Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich
Topics
AI in Basic Science and Materials Discovery
AI Training Data and Evaluation Benchmarks
AI Drug Discovery and Protein Computation
Early-Stage VC Investing Dynamics
Pioneers
Fabian Theis
Summary
- The virtual-cell opportunity is widening as perturbation data catches up with cell measurements. Theis points to ten million transcriptomes measured in one experiment and increasingly broad drug and CRISPR screens; RJ cites Biohub’s announced $1.8 billion collaboration. The virtual-cell question, in Theis’s framing, is predicting cell behavior under all kinds of perturbations. He cautions that equating a virtual cell with perturbation prediction is too simplistic: the ambition includes a cell’s input-output behavior, memory and molecular levels beyond RNA. Pharma needs cell-type-specific screens it cannot yet run at scale.
- Organoids expand the opportunity into multicellular systems, while adding unresolved modeling problems. Theis’s example is “beating little heart cell clumps in a dish”: interactions emerge that a single cell line cannot capture. Missing immune components, limited data and unsettled representations of cellular neighborhoods keep this a research frontier.
- Model complexity needs to earn its place against simple baselines and meaningful metrics. Logistic regression remains competitive for within-organ cell typing, while noisy drug-response prediction can struggle to beat the mean. Theis says ARC’s challenge also exposed exploitable metrics; his instruction is “Do your baselines properly,” then demonstrate value on harder questions.
- Learning a cell’s state and predicting an intervention’s effect are separate modeling jobs. Theis describes a contrastive representation effort that reached 350 million cells across 16 species, with perhaps 100 million parameters. Perturbation prediction adds “the state X plus the action,” using a generative model such as CellFlow rather than assuming representation learning alone solves the problem.
- Unique experimental data offers a clearer commercial moat than a virtual-cell model alone. Theis has repeatedly seen platform companies pivot to “We make a drug,” although the hosts point to emerging structural-biology and reinforcement-learning businesses following other paths. His Cancer Organoid Perturbation Atlas illustrates another buyer proposition: pharma pays for early access to distinctive experimental datasets.
- Europe’s financing weakness appears at company scaling, in Theis’s account, rather than initial formation. Early venture funding and nondilutive support exist, but larger Series A and B rounds often pull companies toward US investors. His Helmholtz center supplies substantial research capacity—around 45 group leaders and 500–600 people—while he pushes each group toward at least one spinout within three years.
- Experiment selection is becoming practical, while biological reinforcement learning remains an ambition. Lab Compass works backward from a desired cellular state to proposed interventions; Theis describes three or four experimental feedback cycles with a Cambridge hematopoiesis collaborator. The aspiration is “we want to RL everything.” RJ characterizes current approaches as “maybe active learning,” a description Theis affirms.
Deep dive
1. Perturbation data makes virtual cells a more actionable proposition
- RJ opens with Biohub’s announced $1.8 billion collaboration with DOE, NIH, Meta and Google to generate open, multiscale biological data. Theis connects that effort to the Human Cell Atlas, while qualifying his perspective: “my perspective is quite academic.” His expectation is that generative modeling and experimental feedback, already progressing at the molecular level, can increasingly move up to cells.
- The basic measurement is a transcriptome: roughly 20,000 dimensions recording messenger-RNA counts in a cell, with many zeros. Latent representations compress that space; clusters suggest cell types, while trajectories can describe differentiation. Brandon’s analogy makes the distinction accessible: DNA supplies ingredients, and “each cell is its own sort of meal or dish,” whose composition transcriptomics attempts to recover.
- What changed is coverage of interventions, not simply the number of observed cells. Theis recalls ten million transcriptomes measured in a single experiment, alongside screens extending to roughly 100 cell lines or batteries of 1,000 drugs. Previously, enormous cell collections often sampled a narrow perturbation space—perhaps control versus disease. Richer intervention coverage creates substantially more demanding prediction problems.
- A perturbation is broader than a CRISPR knockout: neighboring-cell signals, hormones, media conditions, subtle genetic changes and combinations all belong in the input space. Knockouts are useful but unusually drastic. Theis hopes models eventually incorporate human genetics as perturbations; the practical need is pharma’s inability to screen broadly while remaining “cell type context specific.”
- Theis cautions that a virtual cell is not simply synonymous with perturbation prediction. He describes the deeper ambition as modeling the cell’s input-output behavior, memory and different molecular levels—not just RNA. He calls equating virtual cells with perturbation-prediction models somewhat simplistic.
2. Multicellular context makes organoids compelling and difficult
- For Theis, organoids demonstrate why researchers need to move beyond isolated cell lines. Cells can organize into small systems with emergent behavior: his example is “beating little heart cell clumps in a dish.” Such systems offer richer experimental readouts because cells connect and communicate, rather than merely supplying more observations of one cell type under another treatment.
- Access to otherwise scarce tissue is another attraction. Theis describes taking skin cells, differentiating them into small brain-like systems to potentially study neurodegenerative disease genetics. He immediately keeps the limitations attached: organoids can lack immune components and adequate feeding arrangements. They reproduce useful pieces of organs, with substantial gaps.
- How to represent these systems remains unsettled. Cell types provide a useful human categorization—“nature doesn’t care about” those labels—but local cellular niches lack an equally established vocabulary. Spatial omics and pathology therefore raise both conceptual and computational questions. Current models often treat organoids as distributions of cells; graphs might capture subtler relationships, but Theis says there is no agreed optimal representation.
3. Simple baselines survive until the biological question gets harder
- On the hosts’ inductive-bias challenge, Theis argues that imposing priors too strongly can sacrifice predictive power. Perturbation data is nevertheless still limited enough that hybrid models “might be still the thing to go for.” His AlphaFold example is the shift toward learning rotational symmetry from data; longer term, he expects that “data beats human-made additional concepts in these models.”
- His 2019 scGen combined nonlinear encoding with linear shifts in latent space, and simple alternatives remain formidable. For within-organ cell typing, logistic regression tools such as CellTypist remain close to the state of the art. Cross-organ classification introduces subtler tissue-dependent differences; his scTab work found gains from nonlinearity there. The biological question determines whether additional complexity actually contributes.
- On ARC’s Virtual Cell Challenge, Theis says fixed metrics could be exploited; in his account, he thinks the model eventually crowned the winner was essentially based on flow matching. He says that, setting aside exploitation, the actual winner incorporated nonlinearities. RJ adds that a Bo Wang Nature paper traced some apparent linear-model victories to metric choice. Theis’s Open Problems approach asks participants to submit additional metrics, making the definition of success itself part of the scientific work.
- Drug-response prediction supplies the sobering baseline: sometimes the challenge is simply beating the mean because measurements are so noisy. Theis’s instruction to students is “Guys—I mean, guys and girls—I know you want to do the fanciest deep learning thing, but please have the baseline.”
4. A useful virtual cell needs both representations and action dynamics
- Nicheformer illustrates a harder task than assigning a cell label: can one cell’s expression state predict its neighbors’ expression? Communication leaves traces of the surroundings in the measured cell. Theis says stronger nonlinear models become useful when researchers ask for this subtler information. Whether existing data contains untapped signal therefore depends on the question posed, rather than a blanket verdict about foundation models.
- Training objectives remain open territory. The field has borrowed masked-autoencoder approaches from language modeling, while exploring tabular foundation models, JEPA-style methods and contrastive learning. Theis says early biological foundation-model papers often failed to demonstrate scaling laws. Reviewers should require evidence that complexity delivers improving performance across increasing examples and samples, rather than accepting model size as its own justification.
- His contrastive representation work compared against Geneformer at around 30 million cells; the version on Hugging Face reaches 350 million cells, with perhaps 100 million parameters. That larger latent space spans 16 species, potentially supporting cross-species and cross-tissue comparisons. Evolutionary questions become conceivable, although Theis explicitly says they are not his group’s core expertise.
- Intervention prediction adds dynamics to those representations: the model takes a state and an action, then predicts the next state. “I would not claim that pure representation learning is the right way to go,” Theis says of perturbations. His group adds a generative push-forward model using flow matching; he says CellFlow, accepted at Nature Methods, improves perturbation generation but depends heavily on a good latent space.
5. Germany’s research capacity meets a financing and data-access constraint
- Helmholtz’s federal funding supports sustained research programs rather than university-style undergraduate teaching. Theis describes his Computational Health Center as roughly 45 group leaders and professors, with perhaps 500–600 people spanning cellular modeling, DNA, protein design and patient cohorts. He thinks it is Europe’s largest academic center for health AI research. University and clinic affiliations supply training and translational connections; bringing these scales together supports a broader “virtual human” ambition.
- Against a host’s concern about German startup bureaucracy, Theis argues that formation is increasingly competitive, especially in Munich and Berlin. He says the EU creates just as many startups as the US. His diagnosis is that Europe drops off when companies scale: early venture and nondilutive funding exist, but investors taking larger risks are less available. He has repeatedly seen companies seek US funding for larger Series A and B rounds, while acknowledging that this simplifies the picture.
- Health-data integration remains another friction. Theis cites Finland’s FinnGen and Denmark as examples where population sampling and medical records are connected better than elsewhere; in Germany and the US, fragmented records make patient trajectories difficult to build. He points to the European Health Data Space as a mechanism for cross-border combinations, where “maybe in the next 2 years, these things need to start working across borders.” The vision is developing, he says, but implementation is slower than desired.
6. Platform companies still need a buyer and a defensible data advantage
- Theis’s commercialization warning starts with companies promising “We decode biology” without specifying what customers will buy. Platforms are attractive, but consulting does not automatically become a scalable business. He has watched many virtual-cell companies eventually announce “We make a drug,” because investors understand funding an identifiable therapeutic asset more readily than an open-ended modeling platform.
- The hosts push back with a changing market: structural-biology businesses, including Chai, and companies selling reinforcement-learning environments suggest alternatives to owning drug pipelines. They would like comparable activity around systems biology. Theis does not dismiss that possibility; he redirects attention to what those businesses can uniquely generate and provide.
- His example is the Cancer Organoid Perturbation Atlas, a consortium built with multiple pharma partners. Companies buy in because they want early access to perturbation datasets. Coupling modeling to distinctive experimental production is where he sees the “sweet spot”: the model comes with information customers cannot readily obtain elsewhere, giving pharma a concrete reason to participate.
- Theis also questions how many standalone AI platforms the market can sustain. He raises the possibility that “Astra 2.0” might eventually perform parts of the work, adding pressure to differentiate beyond the model.
7. Dual supervision turns biological ambiguity into computable questions
- The Munich School for Data Science embeds a methods supervisor and a domain supervisor in a “Y” structure. Students attend both lab meetings, learning biological or medical language while developing computational methods. Theis’s point is that important domain problems often arrive incompletely formulated; a quantitatively trained student can turn that ambiguity into a metric, a model and a research project.
- After roughly eight years, he sees graduates becoming both capable analysts and model builders, with destinations spanning pharma, research, startups and pure AI. NVIDIA hired two of his people in one month a few months before the conversation. Their advantage, he argues, includes having helped define a difficult real application, beyond demonstrating a small benchmark improvement.
- The approach now spans spatial measurements and single-cell proteomics alongside RNA; Theis mentions a Science proteomics paper from the previous year. Commercial translation is becoming explicit, too: he has asked every group leader to aim for at least one spinout within three years, while saying students increasingly push toward startups themselves.
8. Lab Compass moves the model into repeated experimental decisions
- Theis’s progression is “from mapping, we went to prediction,” followed by steering the next experiment. Conventional academic projects often stop at a model, a concept figure and limited validation before moving on. Repeated feedback would make learning more continuous, but he cautions that genuinely running multiple computation–experiment loops remains uncommon, even where organizations promote the idea.
- Lab Compass addresses the inverse problem: given a forward model and a desired cellular state, what intervention should produce it? Theis describes an inverse generative model intended to answer that question. At the time of recording, he expected the work on bioRxiv “today or tomorrow”; selecting an optimal next experiment, he emphasizes, is itself a difficult modeling problem.
- With a Cambridge collaborator working on hematopoiesis, the group established three or four experimental loops that iteratively improved a model. The enabling factor was a relatively rapid experimental system, alongside close collaboration. Persuading experimentalists to spend substantial time on an unfamiliar model-recommended intervention remains difficult; biological intuition can pull the team toward compromises instead of following the proposed experiment.
- Robotics may eventually help, but Theis thinks it is a little early while assay cycles remain long. “We want to RL everything” expresses the destination, which he calls “a bit out there.” RJ characterizes present approaches as “At the moment it’s more maybe active learning,” and Theis affirms that characterization.