The Virtual Cell: The Foundation Model Biology Needs

Biology is a data-rich discipline, but its most valuable information remains fragmented and locked in disparate silos. Unlike language or vision models that leverage vast, coherent datasets, biological data is multimodal, protocol-dependent, and sparse. This fragmentation limits our ability to make universal discoveries and predict cellular behavior. To overcome this, we need a biology-native foundation model - the Virtual Cell - trained on a unified, standardized data layer. Such a model would be a living, learning representation of cellular state and dynamics, capable of accelerating discovery while keeping experiments and ethics at the forefront.

What Is the Virtual Cell?

The Virtual Cell is a dynamic, AI-powered model of cellular state that learns directly from real experiments across modalities such as microscopy, single-cell omics, proteomics, spatial assays, and perturbation screens.

At a glance, the Virtual Cell should be able to:

• Integrate observations into a shared, calibrated cell-state;
• Predict how that state shifts under perturbations;
• Explain why via pathway‑level attribution; and
• Quantify uncertainty while flagging out‑of‑distribution inputs.

Rather than acting as a static catalog of past observations, it behaves like a flight simulator for cells: a safe and rapid environment where researchers can explore how living systems might respond under different conditions. In practice, observations from diverse assays are translated into a common representation of cell state; that shared space captures identity, context, and perturbation effects. A dynamics component then models how that state changes through time when the cell is exposed to a drug, a genetic edit, or an environmental shift. Finally, modality-specific decoders project the evolving state back into expected measurements - an image, a transcriptomic profile, or a protein signature - so that simulations can be compared to reality and used to guide experiments. Crucially, the model does not just produce point predictions; it communicates how confident it is and recognizes when an input lies outside what it has learned before. In this way, the Virtual Cell becomes both a predictive engine and an instrument for explanation.

Training & Learning Strategy

Building such a model is as much about the learning objectives as it is about architecture.

Core levers include:

• Self-supervision across modalities (masked omics, image↔molecule alignment, temporal consistency);
• Perturbation‑aware objectives over dose and time;
• Explicit modeling of batch/site effects with shared controls;
• Biological priors from pathways, transcriptional regulation, and chemistry; and
• Active learning that proposes the next experiment.

These levers turn diverse, partially labeled measurements into a coherent training signal. Because biology revolves around interventions, the model should learn the difference between pre- and post-states, predict dose-time response surfaces, and respect basic biological shapes such as saturation or monotonicity when appropriate. Site- and batch-effects must be treated as first-class variables rather than nuisances: by modeling them explicitly - and anchoring corrections to shared reference controls - the model retains biology while discounting artifacts. Prior knowledge deserves a seat at the table. Pathway graphs, TF-target relationships, and chemical structure can regularize the latent space and help the system discover mechanisms instead of memorizing correlations. Even modest mechanistic simulators, expressed as differentiable modules, can be distilled into the learned dynamics so that hard-won intuition informs the model without over-constraining it. Finally, an active-learning loop closes the bench↔model cycle: the Virtual Cell proposes measurements where uncertainty is highest or confounding is likely, and fresh data flows back to refine the model.

Evaluating a Virtual Cell

A foundation model for biology should be judged on its generalization, its access to mechanism, and the reliability of its own predictions.

Key criteria:

• Generalization to unseen contexts - new cell types, compounds, edits, doses, and labs;
• Causal alignment of explanations with known interventions and directions of effect;
• Faithful temporal forecasting and cross‑modal inference; and
• Well‑calibrated uncertainty with robust OOD detection.

The most telling tests ask the model to predict outcomes it has not seen before, and to justify those predictions in ways consistent with experimental evidence. Because experiments unfold in time, the model must forecast trajectories rather than single snapshots, interpolating missing intervals and extrapolating beyond the last observation. Throughout, the system should reveal what it knows and what it does not, so scientists can decide when to measure rather than trust. A public, versioned benchmark suite with prospective validation keeps progress honest and comparable.

