Loading
What changed between revision 3 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 3
Import
Re-import of content/virtual-cell-models.md — the file changed after the initial import. Not an edit.
24.0 kB
Revision 3
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: "2026-10-16"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 42–48 → 42–484 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 oflines 72–82 → 72–8223 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." },lines 103–118 → 103–11920 unchanged lines not shown
The [Chan Zuckerberg Initiative](https://aiwiki.ai/wiki/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 scienceThe 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 fiveorganisation, 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 foryears, $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] Inoutside groups, and said the data it generates would be open and freely available.[^biohub2026a] InOctober [Biohub](https://aiwiki.ai/wiki/biohub), the US Department of Energy and the National Institutes of Health announced anOctober 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 measurementexpansion 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 totechnology, 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 DOEdate. The figure combines Biohub's original commitment, more than $500 million of planned DOEspending over five years, NIH datasets built through more than $500 million of earlier federalspending over five years, routed through the federal Genesis Mission for AI in science, NIH datasetsinvestment, and $300 million from Google DeepMind, Isomorphic Labs and Meta, with NVIDIA supplyingbuilt 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: neithercomputing 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.announcement names a system to be built or a test it would have to pass. lines 156–163 → 157–16437 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 192–198 → 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 205–210 → 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 214–253 → 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 brings these questions to the data rather than the models. Its stated rationale isThe 2026 funding raises these questions about the data rather than the models. Its stated rationalethe proponents' side of the argument set out below, that data is the binding constraint, put byis 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 exists today".[^biohub2026a]Biohub's head of science as a need for "orders of magnitude more data than existsThat is a premise, not a finding. Neither announcement names a licence beyond calling the result antoday".[^biohub2026a] That is a premise rather than a finding. Neither announcement names a licenceopen resource, gives a date for releasing data, describes how consent and access will be governedbeyond calling the result an open resource, gives a date for releasing data, describes how consentfor the human donors whose cells and records are pooled, or mentions misuse.[^biohub2026a][^biohub2026b]and access will be governed for the human donors whose cells and records are pooled, or mentionsWhat the corporate funders, one of them Meta, receive was explained to Reuters rather than in themisuse.[^biohub2026a][^biohub2026b] What the corporate funders, one of them Meta, receive wasannouncements: data paid for by commercial funders is held for an embargo period in which they canexplained to Reuters rather than in the announcements: data paid for by commercial funders is heldwork on it before it becomes a public resource, while the government-funded work carries no suchfor an embargo period in which they can work on it before it becomes a public resource, while therestriction. Biohub's head of science also told Reuters he expected a first dataset in about a yeargovernment-funded work carries no such restriction. Biohub's head of science also told Reuters heand accurate predictive models within five years.[^reuters2026]expected a first dataset in about a year and accurate predictive models within fiveThe gaps are the questions that attach to any pooled biomedical resource: who controls access, whetheryears.[^reuters2026] The gaps are the questions that attach to any pooled biomedical resource: whoa donor can withdraw, whether data held in federal repositories stays available as policy changes,controls access, whether a donor can withdraw, whether data held in federal repositories staysand whether models that predict how a person's cells respond could informavailable as policy changes, and whether models that predict how a person's cells respond could[[genetic-discrimination|decisions by insurers or employers]]. The embargo also sharpens theinform [[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 toconcentration described above: the companies that pay for the largest datasets are the first totrain on them.train 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, the premise the 2026 Virtual Biologythat billion-cell corpora will change the picture, the premise the 2026 Virtual Biology InitiativeInitiative commitments rest on; sceptics point out thatcommitments rest on. Sceptics point out that the same argument was made at each previous order ofthe same argument was made at each previous order of magnitude, and that a model trained only onmagnitude, and that a model trained only on transcriptomes may be missing variables rather than data.transcriptomes may be missing not data but variables. 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 279–284 → 299–30725 unchanged lines not shown
[^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.}[^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.}[^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, republished by CNBC. {Reports the embargo arrangement and the timelines as Biohub's head of science described them; the length of the embargo is not given.}[^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.
246 lines are the same in both revisions, and 168 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.