Virtual Organisms and Digital Biology
Virtual-organism claims sit at the highest-risk end of computational biology. A virtual cell asks what happens inside a cellular context after perturbation. A virtual organism must account for development, tissue coupling, physiology, immune state, behavior, microbiome, environment, and time. The phrase is useful only when the modeled organism, decision, input data, uncertainty, and falsifying experiment are explicit.
Use this chapter to:
- Move from cell-level models toward organism-scale representations that include development, tissues, physiology, behavior, and environment.
- A virtual organism is not a larger virtual cell; the organism, genotype, life stage, environment, intervention, and phenotype must be named.
Prerequisites: Systems Biology and Multiscale Modeling for scale transitions; Perturbation Prediction and Virtual Cells for cellular intervention response; Neuroscience AI and Brain Foundation Models for brain and behavior modeling context.
Summary: Move from cell-level models toward organism-scale representations that include development, tissues, physiology, behavior, and environment. Bounded subsystem models and model-organism resources are useful, but general organism simulation is beyond current evidence.
Key point: A virtual organism is not a larger virtual cell; the organism, genotype, life stage, environment, intervention, and phenotype must be named. Open question: whether organism-scale models can predict measured phenotypes beyond the subsystem they were trained on.
Bottom line: Virtual-organism work connects systems biology, neuroscience, aging, plant biology, toxicology, therapeutics, and ecological context.
What is this field trying to solve? Move from cell-level models toward organism-scale representations that include development, tissues, physiology, behavior, and environment.
What is the core idea? A virtual organism is not a larger virtual cell; the organism, genotype, life stage, environment, intervention, and phenotype must be named.
What is the current state of the field? Bounded subsystem models and model-organism resources are useful, but general organism simulation is beyond current evidence.
What do we know, and what remains open? Known reference points include OpenWorm, WormBase, ZFIN, MGI, IMPC, TAIR, PlantCV, zebrafish embryo atlases, mouse atlases, PhysiCell, Morpheus, and whole-cell modeling. What remains open is whether organism-scale models can predict measured phenotypes beyond the subsystem they were trained on.
Why does this matter? Virtual-organism work connects systems biology, neuroscience, aging, plant biology, toxicology, therapeutics, and ecological context.
Introduction
The phrase “digital twin” is easy to use and difficult to earn. In engineering, a digital twin usually refers to a computational representation of a physical system that is updated from measurements and used for a defined decision. In biology, the term is often applied more loosely. A cell-state model, organoid model, animal model, crop model, and patient-specific clinical model all represent different systems and require different evidence.
A virtual organism is a broader biological ambition: a computational representation of a living organism that accounts for enough development, physiology, behavior, and environment to support prediction. It may be reference-based, such as a species or strain model. It may be individual-specific, such as a laboratory animal, crop line, or patient-oriented model. It may also be task-specific, such as a model of zebrafish embryo development after a defined perturbation.
The distinction matters because evidence does not transfer automatically. A virtual cell model that predicts a transcriptional response after CRISPR perturbation does not predict growth, fertility, behavior, host defense, lifespan, field yield, or drug toxicity. A mouse model that predicts a phenotype in one inbred strain does not predict human response. A crop model calibrated under controlled conditions does not predict performance under drought, soil heterogeneity, pest pressure, and farm management unless those variables are measured and tested.
Whole-cell modeling provides a precedent in a minimal organism (Karr et al., 2012). Virtual-cell roadmaps and benchmarks describe the current cell-scale ambition (Bunne et al., 2024; Roohani et al., 2025). Moving from these anchors to organisms multiplies the validation burden. Organisms have bodies, histories, environments, and behaviors. They are not just larger state vectors.
Regulatory and translational communities already recognize that computational models require credibility assessment before decision use. FDA model-informed development materials describe model use as decision-specific, with verification, validation, and applicability assessment rather than generic trust (FDA, Model-Informed Product Development; FDA, Computational Modeling and Simulation Credibility). The same discipline applies before any virtual-organism claim is used in life-sciences research.
Definitions: Virtual Cell, Virtual Tissue, Virtual Organism, Digital Twin
| Term | Appropriate Meaning | Common Overreach |
|---|---|---|
| Virtual cell | A model of cellular state or response to defined perturbations | Claiming whole-cell behavior from one readout |
| Virtual tissue | A spatial or multicellular model of interacting cells in tissue context | Treating simulated pattern formation as validated physiology |
| Virtual organism | A model of organism-level development, physiology, behavior, or phenotype | Inferring organism response from molecular or cellular data alone |
| Digital twin | A data-linked model for a specific system and decision context | Using “twin” as a label for any simulation or dashboard |
| Model-organism simulation | A computational model grounded in a species such as C. elegans, zebrafish, mouse, or Arabidopsis | Generalizing from one species, strain, or life stage to another |
The evidence boundary is practical. A digital twin should name the physical system, data streams, update rule, uncertainty, and decision rights. A virtual organism should name the species or strain, genotype, developmental stage, environment, modeled mechanisms, intervention, and phenotype endpoint. Without those details, the phrase is too broad to evaluate.
