Fabian Theis has been a central figure in various iterations of the Virtual Cell concept. His lab has advanced algorithmic approaches for modeling multi-scale, multi-modal ("multiomics"), time-varying biological data. Theis has shaped large-scale vision and data collection through leadership roles at Helmholtz, involvement in startups, and contributions around Biohub.
What happened: a large bet and new measurement capabilities
Biohub’s recent announcement received wide coverage for the scale of the commitment: $1.8 billion in pledges from Biohub, NIH, DOE, Google, and Meta. The initiative emphasizes new imaging and measurement capabilities that will produce large datasets and make those data available for training AI.
Part of the renewed momentum comes from advances in infrastructure and measurement technologies: DOE and NIH are developing lab-in-the-loop automation at scale, Google is expanding connectomics capabilities, and Biohub has been working with laboratories to develop and refine imaging platforms such as laser phase plate microscopy, light sheet microscopy, and advanced post-processing for conventional microscopes to reveal deep phenotypic data. These technologies are maturing quickly and are likely to be central to Biohub’s plans.
Why Virtual Cell modeling has been challenging
Although multi-scale, multi-modal data are exactly the kind Theis has spent his career modeling, interest in Virtual Cell projects cooled. The 2025 Arc Virtual Cell Challenge highlighted a key issue: several papers showed that constant baselines (sometimes not even linear) often outperform sophisticated models on the task of perturbation prediction. Perturbation prediction asks how a cell’s state will change when inputs or programming are altered (via gene editing, drugs, or environmental factors).
That said, the situation is not entirely bleak: recent work has demonstrated that many evaluation metrics are poorly calibrated and can reward an uninformative mean prediction because most genes change very little. With better-calibrated metrics, deep learning models can outperform mean baselines.
Theis’s perspective: the problem itself is the challenge
Fabian Theis argues that, given current datasets, perturbation prediction is not hard enough to justify very complex algorithms. Instead, he believes attention should shift toward the very problems the Biohub effort aims to address: using large, high-precision, multi-scale, co-registered biological data to predict dynamics from molecules → cells → niches → tissues → organs → organisms, and to understand how these scales influence one another.
This is an opportunistic moment: AI is enabling work on much larger problems, and the problem here is huge. There is a lack of understanding of the "scale interfaces" between different domains of applicability and a lack of integrated data. The Biohub initiative aims to solve the data gap. Fabian Theis has been developing algorithms intended to model these larger-scale problems for years.
Practical research questions and goals
Some of the increasingly complex questions driving this work include:
- How do neighboring cells change my cell’s state?
- What happens to a distribution of cells after a perturbation?
- How can I perturb a cell to bring it into a desired state?
These questions point toward lab-in-the-loop, large-scale experimentation and require increasingly complex models.
Models and precision medicine
Theis’s labs and related initiatives have invested heavily in infrastructure: the Computational Health Center at Helmholtz integrates AI, molecular data, medical imaging, genomics, and clinical data to develop predictive and mechanistic models for precision medicine. The Institute of Computational Biology develops statistical and dynamical models that integrate biological information from single-cell time series up to population-scale omics data.
As a leading figure at this frontier, Theis discussed how he views the Virtual Cell concept at a time when what counts as a difficult problem is changing rapidly. Other topics covered included the business of Virtual Cells, the implications of the "Bitter Lesson" for biology, training the next generation of computational biologists, and closing the lab-in-the-loop.
Conclusion
Interest in Virtual Cell approaches may revive as measurement technologies and scaling infrastructure advance. The large-scale data collection announced by Biohub and partners could address a fundamental shortage: integrated, multi-scale, co-registered biological datasets. Fabian Theis has been working on the algorithms and institutional structures needed to build reliable, scale-aware models that aim to control and predict biological dynamics.
