Virtual cells: progress and challenges from mechanistic modeling to data-driven integration

Literature scope and resource verification methods

This review combines narrative synthesis with structured literature and resource verification using a representative inclusion strategy. It remains a structured narrative review rather than a systematic review and does not claim exhaustive coverage, estimate publication bias, or perform quantitative synthesis. Historical works from 2001 onward trace virtual-cell development; models made public from 2022 through the revision-stage update construct the curated representative landscape in Table 2. Years refer to the version of record when available; otherwise, the year of the first public preprint release is used.

The retained submission records document broad source types, subject areas, and prioritization principles but do not contain a complete log of the original pre-submission searches. During revision, explicit category-specific inclusion criteria were therefore applied consistently to the existing Table 2 landscape, reviewer-suggested SpaRx, and other borderline cases in a retrospective verification and curation audit performed during revision. These revision-stage criteria are not represented as having been predefined before the initial submission. Table 2 is consequently a curated representative landscape rather than an exhaustive inventory. Borderline or cross-category methods are handled according to their primary technical contribution and direct relevance to virtual-cell development. Publication form is recorded separately from validation evidence.

Detailed information on the original records that can be confirmed, the revision-stage search/update process, representative search strings, category-specific inclusion and exclusion criteria, and borderline-case handling are provided in Supplementary Table S1. Supplementary Table S2 provides the model-level curation, role, capability, and evidence audit. Supplementary Table S3 records dated resource verification, including release, revision, model-card, and licensing information where ascertainable. Supplementary Table S4 records dataset provenance and overlap relevant to training/evaluation independence. Recent preprints are retained to capture emerging directions, but their capability claims remain provisional until peer review or independent replication where applicable.

Explicit mechanistic models

Explicit mechanistic models define state variables, parameters, and transitions from known processes. Gene transcription, protein degradation, ligand–receptor binding, and metabolism can be represented by ODEs; intracellular diffusion and membrane transport by PDEs; and low-copy-number reactions by stochastic simulation. A general form is

(6)

Here, x denotes molecular concentration or cellular state, θ denotes parameters such as reaction rates, u denotes an external stimulus or control variable, and t denotes time. With spatial diffusion, the model becomes

(7)

Here, κ is the diffusion coefficient.

Virtual Cell translates experimentally defined cellular geometries, reaction networks, and boundary conditions into solvable mathematical models [2]. Explicit variables and parameters support interpretation, comparison of mechanistic hypotheses, and targeted experiments. However, large gene, protein, and spatial networks require extensive parameter estimation, and multiple timescales and high-dimensional nonlinearities complicate numerical solution.

Whole-cell models extend beyond local pathways. Karr et al. [4] integrated DNA replication, transcription, protein synthesis, metabolism, and cell-cycle submodels to simulate the life cycle of Mycoplasma genitalium. The study demonstrated coordination under a unified cellular state but also the dependence on manual curation, parameter estimation, and module interfaces. PhysiCell instead represents each cell as an agent governed by mechanics, movement, division, death, and phenotypic transitions, coupled to diffusion of nutrients, drugs, and signals [3]. This supports tumor, immune-migration, and morphogenesis simulations but reduces intracellular molecular state to relatively few phenotypic variables.

Between early pathway simulators and current foundation models, systems biology expanded through rule-based biochemical modeling, multicellular cell-based frameworks, reaction–diffusion coupling, and multiscale tissue simulation [57–59]. These approaches improved compositionality, spatial representation, and collective-cell simulation, but remained limited by parameterization burden, incomplete mechanistic knowledge, and computational scaling. Large single-cell atlases and standardized perturbation datasets later shifted part of the bottleneck from manually specifying every interaction toward learning transferable state representations and response maps. Foundation and perturbation models are therefore better viewed as a data-driven continuation of this history than as a discontinuous replacement of mechanistic systems biology.