Model-Organism Evidence Anchors
C. elegans: Anatomy, Connectome, Development, and Behavior
C. elegans is a tractable model for virtual-organism work because its body plan, lineage, genetics, and nervous system are unusually tractable. The adult hermaphrodite nervous system was mapped at synaptic resolution in the classic connectome work by White and colleagues (White et al., 1986). Single-cell atlases added developmental molecular context, including a lineage-resolved molecular atlas of embryogenesis (Packer et al., 2019). WormBase remains the central curated knowledge resource for genetics, strains, anatomy, and literature (WormBase).
OpenWorm is an open project aimed at whole-organism simulation for C. elegans (OpenWorm). The project is valuable as infrastructure and as a public test case for integrative biological simulation. Its limits are equally instructive. Even in a small nervous system, behavior depends on biomechanics, muscles, neuromodulation, sensory context, developmental state, environment, and measurement choices. A connectome is a major input, not a complete organism.
Zebrafish: Development, Imaging, and Embryo-Scale Perturbation
Zebrafish are tractable for organism-scale digital biology because embryos are accessible, development is rapid, imaging is strong, and genetic tools are mature. ZFIN provides the curated organism database layer for genes, phenotypes, anatomy, publications, and community resources (ZFIN).
Single-cell developmental atlases created a foundation for modeling embryogenesis. Farrell and colleagues mapped zebrafish embryogenesis at single-cell resolution (Farrell et al., 2018). Saunders and colleagues then generated a zebrafish single-cell atlas of perturbed embryos with individually resolved developing embryos across genetic perturbations and time points (Saunders et al., 2023). This is close to the kind of evidence that virtual-organism research needs: genotype perturbation, developmental time, organism-level sampling, and cellular readouts.
The limitation is endpoint scope. A developmental atlas with cellular abundance and transcriptomic readouts supports embryo-development claims. It does not automatically support adult physiology, behavior, toxicology, immune response, fertility, or ecological fitness.
Mouse: Mammalian Physiology, Knockout Phenotyping, and Brain Atlases
Mouse biology is closer to human mammalian physiology but far harder to simulate. Mouse Genome Informatics provides the curated genetics and phenotype knowledge layer (MGI). The International Mouse Phenotyping Consortium (IMPC) generates systematic knockout phenotypes and provides a public portal for mammalian gene function (Groza et al., 2023; IMPC).
Large mammalian cell-atlas efforts provide important tissue-level reference maps. The whole mouse brain atlas work illustrates the scale of modern reference building: millions of cells, spatial transcriptomics, and hierarchical cell-type definitions (Yao et al., 2023). These atlases support cell-type mapping and tissue context. They do not create a general mouse simulator.
Mouse extrapolation is especially difficult because strain, sex, age, microbiome, housing, diet, circadian timing, handling, immune history, and laboratory environment all influence phenotype. A model that ignores these variables may reproduce a dataset while failing as an organism-level predictor.
Plants: Genotype, Environment, Development, and Field Conditions
Plant virtual organisms have a different evidence problem. Development and physiology are tightly linked to light, temperature, water, soil, nutrients, pathogens, and management. A plant model that works in a chamber may fail in a field because the environment is part of the organism’s effective state.
Arabidopsis thaliana provides the reference genetics layer for plant biology through The Arabidopsis Information Resource (TAIR). High-throughput phenotyping tools such as PlantCV support image-based plant measurement pipelines (PlantCV). Crop modeling and plant digital twins often combine physiology, phenotyping, weather, soil, remote sensing, and management data. That makes them useful for agricultural decision support when the context is specified, but weak for general claims across geography and growing conditions.
The plant lesson applies broadly: organism models must include environment. A genotype-to-phenotype model without environmental measurement is incomplete for most whole-organism questions.
Development, Physiology, Behavior, and Environment
Organism-scale modeling requires four layers that are often collapsed:
Development
Development is not just a time label. It changes cell types, tissue interactions, regulatory programs, morphology, and sensitivity to perturbation. The same genetic or chemical perturbation may have different effects depending on embryonic stage, postnatal stage, aging state, or regeneration state. A virtual organism therefore needs explicit developmental staging.
