From the Virtual Cell to the Human Digital Clone: A Specification for Whole-Body In Silico Physiology

The phrase "digital clone of a human body" invites two opposite errors. The first is dismissal: the body contains ~3.7×10¹³ cells, each an open thermodynamic system with combinatorially explosive internal state, so a faithful replica is "impossible." The second is hype: because AI models now predict protein structures and transcriptional responses, a whole-body simulation is "imminent." Both errors share a structure — they treat the question as binary rather than as a specification to be met.

This paper takes the specification view. A digital clone does not require simulating every molecule, for the same reason weather forecasting does not require simulating every air molecule: the question is what coarse-graining preserves predictive power for the decisions we intend to make. A cardiac ablation planner needs electrophysiological fidelity at the tissue scale; it does not need ribosome kinetics. A chemotherapy dosing twin needs pharmacokinetics and clonal tumor dynamics; it does not need the patient's gait. "The digital clone" is therefore not one object but a family of models sharing an individual's data substrate, each validated for a declared decision context.

Two fields supply the components. The virtual cell program, transformed since 2023 by single-cell foundation models, supplies the molecular and cellular layer. The digital twin program in medicine — most mature in cardiology — supplies the organ and physiological layer, along with the only regulatory-grade validation methodology currently in existence. Section 2 reviews the first, Section 3 the second, Section 4 formalizes the gap between them, Section 5 states the specification, Section 6 addresses validation and epistemic status, Section 7 enumerates failure modes, and Section 8 proposes falsifiable milestones.

2.1 The mechanistic era

The virtual cell is an old ambition. The Virtual Cell (VCell) platform and E-Cell project of the late 1990s modeled cells as systems of biochemical reactions — ODEs, PDEs, and stochastic kinetics over hand-curated networks. The landmark of this era is the whole-cell model of Mycoplasma genitalium (Karr et al., Cell 2012), which integrated 28 submodules covering every annotated gene function of the smallest known free-living organism and predicted phenotypes of gene knockouts with useful accuracy.

The lesson of the mechanistic era is double-edged. Positive: whole-cell integration is possible — the model made novel, experimentally confirmed predictions (e.g., previously unmeasured kinetic parameters). Negative: it took a decade of curation for an organism with 525 genes. Scaling curation-first modeling to a human cell (~20,000 protein-coding genes, extensive isoform and post-translational complexity, context-dependent regulation) is not a matter of effort; the parameters are largely unmeasured and many are unmeasurable in situ. The mechanistic program did not fail — it hit a data-acquisition wall.

2.2 The foundation-model era

The wall was circumvented rather than climbed. Instead of specifying mechanisms, the field began learning distributions over cell states from massive single-cell omics corpora. Transformer-based models — Geneformer (Theodoris et al. 2023), scGPT (Cui et al. 2024), scFoundation (Hao et al. 2024), and successors such as CellFM and billion-parameter long-context models — are pretrained on tens to hundreds of millions of single-cell transcriptomes and demonstrate transfer to cell-type annotation, gene-function inference, and network biology tasks. Perturbation-response models (GEARS; CPA; the Arc Institute's STATE; flow- and diffusion-based generators such as CellFlow and scDiffusion-class models) directly target the counterfactual question that defines a virtual cell: given this cell state and this intervention, what is the resulting state distribution? Large coordinated data programs — the Human Cell Atlas, CZI's CELLxGENE and Virtual Cells Platform, JUMP-Cell Painting, the Arc Virtual Cell Atlas — supply the substrate, and the framing paper of the field (Bunne et al., Cell 2024, "How to build the virtual cell with artificial intelligence") articulates the target: multimodal foundation models unifying genomics, transcriptomics, epigenomics, proteomics, metabolomics, and spatial context.

2.3 What the benchmarks actually say

The honest reading of 2024–2026 benchmarking literature is sobering and must be stated plainly, because it is the load-bearing caveat for everything downstream:

• —Perturbation prediction remains weak out of distribution. Systematic benchmarks (PertEval-scFM and successors; independent evaluations of post-perturbation RNA-seq prediction) repeatedly find that pretrained single-cell foundation models often fail to beat simple baselines (e.g., mean-shift or linear models) on unseen perturbations, and that "in-the-wild" generalization across cell types, doses, and combinatorial perturbations is far from solved. STATE-class models show measurable cross-cell-type generalization — an advance — but measurable is not clinical-grade.
• —Interpretability has been traded for scale. Retrospectives on the field's shift from mechanistic to learned models note the loss of mechanistic transparency: a black-box distribution over transcriptomes cannot, by itself, tell you why a prediction holds, which matters both scientifically and for regulatory acceptance.
• —The readout is narrow. Most models see the transcriptome. A cell state is transcriptome + proteome + metabolome + morphology + mechanics + spatial neighborhood + history. Multimodal unification is a stated priority, not an achievement.