Mechanistic constraints, structured prediction, and biology-guided evaluation

Mechanistic constraints should be distinguished from biology-guided evaluation in data-driven virtual-cell research. Mechanistic constraints directly affect model structure or optimization, e.g. through reaction-network topology, conservation relations, or dynamical equations. PDCorr, DE Overlap, and pathway-enrichment consistency are biology-guided evaluation metrics when used only to compare predictions; they are not training constraints.

SCALE is discussed here only as a boundary case between structured prediction and mechanistic modeling; its perturbation-prediction architecture is reviewed in the subsection on perturbation-response prediction. Public materials describe PDCorr and DE Overlap as evaluation metrics rather than training constraints [20]. Accordingly, SCALE is treated as a structured, data-driven state-transition model rather than a traditional mechanistic dynamical model.

RegFormer introduces structural priors through gene-regulatory-network hierarchies and uses state-space models for long gene sequences [7]. Such priors constrain dependencies in the representation space without fully specifying the dynamics. Current approaches therefore form a continuum from explicit ODE, PDE, and stochastic models to probabilistic generative and large-scale pretrained models. A mechanistic label is justified only when biological knowledge constrains state variables, model structure, or state transitions, not merely when a model uses biological terminology or attention mechanisms.

General cell-state representation

Cell-state representation models map sparse, batch-dependent expression data to reusable latent spaces that should preserve biological state while limiting technical variation [60, 61]. scGPT, Geneformer, scFoundation, UCE, and TranscriptFormer use distinct pretraining schemes to learn transferable cell or gene representations, including cross-tissue and cross-species settings [15–19]. Because their tokenization and objectives differ, the shared foundation-model label does not imply equivalent biological content or transfer capability.

Cell2Sentence serializes expression as gene sentences to interface with language models, whereas xVERSE learns batch-invariant representations and probabilistic whole-transcriptome generation from single-cell and spatial data [31, 32]. These approaches improve representation or generation, but neither textual fluency nor generative quality establishes that latent variables correspond to regulatory mechanisms.

More generally, large-scale pretraining can support transfer without identifying causal biology. Statistical dependencies and attention weights may mix direct regulation, indirect co-expression, abundance, cell type, and batch effects, so mechanistic claims require intervention-based evidence rather than latent-space interpretability alone.

Perturbation-response prediction

An interactive virtual cell requires perturbation-responsive state transitions in addition to representation. Response models should condition on initial state, intervention, dose, time, and biological context. State emphasizes cross-context perturbation prediction; SCALE learns conditional population transport; X-Pert unifies genetic, chemical, combinatorial, and dose-dependent perturbations; and Tahoe-x1 learns large-scale gene, cell, and compound representations from drug-perturbation data [20–23]. These models move training toward intervention-rich corpora, but matching an endpoint distribution does not identify the temporal path or mechanism that produced it.

X-Cell combines diffusion-based state generation with molecular and phenotypic priors, while UniPert-G2CP transfers phenotype knowledge from genetic to chemical screens [24, 45]. Such transfer remains conditional: knockout and drug treatment differ in dose, target multiplicity, metabolism, and engagement, so mechanism-specific and independent validation remain necessary.

Cross-modal and unified multi-omics modeling

Virtual cells must integrate modalities without treating cross-modal agreement as causal equivalence. GigaTIME predicts multiprotein spatial patterns from H&E images, illustrating the need to evaluate not only image similarity but also cellular composition, neighborhoods, functional readouts, and cross-center stability [62]. EpiFoundation and EpiAgent model chromatin accessibility; CLM-X links RNA and ATAC; HoloCell jointly models epigenomic, transcriptomic, and proteomic states; and CellHermes frames heterogeneous omics tasks as question answering [5, 8, 36–38]. Their cross-modal claims remain dependent on paired-data quality and validation design.

Protein measurements add a functional layer that transcriptomics alone cannot supply. CAPTAIN learns paired RNA–surface-protein representations, whereas ProteinTalks uses perturbation proteomics and neural ODEs to model dynamic protein networks [9, 14]. Evaluation should therefore distinguish transcriptional consistency from prediction of protein abundance, interactions, and functional state.

