Loading
What changed between revision 1 and revision 6, compared over the article's markdown source with its frontmatter — so a change to status, evidence or reversibility shows up as the edit it is.
Older · revision 1
Import· initial corpus import, not an edit
Initial import of content/virtual-cell-models.md — the filesystem corpus, unchanged. Not an edit.
19.5 kB
Revision 1
Revision 6
lines 8–14 → 8–147 unchanged lines not shown
categories: ["foundations", "genetics"]categories: ["foundations", "genetics"]tags: ["machine learning", "systems biology", "simulation", "computational modelling", "single cell", "foundation models"]tags: ["machine learning", "systems biology", "simulation", "computational modelling", "single cell", "foundation models"]summary: "Computational models that simulate a cell's molecular behaviour, from complete cell cycles of minimal bacteria to learned predictions of how human cells respond to perturbation."summary: "Computational models that simulate a cell's molecular behaviour, from complete cell cycles of minimal bacteria to learned predictions of how human cells respond to perturbation."updated: "2026-08-22"updated: "2026-10-30"humanEvidence: "Complete whole-cell models exist only for minimal bacteria; learned models of human cell lines predict part of the transcriptional response to gene silencing, and none is validated as a model of a whole human cell."humanEvidence: "Complete whole-cell models exist only for minimal bacteria; learned models of human cell lines predict part of the transcriptional response to gene silencing, and none is validated as a model of a whole human cell."access: "Most models and datasets are public, including Recon3D, Geneformer, scGPT and Arc's State; commercial world models are proprietary and released as previews to selected academic and industry users."access: "Most models and datasets are public, including Recon3D, Geneformer, scGPT and Arc's State; commercial world models are proprietary and released as previews to selected academic and industry users."reversibility: "context"reversibility: "context"lines 30–37 → 30–3715 unchanged lines not shown
}}`````` **Virtual cell models** are computational systems that represent what happens inside a cell — DNA,**Virtual cell models** are computational systems that represent what happens inside a cell (DNA,RNA, proteins, regulatory networks and the processes they support — closely enough to predict itsRNA, proteins, regulatory networks and the processes they support) closely enough to predict itsbehaviour under a condition nobody has tested. Two lineages carry the name. Mechanistic whole-cellbehaviour under a condition nobody has tested. Two lineages carry the name. Mechanistic whole-cellmodels write down every reaction and integrate them forward in time; learned models are trained onmodels write down every reaction and integrate them forward in time; learned models are trained onmillions of measured single cells and asked to predict the next measurement. A complete whole-cellmillions of measured single cells and asked to predict the next measurement. A complete whole-celllines 42–54 → 42–584 unchanged lines not shown
A mechanistic whole-cell model aims to account for every annotated gene function in an organism atA mechanistic whole-cell model aims to account for every annotated gene function in an organism atonce. The 2012 model of *Mycoplasma genitalium* split the cell into 28 submodels covering 28once. The 2012 model of *Mycoplasma genitalium* split the cell into 28 submodels covering 28processes, spanning 401 genes, 722 compounds and 1,857 reactions, and let each submodel use whateverprocesses, spanning 401 genes, 722 compounds and 1,857 reactions, and let each submodel use whatevermathematics suited it — flux balance for metabolism, stochastic simulation for transcription — withmathematics suited it (flux balance for metabolism, stochastic simulation for transcription), withthe submodels resynchronised at short time steps.[^karr2012] The virtue is interpretability: everythe submodels resynchronised at short time steps.[^karr2012] The virtue is interpretability: everyprediction traces back to a stated mechanism, and a wrong prediction points at a wrong parameter.prediction traces back to a stated mechanism, and a wrong prediction points at a wrong parameter.The cost is that the parameters have to come from somewhere. That model needed roughly 1,800 ofThe cost is that the parameters have to come from somewhere. That model needed roughly 1,800 ofthem, assembled from the published literature on an organism with about 525 genes. A human cell hasthem, assembled from the published literature on an organism with about 525 genes. A human cell hasroughly twenty thousand.roughly twenty thousand. ```figure{"key": "mycoplasma-3d-cell-model", "caption": "A spatial model of the same bacterium the 2012 dynamical model simulated: a molecular census placed in space at one instant. No model of any cell yet combines this structure with dynamics."