The virtual cell in 2026 is therefore best described as a statistical emulator of cellular observables under limited intervention families, improving rapidly, and not yet a general dynamical model of the cell. This is not a criticism; it is a coordinate. Emulators of exactly this kind (learned surrogates constrained by data) are how every mature simulation field — weather, turbulence, materials — handles unresolvable microscale physics. The question is whether the emulator's error is characterized, which brings us to validation (Section 6).

3.1 Definition and the cardiac exemplar

The EDITH consortium (Ecosystem Digital Twins in Health, the EU coordination action) defines a health digital twin as a computer simulation combining knowledge-driven and data-driven models to predict clinically relevant quantities that are difficult to measure directly, with the capacity to forecast the evolution of the patient's state. Three features distinguish a twin from a mere model: individuation (parameterized to one person), assimilation (updated as new measurements arrive), and decision coupling (its outputs feed a specific clinical or engineering decision).

Cardiology is the field's proof of concept, and the depth of that proof is underappreciated outside it. Patient-specific electrophysiological heart models — pioneered by Trayanova's group at Johns Hopkins — are built from contrast MRI, incorporate cell-scale ionic behavior into organ-scale geometry, and are in prospective clinical studies for guiding ventricular-tachycardia ablation targeting and predicting sudden cardiac death risk. On the industrial side, the Living Heart Project (Dassault Systèmes, developed in collaboration with the FDA) produced a validated multiphysics heart model — coupled structural mechanics, electrical activation, and hemodynamics — now used by device manufacturers to evaluate pacemakers, valves, stents, and defibrillators in silico, and its methodology has crossed from research concept to regulatory reality within roughly a decade. Hospitals have used patient-specific twins to plan congenital heart surgeries with outcome predictions clinicians judge more reliable than animal models.

Two things make the cardiac exemplar transferable. First, it demonstrates the hybrid architecture that any whole-body twin will need: mechanistic physics (conservation laws, continuum mechanics, reaction-diffusion) as the skeleton, with data-driven components filling in unmeasurable parameters and accelerating computation. Second, it produced a validation grammar: the ASME V&V 40 standard for assessing model credibility as a function of the decision's risk — the only existing framework under which a simulation's evidence is auditable for regulators.

3.2 Beyond the heart

The Living Heart template has been extended to lungs, liver, brain, eyes, and gut, each developing within its own clinical community, with the explicit long-term architecture being a modular platform in which organ twins plug into a unified virtual human. The EDITH roadmap frames the same trajectory at the European scale: a federated repository of models, datasets, and practices, moving from single-organ systems toward integrated multiscale, multi-organ whole-body twins, backed by dedicated Horizon Europe funding for multiscale patient patho-physiology models. Parallel efforts — physiome-style whole-body physiological models, pharmacometric PBPK (physiologically based pharmacokinetic) models already routine in drug regulation, and "medical world model" architectures that represent patient state and simulate clinical dynamics for treatment policy — populate the intermediate layers.

Whole-body twins today are therefore best described as federations of organ- and system-level models with per-organ validation, not integrated organisms. Which is exactly where the interesting problem lives.

The virtual cell operates at 10⁻⁶–10⁻⁵ m and predicts distributions over molecular states. The organ twin operates at 10⁻²–10⁻¹ m and evolves fields (voltage, stress, concentration) under continuum equations. Between them sits the tissue scale — cellular neighborhoods, microenvironments, vasculature, immune surveillance — where most disease is actually decided. Three coupling strategies are in play:

(a) Parameter passing (weak coupling). Cell-level models supply effective parameters (e.g., ionic conductances, secretion rates, proliferation rates as functions of local conditions) to continuum models. This is how cardiac twins already work: ionic cell models parameterize tissue-scale reaction-diffusion. It is robust, cheap, and validated — and it discards cell-to-cell heterogeneity, which is fatal in oncology and immunology where rare-cell behavior (a resistant clone, an exhausted T-cell subset) drives outcomes.

(b) Agent-based middleware (explicit mesoscale). Tissue simulators (PhysiCell-class agent-based models, virtual tissue and virtual organoid frameworks now being coupled to learned cell models) represent 10⁴–10⁷ discrete cells whose behaviors are supplied by the virtual-cell layer, embedded in continuum fields for oxygen, drugs, and mechanics. Emerging "AI virtual tissue" work — foundation models over spatial transcriptomics, spatial proteomics, and histopathology — attacks the same layer from the data side, learning tissue-state representations directly. This is the most biologically faithful strategy and the most computationally violent one.