Boundaries of model scale and generalization

Increasing data or parameter scale does not by itself guarantee improved biological generalization. Single-cell data retain batch effects, sampling noise, tissue bias, and inconsistent labels, so larger corpora can replicate existing distributions without adding new mechanistic information [63]. Zero-shot representation evaluations and perturbation benchmarks have shown that complex models may underperform highly variable genes, principal components, unchanged predictions, additive models, or linear baselines in particular settings [25, 26]. These findings do not invalidate foundation models; they show that claimed gains depend on splits, preprocessing, baselines, metrics, and whether the task contains genuinely nonadditive or context-specific effects.

OOD evaluation should distinguish unseen cells or states, donors, cell types or tissues, platforms, perturbations, doses, combinations, species, time, space, and modalities. Random splits can preserve cell-line, perturbation, or batch information and therefore do not establish transfer to unfamiliar biology. DenAdel et al. [64] found no clear general scaling law with data size, while VCBench and SCMBench show that model rankings vary across tasks and datasets [65, 66]. Model size, pretraining volume, and the foundation-model label therefore cannot replace explicit distribution-shift evaluation.

For comparison, this review groups methods by primary contribution into mechanistic/structured, transcriptomic foundation, multi-omics, perturbation/dynamic, spatial/multicellular, proteomic, EM/ultrastructure, and system/intervention categories. Figure 3 visualizes this taxonomy, Table 2 provides the representative landscape, Table 1 gives the operational LCM/AIVC comparison, and Fig. 4 summarizes capability emphasis; none is a performance ranking.

Boundaries between large cellular models and AI virtual cells

Recent Reviews, Perspectives, Comments, and other conceptual articles use overlapping but nonidentical criteria for LCMs and AIVCs [1, 67–71]. Across these viewpoints, transferable representation is an important foundation, while stronger virtual-cell claims increasingly emphasize perturbation prediction, state transition, temporal dynamics, multimodal and spatial integration, mechanistic or causal interpretability, experimentally testable prediction, and experimental feedback. The relative weight placed on these properties differs across publications, so no single field-wide boundary should be inferred from terminology alone.

A representation model does not automatically constitute a complete virtual cell, and a perturbation predictor does not automatically constitute a complete virtual cell. Both can nevertheless provide important enabling or predictive components of virtual-cell development. For analytical comparison in this review, LCMs are treated as systems whose central contribution is transferable state representation, whereas AIVCs place additional emphasis on perturbation-responsive transitions, conditional or temporal context, and outputs that can be experimentally tested. This is an operational convention adopted in this review rather than an established field-wide taxonomy.

Table 1 structures this operational comparison using functional emphasis rather than mandatory architectural requirements. Figure 4 provides a representative capability comparison. Its coding describes research emphasis and direct demonstrations in the cited sources rather than a performance ranking or a test of whether a model is a “true” virtual cell.

Dynamic interactions in cell populations

Tissue cells respond to ligands, contact, metabolic competition, and mechanics rather than fixed external conditions. CellTRIP models cell populations as interacting agents with reinforcement learning and self-attention for spatiotemporal prediction [51]. This can represent heterogeneous, context-dependent transitions, but learned actions and interaction weights remain statistical unless linked to measurable communication or biophysical variables and tested by intervention. Combining multi-agent learning with explicit signaling, metabolic, and mechanical constraints may therefore improve interpretability and validation.

Perturbation propagation in tissue space

Spatial foundation models have evolved from simply incorporating coordinate information to explicitly modeling spatial context. Nicheformer is pretrained on both dissociated single-cell and spatial omics datasets, attempting to jointly encode cell identity and the local tissue environment [6]; SToFM decomposes tissue macrostructure, microenvironment, and gene expression into multiscale subslices for pretraining [10]; STARNet combines spatial RNA and ATAC information to infer region-specific gene regulatory networks via heterogeneous hypergraphs [48]. These three approaches represent three complementary directions: spatial representation, multiscale spatial pretraining, and spatial regulatory networks.

