Virtual cell research originated from the development of computational cell biology in the late 20th century. In 2001, Loew and Schaff proposed the Virtual Cell (VCell) platform, constructing a general framework for transforming cell biology problems into mathematical models and numerical simulations^[25]. In VCell, users specify molecular species, reactions, cellular compartments, and, for spatial models, cellular geometry; the software translates these biological descriptions into mathematical equations for numerical simulation^[12]. Compartmental models describe reaction kinetics using ordinary differential equations, whereas spatial models use partial differential equations to describe how concentrations change through local reactions and molecular transport^[12]. This coupling makes cellular geometry part of the mechanistic model, allowing researchers to investigate how molecular localization and diffusion influence signaling dynamics^[12]. However, limited by data scarcity and insufficient computing power at the time, model parameters mainly relied on literature inferences, and research focused on local pathways and processes, failing to cover complete cell behavior.

Subsequent developments extended VCell’s modeling methods and application scope. Blinov et al. introduced compartmental and spatial rule-based modeling, enabling molecular interaction rules to generate reaction networks for deterministic or stochastic simulations across cellular compartments^[26]. Moving-boundary numerical methods developed for VCell were subsequently extended to free-boundary models that explain cell shapes and turning behavior through actomyosin mechanics^[27]. More recently, Lee et al. used VCell to investigate how rapidly accelerated fibrosarcoma (RAF) abundance shapes epidermal growth factor receptor-extracellular signal-regulated kinases (EGFR-ERK) signaling^[28]. These studies illustrate VCell’s continuing role in connecting explicit biological mechanisms with experimentally testable predictions, while emphasizing the need for experimentally constrained parameters and validation^[28].

With the increasing demand for system integration, whole-cell modeling became a key turning point. In 2012, Karr et al. constructed a whole-cell model of Mycoplasma genitalium by integrating 28 submodels representing cellular processes, including metabolism, DNA replication, transcription, and translation^[20]. These submodels used process-specific mathematical descriptions and exchanged shared cellular variables, such as metabolite concentrations and molecular counts, at one-second simulation intervals^[20]. This integration allowed changes in one process to affect other processes through their shared cellular state. For example, simulations of single-gene disruptions predicted effects on cellular growth and enabled comparison with experimentally measured gene essentiality, illustrating how molecular perturbations could be connected to whole-cell phenotypes^[20]. However, its applicability was limited to simple cell systems, model construction heavily relied on human knowledge integration, and insufficient parameters and validation data made it difficult to extend to complex eukaryotic cells.

Following the 2012 model, whole-cell modeling expanded to more complex bacterial systems and became a framework for evaluating experimental data consistency. Macklin et al. developed an integrated E. coli model that cross-evaluated heterogeneous datasets and identified inconsistencies with functional consequences^[29]. Thornburg et al. subsequently constructed kinetic and spatial models of the genetically minimal bacterium JCVI-syn3A, linking metabolism, gene expression, and cellular growth^[21]. Their 2026 study extended this approach to a complete spatially resolved cell cycle, incorporating chromosome replication and segregation, membrane growth, and division^[30]. These advances broadened whole-cell modeling from phenotype prediction to data reconciliation and spatial cell-cycle reconstruction. Nevertheless, inconsistent measurements, omitted mechanisms, and the computational cost of spatial simulations remain important challenges^[29,30].

Breakthroughs at the data level significantly changed the research path. Single-cell transcriptome sequencing technology has achieved large-scale application, enabling precise description of cell states in high-dimensional vector form^[31,32]. The Human Cell Atlas (HCA) project^[33] has systematically built reference cell atlases that provide diverse cellular contexts for model pretraining, annotation, and external evaluation; however, observational data alone cannot determine how cells respond to perturbations. Perturb-seq combines pooled clustered regularly interspaced palindromic repeats (CRISPR)-based genetic perturbations with single-cell RNA sequencing, linking the perturbation assigned to each cell to its transcriptomic readout^[34]. By comparing perturbed cells with controls, researchers can identify affected genes and cellular programs while resolving heterogeneous responses within the population^[16,34]. In the original study, Dixit et al. applied this approach to transcription factors regulating the response of dendritic cells to lipopolysaccharide, demonstrating how targeted interventions could reveal the organization of an immune-response circuit^[34]. Such datasets provide experimentally observed intervention-response relationships for training and evaluating virtual-cell predictors. Spatial omics complements these measurements by preserving positional information and enabling analysis of gene expression within tissue architecture^[35]. These advances collectively construct a comprehensive data system encompassing cell state representation, perturbation response, and spatial environment information, providing a crucial foundation for characterizing dynamic cell changes.