}``` Learned models take the opposite route and assume nothing about mechanism. Geneformer converts eachLearned models take the opposite route and assume nothing about mechanism. Geneformer converts eachcell's expression profile into a sequence of gene tokens ranked by abundance and trains acell's expression profile into a sequence of gene tokens ranked by abundance and trains atransformer over roughly 30 million single-cell transcriptomes;[^theodoris2023] scGPT and itstransformer over roughly 30 million single-cell transcriptomes;[^theodoris2023] scGPT and itslines 68–81 → 72–8613 unchanged lines not shown
```timeline```timeline[[ { "year": "1996–1999", "title": "E-Cell", "text": "Masaru Tomita's group at Keio University builds a simulation environment and assembles a hypothetical self-sustaining cell from 127 genes selected out of the Mycoplasma genitalium genome." }, { "year": "1996-1999", "title": "E-Cell", "text": "Masaru Tomita's group at Keio University builds a simulation environment and assembles a hypothetical self-sustaining cell from 127 genes selected out of the Mycoplasma genitalium genome." }, { "year": "2012", "title": "The first whole-cell model", "text": "Karr and colleagues in Markus Covert's laboratory publish a model of Mycoplasma genitalium covering every annotated gene function, and use it to direct experiments that recovered previously unmeasured kinetic parameters." }, { "year": "2012", "title": "The first whole-cell model", "text": "Karr and colleagues in Markus Covert's laboratory publish a model of Mycoplasma genitalium covering every annotated gene function, and use it to direct experiments that recovered previously unmeasured kinetic parameters." }, { "year": "2018", "title": "Human metabolism reconstructed", "text": "Recon3D assembles a genome-scale account of human metabolic reactions with associated protein structures, extending an effort begun a decade earlier." }, { "year": "2018", "title": "Human metabolism reconstructed", "text": "Recon3D assembles a genome-scale account of human metabolic reactions with associated protein structures, extending an effort begun a decade earlier." }, { "year": "2022", "title": "A minimal cell in silico", "text": "A whole-cell kinetic model of the 493-gene synthetic bacterium JCVI-syn3A reproduces how the cell balances metabolism, gene expression and growth across a cycle." }, { "year": "2022", "title": "A minimal cell in silico", "text": "A whole-cell kinetic model of the 493-gene synthetic bacterium JCVI-syn3A reproduces how the cell balances metabolism, gene expression and growth across a cycle." }, { "year": "2023–2024", "title": "Single-cell foundation models", "text": "Geneformer and scGPT are trained on tens of millions of transcriptomes and released, moving the field's centre of gravity from equations to learned representations." }, { "year": "2023-2024", "title": "Single-cell foundation models", "text": "Geneformer and scGPT are trained on tens of millions of transcriptomes and released, moving the field's centre of gravity from equations to learned representations." }, { "year": "2024", "title": "The AI virtual cell proposed", "text": "A 42-author perspective in Cell sets out what an AI Virtual Cell would have to be: universal representations across scales, queried through virtual instruments, judged against experiment." }, { "year": "2024", "title": "The AI virtual cell proposed", "text": "A 42-author perspective in Cell sets out what an AI Virtual Cell would have to be: universal representations across scales, queried through virtual instruments, judged against experiment." }, { "year": "2025", "title": "Benchmarks and data at scale", "text": "The Arc Institute releases State and launches the Virtual Cell Challenge; the Chan Zuckerberg Initiative starts the Billion Cells Project with 10x Genomics and Ultima Genomics." }, { "year": "2025", "title": "Benchmarks and data at scale", "text": "The Arc Institute releases State and launches the Virtual Cell Challenge; the Chan Zuckerberg Initiative starts the Billion Cells Project with 10x Genomics and Ultima Genomics." }, { "year": "2026", "title": "A bacterial cell cycle in four dimensions", "text": "A spatial and kinetic model simulates the roughly 100-minute cycle of JCVI-syn3A end to end, including growth and division." } { "year": "2026", "title": "A bacterial cell cycle in four dimensions", "text": "A spatial and kinetic model simulates the roughly 100-minute cycle of JCVI-syn3A end to end, including growth and division." }, { "year": "2026", "title": "The Virtual Biology Initiative", "text": "Biohub, the Chan Zuckerberg Initiative's science organisation, commits $500 million in April to open data generation for predictive cell models; in October it, the US Department of Energy and the National Institutes of Health announce an expansion the partners value at $1.8 billion." }]]`````` lines 88–103 → 93–1226 unchanged lines not shown
