI, whose authors include Xingyi Cheng, Pan Li, Le Song, Eric Xing and colleagues. The company has also released open repositories for its broader Artificial Intelligence-Driven Digital Organism, or AIDO, platform and for earlier perturbation-response foundation-model work. (biorxiv.org) (github.com 1) (github.com 2) The paper says the system searched over model architectures and training pipelines and found designs that beat expert-built baselines across multiple perturbation-response benchmarks. The accompanying GenBio AI post says those benchmarks included multiple CRISPR gene-knockdown settings. (biorxiv.org) (genbio.ai) The broader field has been building the data and contests needed for this kind of claim. One example is the Virtual Cell Challenge, hosted by Arc Institute and backed by NVIDIA, 10x Genomics, and Ultima Biosciences, which asked teams to predict gene-expression responses in held-out H1 human embryonic stem cells from CRISPR interference data. (github.com) Researchers outside GenBio AI have been pushing the same direction. A December 2024 Stanford report on the Cell perspective described a virtual human cell as a long-term effort to simulate biomolecules, cells, and eventually tissues and organs with artificial intelligence. (stanford.edu) For now, VCHarness is a preprint, not a peer-reviewed paper, and its strongest claims still need outside reproduction on shared benchmarks. But it adds a concrete idea to the virtual-cell push: let one artificial-intelligence system search for the model, instead of asking researchers to hand-design every part. (biorxiv.org)
