Beyond the Cell Atlas: Fabian Theis on Making Virtual Cells Useful in the Laboratory

1. Guest Background

Fabian Theis leads artificial intelligence and computational biology research at Helmholtz Munich and related institutions. His relevant experience spans both scientific models and the software used to build them. Scanpy, AnnData, and the scverse open-source ecosystem feature in the episode's introduction, alongside work on human cell atlases, virtual cells, and the prediction of cellular responses to perturbations. That combination gives the conversation a practical focus: what should a model enable researchers to do after they have described their data?

Theis also brings an institutional perspective. He estimates that the computational health center he has been building at Helmholtz Munich includes about 45 group leaders and professors and 500–600 people. He describes work across cellular modeling, DNA foundation models, protein design, patient cohorts, and genetics, supported by university and clinical collaborations. Connecting these scales toward a virtual human is an ambition he articulates, rather than an integrated capability demonstrated in the interview.

This roughly 51-minute Latent Space episode examines the transition from observing cells to predicting their behavior and eventually directing experiments. Organoids and generative models supply the scientific examples. Evaluation, commercial incentives, and interdisciplinary supervision explain why progress also depends on how research is organized. Theis is discussing a developing research program, with concrete tools and experimental examples but substantial unresolved questions.

2. What the Episode Covers

The central object in Theis's account is a cell that responds to inputs while carrying memory and multiple layers of molecular state. Predicting those responses is an important test of a model, but he considers the current tendency to equate a virtual cell entirely with perturbation prediction too narrow. The input space itself is much broader than knocking out a gene. Neighboring cells, signaling gradients, hormones, culture media, subtler genetic differences, combinations of interventions, disease, and time all belong in the picture he sketches.

The interview makes this accessible through transcriptomics. Theis approximates a cell's measurements as a sparse vector of about 20,000 messenger-RNA counts. Researchers learn a more manageable latent representation, interpreting clusters as cell types and continuous structures as differentiation trajectories. Those categories are useful human descriptions, he emphasizes, rather than boundaries nature necessarily observes. Organizing cells in a state space helps describe what has been measured; it does not by itself explain where a particular intervention will move them.

His account of expanding data distinguishes two dimensions that headlines can obscure. One is the number of observed cells; the other is coverage of the interventions applied to them. He separately mentions about 10 million transcriptomes measured in one experiment, extensive screens involving roughly 100 cell lines, and batteries of around a thousand drugs. These are illustrative scales from his remarks, not components of one combined experiment. His larger point is that perturbation coverage was historically much sparser than cell sampling, sometimes amounting to little more than control and disease conditions.

Organoids extend the problem into multicellular systems. Theis discusses connected brain organoids, beating heart organoids, and colon and cancer systems. He also describes using skin-derived cells to construct brain-like systems when disease-relevant tissue is difficult to obtain. These examples offer richer biological context than a single cell line, while introducing their own omissions: immune components may be absent, feeding conditions incomplete, and data insufficient. How to represent spatial relationships remains unsettled. Distributions of cells are one current approach; graphs might preserve additional structure, but he does not present a settled solution or a mature vocabulary for local cellular niches.

The modeling discussion then separates representation from dynamics. A foundation model may learn the state x, whereas an action-conditioned model must predict a subsequent state from x and an intervention a. Theis describes CellFlow using flow matching within a learned latent space to generate post-perturbation state distributions. He reports acceptance at Nature Methods and stresses that the method depends heavily on representation quality. The episode supplies no numerical performance gain. Its evidence supports a particular modeling strategy and its dependencies, rather than a general promise of accurate intervention prediction.

3. Core Views: Reasoning, Examples, and Limits

Theis's most useful argument about complexity begins with the task, rather than the architecture. Within an organ, he says, logistic regression remains highly competitive for cell-type classification; CellTypist is his example. Across organs, the problem becomes subtler because a recurring cell type can change with its tissue context. He cites scTab as a case where nonlinearity improved on a baseline. These observations support conditional model selection. They do not establish that deep learning is unnecessary throughout biology, or that every cross-organ task requires a large foundation model.

Nicheformer illustrates a different reason a task might demand more expressive machinery. Its proposed prediction problem uses one cell's expression state to infer the expression of neighboring cells. Theis reasons that ongoing communication can leave traces of the surroundings in a cell's own state. This asks a model to recover environmental information, rather than simply assign a familiar label. Even here, predictive success would need careful interpretation: exploiting a signal associated with neighboring cells does not itself establish the specific causal mechanism of their communication.

His position on inductive bias is similarly conditional. Strong priors can restrict predictive power, he argues, and he expects flexible data-driven models to become more prominent as data grow, with interpretation added afterward. Yet he also says perturbation coverage is not sufficiently rich to dismiss hybrid approaches today. There is no contradiction if data richness is measured against the question being asked. Millions of observations under a narrow range of inputs can still leave the relevant intervention space poorly sampled. Reducing constraints cannot manufacture evidence about conditions the training process scarcely encounters.