(c) Learned surrogates end-to-end (strong emulation). Train models that map patient-level inputs directly to patient-level outputs, letting the network internalize the scale bridge. This is the "medical world model" bet. It maximizes tractability and minimizes mechanistic auditability — and it inherits, at the whole-patient scale, the out-of-distribution fragility already documented at the cell scale.

The plausible architecture is not a choice among these but a stack: (c) as the fast inference layer, (b) as the offline generator of training data and the arbiter when (c) is queried outside its support, (a) as the conservation-law scaffold that keeps everything physical. This mirrors how numerical weather prediction absorbed machine learning: learned emulators run the forecasts; physics defines the state space and catches the emulator when it hallucinates.

The critical unsolved problem in the stack is error propagation across scales. A 10% error in a cell-level secretion-rate prediction may be irrelevant or catastrophic at the organ level depending on feedback topology; without uncertainty quantification that composes across the coupling interfaces, a whole-body prediction has no error bar, and a prediction without an error bar is not a clinical instrument. Negative results are as informative as positive ones here: characterizing which cell-level errors amplify is itself a primary research product.

We propose that "digital clone of a human body" be operationalized as any system meeting the following five requirements, each stated with its falsification-relevant content.

R1 — Molecular grounding. The system's cell- and tissue-level components must be conditioned on the individual's molecular data (genome; where available, transcriptomic/proteomic profiles of accessible tissues) and must predict intervention responses that depend on that conditioning. Test: the twin's predicted drug response must differ between individuals in ways that prospectively match observed pharmacogenomic variation, beyond what demographic covariates predict.

R2 — Multiscale coupling with error control. Cross-scale interfaces must carry calibrated uncertainty, and the composed uncertainty at the decision variable must be reported. Test: empirical coverage — over a validation cohort, the twin's 90% predictive intervals for the clinical endpoint contain the observed value ~90% of the time. A twin that is accurate on average but overconfident fails R2.

R3 — Identifiability from accessible data. The personalization must be achievable from clinically obtainable measurements (imaging, blood work, wearables, standard omics), and the map from data to model parameters must be shown to be well-posed for the decision context — i.e., the data actually constrain the parameters that the prediction is sensitive to. Test: sensitivity-versus-identifiability audit; parameters that are influential but unidentifiable must be flagged and marginalized, not point-estimated. Silent point-estimation of unidentifiable parameters is the field's most common quiet failure.

R4 — Continuous assimilation. The twin updates as new measurements arrive, with a defined forgetting/consistency policy, and its forecast skill must be shown to improve with assimilation (else the "twin" framing is decoration on a static model). Test: forecast-skill curves as a function of assimilation history, benchmarked against the static model.

R5 — Prospective counterfactual validity. The decisive requirement. The twin must demonstrate, in prospective trials, that decisions informed by its counterfactual predictions improve outcomes relative to standard of care — the standard already being applied to cardiac ablation twins. Retrospective fit is inadmissible as primary evidence; biology's dimensionality makes retrospective fit nearly free.

Under this specification, several honest statements become possible. No system on Earth currently meets R1–R5 for the whole body. Cardiac twins meet R2–R5 within their organ domain with R1 partial (genotype-informed variants exist). Virtual cells meet R1 by construction and currently fail R5 entirely — no cell-model prediction has yet been the primary basis of a prospectively validated clinical decision. The specification also clarifies what the clone is not required to be: it is not required to be complete, conscious, real-time at all scales, or unique. It is required to be calibrated, individuated, and prospectively predictive within declared decision contexts — a strict but meetable specification.

6.1 The validation stack

Four evidence layers, in ascending order of cost and authority:

• —Retrospective benchmark performance (cell-level perturbation benchmarks; organ-level held-out patient cohorts). Necessary, weak.
• —Prospective observational forecasting — the twin predicts, sealed, before the outcome is known. This is where forecast-skill and calibration claims become meaningful.
• —Interventional validation — twin-guided decisions in controlled trials (the current frontier in cardiac electrophysiology).
• —Regulatory-grade credibility assessment — V&V 40-style risk-tiered evidence, already accepted for device evaluation via the Living Heart pathway, extended by emerging regulatory thinking on adaptive, software-heavy medical cyber-physical systems, where verification and evidence-documentation norms are themselves expected to evolve.
• —6.2 Claims ledger

Claim Layer Status

C1 Whole-cell mechanistic integration is possible for minimal organisms cell / mechanism Demonstrated (M. genitalium, 2012) C2 Foundation models learn transferable representations of cell state cell / statistical Demonstrated (annotation, gene function) C3 Foundation models predict responses to unseen perturbations at decision-grade accuracy cell / statistical Open — benchmarks show frequent failure vs. simple baselines OOD C4 Organ-scale patient-specific twins can guide interventions with prospective benefit organ / hybrid Partially demonstrated (cardiac EP; trials ongoing) C5 In silico evidence can enter regulatory decisions institutional Demonstrated (Living Heart / FDA pathway; V&V 40) C6 Cross-scale uncertainty composition with empirical coverage multiscale / methods Open — no whole-pathway demonstration C7 Whole-body integrated multi-organ twin of an individual organism Open — roadmap stage (EDITH; modular organ federation) C8 A twin can replace (not augment) clinical trials for novel therapeutics institutional Open, likely furthest — current realistic role is trial enrichment and in silico device testing

The table's purpose is discipline: popular treatments of this subject routinely present C7 in the tone appropriate to C5. The gap between them is the actual research program.