An early use of deep learning was to place dissociated cells in tissue context. Tangram, published in 2021, aligns single-cell or single-nucleus profiles with spatial measurements to infer where cell populations and unmeasured genes are likely to occur^[36]. This is a prediction about location and spatial expression, rather than a prediction of how cells respond to an intervention.

As single-cell transcriptomic datasets expanded, researchers began training models on large collections of cell profiles to learn relationships among genes and cellular states, then explored how these learned features could support tasks such as cell annotation and disease analysis [Figure 2]. Geneformer, published in 2023, orders genes in each cell by expression relative to a reference corpus and learns to recover masked gene identities from context^[10]. Its learned features were also used to prioritize possible therapeutic targets for cardiomyopathy when disease-specific data were limited^[10]. Single-cell generative pre-trained transformer (scGPT), published in 2024, instead pairs gene identities with discretized expression values and learns cell and gene representations while predicting masked expression values^[9]. Both approaches provide features for tasks such as cell annotation. scGPT was also adapted and evaluated for a perturbation task^[9], but pretraining on observed cell profiles alone cannot establish the accuracy of a model’s predictions for new interventions.

The representation approach then expanded in scale and spatial detail. In 2024, scFoundation used masked expression reconstruction on a large human single-cell collection to learn gene and cell features^[37]. Its gene features were also tested within a separate perturbation-prediction pipeline^[37]; that downstream application should not be confused with the data used for pretraining. Nicheformer subsequently combined dissociated and spatial transcriptomes to represent both a cell and its tissue neighborhood^[38], extending the spatial perspective introduced by Tangram. Universal cell embedding (UCE) pursued a related goal of placing cells from different tissues and species in a shared representation^[39]. These models can describe cell states and spatial context, although their success on those tasks does not by itself demonstrate prediction of biological intervention responses.

A parallel research direction asks how cells change after a defined intervention and uses measured responses to train or test that prediction. Graph-enhanced gene activation and repression simulator (GEARS), published online in 2023, learned from single-cell genetic screens and gene relationships to predict expression after single or combined gene perturbations, including combinations not seen during training^[40]. The 2024 agenda for AI virtual cells made experimentally testable responses a central goal^[3]. CellFM, reported in 2025, connects the two directions: it was first pretrained on transcriptomes from around 100 million human cells, then its learned gene features were substituted into GEARS and assessed with Perturb-seq response data^[41]. The authors also adapted CellFM using measured perturbations to predict responses to untested gene combinations^[41]. Thus, its perturbation results rely on a response-prediction step, even though its large-scale pretraining was based on cell profiles. The 2025 Virtual Cell Challenge provided shared tasks for comparing response models^[17]. State later reported predictions across several perturbation types and contexts^[18], while X-Cell focused on genetic responses across cell types^[19]. How well these models transfer to new interventions and biological settings remains an empirical question.

Model evaluation must keep these two lines distinct. For a cell representation, a zero-shot test asks whether pretrained embeddings support an evaluation task without further task-specific training. Kedzierska et al. found that unmodified scGPT and Geneformer embeddings did not consistently outperform simpler methods in cell-type separation and batch integration, and noted overlap between some evaluation and pretraining datasets^[42]. Their result concerns those models and tests; it does not assess every foundation model or gains after fine-tuning. For a response predictor, the withheld unit must be stated: a perturbed gene, a gene combination, a cell type, a donor, or an experimental condition. These splits answer different questions about transfer. A comparison of perturbation-prediction methods also found that deep learning did not consistently outperform simple linear baselines on the examined tasks^[11]. Validation therefore requires clearly defined splits, appropriate baselines, and, where possible, new experimental measurements^[11,17].