Physiology
Physiology connects tissues through circulation, metabolism, endocrine signaling, immune function, neural regulation, and homeostasis. A liver transcriptomic model, cardiac electrophysiology model, or immune-cell model may be valid locally while missing whole-body compensation. Physiological prediction requires measured endpoints such as growth, heart rate, oxygen use, hormone levels, glucose handling, inflammation, fertility, or survival.
Behavior
Behavior adds sensory input, nervous-system state, motor output, learning, motivation, and environment. C. elegans and zebrafish are attractive because behavior is measurable at scale, but behavior remains context-dependent. A neural or connectome model should not be treated as a behavior model unless the behavioral task, stimulus, environment, and measurement are specified.
Environment
Environment is not noise. It is part of the causal structure. Temperature, diet, light, microbiome, cage density, field conditions, infection history, soil, water, and chemical exposures alter phenotype. Virtual-organism work that omits environment is often closer to controlled-condition simulation than organism prediction.
Current Models, Datasets, and Benchmarks
There is no CASP-like benchmark for virtual organisms. Evidence comes from organism-specific resources and task-specific validation.
| Resource or Model Family | Organism or Scope | Evidence Role |
|---|---|---|
| Whole-cell model of M. genitalium | Minimal bacterium | Whole-cell modeling precedent (Karr et al., 2012) |
| OpenWorm | C. elegans | Open integrative simulation project (OpenWorm) |
| WormBase | C. elegans and related nematodes | Curated genetics and biology resource (WormBase) |
| C. elegans connectome | Nematode nervous system | Anatomical wiring reference (White et al., 1986) |
| C. elegans embryogenesis atlas | Nematode development | Single-cell lineage and molecular context (Packer et al., 2019) |
| ZFIN | Zebrafish | Curated organism database (ZFIN) |
| Zebrafish embryogenesis atlas | Zebrafish development | Single-cell developmental reference (Farrell et al., 2018) |
| ZSCAPE perturbed embryos | Zebrafish development | Embryo-scale perturbation and cellular phenotype data (Saunders et al., 2023) |
| MGI | Mouse | Curated genetics and phenotype knowledge (MGI) |
| IMPC | Mouse | Systematic knockout phenotyping (Groza et al., 2023) |
| Whole mouse brain atlas | Mouse brain | Spatial and transcriptomic tissue reference (Yao et al., 2023) |
| TAIR | Arabidopsis thaliana | Curated plant genetics resource (TAIR) |
| PlantCV | Plants | Image-based phenotyping toolkit (PlantCV) |
| PhysiCell and Morpheus | Multicellular systems | Virtual-tissue simulation, not organism replacement (Ghaffarizadeh et al., 2018; Starruss et al., 2014) |
| Virtual Cell Challenge | Cell response prediction | Cell-scale benchmark frame (Roohani et al., 2025) |
The practical benchmark unit is the phenotype. A virtual zebrafish embryo benchmark might ask whether a model predicts cell-type abundance changes after a genetic perturbation at defined developmental stages. A mouse benchmark might ask whether a model ranks knockout phenotypes in held-out genes, strains, sexes, or environments. A plant benchmark might ask whether predictions transfer from controlled conditions to field plots. A behavior benchmark might require held-out stimuli and environments.
Digital Twins Versus Virtual Organisms
Digital twins and virtual organisms overlap, but they should not be used as synonyms.
A digital twin is usually decision-specific. It represents a particular system, receives data updates, and informs a defined decision. In biomedical settings, that decision might be dose selection, device performance, disease progression, manufacturing control, or patient monitoring. FDA model-informed development and computational-model credibility documents are useful because they focus attention on verification, validation, applicability, and the regulatory question rather than labels.
A virtual organism is usually biology-specific. It represents an organism, species, strain, developmental process, physiological subsystem, or experimental condition. It may not be individual-specific. It may not update continuously. It may be used to test biological mechanisms rather than support a near-term operational decision.
The strongest work is explicit about the distinction. A zebrafish embryo model trained on perturbed single-cell data is not an individual digital twin. A patient-specific tumor model is not a whole-organism simulator. A crop growth model linked to weather and sensor data may be closer to a decision-specific digital twin, but only for the agronomic outcome it has been validated against.
