AIDO Cell 1.0

Ask what a kinase inhibitor does to a leukemia cell and three separate models return three separate answers: a binding pose, an expression shift, a change in morphology. None is obliged to agree with the others, and none of them remembers — apply a second perturbation and every predictor restarts from an untouched baseline, not from the cell the first drug already changed.

AIDO Cell is genbio.ai's attempt to replace that pipeline with a simulator. It is built on a world model: one latent state stands for the cell, and interventions act as operators that advance it. Chromatin accessibility, histone marks, transcription factor binding, expression, isoform usage, protein structure and interactions, localization and morphology all decode from that single state, describing one cell at one moment, and reading any of them leaves the state untouched. A control harness tracks that state across interventions and turns commands — knock out a gene, add a compound, clone the cell — into engine steps.

Version 1.0, announced on 18 August 2026 as an early functional preview, ships prototype cells for two human lines: K-562, from a leukemia patient, and Hep-G2, the reference model for liver drug metabolism. Where scGPT, Geneformer and X-Cell model the transcriptome, AIDO Cell treats transcription as one projection of a state that also spans regulation, splicing, structure and phenotype. It shares a brand with, but is not, AIDO.Cell — GenBio's 2024 scRNA-seq encoder, whose checkpoints and parameter count belong to that model alone.

#Key Features

• Stateful, multi-turn simulation: The cell retains everything already done to it, so a resistance mutation followed by two treatments runs as one accumulating trajectory rather than three unrelated predictions. Cells can also be cloned to branch parallel experimental arms from a shared ancestor.
• Coherent multimodal readouts: Nine classes of measurement decode from one shared representation, so they stay consistent with each other rather than disagreeing.
• Multiscale propagation: A molecular intervention propagates through regulation and pathway activity to whole-cell morphology, and a phenotypic change traces back to the genes driving it.
• In-context molecular design: Run backward, the simulator takes a desired phenotype rather than a binding pocket, generates molecules, applies them to cloned cells and profiles each across scales.

#Technical Details

GenBio's formalism for a virtual-cell world model, set out in a May 2026 position paper by Eric Xing and Le Song, couples three components: a multimodal encoder onto a latent cellular manifold, an action-conditioned transition core that evolves that state under an intervention, and a generative decoder constrained to cross-modal consistency. AIDO Cell's public description follows it. The architecture, parameter count, training corpus and compute for version 1.0 have not been published; the technical report is available by request, and no code, weights or preprint have been released.

Evaluation is on Virtual Cell Benchmark 1.0, which GenBio assembled from public atlases: 31 metrics across five task families — small-molecule perturbation, genetic knockout, protein monomer and multimer structure, RNA splicing, and genome regulation. GenBio reports AIDO Cell 1.0 as the only system spanning all five and as state of the art on 24 of the 31 metrics, with AlphaFold 3 retaining an edge on protein–DNA interactions, IsoDDE on protein–ligand docking and AlphaGenome on selected regulation tracks; Boltz-1 is among the structure baselines. Those numbers are GenBio's own, on GenBio's own benchmark, without independent replication — as are the million-perturbation atlases built for each line and an ABL1 design campaign whose molecules recovered 65–88% of imatinib's differentially expressed genes with unrelated chemistry.

#Applications

The intended use is triage: simulating perturbation screens, drug response and resistance in K-562 and Hep-G2 to narrow a hypothesis space before committing bench time, and designing molecules against a cellular phenotype rather than a purified target. Hep-G2's role as the reference line for drug metabolism makes toxicology an early fit. AIDO Foundry adapts the engine to new cell types and modalities as data arrives; AIDO Lab is the interface for running perturbations and browsing readouts. Access is limited to GenBio's team and alpha collaborators, with an early-access program in preparation.

#Impact

AIDO Cell reframes the virtual cell as something to act on over many turns rather than a predictor to query once, and is the first to hold regulation, splicing, structure and morphology in one stateful simulator. It is the opening stage of GenBio's AI-Driven Digital Organism program, whose roadmap Le Song, Eran Segal and Eric Xing set out in a Nature Medicine Perspective. The claims warrant care: the headline case study recapitulates imatinib's known mechanism rather than predicting an unknown one, wet-lab validation of novel predictions is still underway, and the benchmark behind the state-of-the-art claim was built by the same lab — a gap GenBio acknowledges. Coverage is two immortalized cancer lines, the system accepts no viral or bacterial inputs, and its generative scope is confined to therapeutic small molecules, antibodies and nanobodies.