Virtual Cell Abstraction: A Cross-Domain Framework

• Virtual cell abstraction is a framework that replaces traditional cellular units with task-specific surrogates, offering tractability in coordination and simulation.
• In wireless networks, virtual cells enable effective resource allocation and interference management by partitioning base stations and users into optimized clusters.
• In AI and biological modeling, virtual cells serve as executable and generative models that simplify complex cellular dynamics while balancing interpretability and scalability.

Virtual cell abstraction denotes a class of technical constructions in which the ordinary notion of a “cell” is replaced by a logical unit chosen to make coordination, simulation, or evaluation tractable. In wireless networking, the abstraction replaces a conventional physical cell by a neighborhood-level coordination domain or a user-centric serving set for resource allocation and interference management, such as $\mathcal C_v=(\mathcal B_v,\mathcal U_v)$ or $\mathcal V_k\subseteq\mathcal B$ (Yemini et al., 2019, Yemini et al., 2019, Wang et al., 2015). In computational biology and AI virtual cell modeling, it replaces the biological cell by an executable state representation—transcriptomic, morphological, or knowledge-grounded—that supports perturbation prediction, generative simulation, or natural-language evaluation (Jiang et al., 11 Jun 2026, Zhang et al., 26 Mar 2026, Hu et al., 14 Oct 2025, Zhao et al., 26 Feb 2026). In developmental biology and artificial-life simulation, it appears as a reformulation of what a simulated cell is, for example a deformable polygonal unit in VirtualLeaf or a cell-like morphology emerging from virtual molecular interactions (Wolff et al., 2018, Ishida, 2023).

1. Conceptual scope and recurring structure

Across these literatures, the abstraction is defined less by a single ontology than by a recurring operation: a rigid physical or biological unit is replaced by a computationally useful surrogate whose boundary is aligned with the task of interest. In cellular communications, the boundary is chosen to internalize dominant local interference without incurring fully centralized complexity; in AI virtual cell work, it is chosen to capture the level of cellular state actually measured or queried, such as transcriptomes, microscopy phenotypes, or symbolic pathway context (Yemini et al., 2019, Jiang et al., 11 Jun 2026, Zhang et al., 13 Feb 2025, Wei et al., 29 Nov 2025).

The formal objects differ accordingly. Wireless papers define virtual cells as cluster-level or user-centric serving domains. Biological simulation papers define a cell as a polygonal mesh object, a multiset of molecular species on a lattice, or a high-dimensional state conditioned on perturbation. AIVC papers go further and define model-agnostic latent or executable state spaces that connect measurement, perturbation, and decision-making (Wolff et al., 2018, Ishida, 2023, Hu et al., 14 Oct 2025).

This suggests a shared structural logic despite divergent domains: the abstraction is valuable when direct treatment of the underlying system is either too expensive, too noisy, or too weakly identified. The gain is tractability, locality, or interpretability; the cost is an abstraction boundary that omits part of the underlying dynamics.

2. Wireless-network virtual cells as coordination domains

In uplink cellular resource allocation, a virtual BS is a set of base stations, $\{b_1,\ldots,b_n\}$, and a proper clustering partitions both the BS set $\mathcal B$ and the user set $\mathcal U$. A virtual cell is then the pair $\mathcal C_v=(\mathcal B_v,\mathcal U_v)$, so that every BS and every user belongs to exactly one virtual cell (Yemini et al., 2019). This construction sits between isolated per-cell processing and fully centralized cloud-RAN or network-MIMO processing: BSs inside a virtual cell coordinate, while inter-virtual-cell interference is omitted during local optimization and accounted for later in performance evaluation (Yemini et al., 2019, Yemini et al., 2019).

Two closely related but distinct wireless meanings appear. The first is this cluster-centric neighborhood abstraction, typically built by hierarchical clustering with minimax linkage, followed by user affiliation through a closest-BS or best-channel rule (Yemini et al., 2019, Yemini et al., 2019). The second is a user-centric serving-set abstraction in large-scale distributed antenna systems, where each user $k$ forms a virtual cell $\mathcal V_k\subseteq\mathcal B$ consisting of its $V$ nearest antennas, with $|\mathcal V_k|=V$ (Wang et al., 2015). In both cases the virtual cell is the object over which coordination is defined, but the first partitions the network globally whereas the second gives each user its own serving neighborhood.