Limits of Extrapolation
Virtual-organism claims fail when extrapolation is hidden. Common extrapolation gaps include:
- Species: nematode, zebrafish, mouse, plant, and human biology are not interchangeable
- Strain or genotype: inbred lines, wild-type references, mutants, and outbred populations differ
- Developmental stage: embryo, juvenile, adult, aging organism, and regenerative state differ
- Sex: endocrine, immune, metabolic, and behavioral phenotypes may differ
- Environment: temperature, light, diet, microbiome, housing, field conditions, and exposures alter phenotype
- Readout: transcriptomic, imaging, physiological, behavioral, fertility, and survival endpoints test different claims
- Intervention: knockout, knockdown, overexpression, chemical exposure, infection, diet, stress, and injury are different perturbations
- Laboratory or field setting: controlled measurements often fail under real environmental variability
The extrapolation issue is not a reason to reject virtual-organism research. It is the reason to state the context of use. A model may be highly useful for one strain, stage, environment, intervention, and endpoint while being invalid elsewhere.
Evidence Standards for Organism-Scale Claims
Organism-scale claims should be graded by the phenotype and decision:
| Claim | Minimum Evidence | Stronger Evidence |
|---|---|---|
| Cell-state prediction inside an organism | Matched atlas or perturbation data | Held-out genotype, stage, tissue, and independent assay |
| Developmental phenotype prediction | Time-staged embryo or organ data | Prospective perturbation across stages and individuals |
| Physiological prediction | Measured functional endpoint | Independent cohort, strain, sex, environment, and dose |
| Behavioral prediction | Defined task and measurement system | Held-out stimulus, environment, and biological replicate set |
| Plant or crop outcome prediction | Genotype, environment, and management metadata | Field validation across locations and seasons |
| Digital twin for decision support | Data update rule, uncertainty, and decision context | Verification, validation, applicability assessment, and monitored drift |
| Experiment replacement claim | Direct comparison against replaced experiment | Prospective replacement study with error tolerance defined in advance |
Evidence strength increases when the study is prospective, perturbs the system, includes organism-level phenotypes, uses held-out contexts, reports uncertainty, and compares against simple baselines. It decreases when results are only retrospective, only molecular, only one strain, only one environment, or only one readout.
Failure Modes
Treating a Cell Model as an Organism Model
The most common overreach is to forecast organism phenotype from a cellular representation. Cell-state predictions matter, but organisms include tissue architecture, physiology, behavior, and environment.
Treating an Atlas as a Simulator
Atlases are references. They organize cells, tissues, and developmental states. They do not by themselves define intervention dynamics.
Ignoring Environment
Controlled-condition data may be necessary for calibration, but organism phenotypes are environment-dependent. This is especially visible in plant, behavior, immune, metabolism, and aging studies.
Hiding the Endpoint
“Phenotype prediction” is too broad. A model may rank a transcriptomic signature while failing on survival, fertility, growth, movement, or disease severity.
Overgeneralizing From Model Organisms
Model organisms are invaluable because they make biology measurable. They are not smaller humans, generic vertebrates, or universal organisms. Extrapolation needs a biological argument and supporting evidence.
Replacement Claims Without Replacement Studies
Replacement of animal, organoid, field, or clinical experiments is an empirical claim. It requires a direct study showing that the computational approach meets a defined error tolerance for the same context of use.
Practical Validation Checklist
Before accepting a virtual-organism or digital-biology claim, document:
- Organism: species, strain, genotype, sex where relevant, and life stage
- System boundary: cell, tissue, organ, whole organism, population, or field plot
- Environment: laboratory, cage, tank, dish, chamber, field, diet, light, temperature, microbiome, and exposures
- Intervention: genetic, chemical, environmental, infection, injury, diet, management, or no intervention
- Phenotype: molecular, cellular, anatomical, physiological, behavioral, reproductive, survival, yield, or disease endpoint
- Time: developmental stage, exposure duration, sampling time, and follow-up duration
- Data update: static reference, periodic update, continuous sensor stream, or one-time prediction
- Baseline: simple statistical baseline, mechanistic baseline, known genotype-phenotype prior, or domain standard
- Uncertainty: measurement error, parameter uncertainty, biological variability, and extrapolation risk
- Validation: held-out genotype, organism, stage, environment, laboratory, field site, or prospective experiment
- Decision: what action changes because of the output
- Falsifier: the organism-level result that would weaken the claim
The checklist should be completed before discussing replacement, translation, or field deployment.
What is demonstrated?
Demonstrated capability includes bounded organismal subsystems with strong measurement. Examples include C. elegans anatomy and developmental atlases, zebrafish embryo single-cell developmental maps, mouse knockout phenotyping infrastructure, plant image phenotyping pipelines, and virtual-tissue simulation tools. These resources support defined scientific tasks.
