Virtual cells: where does competitive advantage come from?

From Recursion to Rivercell, AI strategies in biology draw on different combinations of data, models and experimental capabilities. Arc Institute’s resources add another dimension: part of the scientific foundation is becoming available to the wider community. For executives, investors and industry partners, how can a costly asset be distinguished from a lasting competitive advantage?

At Recursion, experimental maps of biology support the collaboration with Roche and Genentech to identify relationships between interventions and cellular responses, then guide drug discovery programmes. In a related field, Insitro draws on Lilly’s preclinical data to develop models capable of predicting certain pharmacological properties of molecules. In the first case, experimental mapping contributes to the partnership; in the second, access to a pharmaceutical company’s accumulated data plays a central role.

Alongside these companies, Arc Institute makes STATE available, a model designed to represent cells and predict their responses to perturbations, with specific conditions for commercial use. Rivercell, meanwhile, is developing experimental infrastructure to generate the data required by its own models. These approaches outline several paths: generating observations, using a partner’s data, drawing on shared resources, and deciding which capabilities to retain in-house.

Virtual cells therefore put three assets at the centre of the discussion: data, models and the laboratory. Their value depends on the decisions they enable. An exclusive dataset may cover too narrow a domain; a high-performing model may fail when the biological context changes; an automated laboratory may produce large volumes without reducing the cost of a decision. Three questions provide a useful framework: what information is missing, which prediction meaningfully improves the work, and which experiment still needs to be funded to verify it?

A virtual cell must first define what it predicts

The term readily conjures up a cell fully reconstructed inside a computer. The work considered here addresses more specific tasks: predicting, for example, how gene expression will change after an intervention, or how a population of cells will shift between states. The scope of a prediction depends on the measurements used to train and evaluate the model.

This distinction shapes the entire analysis. An accurate prediction of gene expression does not, by itself, establish a therapeutic effect, the absence of toxicity or efficacy in patients. Each step towards a medical decision introduces further conditions that must be tested.

These activities cover different domains, from cellular responses to the properties of molecules. Comparing them therefore means examining organisational choices and the conditions under which value is created. An experimental map, a model of responses to perturbations and a toxicity prediction must each be assessed against the biological question they claim to answer.

The first question to ask a company is consequently a practical one: which decision should its model improve? Selecting a biological target, ranking molecules, ruling out a hypothesis and choosing the next experiment require different data, validation methods and products.

Data is valuable for the differences it helps explain

The number of cells observed is a convenient metric. It reveals less about competitive advantage than the diversity of the situations studied and the quality of the measurements. A million very similar observations may leave the precise question that matters to an industry partner unanswered.

Learning to predict the consequences of an intervention requires distinguishing its effect from changes that would have occurred without it. Doses, exposure times, cell types and experimental conditions shape this comparison. Controls and repeated experiments help separate a biological response from a measurement problem.

In its account of the Virtual Cell Challenge dataset, Arc Institute emphasises experimental design and quality control. This work illustrates how a dataset’s value begins before model training: with the choice of interventions and the conditions that make their effects interpretable.

For a company, generating its own data allows these choices to follow a strategy. It can investigate a family of mechanisms in greater depth, add a missing measurement or explore situations in which its model is uncertain. Exclusivity becomes useful when it provides information that genuinely improves a prediction.

The company still has to demonstrate that improvement. A relevant comparison measures what proprietary data adds while keeping the architecture and budget comparable. If the advantage disappears when a publicly available resource is used instead, the barrier to entry largely amounts to a temporary head start. If it persists in new and useful contexts, the case becomes stronger.

Preserving protocols, metadata and negative results also matters. A measurement without a documented history can become difficult to reuse. An experiment that disproves a hypothesis may, conversely, prevent investment in another dead end. The quality of a company’s scientific assets depends on what it can preserve and interpret.

Open resources redraw the boundaries of exclusivity

This strategy is unfolding alongside large collective initiatives. On 7 October 2026, Biohub announced an expansion of its Virtual Biology Initiative with the US Department of Energy, the National Institutes of Health and partners including Google DeepMind, Isomorphic Labs and Meta. Its stated objective is to generate and make available biological data suitable for model training, using common standards.

These resources remain a programme to be delivered. Their ambition nevertheless changes the strategic calculation: a startup must be able to explain how its data complements what the wider community will be able to use.

Open access can reduce the cost of getting started, provide common reference points and make comparisons easier. It may also weaken the defensibility of a collection whose main strength was its scarcity. Private advantage would then have to come from measurements that are difficult to obtain, better coverage of a specific problem, or the ability to fill gaps in available data quickly.

For European companies, this raises the question of where to concentrate capital. Replicating an entire global infrastructure would require substantial resources. Using shared resources while developing specialised experimental capabilities offers another strategy. Its relevance will depend on the target market and the value pharmaceutical partners place on that specialisation.

A model must perform beyond the contexts it was trained on

Models are the second candidate for competitive advantage. The difficulty lies in measuring performance that reflects their intended use.

An algorithm can perform well on data similar to what it encountered during training. The test becomes more demanding when it must predict the effect of an unfamiliar intervention, in a different cell type, or using a protocol developed elsewhere. Those changes are precisely what matter to a company hoping to avoid measuring every possible situation.

Scientific publications make a case for scrutinising evaluations. In a study published in Nature Methods in 2025, the authors compared several foundation models and other deep learning methods with simple approaches to predicting changes in gene expression after perturbation. None of the models studied outperformed the simple baselines on the tasks examined.