existing demonstration that the approach works, and it took thirty years and an organism with fewerexisting demonstration that the approach works, and it took thirty years and an organism with fewerthan five hundred genes.than five hundred genes. The learned side is where the money and the human cells are. Arc Institute's State was trained onThe learned side is where the money and the human cells are. Arc Institute's [State](https://aiwiki.ai/wiki/arc_state_model) was trained onobservational data from 167 million cells and perturbational data from over 100 million more,observational data from 167 million cells and perturbational data from over 100 million more,spanning seventy human cell contexts.[^arc2025] Reporting in *Nature* put its performance at roughly a thirdspanning seventy human cell contexts.[^arc2025] Reporting in *Nature* put its performance at roughly a thirdof the genes most strongly affected by a given perturbation in a test set, against 7% forof the genes most strongly affected by a given perturbation in a test set, against 7% forconventional methods.[^eisenstein2026] Xaira Therapeutics has released a perturbation atlas of 25.6conventional methods.[^eisenstein2026] Xaira Therapeutics has released a perturbation atlas of 25.6million cells and a model trained on it; Emma Lundberg's SubCell learns from microscope imagesmillion cells and a model trained on it; Emma Lundberg's SubCell learns from microscope imagesrather than transcriptomes, which is one answer to the objection that expression is not the cell.rather than transcriptomes, which is one answer to the objection that expression is not the cell.The Chan Zuckerberg Initiative's Billion Cells Project, launched in 2025 with 10x Genomics andThe [Chan Zuckerberg Initiative](https://aiwiki.ai/wiki/chan_zuckerberg_initiative)'s Billion Cells Project, launched in 2025 with 10x Genomics andUltima Genomics, exists to supply training data on a scale nobody has had.Ultima Genomics, exists to supply training data on a scale nobody has had. The largest bet on that premise came in 2026. [[biohub]], the Chan Zuckerberg Initiative's scienceorganisation, launched the [Virtual Biology Initiative](https://aiwiki.ai/wiki/virtual_biology_initiative) in April with $500 million over fiveyears, $400 million of it for data generation and new measurement technology and $100 million foroutside groups, and said the data it generates would be open and freely available.[^biohub2026a] InOctober Biohub, the US Department of Energy and the National Institutes of Health announced anexpansion that the partners valued at $1.8 billion in funding, data, computation and measurementtechnology, and described as the largest coordinated commitment to AI-ready biological data todate. The figure combines Biohub's original commitment, more than $500 million of planned DOEspending over five years, routed through the federal Genesis Mission for AI in science, NIH datasetsbuilt through more than $500 million of earlier federal investment, which add data rather than newmoney, and $300 million from Google DeepMind, [[isomorphic-labs|Isomorphic Labs]] and Meta, with NVIDIA supplyingcomputing infrastructure.[^biohub2026b] It is a commitment to measure cells, not a model: neitherannouncement names a system to be built or a test it would have to pass. Commercial systems claim more. In August 2026 GenBio AI, founded in 2024 with headquarters in PaloCommercial systems claim more. In August 2026 GenBio AI, founded in 2024 with headquarters in PaloAlto and laboratories in Paris and Abu Dhabi, previewed AIDO Cell, which the company describes as aAlto and laboratories in Paris and Abu Dhabi, previewed AIDO Cell, which the company describes as aworld model of the human cell spanning DNA, RNA, protein, regulatory networks and whole-cellworld model of the human cell spanning DNA, RNA, protein, regulatory networks and whole-celllines 138–145 → 157–16434 unchanged lines not shown
