🧬 biology

What Makes a Virtual Cell a World Model? Three Gaps, Three Experiments, and a Roadmap

This paper proposes a structured framework for "virtual cell world models" (VCWMs) to address critical gaps between current capabilities and true world modeling by defining formal requirements, presenting empirical evidence of these shortcomings through three experiments, and outlining a staged roadmap with diagnostic capability ladders for developing multiscale, interactive cellular simulators.

Original authors: Chang Yu, Jingbo Zhou, Cheng Tan, Stan Z. Li, Xiaodong Liu, Xiaoming Zhang, Zhaoxiang Zhang, Zhen Lei, Zhongqi Wang

Published 2026-07-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Chang Yu, Jingbo Zhou, Cheng Tan, Stan Z. Li, Xiaodong Liu, Xiaoming Zhang, Zhaoxiang Zhang, Zhen Lei, Zhongqi Wang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine trying to predict the weather. You could take a single, high-resolution photo of the sky right now and say, "It looks cloudy." That's a great description of the present. But if you want to know if it will rain tomorrow, or how a sudden gust of wind will change the clouds, a photo isn't enough. You need a model that understands the rules of the atmosphere—how air moves, how water vapor turns to rain, and how a change in one place ripples through the whole system. This is the difference between a static picture and a "world model."

Now, shrink that idea down from the sky to the tiniest building blocks of life: cells. For years, scientists have been building "virtual cells" using artificial intelligence. Some of these virtual cells are like super-advanced photo albums, taking snapshots of a cell's genes and proteins to describe what it looks like right now. Others are like fortune tellers, guessing what a cell might look like after a doctor gives it a medicine. But there's a big question hanging over this field: Do these virtual cells actually understand how life works, or are they just really good at guessing? If we want to use AI to design new medicines or cure diseases, we need to know if our virtual cells can truly simulate the future, not just describe the present.

This paper, titled "What Makes a Virtual Cell a World Model?", acts like a strict referee for this high-stakes game. The authors argue that just because a computer program uses fancy words like "world model," it doesn't mean it actually is one. They propose a new way to test these systems, pointing out three major "gaps" where current technology is falling short.

The Three Big Gaps: Why "Good Guessing" Isn't Enough

The authors identify three specific ways current virtual cells are tricking us into thinking they are smarter than they are.

The Snapshot vs. The Movie (Representation ≠ Dynamics):
Imagine you have a library of millions of photos of a car. You can train a computer to recognize a red sports car perfectly. But if you ask that computer, "What happens if I slam on the brakes?" it might not know. It knows what the car looks like, but it doesn't know the physics of how the car moves. The paper shows that many current AI models are just taking better photos (representations) of a cell's genes. They are great at describing what the cell is now, but they fail to predict how the cell will change in the future. In one experiment, they found that while these models got much better at describing the current state (a gain of about +0.21), they barely improved at predicting the future fate of the cell (a gain of only +0.019). It's like having a crystal ball that tells you exactly what you're wearing, but can't tell you if you'll trip over your shoelaces in five seconds.
The Crystal Ball vs. The Steering Wheel (Prediction ≠ Intervention):
Now, imagine a weather app that can tell you, "If it rains, the ground gets wet." That's a prediction. But a true "world model" should let you ask, "What if I make it rain by seeding the clouds?" and then show you the entire new chain of events that follows. The paper tests this by asking AI models to simulate a sequence of events: "If we change Gene A, then change Gene B, what happens?" They found that while the models could guess the final result of a single change, they failed when asked to use that result as a starting point for the next change. In their tests, when they tried to chain predictions together, the error grew significantly (the difference jumped from 295.67 to 320.12), and the models completely failed to predict the outcome of combined changes (0% success rate). This means the AI is just guessing the end of the story, not understanding the plot twists in the middle.
The Multitool vs. The Symphony (Multimodality ≠ Multiscale):
Scientists are now feeding AI models all kinds of data at once: RNA (the cell's instructions), proteins (the workers), and chromatin (the filing cabinet). It's like giving the AI a multitool with a screwdriver, a knife, and a corkscrew. But having all the tools doesn't mean the AI knows how to build a house. The paper shows that while these models can find connections between the different tools (like noticing that a specific screwdriver often goes with a specific screw), they don't actually understand how changing one part affects the whole machine. In an experiment, they tried to see if changing the "filing cabinet" (chromatin) would change the "instructions" (RNA). Even though the model knew they were related, the actual change in the instructions was tiny (around 10⁻³), proving that the model hasn't truly learned how the different levels of the cell work together.