Representative results illustrate why performance claims must be interpreted within their specific evaluation settings [Figure 3 and Supplementary Table 1]. In the original GEARS study ^[40], GEARS reduced the mean squared error (MSE) for the top 20 differentially expressed genes relative to the next-best baseline by 54.2% on the Norman 2019 dataset and by 17.4% on the Replogle 2020 dataset, both of which contain K562 double-perturbation data [Figure 3A]. In a separate zero-shot evaluation of cell embeddings, frozen pretrained embeddings from scGPT-human and Geneformer did not consistently outperform highly variable genes (HVG), Harmony, or single-cell variational inference (scVI) across five datasets. This comparison must be interpreted in light of differences in training access, because Harmony and scVI were fitted to each evaluation dataset, whereas scGPT-human and Geneformer were evaluated without task-specific training^[42] [Figure 3B]. In an independent double-perturbation benchmark, the additive baseline achieved the lowest mean L2 prediction error (2.77), compared with GEARS (4.10), scGPT (4.62), scFoundation (4.72), and a no-change baseline (5.62), across 62 held-out gene pairs per split and five splits^[11] [Figure 3C]. Because Figure 3A-C differ in task, training access, data splits, comparators, and metrics, their values should not be directly compared or combined into a single model ranking. Figure 3D summarizes the progressively stronger forms of evaluation required to support claims about representation quality, perturbation prediction, cross-context transfer, prospective experimental validity, and mechanism.

Predicting the effect of an intervention does not by itself explain its mechanism. A controlled perturbation experiment can establish an effect on a measured outcome in the system studied; a model may learn to predict that effect for untested perturbations. Agreement with held-out responses, however, does not identify the intermediate regulatory interactions that produced them or show that the response will recur in other tissues. Mechanistic claims require targeted follow-up, such as independent perturbations or rescue experiments, while claims of transfer require data from cell types or conditions excluded from training. Table 2 therefore lists the questions addressed by representative models alongside their reported evaluation evidence and limits^[11,40].

AI agents can assist with experiment design and with the construction or assessment of virtual-cell models, but an agent coordinating these tasks is distinct from a model that represents or predicts cellular behavior. BioDiscoveryAgent, presented at International Conference on Learning Representations (ICLR) 2025, uses a large language model (LLM) with literature-search and code tools to prioritize genetic perturbations and was evaluated using existing perturbation datasets^[45]. More recent preprint studies illustrate model-building workflows: CellForge coordinates agents to design architectures and generate executable code^[46], while VCHarness uses an AI coding agent to construct and refine perturbation-response models^[47]. The reported evaluations concern experiment prioritization or the performance of the resulting models, not the agent as a cellular model.

Other preprint studies focus on assessing explanations or existing model repositories. VCR-Agent uses retrieved biological knowledge and verifier-based filtering to produce candidate mechanistic explanations^[48], whereas AblateCell reproduces model repositories and performs ablations to assess contributions to benchmark performance^[49]. These computational checks can support research and guide follow-up experiments, but they do not independently confirm a proposed mechanism or cellular response in biological experiments.

Virtual-cell research has therefore expanded in several complementary directions. Mechanistic platforms and whole-cell models specify processes to simulate; single-cell foundation and spatial models characterize measured cellular states; perturbation models forecast responses; and agents help coordinate model development and testing. These approaches may be combined, but their value depends on whether they answer the biological question at hand and hold up under appropriate tests. Table 2 summarizes what representative methods model, what evidence supports their performance, and where validation remains limited. A key next step is to connect these capabilities through experiments that test predicted responses and, where mechanisms are claimed, the proposed explanations for them.

In basic research, virtual cells play a crucial role in providing quantitative explanations of cell states and their transformation processes^[50]. The VCell platform has long supported the construction of a complete lineage from compartmentalized models to three-dimensional reaction-diffusion models, enabling mechanistic numerical simulations of calcium signaling, cell polarization, membrane transport, and signal transduction networks^[50]. Subsequent versions (VCell 7.0, https://vcell.org/vcell-7-0-with-kinematics-released) have further introduced reaction-diffusion simulation capabilities under moving boundaries, providing new computational power for explaining the mechanisms of dynamic cell morphological changes such as cell migration and cell shape remodeling^[27]. Simultaneously, AI methods are expanding the scope of explanation from specific pathways to data-driven learning of regulatory relationships. CellProphet can infer dynamic gene regulatory networks from single-cell differentiation time-series data, thereby explaining how transcription factors regulate cell fate determination over time^[51]. Squidiff can reconstruct cell fate trajectories under developmental and interventional conditions^[52], while VCWorld combines structured knowledge with large language model reasoning to generate interpretable mechanistic hypotheses^[53]. These methods have further advanced the research focus from static state description to the computational explanation of state transition mechanisms^[54].