Demonstrated capability also includes organism-aware perturbation datasets. Zebrafish perturbed-embryo single-cell data and mouse knockout phenotype resources are closer to organism-scale evidence than observational atlases because they connect genotype or perturbation to measured phenotype.
Demonstrated capability includes digital models for local experimental planning. A model may prioritize which gene, stage, tissue, condition, or phenotype to measure next. That is a valuable use case when the next experiment remains part of the workflow.
What is theoretical?
Theoretical capability includes hybrid organism models that connect cell atlases, perturbation data, developmental trajectories, physiology, behavior, and environmental measurements. This is a plausible research direction in tractable model organisms, especially when experiments are high-throughput and phenotypes are measured repeatedly.
Theoretical capability also includes decision-specific digital twins for experimental organisms, crops, organoids, or preclinical models. The key requirement is a narrow context of use: a defined organism, data stream, update rule, endpoint, and decision.
Theoretical capability includes preclinical triage systems that use organism-scale data to rank compounds, genetic targets, environmental exposures, or experimental conditions. Evidence should come from prospective tests, not only retrospective fit.
What is beyond current capability?
Beyond current capabilities includes a general virtual organism that forecasts phenotype from genome and environment across arbitrary contexts.
Beyond current capabilities includes replacement of animal, organoid, field, or clinical experiments without a direct replacement study in the same context of use.
Beyond current capabilities includes reliable human translation from model-organism simulation alone. Model organisms provide evidence, mechanisms, and experimental tractability. They do not remove the need for human, clinical, or field validation when those are the target contexts.
Beyond current capabilities includes patient-specific whole-body biological twins for treatment selection without clinical validation. That claim belongs partly to clinical AI, but the biological-modeling burden begins here.
What would make this more promising?
A virtual-organism claim moves upward when it survives organism-level perturbation and context shift. The strongest evidence would include:
- Prospective prediction before the organism experiment is run
- Held-out genotypes, strains, sexes, developmental stages, or environments
- Organism-level endpoints, not only molecular intermediates
- Time-course validation across development or response trajectories
- Independent laboratory or field validation
- Direct comparison against the experiment or assay proposed for replacement
- Clear error tolerance linked to the decision
- Calibrated uncertainty under distribution shift
- Public benchmark tasks with held-out organism contexts and simple baselines
The claim weakens when performance depends on one strain, one laboratory, one developmental stage, one environmental condition, one endpoint, or one retrospective split.
What should researchers, biotech teams, funders, and program leaders do with this?
Use virtual-organism work to make experiments sharper. The best near-term role is prioritization: which organism, genotype, stage, tissue, condition, or phenotype should be tested next.
Define the context of use before building. The model specification should name the organism, environment, intervention, endpoint, and decision. Broad organism language should be replaced with the exact biological claim.
Build the validation plan early. Program budgets should include prospective perturbations, held-out environments, organism-level measurements, independent replication, and uncertainty analysis. A virtual-organism effort without organism validation is a data-integration project.
Separate discovery from replacement. Discovery tools nominate mechanisms and experiments. Replacement tools require direct evidence that the replaced assay, animal experiment, field trial, or clinical measurement is no longer needed for the defined decision.
Further Reading and Related Chapters
- Systems Biology and Multiscale Modeling: scale transitions and causal evidence
- Perturbation Prediction and Virtual Cells: cell-scale perturbation evidence
- Spatial Omics and Tissue Models: spatial context in tissues
- Neuroscience AI and Brain Foundation Models: brain, neural activity, and behavior
- Aging and Longevity Biology AI: organism-level biomarkers and lifespan claims
- Plant, Crop, and Agricultural AI: plants, crops, field conditions, and environment
- Environmental and Ecological AI: organism-environment and ecosystem context
- Benchmarks for Bio AI: benchmark construction and evaluation limits
FAQ
What is a virtual organism?
A virtual organism is a computational representation of organism-level biology. It should name the organism, stage, environment, intervention, endpoint, uncertainty, and validation plan.
Is a digital twin the same thing as a virtual organism?
No. A digital twin is usually tied to a specific system, data update process, and decision. A virtual organism may be a reference biological model rather than an individual-specific decision tool.
Why are model organisms still difficult to simulate?
Model organisms simplify measurement, but they remain living systems with development, physiology, behavior, environment, and biological variability. Tractability is not the same as completeness.
What is the strongest near-term use case?
The strongest near-term use is experimental prioritization: selecting perturbations, stages, tissues, environmental conditions, or phenotypes for measurement.
When is replacement of experiments credible?
Replacement is credible only after direct comparison with the experiment being replaced, using a prespecified error tolerance and the same context of use.