AIDO Cell: A General-Purpose Simulator for Cell Biology

Engineering is the practice of applying rational design frameworks to solve real-world problems. Electrical engineers design circuits for voltage and current, chemical engineers compute kinetics and transport, mechanical engineers model strain and stress, and aerospace engineers simulate aircraft flows and trajectories. In each case, designs are simulated in math and software long before they enter the real world. Rational engineering of biological systems has been a long-standing goal in medicine and biotechnology, but until now biology has had no equivalent simulation framework.

The stakes of closing that gap are enormous. Of every 10,000 compounds that enter the drug development pipeline, on average only one reaches the clinic: a process that takes more than a decade and costs over $2 billion per approved drug. Most of that cost is the years spent discovering, often too late, how a promising molecule behaves inside the staggering complexity of real living systems.

The natural starting point for biological simulation is the cell: the smallest unit of life. Cells are complex entities containing several different types of molecules that underpin life. Despite this importance, to date engineering simulations have not been successfully applied to the entire cell. While some components within cells (proteins, expression) have been successfully studied using ML and AI simulations, these models are built around a single assay, modality, or prediction task, fragmenting what is fundamentally one biological system into many separate pieces. Stitching them into pipelines does not fix the fundamental problem: predictions using isolated components do not account for the interplay between all these components within a living cell. Simulating a biological experiment requires a system that holds a cellular state that persists across manipulations, accepts interventions that modify it, and reads out multimodal observations on demand. 

To address this, we developed AIDO Cell, a general-purpose simulator for cell biology.

What Is AIDO Cell?

AIDO Cell simulates how a cell would change in response to experiments, and how those changes appear across various experimental measurements, i.e. a virtual cell. This requires an AI architecture that can be responsive to user commands, but most established AI architectures for image, video, and text modeling are only suitable for prediction, representation, or generation. Instead, AIDO Cell is built on a world model. While other AI models perform predictions all in one go, world models are unique in that they perform predictions on a unified world state, allowing users to take actions midway through a prediction and decode how observations in the real world would unfold in response to those actions. This internal state can be perturbed, redirected, and decoded into experimental readouts from the current cell context. 

Our world model serves as our cell simulation engine. On its own, the world model simulates how the whole cell changes, but turning that capability into a usable research tool requires a control layer wrapped around the model. To effectively apply this engine, AIDO Cell also ships with a world model harness, which keeps track of the cell’s state across every scale over consecutive interventions and translates user commands into steps the engine can execute. Together, the engine and the harness provide an interface for cell biology experiments to be programmed, observed, replayed, and debugged. 

The world model system design has three main benefits:

Coherent multimodal readouts. Different modalities decode from a single shared state. Chromatin accessibility, histone marks, transcription factor binding, expression, isoform usage, protein structure and interaction, localization, abundance, and morphology all describe the same cell at the same moment. Each can be queried without modifying the underlying state. Because they share one underlying representation, the readouts stay consistent with each other, unlike independent predictors. 

Multiscale simulation. The same state can be read across the full hierarchy of scales – from DNA and RNA through regulatory networks to the whole-cell level. Morphology and gene expression change simultaneously, allowing users to zoom in on the gene pathways regulating this change. Users can also zoom out from molecular interventions, propagating their effects through regulation to pathway activity and cell phenotypes.

Continuous multi-turn experimentation. Every state is a starting point for the next operation. There are no limits on how many perturbations can be applied. Cells can be cloned to branch an experimental trajectory or simulate experimental replicates. A resistance mutation and two follow-up treatments can be run as separate arms from a shared ancestor and compared directly. This is closer to how real biology unfolds over multi-step bench experiments, rather than making batches of isolated predictions.

Virtual Cell Benchmark 1.0

Measuring a general system requires a general benchmark. For this, we developed the Virtual Cell Benchmark 1.0, which aggregates public biological atlases into 31 metrics across five task families: small-molecule perturbation, genetic knockout, protein multimer and monomer structure, RNA splicing, and genome regulation.

As the table below shows, our simulation engine, AIDO Cell 1.0, is the only system that spans all five families. All other methods — whether structure predictors like AlphaFold 3 and Boltz-1, regulatory genomics models like AlphaGenome, or single-cell foundation models like scGPT and Geneformer — focus on individual components and so address only one or two of these. 