**Parameters nobody has measured.** The mechanistic route needs rate constants, concentrations and**Parameters nobody has measured.** The mechanistic route needs rate constants, concentrations andlocalisations for every reaction. For bacteria these were assembled from decades of literature. Forlocalisations for every reaction. For bacteria these were assembled from decades of literature. Fora human cell most of them have never been measured in any cell, let alone the one being modelled —a human cell most of them have never been measured in any cell, let alone the one being modelled.the same constraint that limits [[human-digital-twins|person-scale models]] one level up.The same constraint limits [[human-digital-twins|person-scale models]] one level up. **A noisy readout.** Single-cell RNA sequencing is the field's main training signal and it is noisy.**A noisy readout.** Single-cell RNA sequencing is the field's main training signal and it is noisy.Differences a model learns may be biology or may be artefacts of how the sample was handled, andDifferences a model learns may be biology or may be artefacts of how the sample was handled, andlines 174–180 → 193–19928 unchanged lines not shown
would have to demonstrate: representations that transfer across scales, a way to interrogate themwould have to demonstrate: representations that transfer across scales, a way to interrogate themthat resembles an instrument, and evaluation against experiment rather than against held-outthat resembles an instrument, and evaluation against experiment rather than against held-outdata.[^bunne2024] Much of what is marketed under the name meets none of those conditions, and thedata.[^bunne2024] Much of what is marketed under the name meets none of those conditions, and thelabel has spread faster than the capability behind it — a pattern already familiar fromlabel has spread faster than the capability behind it. The same pattern is familiar from[[ai-drug-discovery]], where the category has no regulatory definition and therefore no denominator.[[ai-drug-discovery]], where the category has no regulatory definition and therefore no denominator. The plausible near-term uses are narrower than the name suggests and real. Triaging whichThe plausible near-term uses are narrower than the name suggests and real. Triaging whichlines 187–192 → 206–2316 unchanged lines not shown
against. [[altos-labs]] took the generalist prize in the 2025 challenge, which indicates where someagainst. [[altos-labs]] took the generalist prize in the 2025 challenge, which indicates where someof the interest is coming from.[^arc2025]of the interest is coming from.[^arc2025] A neighbouring line of work uses language-model agents to interpret perturbation data that alreadyexists rather than to predict it. Xin Jin's laboratory at Scripps Research posted a whole-brainatlas as a preprint in March 2026: an in vivo Perturb-seq screen in 74 juvenile mice thatdisrupted 1,947 disease-associated genes and profiled more than 7.7 million nuclei.[^shi2026] InOctober the laboratory, OpenAI and PerturbAI reported giving each of 1,946 perturbations from itto its own agent running OpenAI's GPT-Rosalind model, with the atlas manuscript withheld. Theagents matched or refined 48 of the 55 findings in that manuscript that the supplied data couldsupport, and the run took 32 minutes; the comparison figure of months to a year of human work isan illustrative estimate built on assumed minutes per perturbation and per paper, not ameasurement. In a blinded review by 30 doctoral-level biologists, human and agent novelty scoresfor 60 reports correlated at a Spearman coefficient of 0.61, and in a separate comparison ofmatched pairs the experts preferred the distributed reports to single-agent ones in 45 of 60judgements.[^sun2026] Among the leads were a response in cerebellar neurons to loss of Tsc2 thatloss of its partner Tsc1 did not produce, and raised expression of cholesterol-synthesis genesafter Dctn1 disruption, which OpenAI's announcement presented as possible clues to braindisease.[^sun2026][^openai2026] Both are hypotheses from mouse data that no experiment has yettested. The authors say independent datasets and prospective experiments are needed to showwhether the method generalises. Eight of the 14 authors work for OpenAI or PerturbAI, and thepreprint had not been peer reviewed as of October 2026.[^sun2026] Risks are mostly of the ordinary kind. A confident wrong prediction that redirects a researchRisks are mostly of the ordinary kind. A confident wrong prediction that redirects a researchprogramme costs years, and a model that cannot be inspected is hard to argue with. Training corporaprogramme costs years, and a model that cannot be inspected is hard to argue with. Training corporaare drawn from immortalised cell lines and from donors who are not representative, so predictionsare drawn from immortalised cell lines and from donors who are not representative, so predictionslines 196–216 → 235–2733 unchanged lines not shown