High‑Impact Use Cases

From drug discovery triage to guided reprogramming and differentiation, the virtual cell touches many corners of the life sciences. In drug discovery, it can estimate efficacy and toxicity across cell types, shrinking the search space before wet‑lab screening. For target identification, it performs pathway attribution and proposes combination design to overcome resistance. Genome‑editing programs turn to it for CRISPR screen design, concentrating measurements where they will be most informative by choosing doses, timepoints, and perturbations that reveal pathway structure. Safety teams benefit from translational safety read‑across, gaining an early warning system for adverse morphologies and signatures across tissues. Day‑to‑day practice improves as an experimental copilot recommends controls, flags batch risks, estimates power, and previews likely failure modes before a plate ever hits the incubator.

Why Generic Foundation Models Fall Short in Biology

Generic models for text or images succeed because they rely on a small number of dominant, interoperable formats and are trained on massive, naturally pooled data from the web. Biology, however, is fundamentally different. It is multimodal, pulling from microscopy, omics, proteomics, and other assays, each with its own characteristics. It is protocol-dependent, so results vary with reagents, instrument calibration, and even the operator. The data is sparse and counterfactual, with countless unmeasured combinations of cell types, perturbations, doses, and times; causal inferences require careful design. And it is metadata-sensitive - without rich, standardized context, batch effects can overwhelm true signal. A generic model cannot simply be repurposed to handle these complexities. A biology-native model must be multimodal, causal-aware, metadata-rich, and uncertainty-calibrated by design.

The Data Silo Problem (and Why It Blocks Discovery)

In practice, biology’s most valuable datasets are scattered across labs, file systems, and ad‑hoc pipelines. Inconsistent semantics - from channel names and plate maps to QC flags - mean different things in different places. Heterogeneous pipelines for segmentation, normalization, and feature extraction make outputs hard to compare even when the raw images look similar. Missing metadata about cell provenance, passage number, dose and time, instrument settings, or environmental conditions is common. The result is wrangling over insight: models overfit to local artifacts and fail to travel, and cross‑study synthesis becomes brittle enough to prevent testing universal hypotheses. The virtual cell cannot learn genuine regularities until the underlying data is expressed in a coherent, standardized form.

Solution: CellDAO and the Open Morphology Standard (OMS)

CellDAO addresses both semantics and incentives so that high-value data can be pooled without erasing lab identity. At its core is the Open Morphology Standard (OMS), a shared schema that travels with the data and spells out what was measured, how it was measured, and under what conditions. Rather than listing every possible field, OMS focuses on essentials: provenance of the biological sample (cell line or primary material, including passage and culture context when available), precise perturbation descriptions (chemical structures and doses; genetic edits such as CRISPRi/a or knockouts; environmental shifts like hypoxia or temperature), details of acquisition (instrument, objective, channels and stains, exposure, plate layout), and the processing steps that turned raw measurements into features (segmentation, feature extraction, normalization, and quality flags). The dataset also carries its license and clear attribution.

CellDAO pairs this common language with practical tooling and aligned incentives. Validators and converters reduce friction at ingestion time and help labs become OMS-compliant without rewriting their stack. Open SDKs make it straightforward to read, validate, and transform datasets, while reference controls and well-characterized synthetic data keep benchmarks fair. Contributors receive transparent credit and lineage tracking records how shared data flows into downstream results. Together, OMS and CellDAO provide the coherent corpus and the motivated community required to train, evaluate, and continually improve the Virtual Cell.

Conclusion

A biology-native foundation model will not emerge from algorithms alone. It requires a shared language for data, incentives that reward contribution, and a culture of reproducibility. The Virtual Cell - trained on OMS-compliant datasets coordinated through CellDAO - offers a path from fragmented observations to predictive, mechanistic understanding. By building the commons alongside the model, we can compress cycles of discovery and deliver safer, more effective therapies faster.