Fabian Theis on Scaling Virtual Cell Models with Biohub’s $1.8 B Data Push

The $1.8 B data bet

Biohub announced a $1.8 B commitment from NIH, DOE, Google, Meta and its own funds to build “lab-in-the-loop” automation and new imaging platforms such as laser phase-plate and light-sheet microscopy. The initiative aims to generate co-registered, multi-scale datasets that span molecules, cells, tissues and whole organisms. The source links the timing of the investment to the maturation of these measurement technologies, which now can produce the volume and precision required for training large AI models.

Why current virtual-cell benchmarks fall short

The 2025 Arc Virtual Cell Challenge revealed that “constant baselines (not even linear!) often beat sophisticated models for ‘Perturbation Prediction’.” Papers cited in the interview note that many genes barely change after perturbation, so a naïve mean prediction can score highly under poorly calibrated metrics. A re-evaluation using better-calibrated metrics shows deep-learning models can surpass the mean baseline, but the underlying task remains insufficiently challenging.

Shifting the problem to scale interfaces

Theis argues that the real scientific frontier is predicting how changes at one biological scale propagate to others—molecules → cells → niches → tissues → organs → organisms. He contends that the new Biohub data will enable models to learn these “scale interfaces,” a problem that current datasets cannot support. His labs at Helmholtz’s Computational Health Center and the Institute of Computational Biology are already integrating single-cell time series, population-scale omics, imaging and clinical data to build mechanistic, predictive models for precision medicine.

Our read: The interview reframes the virtual-cell narrative from incremental benchmark improvements to a data-centric ambition that matches the scale of the $1.8 B investment. By exposing the limits of current perturbation-prediction tasks and highlighting the need for multi-scale, co-registered data, Theis makes a compelling case that the next wave of AI-biology breakthroughs will depend less on algorithmic tricks and more on the breadth and fidelity of the biological measurements fed into them.