SpatialProp focuses on the propagation of perturbations at the tissue scale. This method represents spatially resolved single-cell transcriptomes as a cellular graph, predicts gene expression based on the local microenvironment, and propagates the effects of local perturbations to other cells within the tissue [49]. Its CausalInteractionBench utilizes cell–cell interactions with known directions to assess whether the model recovers directionally consistent communication relationships. This evaluation is closer to causal inference than a simple comparison of expression correlations, but it still does not constitute causal proof. Rigorous validation requires controlled perturbations, temporal sequence, spatial blocking, or independent tissue experiments.

SpaRx provides a complementary spatial intervention example. It uses graph-based domain adaptation to transfer pharmacogenomic drug-response knowledge to single-cell spatial transcriptomic data and predicts spatially heterogeneous drug sensitivity within tissue [50]. It was retained in the revised representative landscape because spatially resolved intervention-response prediction is a central technical contribution under the same revision-stage category criterion applied to other spatial and perturbation methods.

VirTues further organizes the spatial proteome into hierarchical representations such as proteins, cells, niches, and tissues, and supports cross-panel transfer, spatial biomarker discovery, and patient stratification [53]. This work provides a relatively clear example of a multiscale representation, but whether its cross-level state transitions are interpretable still requires verification through interventions at different scales and longitudinal data.

Predictive virtual embryos

Predictive virtual embryos extend virtual-cell ideas to development by linking cell state, lineage, spatial position, movement, and tissue morphology. Cao et al. [52] describe this as a research vision rather than an implemented, systematically benchmarked model. A testable system would need to model division, migration, fate, morphogens, mechanics, and tissue boundaries with aligned timescales and uncertainty; transcriptional prediction alone cannot establish correct developmental structure.

World models, cross-scale modeling, and system integration

CellOS uses joint-embedding prediction to learn multiple views of cellular state and is presented as a cellular world-model direction [56]. This may support unified representation, querying, and state prediction, but the current preprint still requires independent verification of scope and reproducibility.

Multimodal input alone is not equivalent to cross-scale modeling. A conceptual bidirectional hierarchy can be written as

(8)

and

(9)

A cross-scale virtual system must specify interpretable interfaces through which molecular changes alter cellular behavior and tissue structure, and through which tissue context feeds back to lower levels. These mappings, not merely the presence of several modalities in one network, require joint experimental validation.

Volume electron microscopy and subcellular ultrastructure

Volume EM contributes 3D organelle geometry, membrane contacts, and network topology that are not captured by transcriptomic state alone. Current EM foundation models remain enabling components rather than complete virtual cells: EM-DINO/OmniEM support restoration and segmentation, EMCFsys combines restoration, segmentation, and sparse 3D reconstruction, and DF5T unifies several restoration tasks [11–13]. An ultrastructure-aware virtual cell would need to align these structural features with molecular and metabolic states and allow them to participate in state transitions, rather than treating EM as static imagery.

System-level modeling and intervention design

The Virtual Yeast Perspective organizes mechanistic knowledge, ultrastructure, and dynamic omics as data pillars for a future system-level model, but it does not report a complete implemented virtual cell [54]. GEMGen generates molecules for target cellular phenotypes, whereas SequenTx uses cellular-state models and reinforcement learning to design sequential drug interventions with in vitro and in vivo testing [46, 55]. These examples have different evidence profiles: candidate generation, model-guided intervention selection, independent experimental validation, and closed-loop updating are distinct claims and are evaluated separately in the later discussion of experimental validation and feedback.

Data resources, prior knowledge, and coverage boundaries

Data coverage remains a major constraint. More cells improve statistical stability but cannot replace diversity in perturbations, contexts, dose, time, spatial location, and measurement platforms. For example, Stack uses scBaseCount together with CELLxGENE and Parse PBMC resources, while the Arc Virtual Cell Atlas integrates scBaseCount, Tahoe-100M, Virtual Cell Challenge data, and other resources [43, 72, 73].

These resources have different provenance roles. Primary measurements, standardized compilations, knowledge bases, derived benchmarks, model weights, and reproduction packages should not be counted as independent experimental data. Overlap among CELLxGENE, scBaseCount, Arc Atlas, Tahoe resources, and model-specific corpora can also inflate apparent training scale or external validation when the same parent studies or cells are reused.