Inside a cluster-centric virtual cell, the communication model can be uplink CoMP joint decoding with infinite-capacity backhaul among BSs in the same virtual BS. The resulting optimization is the sum-capacity of a multiple-access channel with multiple receiving antennas, solved per virtual cell by cyclic coordinate ascent and generalized water-filling (Yemini et al., 2019). A different physical-layer choice is made in the companion formulation with single-user detection, where BSs jointly allocate channels and power but decode users separately; that problem is mixed-integer and NP-hard, leading to a continuous reformulation and an alternating channel-allocation/power-allocation method (Yemini et al., 2019).

The central systems insight is a cooperation-complexity continuum. In the joint-decoding formulation, average system sum rate increases monotonically as the number of virtual cells decreases, because more interference becomes internal to the cluster and is handled cooperatively (Yemini et al., 2019). In the single-user-detection formulation, the penalty of fully distributed optimization relative to fully centralized optimization can be as high as $\mathcal V_k\subseteq\mathcal B$0, while a few BSs within a virtual cell can achieve almost the same performance as fully centralized optimization if the virtual cells are designed properly (Yemini et al., 2019). In the user-centric DAS formulation, the no-grouping MRT regime favors small virtual-cell size, with the rule of thumb $\mathcal V_k\subseteq\mathcal B$1, whereas grouping users with overlapping virtual cells and applying ZFBF reverses the trend and makes larger $\mathcal V_k\subseteq\mathcal B$2 beneficial at higher complexity (Wang et al., 2015).

A related virtualization-oriented variant appears in small-cell networks with FD self-backhauls, where radio spectrum, MBSs, SBSs, time slots, and backhaul resources are virtualized as allocatable resources under a Virtual Resource Manager (Chen et al., 2015). Although the paper does not use the exact term “virtual cell abstraction,” it constructs an equivalent logical access-and-backhaul entity through variables for association, access time-share, backhaul time-share, and spectrum splitting, solved by an ADMM-based distributed algorithm (Chen et al., 2015).

3. Simulated cells in developmental biology and artificial chemistry

In VirtualLeaf, the abstraction shift concerns the internal representation of a simulated biological cell rather than a coordination domain. A cell is formalized as $\mathcal V_k\subseteq\mathcal B$3, where $\mathcal V_k\subseteq\mathcal B$4 is a set of vertices, $\mathcal V_k\subseteq\mathcal B$5 a set of edges, and $\mathcal V_k\subseteq\mathcal B$6 a set of cell attributes; tissue is $\mathcal V_k\subseteq\mathcal B$7, with shared vertices and edges across adjacent cells (Wolff et al., 2018). The paper’s contribution is to redefine the virtual cell from a plant-style, topology-fixed polygonal unit into a deformable and topologically reconfigurable object by introducing a sliding operator that allows cell-cell shear. T1 and T2 transitions then emerge naturally from the same Hamiltonian-plus-Metropolis dynamics that governs shape change, rather than from separate threshold-triggered graph edits (Wolff et al., 2018).

This version of virtual cell abstraction is geometric and mechanical. It retains polygonal tessellation but upgrades the interface from a single straight edge to a multi-segment boundary with 2-connected membrane nodes. That change allows curved or buckled interfaces, makes junction sliding possible, and places VirtualLeaf between standard vertex models and the Cellular Potts Model as an off-lattice generalization (Wolff et al., 2018). In this setting, what is abstracted is not only the cell but also the interface: the model’s expressive power depends on how much geometric and topological freedom is granted to cell boundaries.

A different biological meaning appears in the Multi-set chemical lattice model for emergence-of-life simulation. Here the primitive object is not a predeclared cell at all, but a multiset of molecule types in each site of a $\mathcal V_k\subseteq\mathcal B$8 hexagonal lattice. Reactions are symbolic multiset rewrites such as $\mathcal V_k\subseteq\mathcal B$9, augmented by diffusion, polymerization, decomposition, and explicit energy-resource flow (Ishida, 2023). Polymerized molecule 2 acts as a boundary-like structure, polymerized molecule 1 stores morphology-related and catalytic information through the arrangement of molecules 6 and 7, and added energy molecules 20, 25, and 19 implement a crude metabolism (Ishida, 2023).

In that artificial-chemistry setting, the abstraction is a minimal evolving cell-like system rather than a realistic biochemical cell. The paper reports the emergence of cell-like shapes with the four minimum cellular requirements named in the abstract—boundary, metabolism, replication, and evolution—based solely on virtual molecular interactions (Ishida, 2023). A plausible implication is that, in this lineage of work, virtual cell abstraction functions as a middle level between cellular automata and molecular dynamics: sufficiently symbolic to remain tractable, but sufficiently structured to let morphology, metabolism, and selection co-emerge.

