GenBio AI Unveils a Breakthrough World Model for Simulating Human Cells

On August 18, 2026, Palo Alto-based GenBio AI announced the preview release of the first general-purpose simulator and world model of the human cell called the AIDO Cell. Ultimate aim is to reduce the dependence on clinical testing of drugs, or at least reduce a step.

Virtual Cells for Virtual Experiments

The system aims to enable researchers to create virtual cells, apply genetic or chemical interventions i.e best 3d drug candidates, and observe responses across multiple biological scales in a single, coherent framework. Unlike specialized AI tools that handle isolated tasks such as protein structure prediction, binding or gene expression analysis, AIDO Cell integrates DNA, RNA, proteins, regulatory networks, and whole-cell behavior into a singular unified model. Current preview suggest that it has a persistent cellular state, allowing sequential experiments: scientists can intervene, observe changes, then apply further perturbations to the updated state.

This multi-turn capability mirrors real laboratory workflows more closely than one-shot predictions from separate models and adds another step to drug discovery where predicted models can be vetted further before wet lab production. The initial release only seems to focuses on two well-studied human cell lines: K-562, derived from leukemia, and Hep-G2, from liver cancer. These lines have decades of supporting multi-omics data, making them ideal prototypes for validation.

Users can simulate perturbations such as gene knockouts, knockdowns, overexpressions, or small-molecule treatments, then query consistent multimodal readouts—including chromatin accessibility, expression profiles, protein structures and interactions, localization, and morphology—without altering the underlying state. In benchmarks spanning 31 metrics across five task families (small-molecule perturbation, genetic knockout, protein structure, RNA splicing, and genome regulation), AIDO Cell 1.0 claims to have achieved state-of-the-art results on 24 of them.

On the remaining metrics where specialists retain a modest edge, AIDO Cell remains competitive for now:

• AlphaFold 3 holds an advantage on certain protein–DNA interactions.
• isoDDE leads on some protein–ligand docking metrics.
• AlphaGenome performs better on select genome regulation tracks.

It is the only system reported to cover all five families comprehensively, while remaining competitive with specialized models like AlphaFold 3 or single-cell foundation models on individual tasks. Beyond forward simulation, the platform supports “in-context molecular design.” Researchers can potentially define a desired cellular phenotype, generate candidate molecules, apply them to cloned virtual cells, and evaluate effects across scales.

AIDO Cell forms part of GenBio’s broader AI-Driven Digital Organism (AIDO) vision. Supporting tools include AIDO Lab for interactive experimentation and AIDO Foundry for adapting the model to new cell types or data modalities as fresh experimental results become available. Early signs of cross-context scaling suggest the system can leverage data from multiple cell lines to improve predictions in previously unseen settings. The next steps could possibly and hopefully mimic entire organs or nervous system or the entire human body.

GenBio AI, co-founded by Nobel laureate David Baker, Eric Xing, and other leading researchers from institutions including Stanford, Carnegie Mellon, and the University of Washington, positions AIDO Cell as an early but functional demonstration rather than a complete solution. “What’s exciting here isn’t that we’ve solved cellular biology—we haven’t, at least not yet,” Baker noted. Xing even went as far as comparing AIDO Cell to the current stage to a GPT-1 or GPT-2 moment for biological simulation.

We at Labcritics are skeptical about the model’s ability to mimic cells at the atomic level (which will have real usecases and provide a full understanding), at best with currently available computation, it should be possible to virtually mimic at the molecular levels. We aren’t certain if the AIDO virtual cell even holds the capability to virtually multiply or die or any other real-world scenario’s. Access is currently limited to the company’s team and alpha collaborators, with an early-access program planned for academic, biotech, and pharmaceutical researchers. We will ensure we are in the front row to play with this model and get you first hand reports.

A technical report detailing architecture, benchmarks, and case studies is available upon request. Further versions with expanded cell types and improved accuracy are expected later in 2026 and beyond. By offering a stateful, multi-scale virtual cell, AIDO Cell seeks to accelerate drug discovery, toxicology, and disease research, potentially reducing the high failure rates and costs that have long characterized biological innovation.