Table 3 therefore separates experimental data, prior knowledge, and derived evaluation resources and adds concise provenance flags. Known overlap is reported only for traceable parent/derived relationships; potential overlap requires a source-level rationale; otherwise the relationship is recorded as not ascertainable. Supplementary Table S4 provides the corresponding provenance audit and implications for training/evaluation independence.

Model openness and licensing

Reproducibility must be assessed separately for papers, code, weights, licenses, data versions, vocabularies, preprocessing, checkpoints, and inference settings. Code release does not imply released weights or unrestricted use, and a project page is not equivalent to a complete reproduction package.

Because repositories change rapidly, revision-stage mutable-resource verification was performed on 3 September 2026, and current publication and resource status was rechecked on 11 September 2026. Release/tag, revision, model-version, and licensing details are recorded where ascertainable in Supplementary Table S3. Statements such as “no official author release identified as of 11 September 2026” are deliberately time-bounded and do not establish that a resource does not exist. Table 4 summarizes the availability of code and model weights and the licensing status of representative virtual-cell models.

Parameter scale, training data, and compute disclosure

Table 5 separates architecture from primary training objectives while summarizing parameter scale, training data, hardware, and duration. These quantities are reported unevenly and are architecture-specific; training-cell counts may also refer to raw profiles, unique cells, processed examples, or repeated sampling. Parameter count is therefore neither a reproducibility metric nor a substitute for performance, OOD generalization, or experimental validation.

Training objectives further delimit what a model is optimized to learn. Representation models use masked prediction, reconstruction, or autoregressive prediction; multimodal systems use contrastive alignment, cross-modal reconstruction, or generation; perturbation/dynamics models use supervised state prediction, optimal transport, flow matching, or diffusion-related objectives. These objectives cannot be converted directly into capability claims: reconstruction does not establish perturbation causality, masked prediction does not imply temporal dynamics, and in-distribution performance does not establish OOD generalization. Model-level objectives are reported in Table 5 without inferring unavailable losses.

Multilevel evaluation framework

Virtual-cell evaluation requires multiple levels. Numerical reconstruction metrics, including mean squared error, correlation, negative log-likelihood, and distribution distance, measure agreement but not recovery of key biological processes. Biological-effect metrics include differentially expressed gene overlap, perturbation-effect correlation, pathway enrichment, transcription-factor activity, and cell composition. PDCorr and DE Overlap assess perturbation effects [20], but remain sensitive to thresholds, sample size, and statistical procedures.

External and OOD evaluation asks whether performance transfers across a genuine distribution shift such as individuals, cell types, tissues, perturbations, platforms, species, time, or spatial contexts. Independent experimental validation is different: a prediction is tested using a newly generated experiment rather than only a held-out computational test. Closed-loop experimental feedback is narrower still and requires model predictions to influence experiment selection, followed by use of the experimental result for model update, recalibration, or a subsequent decision.

To avoid implying a single maturity ladder, the revised review assesses five complementary, non-ordinal evidence dimensions. These dimensions are used as a practical analytical framework to separate five distinct questions that are often conflated in virtual-cell evaluation: whether a model or claim is sufficiently documented to be traced and examined; whether empirical performance has been demonstrated for the stated task; whether that performance transfers beyond the original evaluated distribution; whether mechanistic or causal claims receive direct support beyond predictive association; and whether model predictions have been tested in independently generated experiments. These questions correspond, respectively, to citable and retrievable documentation (DOC), task-specific empirical evidence (TASK), external and OOD generalization evidence (GEN), mechanistic and causal evidence (MECH), and independent experimental validation (EXP). The dimensions are not successive validation levels: TASK does not imply GEN; GEN does not imply MECH or EXP; strong MECH evidence does not require completion of every OOD test; and EXP does not imply closed-loop updating. Closed-loop experimental feedback remains reported separately. Table 6 defines the dimensions and their limitations, while Supplementary Table S2 applies the same evidence coding across the representative landscape. Publication form, model scale, and the technical taxonomy are reported separately from these dimensions.

