Rivercell Manifesto

Section 01We aim to learn how perturbations change cells

Observational datasets describe cellular states and associations, but cannot reliably isolate the causal effect of an intervention. To estimate that effect, we apply perturbations—deliberate interventions such as drug exposures, genetic changes or changes in the cell’s environment—and compare responses with appropriate controls.

Single perturbations are not enough to capture the full complexity of cellular responses. Combinatorial genetic screens show that two interventions can produce a response that differs from the sum of their individual effects. ^[2] Changing their order can further alter the outcome.

The number of possible combinations grows rapidly as we vary interventions, doses and sequences. Rivercell will generate large-scale combinatorial perturbation datasets to train the virtual cell to predict these interactions, including how one intervention changes the response to the next.

Section 02We measure perturbation responses across scales and over time in the same cell

Cellular response is fundamentally a multi-scale and dynamic process. A treatment acting on a molecular target can change gene expression, cell shape and behaviour. Similar endpoints can follow a transient response, gradual adaptation or a delayed effect. Understanding these differences therefore requires complementary measurements across scales, linked over time in the same cell.

We will image living cells through perturbation experiments, then collect their endpoint transcriptomes. Linking the two readouts connects changes in morphology to gene expression in the same cell. Scaling this workflow requires hardware designed for programmable perturbation experiments that does not exist today. Our platform will provide that capability.

Linked measurements over time address the monomodal and static limitations of existing data. They still provide only partial evidence of the cell’s underlying biological state: the molecular activities and organisation that determine its response. Protein activity and metabolism, for example, remain partly unobserved. The virtual cell must infer the cell’s state from the available measurements and the history of interventions, while representing what remains uncertain.

Section 03We scale experiments across biological contexts

A useful virtual cell must generalize outside of its training set to predict how an intervention will work in biological contexts it has not yet seen. This limits the range of biology over which a model can predict, even when it has been trained on millions of cells.

Our high-throughput platform will test the same perturbations across cell types, genetic backgrounds and environments. Responses learned from one cell type cannot be assumed to hold in another. By making context an experimental variable, we can measure how it changes a response and train the model on those differences.

Landmark perturbation atlases have greatly expanded the data available to train virtual-cell models. Tahoe-100M contains 100 million transcriptomic profiles across 50 cancer cell lines. ^[5] Liu and colleagues’ multimodal atlas includes approximately 65 million single-cell profiles across 1,000 CRISPR knockouts in one cell line, A549, with live-cell imaging at single timepoints. ^[6] By measuring responses to deliberate interventions, these atlases address the observational gap and provide a foundation for a new generation of virtual-cell models. Broader coverage of cell types, genetic backgrounds and environments is the next step towards reliable prediction in unseen contexts. ^[3]

Rivercell will use scale to broaden biological coverage, testing the same interventions in diverse contexts while retaining enough replication to measure variability. We will evaluate predictions in contexts withheld from training to establish where they generalise and where further experiments are needed.

Section 04We build a virtual cell that can design the next experiment

Even at massive experimental scale, we can test only a fraction of the possible perturbations, sequences and contexts. Choosing the next experiment is therefore critical. Language-based reasoning can generate hypotheses; establishing how cells will respond requires experimental evidence. We lack general equations to derive those responses, and testing them in living cells takes time. Our virtual cell must learn from each result, refine its predictions and select experiments that resolve the most consequential uncertainties.

In Human Compatible, Stuart Russell explains how an AI agent makes decisions when it can observe only part of its environment. It maintains a belief state, an uncertain estimate of what is happening, and updates it using past actions and new observations. ^[4] A virtual cell must do the same for the living cell it models. Our world model will learn to predict how cellular state evolves after a perturbation and revise its estimate as linked measurements arrive.

The model will select experiments expected to resolve uncertainties that matter for prediction: which perturbations to apply, in which order, in which context and when to measure. An experiment is informative when competing explanations predict different outcomes. Selection must account for what the platform can execute and what the result could teach us.

Section 05We build a programmable platform that executes the experiments at scale

Generating longitudinal, multimodal perturbation data across diverse biological contexts requires a wet lab built for programmable experiments at scale that does not yet exist. Our platform will be the experimental component of the loop, producing the evidence the virtual cell requests. A question about an untested sequence or context must translate into an experiment that supplies the missing evidence.

In an ideal experiment, imaging follows individual cells continuously through a controlled perturbation. The virtual cell uses these observations to decide whether to deliver a second intervention, and at what dose and time. An endpoint transcriptome links the cell’s molecular response to its observed history. Controls and replication distinguish treatment effects from variability. Repeating this design across diverse contexts addresses the four data limitations together.

The model’s choices must translate directly into executable protocols. A common interface for experimental instructions and results will let it request new evidence and revise the next round as that evidence arrives.