Despite simulating a much broader range of capabilities, AIDO Cell 1.0 reaches state of the art on 24 of those 31 metrics, matching or exceeding the strongest specialized baseline across nearly every family. On the few metrics where a task-specific model retains a modest advantage—AlphaFold 3 on protein–DNA interactions, isoDDE on protein–ligand docking, AlphaGenome on select regulation tracks—AIDO Cell remains competitive while outperforming those same models across the remainder of their respective domains.

Prototype Virtual Cells

The AIDO Cell 1.0 release is focused on simulating two human cell lines: K-562 and Hep-G2. Over fifty years of scientific literature, these well researched cell lines have served as global reference points for mapping human gene regulation and drug response. K-562 originates from a patient with leukemia and has been a foundation for precision cancer medicines, while Hep-G2 originates from a liver cancer patient and has long served as the standard model for liver drug metabolism and toxicity. The volume of existing multi-omics data and prior literature provides a suitable testing ground to verify simulation accuracy before extending AIDO Cell to other cell types. 

While these two cell lines serve as initial prototypes, they already enable new forms of exploration. For example, we used AIDO Cell for K-562 and Hep-G2 to simulate one million randomized five-plex perturbations per line, spanning gene knockouts, knockdowns, overexpressions, and small molecules. Each node in the resulting atlas is a simulated state and each arrow a perturbation, so a real multi-step experiment plots as a trajectory across the atlas rather than as a scatter of unrelated predictions.

In-Context Molecular Design

Because readouts and perturbations act on one shared state, simulations can be run in both directions. Forward: apply a drug and read the cellular response. Backward: specify a desirable cell state and design molecules that reach it. This second mode turns the virtual cell into a phenotypic molecular design platform, generating molecules for their effect within a specific cell context rather than simply finding a molecule to bind to a target. The user defines a target not as a protein pocket or a binding affinity but as a cellular phenotype, and the system generates candidate molecules, applies them to cloned cells, and characterizes each design across scales in a single loop. We call this capability in-context molecular design.

As a concrete example, we use the K-562 prototype to run a functional analog design campaign against the ABL kinase inhibitor imatinib. Imatinib is applied to one clone to define the target state, and then structurally novel molecules are generated, applied to fresh clones, and profiled. The entire campaign runs in a simple loop:

# Create the reference perturbation
imatinib = "CC1=C(C=C(C=C1)..."
cell = Cell("K-562")
cell_ref_perturbed = cell.clone()
cell_ref_perturbed.small_molecule_perturbation(imatinib)
cell_ref_perturbed.get_expression()

# Generate target-binding molecules, characterize them in cells
for _ in range(N):
  molecule = design_molecule_for_target("ABL1")
  cell_perturbed = cell.clone()
  cell_perturbed.small_molecule_perturbation(molecule)
  cell_perturbed.get_protein_structure("ABL1", molecule)
  cell_perturbed.get_protein_ligand_interactions(molecule)
  cell_perturbed.get_expression()
  cell_perturbed.get_cell_age()

Each candidate is characterized across scales from the same underlying state: 2D chemistry, predicted target co-fold and binding contacts, transcriptional response, and predicted biological age shift. Of the generated candidates, several land near imatinib in the perturbation atlas, recovering 65 to 88 percent of its differentially expressed genes with very different chemistry. The multimodal readouts give a richer characterization of each molecule: designs that match imatinib transcriptionally can diverge on protein binding or have off-target effects, so the screen presents many different ways of ranking molecules. 

This design loop is one of several end-to-end case studies in the detailed technical report, which also covers drug response prediction, combinatorial perturbation screens, structural mechanisms, and disease resistance analysis across the K-562 and Hep-G2 prototypes.

Customizing Virtual Cells with AIDO Foundry

In AIDO Cell 1.0, we demonstrate the engine with prototype virtual cells for K-562 and Hep-G2, two of the most well-characterized and validated human cell lines. In practice, many important biological questions lie in the domains that we have the least data on (i.e. rare diseases), or are fundamentally unable to sample (e.g. living brain tissue). Biological simulators are critically needed in situations where obtaining enough samples is difficult or impossible. This is why we built AIDO Cell as an adaptable foundation that can be rapidly customized to new cells, both cell lines and primary cells, as well as new data modalities and interventions in AIDO Foundry.

AIDO Foundry is an automated research system that updates the virtual cell as new experimental data comes in, so its predictions keep improving with each round of experiments. We already use it to train on genetic and small-molecule perturbations, transcriptomics, protein abundance and localization, and imaging, but it enables adaptation to new data modalities as well. 

AIDO Lab is the interface where the researchers can interact with, manipulate and analyze virtual cells: browsing multimodal readouts, running perturbations in sequence, or designing binders in a specific cellular context.