Design of out-of-distribution evaluation

Perturbation prediction benchmarks should clearly distinguish between different OOD tasks. Random cell splitting primarily tests in-distribution performance under closely matched conditions; unseen cell types test cross-context transfer; unseen perturbations test the extrapolation of perturbation representations; and unseen combined perturbations test whether the model can recognize nonadditive interactions; cross-laboratory and cross-platform testing, meanwhile, assess robustness to variations in batches and measurement mechanisms.

To separate additive from nonadditive effects, let , , and denote the observed effects of perturbations A, B, and their combination:

(10)

Here, is the nonadditive interaction. A metric dominated by may reward a simple additive model; interaction prediction should therefore be judged primarily by recovery of .

A rigorous benchmark should also prevent information leakage. If different doses of the same drug, different compounds targeting the same target, or highly similar genes in the same pathway appear separately in the training and test sets, the test perturbations may still closely resemble the training data. A more rigorous evaluation should group data by target, pathway, chemical scaffold, cell lineage, and experimental batch, and report any potential overlap between the pretraining corpus and the test set.

The term zero-shot is not used consistently across studies, so Table 7 reports the biological axis of generalization rather than a single OOD label. The comparison distinguishes unseen cell states, donors, cell types or tissues, perturbations, doses and combinations, platforms or batches, species, temporal or spatial extrapolation, and cross-modal or cross-assay transfer where reported. It also records the evaluation setting or holdout, the main reported baseline, mechanistic or causal evidence, independent experimental validation, and uncertainty reporting. Simple baselines remain important because a complex model should establish biological gain beyond context or perturbation means and other low-complexity predictors. A same-dataset held-out test is not treated as independent experimental validation, and claims of external evaluation must be interpreted together with pretraining/evaluation provenance. Uncertainty, calibration, OOD detection, and abstention remain unevenly reported across the field.

Table 7 includes predictive models selected by the same revision-stage Table 2 inclusion criteria plus direct relevance to prediction, transfer, perturbation response, or generalization. Selection does not depend on favorable results, publication form, or how extensively a model was discussed in the original manuscript.

From directional association to causal validation

Most single-cell perturbation models lie between associational and interventional prediction. They learn conditional distributions after implemented interventions and can predict states following a specified perturbation under similar conditions, but they do not provide complete counterfactual reasoning. True counterfactuals for the same initial cell cannot be observed under both treatment and control. Population matching, latent alignment, or optimal transport therefore approximate the missing comparison [41]. Their validity depends on assumptions about unmeasured confounding, matching quality, and whether the observed population adequately represents the relevant initial states.

Directional interaction benchmarks such as CausalInteractionBench can test whether predictions align with known cell communication directions, but cannot independently confirm specific causal pathways [49].

Cross-modal transfer between genetic and chemical models can test whether perturbation representations are consistent but cannot prove a shared mechanism. Causal evaluation should consider temporal order, dose–response, intervention specificity, mediators, and replication in independent experiments.

Unified benchmarks and community evaluation

VCBench evaluates unseen cellular backgrounds, perturbations, and merged datasets and shows that rankings depend on tasks and metrics [66]. SCMBench similarly finds that foundation models do not dominate every multi-omics integration task [65]. The Virtual Cell Challenge promotes specialized datasets and open, reproducible evaluation [82]. MVCBench extends standardized comparison to drug-induced multimodal phenotypes and explicitly evaluates representation methods under both in-distribution and OOD settings spanning unseen compounds, cell lines, assay plates, and datasets [110]. It is treated here as a benchmark and generalization-evidence resource rather than as a Table 2 virtual-cell model. Together, these efforts support reporting global agreement, perturbation-specific effects, cell-background transfer, and cross-dataset robustness.