Beyond helping to interpret biological systems, virtual-cell models can prioritize candidate interventions for experimental testing, extending their potential value in preclinical drug discovery^[55]. In drug development and disease research, a crucial task of virtual cells is predicting cellular responses after interventions such as drugs, cytokines, or gene editing^[55]. Recent large-scale AI-driven virtual cell models have converged on this direction, although they differ in emphasis. Xaira Therapeutics’ X-Cell predicts transcriptomic responses to genetic perturbations across cellular contexts, including primary human T cells^[19]. Alibaba DAMO Academy’s Lingshu-Cell takes the form of a generative cellular world model capable of simulating transcriptome states and predicting cellular responses to genetic perturbations or cytokine stimulation, with potential implications for drug development, disease mechanisms, and personalized medicine^[43]. XPert offers a more focused example in drug-perturbation prediction. By modeling responses across drug doses and time points, XPert captures transcriptional pharmacodynamic patterns and highlights genes associated with drug response. After pretraining on large-scale perturbation data, XPert improved prediction of treatment-associated transcriptomic changes in held-out breast cancer patients by 15.04% relative to training from scratch; its attention analysis also highlighted genes previously linked to drug resistance^[56]. PRiMeFlow uses flow matching to model heterogeneous single-cell responses to genetic and small-molecule perturbations, improving the recovery of differentially expressed genes in PerturBench, providing a new technical solution for cell state prediction in high-throughput drug screening^[57].

The aforementioned models all point to a trend: virtual cells are gradually evolving from academic demonstrations into actionable tools in preclinical research and drug discovery processes. The advantages of virtual cells are particularly pronounced for highly heterogeneous systems such as tumors, immune systems, and stem cells. The cell states in these systems are extremely diverse, making it difficult for traditional population averaging experiments to fully capture them, while virtual cell models naturally possess the ability to handle heterogeneity.

From an application perspective, AI agents are particularly relevant to virtual screening and perturbation prioritization. Instead of passively scoring a fixed candidate list, an agent can formulate a biological objective, query prior knowledge, select perturbations to test, call a virtual-cell model to estimate responses, analyze failures, and propose the next experimental round. This is consistent with the broader goal of in silico phenotypic screening, for which AssayBench has recently been proposed as an assay-level benchmark for evaluating LLMs and agents on CRISPR screen prediction across phenotypic endpoints^[58]. In future drug discovery pipelines, such agents may serve as a control layer that links virtual-cell prediction with wet-lab validation, while human experts remain responsible for biological plausibility, safety constraints, and final experimental decisions.

The ability to design virtual cells has the most direct application value in the field of synthetic biology, namely, the ability to simulate a new cell system in a computer before physically realizing it. In 2022, Thornburg et al.’s dynamic simulation of the JCVI-syn3A minimal cell was able to capture interactions among metabolism, genetic information processes, and growth over the cell cycle, providing computational guidance for genome simplification^[21]. The JCVI-syn3A four-dimensional whole-cell model (4DWCM) takes this capability to a new level^[30]. This model accurately reproduces experimental measurements such as 105-minute doubling time, copy number ratio of replication origin to endpoint, mRNA lifetime, and ribosome number, and reveals spatial heterogeneity, random allocation, and metabolic-genetic coupling^[30].

AI methods are also entering the synthetic biology design process. Shcherbakova et al. proposed a proof-of-concept framework based on a variational autoencoder, using binary gene presence vectors to model the composition of the bacterial genome, thereby generating candidate reduced E. coli gene sets^[59]. In the same year, Gherman et al. trained a machine learning agent model to predict cell division with high accuracy by learning from data from the whole-cell model. They eventually designed a simplified E. coli cell with about 40% of the genes deleted from the whole-cell model and validated it using the whole-cell model itself^[60].