Together, they are our simulation ecosystem: AIDO Cell is the simulator, AIDO Foundry is how you adapt it to a new biological context, and AIDO Lab is the command center where you experiment and observe the result.

The system is also showing early signs of scaling with data. Holding out Hep-G2 entirely and testing only on drugs only seen in other cell lines, we see improvement on some tasks that scales with the number of training cell lines. This type of cross-context scaling means the system isn’t simply getting better within a single cell line as it sees more of that cell line’s data; it’s using information from other contexts to update biological priors and model weights, improving predictions in contexts it has never encountered.

In practice, cross-context scaling would mean AIDO Cell can be extended to primary cells and disease cells that are not amenable to massive-scale data generation, drawing on what it has learned across other contexts rather than requiring that scale be built up from scratch for each new one. This hints at new types of scaling laws for biological models built on breadth across biological contexts rather than just depth within a single one, an exciting direction that we’re continuing to characterize.

A First Step, With a Clear Path Ahead

AIDO Cell 1.0 is competitive with specialized SOTA models on perturbation, structure modeling, and genome regulation tasks, but it’s also pushing into territory that no existing model attempts. Simulating the multi-scale, layer-spanning dynamics of a cell remains a challenge to both predict and validate. Structure prediction was a fertile testbed for AlphaFold partly because of the CASP competition, a blind, community-run benchmark that let the field check results independently. Once a model cleared the bar for experimental accuracy on CASP, researchers could trust a predicted structure to usually be faithful to a real X-ray crystallography experiment.

Multi-scale simulation needs its own version of this: a benchmark that tells us when a simulation is faithful enough to substitute for a real experiment. Building it will require many combined experimental technologies, such as pooled perturbation screens and multimodal assays. This standard is still nascent, but pushing toward a ‘CASP for holistic cell simulation’ is part of our ongoing work. We built Virtual Cell Benchmark 1.0 as a first version using existing resources, but we think the standard should ultimately be owned by the field rather than any one lab. 

We also validate the multi-step multi-scale simulations against established biology literature in case studies, but wet-lab validation on novel predictions is the ultimate test. These experiments are already underway, but advancing the system’s remaining capabilities, building a shared evaluation standard, and validating at scale are tasks we’re actively looking to partner on with ambitious and bold players at the forefront of AI and biology.

What Comes Next

We’re building toward a full virtual cell bank, a commons where researchers can download custom cells from many different biological contexts and run open-ended experiments on them. However, we cannot build it alone. We are looking for partners on three main tracks: 

1. Using AIDO Foundry to enable simulation of additional readouts in AIDO Cell
2. Expanding the type of cells that the engine can simulate.
3. Using our simulator for K-562 and HepG2 to prioritize experiments, triaging perturbations or narrowing a hypothesis space before committing bench time.

Inside the system, we are working toward a more comprehensive simulation including post-translational modification, degradation and turnover, signaling cascades, metabolites, and temporal dynamics. 

Beyond the system, the cell is a composable unit: something that can be engineered directly into larger systems. A simulator that operates at cellular scale is the first step toward doing this at organism scale. This work serves our broader mission toward creating an AI-Driven Digital Organism (AIDO), a platform where biology can be programmed, observed, replayed, and debugged.

We are far from solving cell biology, but we have built a system that can be interrogated as though it were a cell, that ingests many different types of multimodal data into a single shared world model, and that improves as it observes new biological contexts.

Responsible Development

As with any powerful AI system for biology, AIDO Cell may raise dual-use questions. At this stage, AIDO Cell does not allow viral or bacterial inputs, and so cannot be directly used to optimize viral or bacterial design or infer their impact on human cells. Its generative capabilities are limited to designing therapeutic molecules — small molecules, antibodies, and nanobodies – to rescue diseased cell states and functions. While we hope one day to provide a platform for engineering all of biology, this version is strictly scoped for disease understanding, drug design and disease treatment, not for creating new living organisms.

Even so, access today is granted on a controlled, case-by-case basis under terms that expressly prohibit use in developing any application intended to cause harm. We expect biosecurity to become an increasingly important part of this field and we’re engaging with it directly. We support the development of shared industry standards and GenBio AI’s scientific leadership participates in the responsible-AI-for-biology conversation in their own academic capacities.

Access

AIDO Cell is running today for our own team and a group of alpha collaborators. An early-access program is in preparation for researchers across academic, biotech and pharma. 

Whitepaper

The full whitepaper – architecture, benchmark results and case studies in detail – is available by request below