Attention Deficit · virtual cell
Biology · simulation · evaluation

The Virtual Cell Is an Experiment Engine

The new claim is learned biological state: models trained across perturbations, omics, structures, and images that can be queried like an experiment engine.

The practical test is whether the model helps choose the next real experiment.
perturb
clone
read out
compare
next test
Zoom out

How did we get here? Scientific computing already modeled parts of biology; AI is trying to learn the shared state underneath them.

01 · Why now

The field is moving from one-off predictions to experiment loops.

The older pattern was a model for one assay, one modality, or one target. The newer ambition is a learned cell state that can be perturbed, read out, and used to plan the next test.

interventionPerturb

Apply a drug, knockout, knockdown, overexpression, or combination.

observationRead out

Ask for expression, structure, regulation, morphology, or other measurements.

decisionChoose next

Use the prediction to prioritize the next wet-lab experiment.

The useful outcome is a better next experiment.
Sources: GenBio AI-Driven Digital Organism Cell; AI virtual-cell priorities paper.

Decision-making is the anchor.

The clear way to talk about a virtual cell is as an experiment loop. The model proposes or evaluates a perturbation, produces readouts, and helps decide what evidence should be collected next.

02 · What changed

The difference now is learned biological state.

AI virtual cells are being built from large biological data across measurements and scales. The model learns representations from data instead of only executing hand-written equations.

perturbation
drug & gene effects
omics
cell state
structure
protein context
imaging
morphology
benchmarks
transfer tests
The model improves when new experimental contexts update the representation.
Sources: AI virtual-cell priorities paper; GenBio AI-Driven Digital Organism Cell release; Arc 2026 Virtual Cell Challenge.

Scientific computing is the baseline.

Computational modeling is already part of biology. The new claim is that large neural models can learn a transferable representation of cellular state from growing experimental data.

03 · What kind of AI

A world model keeps state between interventions.

GenBio describes AIDO Cell as a world model: a unified latent cell state that can be updated by an intervention and decoded into several predicted readouts.

Apply intervention
latent statebaseline
Decode predicted readouts
expression
baseline
regulation
baseline
morphology
baseline

Illustrative interaction: the intervention updates one model state, then decoders produce several kinds of output.

The latent state updates once, then decoders turn it into several predicted readouts.
Source: GenBio AIDO Cell release, world-model and multi-turn experiment description.

One latent state supports several readouts.

AIDO Cell is described as a learned latent state. An intervention updates that state, decoders produce the requested readouts, and the control layer preserves the state across consecutive interventions.

04 · Outcome

Drug discovery is the pressure point.

The useful direction is experiment triage and molecular design: test many candidates in silico, then spend lab time on the ones most worth checking.

10,000 to 1

GenBio says roughly one compound reaches the clinic for every 10,000 entering the drug-development pipeline.

65-88%

In a K-562 imatinib case, GenBio says several candidates recovered this share of imatinib's differentially expressed genes.

AI helps when it narrows which wet-lab paths deserve time.
Source: GenBio AIDO Cell release and K-562 imatinib example.

Triage is the business case.

A useful model can reduce the number of unpromising wet-lab paths or prioritize better ones earlier. That changes cost, time, and scientific attention.

05 · The hard part

The test is transfer to new biology.

Benchmark papers warn that models can look strong under familiar splits and lose performance under new cell contexts, new perturbations, or cross-dataset evaluation.

gate 1Unseen cell context

Does it generalize beyond the cells it already knows?

gate 2Unseen perturbation

Does it handle a new intervention?

gate 3Cross-dataset transfer

Does performance hold across labs and platforms?

gate 4Wet-lab result

Does the prediction survive a real experiment?

GenBio
Arc
Biohub
benchmarks
The proof is a prediction that holds in a context the model has never seen perturbed.
Sources: virtual-cell benchmark papers, Chan Zuckerberg Biohub, Arc State, Virtual Cell Challenge.

Validation is the story.

The field is building models, data infrastructure, and benchmarks at the same time. The honest question is whether the systems transfer to new biological contexts and change wet-lab decisions.