applies weakly here, since a model of how a cell responds to a perturbation is not a design tool forapplies weakly here, since a model of how a cell responds to a perturbation is not a design tool forpathogens, though the distinction narrows as the models grow.pathogens, though the distinction narrows as the models grow. The 2026 funding raises these questions about the data rather than the models. Its stated rationaleis the proponents' side of the argument set out below, that data is the binding constraint, put byBiohub's head of science as a need for "orders of magnitude more data than existstoday".[^biohub2026a] That is a premise rather than a finding. Neither announcement names a licencebeyond calling the result an open resource, gives a date for releasing data, describes how consentand access will be governed for the human donors whose cells and records are pooled, or mentionsmisuse.[^biohub2026a][^biohub2026b] What the corporate funders, one of them Meta, receive wasexplained to Reuters rather than in the announcements: data paid for by commercial funders is heldfor an embargo period in which they can work on it before it becomes a public resource, while thegovernment-funded work carries no such restriction. Biohub's head of science also told Reuters heexpected a first dataset in about a year and accurate predictive models within fiveyears.[^reuters2026] The gaps are the questions that attach to any pooled biomedical resource: whocontrols access, whether a donor can withdraw, whether data held in federal repositories staysavailable as policy changes, and whether models that predict how a person's cells respond couldinform [[genetic-discrimination|decisions by insurers or employers]]. The embargo also sharpens theconcentration described above: the companies that pay for the largest datasets are the first totrain on them. ## Outlook## Outlook The test the field has set itself is clear enough to fail. A virtual cell would predict aThe field has set itself a test that a model can clearly fail. A virtual cell would predict aperturbation's effect in a cell type it was not trained on, and the prediction would hold whenperturbation's effect in a cell type it was not trained on, and the prediction would hold whensomeone ran the experiment. Nothing published meets that standard, and the honest reading of thesomeone ran the experiment. Nothing published meets that standard, and the 2025 benchmarks indicate2025 benchmarks is that scale alone has not yet produced it. Proponents argue that data is thethat scale alone has not yet produced it. Proponents argue that data is the binding constraint andbinding constraint and that billion-cell corpora will change the picture; sceptics point out thatthat billion-cell corpora will change the picture, the premise the 2026 Virtual Biology Initiativethe same argument was made at each previous order of magnitude, and that a model trained only oncommitments rest on. Sceptics point out that the same argument was made at each previous order oftranscriptomes may be missing not data but variables.magnitude, and that a model trained only on transcriptomes may be missing variables rather than data. Two harder questions sit behind the benchmark. Whether cells are predictable at all in the relevantTwo harder questions sit behind the benchmark. Whether cells are predictable at all in the relevantsense is not settled: a cell is a stochastic system, and the useful prediction may be a distributionsense is not settled: a cell is a stochastic system, and the useful prediction may be a distributionrather than a state. And a model of a cell in isolation says little about cells in tissue, whererather than a state. And a model of a cell in isolation says little about cells in tissue, wheresignalling between neighbours does much of the work — the reason a virtual cell is not a step on asignalling between neighbours does much of the work. For that reason a virtual cell is not a step onpath that ends at [[whole-brain-emulation]], despite the shared vocabulary. Ona path that ends at [[whole-brain-emulation]], despite the shared vocabulary. On[[technology-readiness-level|readiness]] the mechanistic branch is furthest along and applies to[[technology-readiness-level|readiness]] the mechanistic branch is furthest along and applies tobacteria; the branch that applies to human cells is earlier than its funding implies. The claim thatbacteria; the branch that applies to human cells is earlier than its funding implies. The claim thatprogress toward [[artificial-general-intelligence|more general AI systems]] will carry biology alongprogress toward [[artificial-general-intelligence|more general AI systems]] will carry biology alonglines 240–244 → 297–30723 unchanged lines not shown