Other Companies Working On the Same virtual cell challenge and their current status:

Organization	Key Project / Model	Progress: papers & code
Chan Zuckerberg Biohub / CZI	Virtual Cells Platform, TranscriptFormer, rBio, Billion Cells Project	Most open of the group. TranscriptFormer (Pearce et al., 2025) is a cross-species generative family trained on up to 112M cells across 12 species spanning ~1.53B years of evolution, ~300M transformer parameters, with code at github.com/czi-ai/transcriptformer. Cziscience rBio, a reasoning model distilled from TranscriptFormer’s biology, shipped on the Virtual Cells Platform in Aug 2025 and is usable as a training framework or benchmark. Chan Zuckerberg Initiative The Virtual Cell Challenge framing paper ran in Cell. Cell Press An Oct 2025 NVIDIA collaboration added cz-benchmarks, an open Python evaluation package, plus RAPIDS-singlecell tooling and NVIDIA Clara models (MONAI, CodonFM) on the platform. R&D World Billion Cells Project (with 10x Genomics and Ultima Genomics, launched Feb 2025) is still a data-generation program — the 1B-cell target is not yet delivered.
Arc Institute	STATE model, Arc Virtual Cell Atlas, Virtual Cell Challenge	Code out, peer review not. STATE was released June 2025 under a non-commercial license and described in a preprint on Arc’s site that has not been peer reviewed; Arc reported >50% improvement in perturbation-effect discrimination and >2× accuracy on differentially expressed genes, with the State Transition module trained on 100M+ perturbed cells across 70 contexts. Genengnews Weights and code are on GitHub (ArcInstitute/state) under the Arc non-commercial license, GitHub now at v1.0. Arc Institute Challenge results are the honest headline: 5,000+ registrants across 114 countries, 1,200+ teams submitting and 300+ final entries — but Arc’s own wrap-up states that perturbation prediction models are not yet consistently beating naive baselines across all metrics. Arc Institute Altos Labs took a new Generalist Prize at NeurIPS 2025 with a flow-matching generative model. Genengnews
Xaira Therapeutics	X-Cell	Newest and largest; release still partial. Launched 17 March 2026, trained on X-Atlas/Pisces — 25.6M perturbed single-cell transcriptomes across seven cellular contexts — reaching 4.9B parameters. Business Wire Posted as a bioRxiv preprint, not peer reviewed; reported zero-shot generalization to iPSC-derived melanocyte progenitors and primary human T cells. Genengnews Architecture is diffusion-based with cross-attention over prior biological knowledge; co-senior author Bo Wang previously led scGPT. BiopharmaTrend The GitHub repo (Xaira-Therapeutics/X-Cell) lists X-Cell Mini (55M) on HuggingFace under CC BY-NC-SA 4.0, but states weights and inference code are still in active development. GitHub Predecessor dataset X-Atlas/Orion (8M cells, all human protein-coding genes) was released publicly in June 2025. Business Wire
Insilico Medicine	Virtual Aging Cell (VAC) platform	Announcement stage — weakest evidence in the table. Insilico launched a VAC webpage and previewed a multi-agent platform in August 2026, describing six biological hierarchies with dual top-down/bottom-up Master Agents and Specialist Agents, fused with its PreciousGPT foundation models. PR Newswire Precious3GPT, developed with Harvard Medical School, was open-sourced and feeds VAC’s agentic layer. TipRanks No VAC preprint, benchmark, or code release has accompanied the announcement — the architectural claims are vendor self-reported.
Recursion Pharmaceuticals (Valence Labs)	TxPert and related imaging/transcriptomic models	Only peer-reviewed model here. TxPert was published in Nature Biotechnology in 2026 (Wenkel, Tu, Masschelein et al., doi:10.1038/s41587-026-03113-4); it uses multiple gene–gene knowledge graphs for perturbation prediction and approaches split-half experimental reproducibility on unseen single perturbations across K562, RPE1, HepG2 and Jurkat. Nature Code is public at github.com/valence-labs/TxPert, Valence Labs though only a subset of models is included, since the best-performing ones depend on proprietary knowledge graphs. GitHub Notably small — 8M total parameters. Substack TxFM, a companion transcriptomics foundation model, was presented at ICLR 2026. Investing.com
Noetik	OCTO-VirtualCell	Technical report only, no public code. Announced Dec 2024 as a technical report (not a journal paper) alongside Celleporter, an interactive tool for viewing virtual cell simulations across spatial locations; OCTO-vc is a multi-scale multimodal transformer trained via masked token modeling on nearly 40M cells of proprietary spatial transcriptomics from 1,000+ patient tumor samples. Yahoo Finance Progress since has come through conference presentations — a SITC 2025 poster on genotype-specific fibroblast–macrophage crosstalk in NSCLC, plus an invited talk on predicting treatment outcomes. Noetik No weights or code released.

A few things worth flagging for editorial framing:

Primary Refernece: GenBio AI