What If We Framed the Problem Wrong from the Start? AI, Scientific Leaps, and the Virtual Cell
Changing the Map
In 1907, Einstein imagined a person falling from a roof. Nearby objects would appear weightless. He was not solving Newton’s equations faster; he was asking whether gravity’s disappearance in free fall revealed a deeper equivalence with acceleration.
That thought experiment became a route toward General Relativity. It feels like a leap because Newtonian gravity still worked well: no error gradient said to recast gravity as spacetime geometry. The hard move was choosing a new starting point that gave old facts new meaning.
Two essays turn that moment into an AI question. In LLMs can’t jump, Tom Zahavy argues that LLMs can induce and deduce but cannot create a genuinely new explanatory premise. He describes discovery as experience E, a Jump, axioms A, then test; AI may therefore need a “synthetic laboratory” for making counterfactual experience.
In Hot Take: LLM can ‘jump’, Yong Zheng-Xin asks why one theory must have one discovery path. A Feynman-style argument can start from a massless spin-2 field, impose consistency requirements, and reconstruct General Relativity. Connected knowledge and experimental feedback may make some “inspiration” searchable.
But Feynman knew the destination and used tools developed after Einstein. His route proves reconstruction is possible, not that a model in 1907 would invent the premise.
I therefore separate two Jumps. A weak Jump finds a distant connection within existing knowledge; a strong Jump creates a new variable, axiom, or viewpoint. The strong version is not magic: data accumulate, anomalies create tension, and old explanations become awkward before reorganization. Still, accumulation and representational change are different operations. One fills blanks on a map; the other asks whether the map should be replaced.
The Courage to Reject the Map
Once the difference lies in how a problem is represented, the bottleneck is no longer only data or compute. It is also the willingness to abandon a useful framework that may support an entire research program.
Maybe the whole way we are representing this problem is wrong?
People rarely stop to ask this. Admitting a representation may be wrong means reinterpreting earlier data, methods, and measures of success. Filling gaps on the old map is safer; most proposed replacements will be wrong. Science needs caution, but caution can harden into path dependence.
This is my deeper concern about an AI Scientist. If humans predefine its objectives, variables, units, and metrics, a capable system may only fill our gaps faster. Beyond an Explorer that searches and a Theorist that builds explanations, it needs a Representation Critic asking why this quantity is predicted, why these objects are basic, and what evidence would force a redefinition.
A controllable world model—or a physical body—does not solve this automatically. Concepts can arise from experience, mathematical consistency, or analogy, while a simulator trained on old theory may freeze old assumptions into “reality.” We still lack a reliable mechanism and test for strong Jumps.
In biology, this philosophical problem becomes an engineering choice: what do we treat as a cell state?
When a Cell Becomes an RNA Vector
“Virtual cell” suggests a cell reconstructed in a computer—metabolizing, sensing, and changing over time. Many current systems tackle a narrower, measurable task: given a starting transcriptome and a genetic, chemical, or cytokine perturbation, predict the change in RNA expression.
Representation here is not merely philosophical. Inputs, outputs, and loss functions define what counts as a cell and what matters; an RNA-only target pushes other differences aside.
That task is valuable. Arc Institute’s STATE moves beyond one cell line, using nearly 170 million observational cells and more than 100 million perturbed cells across about 70 human contexts. Intervention data let it learn not only which genes co-occur, but where expression may move after a change.
Yet the output remains primarily RNA. Is the transcriptome the cell state, or one measurement of it? RNA is rich, but it is not chromatin, protein activity, metabolism, morphology, spatial neighborhood, microenvironment, or history. Cells with similar RNA profiles may respond differently because they arrived there differently or occupy different tissues. Accuracy on the RNA projection alone does not prove that a model has found the mechanism generating it.
The 2025 Virtual Cell Challenge made this concern testable. Competitors adapted on some H1 human embryonic stem-cell perturbations, then predicted held-out H1 perturbations. Models did not reliably beat naive baselines on every metric, and leaders still combined deep learning with statistical features. This makes neither AI useless nor STATE a failure. It marks the frontier: broader human transcriptomic prediction is not yet a complete model across species, tissues, and patients.