[^roohani2025]: `paper` Roohani, Y. H. et al. "Virtual Cell Challenge: Toward a Turing test for the virtual cell." *Cell*, 2025. {Describes the benchmark and its dataset of around 300 CRISPR-interference perturbations in H1 human embryonic stem cells.}[^roohani2025]: `paper` Roohani, Y. H. et al. "Virtual Cell Challenge: Toward a Turing test for the virtual cell." *Cell*, 2025. {Describes the benchmark and its dataset of around 300 CRISPR-interference perturbations in H1 human embryonic stem cells.}[^eisenstein2026]: `news` Eisenstein, M. "Can biology move into the Matrix?" *Nature*, vol. 654, 4 June 2026, pp. 286–288. {A reported feature; the State and challenge figures in it are attributed to the researchers involved rather than independently verified.}[^eisenstein2026]: `news` Eisenstein, M. "Can biology move into the Matrix?" *Nature*, vol. 654, 4 June 2026, pp. 286–288. {A reported feature; the State and challenge figures in it are attributed to the researchers involved rather than independently verified.}[^genbio2026]: `statement` GenBio AI. "AIDO Cell: A General-Purpose Simulator for Cell Biology." Preview announcement, 18 August 2026. {A company describing its own unreleased system; no peer-reviewed description or prospective experimental validation accompanied it.}[^genbio2026]: `statement` GenBio AI. "AIDO Cell: A General-Purpose Simulator for Cell Biology." Preview announcement, 18 August 2026. {A company describing its own unreleased system; no peer-reviewed description or prospective experimental validation accompanied it.}[^biohub2026a]: `statement` Biohub. "Biohub Launches the Virtual Biology Initiative to Galvanize a Global Effort to Create the Open Data Foundation for AI-Accelerated Biology." Press release, 29 April 2026. {The funder announcing its own programme; it promises open data but names no licence, release date or validation target, and does not address donor consent or misuse.}[^biohub2026b]: `statement` Biohub. "International, cross-sector collaboration commits nearly $2 billion to build foundational data for AI models to predict and treat disease." Press release, 7 October 2026. {The partners' own valuation of a total that combines new money, planned agency spending, earlier federal investment and in-kind data and computing; it does not say what the corporate funders receive.}[^reuters2026]: `news` Hu, K. "US government, Google join Zuckerberg-backed Biohub in $1.8 billion push for AI biology data." Reuters, 7 October 2026. {Reports the embargo arrangement and the timelines as Biohub's head of science described them; the length of the embargo is not given.}[^arc2025]: `statement` Arc Institute. "Arc Institute's first virtual cell model: State," June 2025; and "Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections," December 2025. {The institute's own accounts: the State training figures and the challenge prize results are reported by the party that built one and ran the other.}[^arc2025]: `statement` Arc Institute. "Arc Institute's first virtual cell model: State," June 2025; and "Virtual Cell Challenge 2025 Wrap-Up: Winners and Reflections," December 2025. {The institute's own accounts: the State training figures and the challenge prize results are reported by the party that built one and ran the other.}[^shi2026]: `preprint` Shi, T. et al. "Genome-scale functional mapping of the mammalian whole brain with in vivo Perturb-seq." bioRxiv, 2026. doi:10.64898/2026.03.16.711480. {Loss-of-function edits in juvenile mice at a single timepoint with RNA readout only; the authors note AAV tropism leaves out some cell types.}[^sun2026]: `preprint` Sun, A. et al. "Distributed AI agents enable comprehensive discovery of biological insights at scale." bioRxiv, 2026. doi:10.64898/2026.10.15.759436. {Evaluated on one mouse atlas built by the same laboratory; none of the proposed mechanisms was tested experimentally in the study.}[^openai2026]: `statement` OpenAI. "Using thousands of AI agents to uncover new leads in biological data." Company blog post, 21 October 2026. {A co-author's own account; it reports the distributed agents' literature retrieval as more than 700 references, where the preprint reports about 711 times the single agent's 11.} removed, struck through added, underlinedLine numbers count the serialised markdown of each revision, frontmatter included.
224 lines are the same in both revisions, and 139 of them are not shown. Neither revision here is necessarily the current article — read Virtual cell models for the text this wiki stands behind today.