Unified benchmarks should provide raw and normalized data, sample and cell metadata, perturbation dose and timing, negative controls, failure flags, split definitions, simple baselines, and executable evaluation code. Pretraining sources must be disclosed to identify direct or indirect overlap with test data. Results should be reported across replicates and biologically meaningful subgroups, not only as one pooled average. Application priorities differ: drug screening emphasizes false positives and uncertainty; mechanistic studies emphasize pathways and regulatory relationships; patient stratification emphasizes cross-individual transfer; and development emphasizes trajectories and morphology. Evaluation should therefore remain multidimensional rather than collapse capabilities into one score, because the same model can be credible for one use and unsupported for another.

From prediction to actionable experimental feedback

Independent experimental validation and closed-loop experimental feedback should be distinguished. Independent experimental validation occurs when model predictions are tested in independently generated experiments. Closed-loop experimental feedback requires an additional system-level step: model predictions influence experiment selection, experimental results feed back into model updating or recalibration, and the updated information affects a subsequent decision. Independent experimental validation therefore does not by itself imply closed-loop updating.

Fully reproducible end-to-end closed-loop virtual-cell systems remain rare. SequenTx demonstrates model-guided sequential intervention design with substantial independent in vitro and in vivo experimental validation, but the reported workflow does not establish a complete experimental-feedback cycle in which new experimental observations update or recalibrate the model before a subsequent decision [46]. ProteinTalks contributes a dynamic perturbation-proteomics component that can support experimentally testable state-transition hypotheses [9], whereas Virtual Yeast is a framework-oriented Perspective describing an active experimental cycle rather than a fully validated implemented closed-loop system [54]. These examples should therefore not be interpreted as having equivalent evidentiary maturity. Evaluation of future closed-loop systems should report experimental hit rate, learning from failed experiments, information gain per round, cross-round calibration, and experimental cost savings relative to random or expert strategies.

Data heterogeneity and missing negative results

Single-cell and spatial data vary across tissues, individuals, platforms, doses, and time, while coverage is much richer for transcriptomics than for perturbation proteomics, temporal epigenomics, or 3D ultrastructure [69, 70, 95]. RNA alone cannot directly represent protein abundance, phosphorylation, complex assembly, or organelle geometry [69, 92, 93, 95].

Perturbation resources also remain concentrated in common cell lines and transcriptomic endpoints. Weak, invalid, or failed perturbations are often filtered or underreported, which can bias models toward predicting meaningful effects for every intervention. Future datasets should retain negative controls, low-efficacy and failed experiments, variable delivery efficiency, and evidence that perturbations reached their intended targets.

Out-of-distribution generalization and uncertainty quantification

Beyond appropriate benchmark design, practical deployment requires calibrated uncertainty estimates, OOD detection, and prediction-abstention mechanisms. Predictive uncertainty includes aleatoric uncertainty from measurement noise and biological stochasticity and epistemic uncertainty from incomplete training coverage. Models should output state distributions, prediction intervals, or multiple plausible trajectories rather than a single expression matrix.

For inputs far from the training distribution, models should report both OOD status and calibrated confidence and should abstain or reject predictions beyond defined applicability boundaries. Calibration should be evaluated separately across cell types, perturbation classes, doses, and platforms because a single aggregate confidence score can conceal poorly covered subgroups. These safeguards support more reliable experimental decisions and make explicit where model recommendations should not be acted upon.

Continuous-time dynamics

Most perturbation models deal with single-step mappings between the control state and the perturbation endpoint, ignoring the intermediate temporal processes. The same endpoint may result from different dynamical pathways, and short-term stress responses may differ from long-term cellular fate. Methods such as RNA velocity use the ratio of unspliced to spliced RNA to infer the direction of local state changes, providing partial information for dynamic modeling [111]. Continuous state evolution can be described by neural ODEs [112]:

(11)

When stochastic transcription and environmental fluctuations are considered, a stochastic differential equation can be used:

(12)

Here, and denote drift and diffusion, and is a Wiener process. State-space models, dynamic optimal transport, neural processes, and time-stamped generative models may also recover continuous trajectories. Reliable dynamics require genuine time-series data; pseudotime ordering of endpoints is not a substitute.