The Roadmap: Building a Real Virtual Cell

So, what do we do? The authors don't say we should stop building these models; they say we need to be honest about what they can do. They propose a "ladder" of capabilities to grade these systems, rather than just giving them a single grade.

Level 0: Just a snapshot. It describes the cell but can't predict the future.
Level 1: It can guess the next step, but only once.
Level 2: It can keep going, simulating a sequence of changes, but only if the changes are simple.
Level 3: The "Holy Grail." This is a true World Model. It can simulate long, complex sequences of changes, understand how different parts of the cell (from molecules to tissues) talk to each other, and tell you how sure it is about its predictions.

The paper outlines a roadmap to get there. First, we need to build "Candidate" models that can actually roll out a sequence of events, not just guess the end. Next, we need to connect the different scales of biology so the model understands how a tiny molecular change ripples up to the whole tissue. Finally, we need to build "Interactive" models that can work with scientists in a loop: the AI suggests an experiment, the scientist runs it in a real lab, and the AI learns from the result to get better.

The Takeaway

The main message is a gentle but firm reality check. We are currently very good at building AI that describes cells and guesses the outcome of a single experiment. But we are not yet at the point where we have a "Virtual Cell" that truly understands the rules of life and can simulate the future. The authors aren't saying these tools are useless; they are saying we need to stop calling them "World Models" until they pass the tests. By setting these clear standards, they hope to guide scientists toward building AI that can truly help us design better medicines and understand the complex, dynamic dance of life.

Technical Summary: "What Makes a Virtual Cell a World Model? Three Gaps, Three Experiments, and a Roadmap"

Problem Statement
The field of virtual cell research has rapidly integrated mechanistic simulators, single-cell foundation models, perturbation predictors, and multimodal integrators. However, the term "world model" is increasingly applied to systems with fundamentally different capabilities, creating terminological ambiguity. Current systems often conflate representation learning with dynamics, endpoint prediction with intervention, and multimodal observation with multiscale modeling. The authors argue that without a structured framework, claims of "world modeling" in cellular systems are unjustified, as representation capacity, endpoint accuracy, and multimodal coverage are insufficient on their own. The core problem is the lack of a formal definition and evaluation criteria that distinguish a true Virtual Cell World Model (VCWM) from a standard predictor or encoder.

Methodology and Framework
The paper proposes a structured framework to define and evaluate VCWMs, moving away from settled nomenclature toward a set of necessary conditions and diagnostic axes.

Formal Definition:
Drawing from general world model definitions in reinforcement learning ( $s_{t+1} \sim T(s_t, a_t)$ ), the authors formalize a VCWM as a tuple:
$VCWM = (O, S, C, A, G_t, T_{\Delta t})$
Where:
- $O$ : Cellular observation space.
- $S$ : Maintained, updateable cellular state.
- $C$ : Context (split into biological $C_{bio}$ and observational $C_{obs}$ ).
- $A$ : Intervention space.
- $G_t$ : Time-indexed biological structure (e.g., regulatory graphs).
- $T_{\Delta t}$ : A reusable transition operator.
Five Necessary Conditions (NC1–NC5):
To qualify as a candidate VCWM, a system must satisfy:
- NC1: Explicit, updateable state representation.
- NC2: An explicit, callable state transition mechanism (not independent endpoint inference).
- NC3: Intervention-conditioned transitions (actions modulate $T_{\Delta t}$ itself, not just the output head).
- NC4: Support for future and counterfactual evaluation from a shared starting state.
- NC5: Operational uncertainty and failure boundaries (calibrated abstention).
Three Diagnostic Axes and Ladders:
The authors introduce three independent axes, each with an L0–L3 capability ladder, to avoid collapsing distinct capabilities into a single grade:
- Dynamics: Ranges from L0 (snapshot representation) to L3 (calibrated long-horizon rollout with controlled error growth).
- Intervention: Ranges from L0 (action as metadata) to L3 (validated counterfactual rollouts under unseen interventions).
- Scale: Ranges from L0 (single-scale) to L3 (bidirectional cross-scale closure where fine-scale changes propagate upward and coarse-scale context constrains lower scales).