Identifiability, not compute, is the binding constraint. The whole-body state space is vast; the clinically accessible measurement channel is narrow (a blood draw is a ~10⁻⁵ sample of one fluid compartment). Many parameter configurations are observationally equivalent until an intervention distinguishes them — and interventions on humans are ethically rationed. Expect twins to be sharply better in organs with rich imaging channels (heart, brain, eye) than in diffuse systems (immune, endocrine), and treat that asymmetry as a prediction of this framework.

Distribution shift is biological, not incidental. A twin calibrated on a patient's Tuesday physiology meets Wednesday's infection. Aging, microbiome drift, and disease progression guarantee that every twin operates under perpetual mild distribution shift; assimilation (R4) is the countermeasure, and its failure modes (over-forgetting, chasing noise) need the same scrutiny as the model itself.

The interpretability–scale trade is not free. Regulatory and clinical acceptance of black-box components will lag their benchmark performance; the hybrid stack of Section 4 is partly an epistemic architecture — mechanistic layers exist to make learned layers auditable.

Privacy and misuse are structural, not peripheral. A whole-body twin is the most complete dossier on a person ever assembled — genome, physiology, behavior, prognosis. Federated architectures (as in the EDITH design) and health-data regulation (EHDS-class regimes) are load-bearing components of feasibility, because without trusted custody the data substrate will not be contributed at the required scale. Insurance and employment discrimination on twin-predicted risk is the obvious failure trajectory and requires statutory, not merely technical, mitigation.

What is not a constraint. No known physical or computational law prohibits a decision-grade digital clone. The system need not simulate quantum degrees of freedom (decoherence timescales in warm tissue make classical coarse-graining adequate for every clinical decision context so far identified); it need not be real-time everywhere; it need not resolve consciousness. Claims that the clone "captures the person" — cognition, identity, experience — belong to a different problem class entirely and should be explicitly firewalled off: a physiological twin makes no phenomenological claims, and conflating the two is the single most reliable marker of an unserious treatment.

M1 (cell, ~2–4 yr): A virtual-cell model outperforms all simple baselines on combinatorial, unseen perturbations across ≥3 held-out cell types on public benchmarks, with calibrated predictive distributions. Falsifier of optimism: continued baseline-parity through 2028 despite 10× data would indicate the transcriptome-only observable is insufficient and force the multimodal pivot.

M2 (tissue, ~3–6 yr): A learned-cell-model-driven tissue simulation prospectively predicts a measurable tissue-level outcome (e.g., organoid drug response, immunotherapy response in patient-derived cultures) with pre-registered accuracy targets.

M3 (organ, ongoing–5 yr): Twin-guided cardiac ablation demonstrates superiority in randomized trials; a second organ system (respiratory or hepatic) reaches interventional validation.

M4 (multi-organ, ~5–10 yr): Two or more organ twins of the same individual, sharing one assimilation loop, produce a coupled prediction (e.g., cardio-renal drug response) with composed uncertainty meeting R2 coverage — the first genuine whole-body-fragment clone.

M5 (institutional, ~7–12 yr): A regulatory submission in which twin-cohort in silico evidence substitutes for a portion of a clinical trial arm is accepted under a V&V 40-descended framework.

Each milestone is stated so that its non-arrival is informative. If M4 stalls specifically on uncertainty composition while single-organ twins flourish, the field will have learned that the scale-bridging problem (Section 4), not data volume, is rate-limiting — which is this paper's central conjecture.

The virtual cell and the medical digital twin are two halves of one instrument, currently being built from opposite ends. The cell program supplies molecular individuation and is strongest exactly where the twin program is weakest (mechanism below the imaging scale); the twin program supplies validation grammar, regulatory pathway, and prospective clinical evidence — everything the cell program lacks. The human digital clone is the closure of this pincer, and it should be discussed as a specification with five requirements rather than as a metaphor with a release date. Nothing in physics or biology forbids it. What it demands is paid in identifiability, calibrated uncertainty, and prospective validation — strict currencies, but ones we know how to earn, because one organ has already earned them.