This finding applies to the models and tests in that study; it does not settle the performance of every subsequent generation. It highlights the usefulness of a comparison that commercial presentations can easily leave in the background: how much is gained over a far less expensive method?

The research introducing Systema addresses another difficulty. According to the authors, systematic differences between experimental groups can inflate apparent performance and obscure the challenge of predicting the effects specific to a perturbation. Evaluation must therefore establish what the model has actually learned.

Arc Institute has incorporated this requirement into its Virtual Cell Challenge 2026. Participants must predict responses to genetic perturbations in six cell lines whose perturbed responses are not provided to them. The competition, which is still under way, tests transfer to new cellular contexts.

For an investor, the practical implication is clear: ask for results on data held out from training, a documented evaluation protocol and a comparison with simple methods. For a customer, a prospective test must be added: make the prediction first, then conduct the experiment that will verify it.

The laboratory can build a lead and tie up capital

Rivercell makes control over experimentation part of its strategy. According to the company, existing instruments do not fully meet its data-generation needs, so it is developing a purpose-built platform. This approach puts measurement engineering on the same strategic footing as model training.

The laboratory can shorten the path from a question to a usable answer. An integrated team can modify a protocol, investigate an anomaly and generate the observations needed for a new hypothesis. The expertise accumulated through that work can be difficult to replicate, even when the general methods are published.

This integration comes at a cost. Equipment, consumables, maintenance, scientific teams, quality controls and experiments that need to be repeated all contribute to the platform’s economics. Automation changes how some tasks are performed; it does not guarantee that each additional data point will be cheaper or more useful.

The risk emerges when every contract requires a largely new experimental setup. The company may develop a high-quality scientific business while reusing relatively little of its investment. The promise of a platform must be tested against this reality: how many protocols, datasets and software components actually carry over to the next project?

Owning the laboratory is not the only way to control this chain. A partner can generate data using a jointly defined protocol. The choice depends on confidentiality requirements, the frequency of adjustments, the expertise needed and the expected utilisation of the equipment. Building, leasing or outsourcing becomes a measurable strategic decision.

Cost per cell provides only part of the information needed to assess that choice. The cost of a valid experimental campaign, the time required to confirm a prediction and the share of results that can be reused bring the analysis closer to what the customer is buying.

Contracts reveal who retains the value

Data ownership adds a dimension that is often less visible. A pharmaceutical company may hold a substantial body of experimental data and seek a business capable of exploiting it. In this configuration, the startup provides the model and expertise, while the critical data remains with its partner.

The partnership announced by Insitro and Lilly in September 2025 illustrates this division. It aims to build models of pharmacological properties, including absorption, metabolism and toxicity, using Lilly’s proprietary preclinical data. The companies describe an architecture designed to preserve data confidentiality. The announcement sets out a development objective rather than demonstrating savings already achieved.

The contractual stakes extend beyond initial access. Who can reuse the observations? Who retains the adapted models? Can the results of one partnership improve a service offered to other customers? The answers determine whether an advantage can accumulate from one contract to the next.

The distinction between scientific access and commercial use also applies to open resources. Arc Institute makes STATE available, with tools, models and a route for commercial use. The model licence terms must be examined separately from access to the code. Downloading a resource does not automatically establish the right to incorporate it into a product for sale.

The choice of business model follows from those rights and the work required. Selling software access, charging for an experimental campaign and sharing in the value of a therapeutic programme involve different costs, timelines and risks. The same scientific demonstration can support several businesses with very different economics.

Competitive advantage is measured in the next decision

Recursion provides a commercial point of comparison. In its full-year 2025 results, published in February 2026, the company reported biological maps accepted by Roche and Genentech under their collaboration, along with payments linked to milestones. These developments document a partner’s interest in scientific deliverables. They do not establish clinical efficacy.

This distinction helps clarify the stages of value creation. A model may first help select an experiment; a biological map may then guide a programme; a molecule must subsequently pass its own validation steps. Advantage must be assessed at the stage where the company promises to deliver it.

The most compelling hypothesis for virtual cells is cumulative learning: the model identifies experiments likely to resolve an uncertainty, the laboratory carries them out, and the observations improve subsequent predictions. To build a lasting lead, this loop must produce an observable gain and retain reusable results.

That gain can take several forms: a hypothesis ruled out earlier, a more useful selection of molecules, or fewer experiments needed to answer a defined question. It must be compared with the usual process, including the costs of modelling, measurement and validation.

Data, models and the laboratory can therefore contribute jointly to competitive advantage. None is sufficient to demonstrate it on its own. The next experiment offers a more concrete test: why was it selected, what did it teach, and which decision can now be made?

Five questions to ask before investing or entering a partnership

1. Which specific decision does the model improve? Ask for a use case, a baseline method and a measured improvement in time, cost or the quality of selection.
2. What can it predict in a genuinely new context? Examine the data excluded from training, the test conditions and prospective experimental validation.
3. What advantage does proprietary data provide? Compare performance with and without it, then identify what a competitor would need to fund or learn to reproduce the result.
4. How much does a usable experimental answer cost? Include controls, repeated experiments and scientific work; measure what can be reused in the next project.
5. Which rights and results remain with the company after a contract? Check data reuse, licences, exclusivity and ownership of improvements to the model.