That is why borrowing the appearance of language models is insufficient justification for single-cell foundation models. Theis describes an unsettled landscape of masked autoencoders, tabular foundation models, JEPA approaches, and contrastive learning. He wants demonstrated scaling behavior to show that additional complexity earns its place. For scConcept, he reports comparisons up to about 30 million cells in the paper and a Hugging Face version using about 350 million cells and roughly 100 million parameters, with scaling behavior beginning to emerge. These are approximate speaker-reported figures. No numerical improvement curve is supplied, so the interview cannot establish the return on a particular increase in training scale.

The boundaries become especially visible in his discussion of species. The larger latent space, he says, spans 16 species and opens possibilities for cross-species and cross-tissue questions. He immediately qualifies the evolutionary ambition: it is outside the team's core expertise, and he does not claim they already do it well. Likewise, CellFlow's dependence on a good latent space means representation scale cannot simply be converted into intervention-prediction capability. The representation and the action-conditioned transition each require evidence appropriate to their roles.

Simple models remain an essential diagnostic. Theis recalls scGen around 2019 combining nonlinear encoding with linear shifts in latent space. That is more specific than calling it a wholly linear model, and he says such approaches can remain difficult to beat on simple tasks. For subtle knockout effects in cell lines, additional complexity may overfit. In noisy drug data, even beating a mean prediction can be difficult. A baseline victory therefore raises questions about signal strength, task construction, and generalization; it is not automatically a verdict against an entire model family.

Metrics further complicate the meaning of victory. Theis criticizes fixed metrics in the Arc Virtual Cell Challenge as exploitable and says the ultimately recognized winner used flow matching. He contrasts this with Open Problems inviting additional metrics from participants. These are his explanations of the competitions, rather than an independent audit supplied by the episode. The discussion also identifies metric choice as a reason model comparisons can mislead without deliberate gaming. Rankings need to be read alongside holdouts and the biological outcome the evaluation actually rewards.

His broader prediction is that problem and metric formulation may become more important as automated systems search more model combinations. The reasoning is straightforward: search can optimize a specified objective, but it does not guarantee that the objective captures the scientific question. Better optimization of an ill-chosen target can deepen the mismatch. Across these examples, the interview's standard of proof is consistent: scale, nonlinearity, and leaderboard position become persuasive only when tied to a defensible task and an appropriately limited claim.

4. Learning and Application

A practical research plan can begin by naming the required generalization: classification within an organ, transfer across tissues, neighborhood inference, or prediction under an unseen intervention. This is an application of the interview's distinctions, rather than a procedure Theis claims to have validated universally. Record cell counts separately from cellular contexts, intervention types, combinations, and observation times. Then check whether the evaluation actually withholds the conditions the project hopes to predict. Mean predictions, linear models, and simple latent-space transformations provide useful starting comparisons before committing to more expensive approaches.

For organoid work, define the experimental system's scope alongside the model's scope. Specify which cellular relationships are represented and which immune or feeding conditions are missing. If the intended question depends on an absent component, narrow the claim or arrange additional validation. A cell-distribution representation may be easier to connect to existing workflows, while a graph may preserve relationships relevant to the task. The richer representation should earn its cost through the question and available data; complexity alone does not resolve an incomplete experimental system.

Lab Compass provides the episode's concrete bridge from prediction to action. Theis describes an inverse generative model that proposes actions for reaching a desired state, based on a forward world model. He reports three or four experiment–computation iterations with Cambridge hematopoiesis collaborators using a rapid-cycle system. A team attempting a comparable workflow could first agree on the target state, feasible operations, measurements, and how results will update the next proposal. The reported example makes repeated iteration tangible, but supplies neither a universal protocol nor a quantified guarantee of improvement.

The operational boundaries deserve equal attention. Theis says repeated loops remain uncommon. Selecting the next experiment is difficult, experimental collaborators must commit time, unusual model suggestions can conflict with intuition, and assays may take too long for rapid feedback. Treat inverse-model outputs as candidates for joint review, with experimental turnaround and discussion included in the plan. Three or four iterations are meaningfully different from demonstrating a continuously autonomous reinforcement-learning laboratory, which remains a vision in this conversation.

Commercial planning requires another explicit choice: what is the customer buying? Theis observes platform companies moving into drug development, while a host argues that saleable assets can be easier for investors to understand than tool sales to pharma. That observation does not imply every modeling team should become a drug company. The cancer organoid perturbation atlas offers a more specific mechanism: pharmaceutical participants pay for early access to data. A project can therefore distinguish software capability, privileged data access, and drug assets as different deliverables. Paid access illustrates a purchasing arrangement; it does not establish profitability or successful drug discovery.

The Y supervision model suggests how to connect these questions inside a research team. A methods PI and a domain PI jointly propose work, with students ideally participating in both groups' meetings. The useful mechanism is sustained translation of domain uncertainty into measurable problems, not an extra supervisory name. Learning both vocabularies takes time, so that participation needs an actual place in the workload. Finally, verify data availability before designing a broader validation program. Theis contrasts linked-data examples in FinnGen and Denmark with fragmented patient trajectories in Germany and the United States. His suggestion that the European Health Data Space would advance linkage in roughly two years is an interview-time expectation. A project should plan around access it can establish, rather than assume that anticipated interoperability is already available.

Source

• Original episode: Mini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich