Fabian Theis on Virtual Cells, Organoids and Closing the Lab Loop

RJ Honicky and Brandon Anderson introduce Fabian Theis, who leads computational health research at Helmholtz Munich. They point to his group’s contributions to single-cell analysis tools including Scanpy, AnnData and scverse, as well as models of cell behavior.

The hosts ask why virtual cells are attracting attention as a large Biohub collaboration plans to generate multiscale biological data. Theis cautions that his perspective is academic. He expects advances in data collection to help move laboratory-guided modeling from molecules toward cells.

His lab began modeling perturbation responses in 2019 with scGen. Theis says relatively simple models remain competitive partly because many available datasets and questions have been limited. A fuller virtual-cell problem asks how a cell with memory and multiple molecular layers responds to many kinds of inputs.

Theis says pharmaceutical screening needs models sensitive to cell type and context. He hopes current commercial interest will be driven by that need and better experiments, rather than simply attaching an AI label to existing work.

He recalls the growth of single-cell measurements from thousands of observations to experiments with millions of transcriptomes. Historically the range of perturbations was much sparser than the number of cells measured; larger genetic and drug screens are beginning to make that input space richer.

To explain transcriptomics, Theis says cells generally share a genome but read out different genes. A cell’s measured messenger-RNA counts can be represented as a high-dimensional vector; groups and trajectories in that space can help researchers describe cell types and differentiation.

Anderson compares the genome to a recipe book and transcriptomics to identifying the ingredients used in a dish. He describes a CRISPR knockout as one way to perturb a system and observe its effects. Theis agrees but emphasizes that a complete knockout is only one, relatively drastic kind of perturbation.

Cells also receive signals from neighbors, hormones, gradients and their growth medium. Theis treats these as part of a broader input or perturbation space, including combinations and changes over time. A useful virtual-cell model would eventually have to account for more than one isolated gene edit.

Theis describes organoids: cells grown into small multicellular systems, including neural structures whose cells connect and heart-cell clusters that beat. Such systems can help researchers study conditions for which tissue is hard to obtain, although they remain incomplete representations of organs.

For example, an organoid may lack immune components or other supporting systems. Modeling cell neighborhoods is also an open problem: the field does not yet have a universally accepted way to represent spatial relationships, and data remain limited.

The hosts ask about the tension between the “bitter lesson” of scaling data and the scientific need for interpretable, simplified models. Theis says strong prior assumptions can improve interpretability or robustness but may reduce predictive power if they constrain the model incorrectly.

Because perturbation datasets are not yet as rich as ordinary cell observations, he thinks hybrid approaches still have a role. Over time, more flexible data-driven models may learn regularities that researchers previously built in, with interpretation added afterward.

Theis distinguishes easy and difficult benchmarks. Logistic regression can perform well on some cell-type classifications, while a task such as predicting neighboring cells’ expression from one cell’s state asks for subtler information and may reward stronger models.

He says the field is still testing representations and training objectives rather than having settled on a best method. Researchers have borrowed masked modeling from language models, explored contrastive approaches and are increasingly expected to demonstrate meaningful scaling behavior.

Theis describes scConcept as a representation model with observed scaling across larger datasets, including a version trained on hundreds of millions of cells from multiple species. He stresses that representation learning alone is not the whole perturbation-prediction problem.

He frames a perturbation model as taking a cell state plus an action and predicting a new state. His group combines learned representations with a generative approach such as flow matching; he says CellFlow performed better at generating perturbation effects when supported by a good latent space.

The hosts return to a basic machine-learning lesson: compare against strong simple baselines. Theis says logistic regression can remain near state of the art for some within-organ cell-type tasks, while cross-organ generalization is harder and may benefit from nonlinear models.

For perturbation benchmarks, Theis says dataset splits, holdouts and metrics require care. Discussing the Arc Virtual Cell Challenge, he says some fixed metrics could be exploited and that the ultimately recognized winning approach used nonlinear flow matching. The lesson is to ask whether the metric captures the biological question.

Even for current models, he tells students to compare with simple predictions, including a mean baseline for noisy drug-response data. The hosts add that metric choice can change the apparent ranking of methods; a more elaborate model is not evidence of progress by itself.

The conversation moves to research funding and Helmholtz Munich. Theis describes a federally funded research organization with university and clinical connections. Its computational health center brings together groups working at molecular, cellular and patient scales, without the same undergraduate teaching role as a university department.

He sees increasing interest in moving academic discoveries into startups. Asked about Germany, he says bureaucracy and digital processes can be obstacles, but believes the larger gap is scaling: early funding exists, while larger later-stage rounds may lead companies to seek US investors.

Theis also discusses health-data access. He points to Nordic examples of connecting population samples and medical records, while noting that interoperability remains difficult in both Germany and the United States. Cross-border health-data initiatives could help, but implementation takes time.

Building a sustainable virtual-cell business is not straightforward. Theis observes that some platform companies eventually emphasize drug assets because those offer a clearer path to monetization. The hosts discuss companies trying alternative models and the difficulty of selling tools to large pharmaceutical buyers.

Theis sees value in combining modeling with distinctive experimental data. He describes a cancer-organoid perturbation atlas consortium in which pharmaceutical partners support data generation and receive early access. In his view, data creation and modeling together can offer more value than a model alone.

The hosts ask about his “Y model” for training students. At the Munich School for Data Science, two principal investigators—one focused on methods and another on a domain problem—jointly supervise a project. Students ideally attend both groups’ meetings and learn each side’s language.

Theis says quantitatively trained students can translate a biological question into a measurable problem and develop methods with experimental partners. Over years, the work has extended from RNA to areas including single-cell proteins and spatial measurements. Graduates have gone into research, pharma, startups and AI.

Theis describes a progression from mapping cells to predicting responses and then choosing the next experiment. Papers often stop after a model and limited validation; the opportunity is to feed experimental results back into the model repeatedly.

He introduces LabCompass as an inverse generative approach: given a forward model and a desired state, it proposes what intervention might reach that state. He describes several iterative lab cycles with a Cambridge collaborator working on blood-cell development.

Closing that loop is also organizational. Experimental work takes time, so computational researchers must persuade lab colleagues that a proposed follow-up is worth doing. Theis sees active learning as a present approach and reinforcement learning or more automation as longer-term possibilities, not solved problems.