Biohub and Partners to Direct $1.8 Billion Toward Data for Virtual Cells

Biology has repeated a story familiar to every machine learning team. You build a complex model, spend compute, assemble a polished demo, and then the simple rule "predict the average" performs just as well. This is exactly what virtual cell projects have run into: systems meant to predict how a cell will respond to a drug, a gene edit, or a change in environment.

After the results of the 2025 Arc Virtual Cell Challenge, it might have seemed that large models simply don't work here. But the problem turned out to be more interesting. Some of the tests were designed so that a safe, averaged answer earned a high score: most genes barely change after an intervention, so a model could ignore rare but biologically important shifts.

Fabian Theis of Helmholtz Munich believes the debate over whether a neural network beats the average is steering the field off course. If a dataset describes a task that is too simple and narrow, a complex model truly has no room to show its advantage. The next step is not to add layers for the sake of layers, but to collect data that include time, space, a cell's environment, and transitions between levels of the organism.

Why the average suddenly became a strong rival

Consider a set of 10,000 genes. After an intervention, only a few hundred change noticeably, while the rest stay almost where they were. A model that outputs a typical value for each gene will often land close to the right answer. The overall error will be small, even though it misses precisely the interesting changes.

For a developer, this is the ordinary problem of imbalanced data. A fraud detection system can reach high accuracy by always answering "transaction is normal" if fraudulent payments are rare. A high overall score does not mean the product solves the task at hand.

That is why virtual cell models should be evaluated separately on substantially changed genes, new types of interventions, and unseen combinations of conditions. It is also useful to check whether a prediction preserves biological relationships, rather than merely reducing the average numerical error.

From a cell snapshot to a model of a process

Theis suggests looking beyond a single measurement. A cell's state depends on neighboring cells, tissue, time, and the history of interventions. A model should answer not only the question "what will change after the drug," but also predict the distribution of possible states, the system's further development, and a way to steer it toward a desired state.

This is where the initiative from Biohub and its partners, with stated commitments of about $1.8 billion, makes practical sense. The money is planned to go, among other things, toward new methods of imaging, measurement, and automated laboratories. Such setups can close the loop: an algorithm chooses the next experiment, a lab runs it, the results return to the model, and the model refines the next step.

This is an important shift. In a language model, a new training example is relatively cheap. In biology, each data point may require growing cells, equipment, reagents, and time. So value is created not only by architecture but also by the ability to choose the most informative experiment.

What teams should do about it

The first takeaway is universal: before training an expensive model, fix simple baselines: the average, the last known value, a linear model, and a nearest neighbor. If the neural network does not beat them, check the data, the train/test split, and the metric first.

The second takeaway is that the test should reflect the decision that will be made based on the forecast. For selecting a therapy, what matters is not the average error across thousands of nearly unchanged features, but the ability to find a small group of changes that affect the outcome.

Finally, complexity is justified where it matches the complexity of the observed process. Multiscale models are needed if the data truly link molecules, cells, tissues, and time. Without such data, a "virtual cell" risks remaining an expensive interpolator. With them, it can become a working tool for choosing experiments and developing precise therapy.