Continuous-time inference is additionally limited by sparse temporal sampling and the absence of one-to-one correspondence between cells at successive measurements. Snapshot populations can support multiple plausible trajectories, creating an identifiability problem that cannot be resolved by biological plausibility alone. Biological stochasticity must also be separated from measurement and model uncertainty. A biologically plausible trajectory does not by itself establish that the underlying dynamics have been correctly identified. Internal temporal generalization can be assessed with held-out time points, held-out perturbation–time combinations, or held-out trajectories where applicable. Independent temporal validation requires separately generated perturbation time series, lineage tracing, live-cell imaging, or other new temporal experiments. Held-out time points from the same dataset remain internal evaluation rather than independent validation.

Aligning mechanistic knowledge with latent representations

Latent representations should be connected to regulatory networks, signaling pathways, metabolic reactions, and organelle functions rather than interpreted from attention alone. Regulatory modules, graph priors, conservation relations, and reaction constraints can provide intermediate biological structure [7, 16, 29], but their value must be tested by interventions rather than reconstruction quality or plausible attention maps. Because prior knowledge is incomplete and context-dependent, known networks are better treated as updatable soft constraints than fixed truth.

Joint cross-scale modeling

Linking molecular expression to cellular phenotype and tissue structure remains difficult because single-cell, agent-based, and tissue-mechanics models use different variables and timescales [52–54, 71]. The following conceptual framework abstracts cross-scale transfer and is not a specific existing model. For simplicity, population effects are included in tissue aggregation; an explicit population state may be inserted when required:

(13)

(14)

(15)

Here, , , and are molecular, cellular, and tissue latent states; is the state of cell i; N is the number of cells; and is spatial and tissue context. Ultrastructure can be included in or modeled as an intermediate layer. A hierarchical model should propagate information upward and allow tissue context to regulate lower-level states.

Concrete cross-scale interfaces are required for this hierarchy to become operational. Molecular-to-cell mappings may translate pathway activity, protein abundance, metabolic state, receptor or signaling activity, and cell-cycle state into probabilities or rates of proliferation, death, differentiation, migration, and secretion. Cell-to-tissue interfaces may aggregate cell density, interaction forces, ligand fields, extracellular matrix state, and spatial neighborhoods. Tissue-level feedback can in turn modify cellular states through hypoxia, nutrients, mechanical stress, cytokines, and cell–cell contact. Because these processes evolve on different timescales, implementations may require multirate updating, event-based synchronization, or explicit coarse/fine temporal alignment. Validation must therefore test not only outputs at each scale but also the cross-scale mapping itself.

Cross-scale models should be validated layer by layer: transcriptomics and proteomics at the molecular level, morphology and fate tracking at the cellular level, and spatial omics and imaging at the tissue level. The mapping between levels also requires validation. Correct molecular predictions do not guarantee correct cell behavior, and correct local cell states do not guarantee tissue morphology. Simultaneous agreement across levels, preferably under the same intervention and over time, provides stronger evidence for reliable inter-level relationships.

Active experimental design and sequential learning

Virtual cells should improve experimental design rather than replace experiments. A simple acquisition objective can balance expected information gain against experimental cost [1, 46, 70]:

(16)

Selection should favor perturbations that distinguish mechanisms or reduce uncertainty, not merely those predicted to have large effects. Sequential learning also requires records of data versions, model configuration, confidence, rejected recommendations, failed experiments, and what changed between rounds. Drug-screening or precision-medicine claims should retain a validation ladder from computational prediction to appropriate in vitro, organoid, animal, or clinical testing rather than equating expression prediction with therapeutic efficacy.

Open standards and reporting guidelines

Virtual-cell studies remain difficult to compare because data formats, preprocessing, perturbation encoding, and evaluation splits differ. Reports should specify state variables, spatial context, perturbation, dose, time, uncertainty, preprocessing, baselines, code/weights where possible, and applicability limits. Rapidly updated models and preprints should additionally record publication form, versioned model cards, training-corpus or weight changes, and benchmark versions so that later evaluations can be tied to the exact artifact tested.