Key Contributions and Empirical Results
The paper provides three "motivation experiments" to empirically demonstrate that current state-of-the-art systems exhibit the identified gaps, failing to meet VCWM criteria despite advanced capabilities.

Gap 1: Representation $\neq$ Dynamics
- Experiment: Using the Weinreb LARRY hematopoiesis dataset, the authors tested whether fine-tuned Geneformer representations (via LoRA) improved current-state readouts versus future-fate predictions.
- Result: While current-state gain was significant (+0.2117), future-fate gain was negligible (+0.0192).
- Conclusion: Rich representations do not inherently encode the dynamics required for future state prediction; foundation models remain L0 on the dynamics ladder unless they expose a maintained state and transition.
Gap 2: Prediction $\neq$ Intervention
- Experiment: Using Norman K562 Perturb-seq data and the GEARS model, the authors tested "state-space closure" by iteratively applying perturbations (feeding the predicted output of one perturbation as the input for the next).
- Result: Iterative reuse resulted in a mean manifold L2 error increase of +24.46 compared to direct prediction, with a 0% positive-combination rate.
- Conclusion: Predicting endpoints does not guarantee that the output constitutes a valid, reusable state for subsequent interventions. The transition operator is not closed under iteration.
Gap 3: Multimodality $\neq$ Multiscale World Modeling
- Experiment: Using SHARE-seq mouse skin data and the MultiVI model, the authors evaluated whether learned RNA-ATAC associations could drive intervention effects.
- Result: While the model learned a strong association (2.78 $\times$ random controls), linearized interventions on chromatin (opening/closing DORC-linked peaks) produced negligible changes in predicted RNA ( $\sim 10^{-3}$ ).
- Conclusion: Learning cross-modal associations does not equate to learning operational, cross-scale transition relations capable of simulating intervention effects.

Significance and Roadmap
The paper claims significance not as a benchmark or a final solution, but as a "disciplined basis for evaluation" to prevent the premature classification of useful components as world models.

Clarification of Ambiguity: It distinguishes between the ambition of the AI Virtual Cell (AIVC) and the criteria for a VCWM candidate, clarifying that current systems often lack the necessary commitments to state, transition, and intervention.
Staged Roadmap: The authors propose a phased roadmap for progress:
- Phase 1 (Candidate VCWM): Focus on establishing shared state and callable transitions with uncertainty reporting.
- Phase 2 (Multiscale & Structure-Aware): Integrate structured constraints (regulatory graphs, spatial neighborhoods) that constrain the transition operator $T_{\Delta t}$ .
- Phase 3 (Generalizing & Interactive): Achieve transfer to held-out biological contexts and enable closed-loop scientific planning where an agent queries the VCWM for counterfactuals to prioritize wet-lab experiments.

The authors emphasize that this framework does not assume biology-verified causal closure or a unified benchmark ecosystem exists yet. Instead, it serves to ensure that claims of "world modeling" are backed by evidence of state maintenance, transition reusability, and intervention-conditioned dynamics before moving to complex, interactive applications.

Technical Summary: "What Makes a Virtual Cell a World Model? Three Gaps, Three Experiments, and a Roadmap"

More like this