When context changes, the first thing to question may not be parameter count alone, but the coordinate system itself.
Crossing Contexts, Not Just Reproducing Them
If HeLa, HEK293, and T cells each need a model built mainly from their own background, we may gain excellent local simulators without a general account of cells. My stronger hypothesis is a kind of cellular physics: shared mechanisms provide rules, while lineage, genotype, individual, tissue environment, history, and intervention provide boundary conditions.
“General” therefore does not mean putting every cell into a larger foundation model or erasing biological differences. It means separating what transfers from what must be conditioned on context. On this hypothesis, HeLa, HEK293, and T cells would be constrained realizations of shared mechanisms rather than sealed worlds.
Context cannot remain merely training data; it must become an explicit condition tested in combinations absent from training.
This is a research direction, not an achieved system. A 2024 Cell AI Virtual Cell roadmap calls for representations spanning species, modalities, and contexts while linking molecular, cellular, and tissue scales. But my hierarchy—shared mechanisms, then lineage, individual, environment, and intervention—is a synthesis, not STATE’s architecture. No single model has been broadly validated across all these scales and contexts.
Generalization cannot be compressed into one score. Mouse → human and HeLa → patient are different extrapolations: one crosses species; the other also crosses cell-line genetics, tissue organization, disease stage, patient background, and treatment history. A shared representation does not make either extrapolation valid by default. The Challenge tested held-out perturbations after partial H1 adaptation; it tested neither extrapolation directly.
Each step changes more than the test set; causal relationships may change with context.
Every disease claim therefore needs a boundary label: which cells, tissues, species, patient contexts, interventions, and time horizons were tested? Accuracy supports conclusions only within that background. A shared representation is not a shared intervention effect; without separating invariant rules from boundary conditions, new tools may repeat old failures from animals to humans and dishes to patients.
Ask the Model When It Will Fail
An RNA vector is not an inherently wrong representation of a cell, nor are rapid killing and tumor shrinkage inherently wrong measures of drug efficacy. They are boundaries: they decide which changes become visible. If benefit lies in a long-term disease trajectory or treatment timing, a narrow readout can turn “does this work?” into the wrong question.
Vorasidenib is a counterintuitive example. The FDA review records that the applicant supplied no in-vitro or in-vivo tumor-model data demonstrating antitumor activity or growth inhibition. That describes the submitted evidence; it does not prove that every cell and animal experiment failed or that preclinical models lack value.
INDIGO enrolled 331 patients with residual or recurrent, predominantly non-enhancing, IDH1/2-mutant WHO grade 2 astrocytoma or oligodendroglioma whose only prior anticancer therapy was surgery. Independent review found median progression-free survival of 27.7 months with vorasidenib versus 11.1 with placebo (HR 0.39). In August 2024, the FDA approved it for postsurgical patients aged 12 or older with susceptible IDH1- or IDH2-mutant grade 2 astrocytoma or oligodendroglioma.
This establishes neither efficacy across all gliomas, cure, nor an overall-survival benefit. Its narrower lesson is that clinical benefit can first appear as a delay in disease progression rather than immediate lesion removal. If “efficacy” includes only rapid cell death or dramatic shrinkage, a model may never be allowed to see the relevant answer.
The virtual cell’s test is therefore not merely reproducing RNA. Can it predict an unseen state, intervention, and context together? Can it name the cells, tissues, species, and time horizons its conclusion covers—and say “I don’t know” outside that range?
Its AlphaFold moment may begin not with one architecture winning a leaderboard, but with a reframing: cells and species as instances of shared mechanisms under different boundary conditions. A universal virtual cell in this strong sense remains a direction, not a finished system. Data and architecture matter; so does choosing the right representation of life.
Scientific courage is not always about answering faster. Sometimes it means admitting the question was wrong from the start.
References
- Tom Zahavy, LLMs can’t jump.
- Yong Zheng-Xin, Hot Take: LLM can ‘jump’.