4. AI virtual cells as executable models of perturbation response

In the AIVC literature, a virtual cell is usually an executable, decision-relevant model of cell state learned from data rather than hand-written mechanistic rules. The most explicit high-level formulation introduces a model-agnostic Cell-State Latent $\{b_1,\ldots,b_n\}$0, together with measurement operators $\{b_1,\ldots,b_n\}$1, lift/project operators for cross-scale coupling, an intervention operator $\{b_1,\ldots,b_n\}$2, and a temporal propagator $\{b_1,\ldots,b_n\}$3 (Hu et al., 14 Oct 2025). This does not specify a single architecture; it specifies an operator grammar for what an AIVC should support: modality alignment, perturbation, dynamics, and decision-aligned readout.

Transcriptome-centered models instantiate this abstraction by treating the cell operationally as a high-dimensional gene-expression state plus perturbation and context. OCOO-T is a minimalist flow-matching model that uses a vanilla Transformer stack directly on continuous gene expression profiles, with perturbation, dosage, and cell-line or cell-type information injected through adaptive layer normalization and in-context tokens (Jiang et al., 11 Jun 2026). Its formulation makes the abstraction boundary explicit: it models transcriptomic state and perturbation covariates, while omitting explicit signaling cascades, chromatin dynamics, protein turnover, morphology, and metabolism (Jiang et al., 11 Jun 2026). On Tahoe100M it reports PDCorr $\{b_1,\ldots,b_n\}$4, MSE $\{b_1,\ldots,b_n\}$5, MAE $\{b_1,\ldots,b_n\}$6, and strong PDS and DE metrics, supporting the claim that a transcriptome-level virtual cell can be practically predictive at scale (Jiang et al., 11 Jun 2026).

Lingshu-Cell pushes the transcriptomic abstraction toward a generative “cellular world model.” It models a full discrete transcriptome over about 18,080 genes using masked discrete diffusion, rather than restricting to HVGs or rank-ordered panels, and conditions generation on donor, cell type, or perturbation tokens (Zhang et al., 26 Mar 2026). The model represents a virtual cell as a sampled transcriptomic state $\{b_1,\ldots,b_n\}$7 from $\{b_1,\ldots,b_n\}$8 or $\{b_1,\ldots,b_n\}$9, where $\mathcal B$0 encodes identity and intervention. It reports strong unconditional generation across tissues and species and top overall performance on the Virtual Cell Challenge H1 perturbation benchmark by average rank (Zhang et al., 26 Mar 2026).

VCWorld offers a different, explicitly white-box strand. It reframes perturbation-response prediction as reasoning over a biological knowledge graph $\mathcal B$1, retrieves analogue and contrast cases for a gene-centric query $\mathcal B$2, and uses an LLM to infer DE or directional change through explicit mechanistic hypotheses (Wei et al., 29 Nov 2025). Here the virtual cell is neither a transcriptome-only generator nor a latent dynamical system, but a symbolic biological world model in which drug targets, pathways, PPIs, and gene programs serve as the state substrate for stepwise inference (Wei et al., 29 Nov 2025).

Image-based virtual cell models instantiate the abstraction at the level of morphology rather than transcriptomics. CellFlux defines a perturbation-conditioned distribution $\mathcal B$3 over microscopy images and learns a continuous flow in image space from control morphology to perturbed morphology using flow matching (Zhang et al., 13 Feb 2025). Evaluated on BBBC021, RxRx1, and JUMP, it reports a 35% improvement in FID scores and a 12% increase in mode-of-action prediction accuracy over existing methods (Zhang et al., 13 Feb 2025). CellFluxRL then adds post-training with reinforcement learning and seven biologically motivated rewards spanning biological function, structural validity, and morphological correctness, explicitly shifting the abstraction from “visually realistic” to “biologically meaningful” generated cells (Wu et al., 23 Mar 2026).

5. Evaluation, benchmarking, and refinement frameworks

As virtual cell models diversified, several papers shifted the abstraction from simulation to evaluation. SC-Arena defines a virtual cell as an instance of a “Knowledge Cell” class with attributes and methods: attributes include expression-based features, text-based descriptions, and ontology-grounded identities, while methods represent cell-to-environment processes and environment-to-cell responses (Zhao et al., 26 Feb 2026). On that basis it unifies five natural-language tasks—cell type annotation, cell captioning, cell generation, perturbation prediction, and scientific QA—under a single virtual-cell evaluation object, scored with knowledge-grounded matching against resources such as Cell Ontology, CellMarker, NCBI, UniProt, GO, and PubMed (Zhao et al., 26 Feb 2026).