The combination of spatial omics and AI methods is expanding the application of virtual cells from single-cell level prediction to perturbation response modeling in tissue microenvironments. Celcomen is one of the important advances in this direction^[44]. This model utilizes a mathematical causal framework and generative graph neural networks to decouple intracellular and intercellular gene regulatory programs from spatial transcriptome data, generating perturbed counterfactual spatial transcriptome atlases, providing researchers with samples that are difficult to obtain experimentally^[44]. Another important tool is BioInformatics WalkThrough (BIWT), a software that can directly initialize agent-based spatial models from single-cell and spatial molecular data, transforming static omics snapshots into evolvable multicellular dynamics simulations^[61]. Spatial context is represented differently in these two approaches. Celcomen constructs a graph of neighboring cells or spots from measured spatial coordinates and models gene interactions within cells and across their neighborhoods. Its simulated perturbations yield testable spatial transcriptomic predictions: predicted responses to Tgfbr2 and Jak2 knockouts were compared with measured knockout lesions in mouse lung tumors^[44]. BIWT instead preserves observed cell locations and identities when initializing an agent-based tissue model. Its reported results show how alternative spatial initializations affect simulated tumor-immune dynamics and clustering; independent validation of prospective intervention outcomes was not reported^[61].

This application direction corresponds to a significant upgrade in the concept of virtual cells, moving from virtual cells to virtual cells in tissue. Future clinical applications, such as predicting the response of immune cells to immunotherapy in the tumor microenvironment and predicting the mobilization and differentiation pathways of stem cells after tissue damage, will inevitably require models to simultaneously understand both the intrinsic cellular state and the extrinsic cellular environment. As Celcomen demonstrated, models capable of generating counterfactual spatial transcriptome maps enable researchers to explore intervention scenarios that are not directly observable in real-world experiments due to ethical or technological limitations.

Among all application scenarios, personalized medicine represents the most ambitious and long-term goal of virtual cells. Its ideal vision is to build a continuously updated computational model for each patient, integrating biological data such as genomics, transcriptomics, and proteomics to predict the patient-specific cellular responses to candidate drugs, thereby guiding optimal treatment strategies.

Bhardwaj et al. define virtual cells as digital cell twins, considering them a core paradigm for integrating multimodal biological data and achieving predictive and mechanistic understanding of cell function^[62]. Vallée further clarifies that the next generation of digital twins must adopt a multi-scale framework, covering the entire chain of data from molecules to cells, tissues, organs, and even clinical behavior and the environment^[23]. Lyu et al.’s review of organ digital twins systematically reviews the methodological framework for constructing digital twins of human organs, particularly emphasizing the crucial role of physical information AI in enhancing model fidelity, scalability, and personalization capabilities^[24].

In specific disease areas, conceptual exploration of digital twins has already begun in chimeric antigen receptor T cell (CAR-T) cell therapy. Related reviews define digital twins as a transformative approach that leverages high-dimensional patient data to provide virtual representations of the treatment-tumor trajectory and outline its development path in optimizing CAR-T therapy^[22]. Noetik’s OCTO-virtual cell platform (https://www.noetik.ai/aacr-2025), showcased at American Association for Cancer Research (AACR) 2025, represents a more practical multimodal discovery engine, and was presented alongside Perturb-map, a separate high-throughput in vivo platform for therapeutic target identification and immune phenotype modeling.

However, it is crucial to recognize that true patient-level cell digital twins are still a significant distance away. Current virtual-cell studies draw on cell lines, model organisms, primary human cells, and patient-derived data, but integrating these sources into continuously updated models for individual patients remains challenging. While X-Cell has achieved breakthroughs in zero-shot generalization - generalizing to human primary T cells from multiple donors - this is still at the cell type level rather than the individual patient level. XPert’s evaluation on patient transcriptomic data provides an initial example of transfer from preclinical perturbation datasets, while prospective testing would be needed to assess its use in treatment decisions. To truly realize personalized medical applications of digital cell twins, multiple engineering and scientific challenges need to be overcome, including data assimilation, real-time updates, cross-scale coupling, and clinical validation. Bhardwaj et al. also pointed out that computational scalability, parameter inference, and ethical considerations remain key bottlenecks in the development of digital cell twins^[62].