A broader benchmarking paper argues that current virtual cell models are often overestimated under standard setups and should instead be tested under in-the-wild conditions: unseen cell contexts, unseen perturbations, and cross-dataset generalization (Mao et al., 30 Apr 2026). Its main conclusion is sharply diagnostic: under strict conditions, performance drops markedly; simple linear approaches can capture broad transcriptional trends; and different metrics emphasize different biological properties and can reorder model rankings (Mao et al., 30 Apr 2026). This places evaluation itself inside the virtual cell abstraction: what counts as a good virtual cell depends on whether one values global transcriptomic trend prediction, perturbation-specific differential expression, or transport across context shifts.

Another line of work treats virtual cell abstraction as a control problem over modeling hypotheses. CellScientist separates a high-level hypothesis space $\mathcal B$4 from a low-level implementation space $\mathcal B$5, formalizing model refinement as a bilevel loop in which structured hypotheses are realized as admissible programs and execution discrepancies are routed back to specific hypothesis edits (Li et al., 8 May 2026). The framework maintains a model state $\mathcal B$6 and implements a closed Hypothesis $\mathcal B$7 Implementation $\mathcal B$8 Hypothesis cycle with auditable refinement traces (Li et al., 8 May 2026). This extends the abstraction from “what a cell model represents” to “what object is actually being revised when a virtual cell model fails.”

A survey of LLMs and virtual cells proposes a related taxonomy: LLMs as Oracles for direct cellular modeling and LLMs as Agents for orchestration of literature access, tool use, workflow automation, and scientific reasoning (Li et al., 9 Oct 2025). That framing does not define a new technical object, but it situates recent virtual cell abstractions within two complementary roles: a biological world model and a scientific control plane.

6. Recurrent design tradeoffs and limitations

Several tensions recur across the literature. In wireless systems, the abstraction trades fidelity for scalability: virtual-cell decomposition avoids global optimization but ignores or externalizes cross-boundary interference during local resource allocation, so performance remains below full centralization unless clusters are sufficiently large (Yemini et al., 2019, Yemini et al., 2019). In biology, transcriptomic virtual cells gain scale and predictive power by abstracting away mechanism, while image-based virtual cells gain phenotypic realism but remain limited to morphology and perturbation-conditioned appearance rather than causal intracellular dynamics (Jiang et al., 11 Jun 2026, Zhang et al., 13 Feb 2025, Wu et al., 23 Mar 2026).

Cross-scale coupling is another persistent limitation. The Cell-State Latent framework explicitly notes that anchors linking molecular, cellular, and tissue scales are sparse, and that dose, time, combination effects, and transport across laboratories and platforms are still inadequately handled (Hu et al., 14 Oct 2025). Benchmarking work echoes this concern, showing weak transport under unseen-cell or cross-dataset conditions and warning that naive dataset aggregation can reduce performance (Mao et al., 30 Apr 2026). Transcriptome-only and image-only models make this tradeoff explicit: they are useful virtual cells precisely because they operate at the level of available measurements, but that same choice constrains biological completeness (Jiang et al., 11 Jun 2026, Zhang et al., 26 Mar 2026, Zhang et al., 13 Feb 2025).

Interpretability is also heterogeneous. White-box or symbolic systems such as VCWorld and CellScientist expose evidence, reasoning paths, and refinement traces, whereas most generative transcriptomic and morphological models remain predictive rather than mechanistic (Wei et al., 29 Nov 2025, Li et al., 8 May 2026). SC-Arena and the AIVC evaluation blueprint both respond to this by emphasizing function-space readouts, ontology-aware scoring, leakage-resistant partitions, and decision-relevant outputs rather than reconstruction metrics alone (Zhao et al., 26 Feb 2026, Hu et al., 14 Oct 2025).

Taken together, these works suggest that virtual cell abstraction is best understood not as a single model family but as a design principle. A “cell” becomes a task-specific surrogate: a coordination neighborhood in wireless systems, a deformable polygon or artificial chemistry aggregate in simulation, a transcriptomic or morphological state generator in AI modeling, or an executable hypothesis object in evaluation and refinement. The abstraction succeeds when that surrogate preserves the causal or operational structure needed for the intended query, and fails when omitted interactions, scales, or constraints become dominant.