Skip to content
ContentLora

    Tip: press / anywhere to search.

    technology

    Virtual cell models: AI that predicts how cells respond

    Also known as virtual cell, AI virtual cell, AIVC, cell foundation models, perturbation prediction models

    Virtual cell models aim to be AI simulations of cells, learned from biological data, that predict how cells respond to perturbations such as switching off a gene.[1][2] Large single-cell foundation models exist,[3][4] but in a 2025 study none beat simple baselines at predicting perturbation effects,[5] and the first Arc Institute Virtual Cell Challenge found models not yet consistently ahead of naive baselines.[6]

    Editor reviewedUpdated Wiki14 claims · 100% primaryAI for science, Life sciences, Artificial intelligence
    Key facts

    The idea

    A 2024 perspective by Charlotte Bunne and co-authors defines AI virtual cells as high-fidelity simulations of cells under different conditions, learned directly from biological data across measurements and scales.[1] Its authors hope such models will help identify drug targets, predict how cells respond to perturbations and scale up hypothesis testing.[2] For an earlier AI-for-biology tool, see AlphaFold and how AI predicts protein structures.

    The practical motive is drug discovery. The Arc Institute notes that about 90% of drugs fail clinical trials because of poor efficacy or unintended side effects, and says a highly predictive virtual cell model could help researchers discover drugs on a computer first.[7]

    Early models

    Geneformer, published in 2023, was pretrained on about 30 million single-cell transcriptomes to make predictions where data are scarce.[3] scGPT, published in 2024, was trained across more than 33 million cells and adapted for tasks including cell type annotation and perturbation response prediction.[4]

    How well they predict

    A 2025 Nature Methods study compared five foundation models and two other deep learning models with deliberately simple baselines for predicting gene-expression changes after one or two genetic perturbations, and none outperformed the baselines.[5] The 2024 perspective itself lists evaluation strategies, data needs and community standards among the requirements still to be met.[8]

    The Virtual Cell Challenge

    In June 2025 the Arc Institute launched a public competition, with a grand prize worth $100,000, for the model that best predicts how cells respond to genetic perturbations.[9] Arc framed it as a test of whether a model can stand in for a real Perturb-seq experiment.[10] The benchmark used about 300,000 single-cell profiles from human embryonic stem cells with 300 CRISPRi perturbations, chosen to differ from common training data.[11]

    More than 5,000 people from 114 countries registered, and more than 300 teams made final submissions.[12] Arc said the results showed models were not yet consistently beating naive baselines across all metrics, although they improved at telling perturbations apart and at finding differentially expressed genes.[6] The winners combined deep learning with classical statistical features, and almost all models did worse than the baseline on mean absolute error.[13]

    The 2026 round is harder: models must predict gene knockdown responses in cell lines they have not seen before.[14] For other AI-in-biology claims and how well they hold up, see the evidence on AI-designed drugs.

    More in AI for science

    Developing in AI & Computing

    Questions readers ask

    What is a virtual cell?

    A proposed AI model that simulates cells under different conditions, learned directly from biological data. Its backers hope it could help find drug targets and predict how cells respond to perturbations.[1][2]

    Do virtual cell models work yet?

    Not reliably for their main task. A 2025 study found no deep learning model beat simple baselines at predicting perturbation effects, and the 2025 Virtual Cell Challenge found models were not yet consistently ahead of naive baselines.[5][6]

    What is the Virtual Cell Challenge?

    A recurring public competition from the Arc Institute, launched in June 2025 with a grand prize worth $100,000, for the model that best predicts how cells respond to genetic perturbations.[9]

    Why do researchers want virtual cells?

    Arc points to drug development, noting that about 90% of drugs fail clinical trials because of poor efficacy or side effects, and says a highly predictive model could help researchers discover drugs on a computer.[7]

    Sources

    · 14 claims from 5 publishers

    Each numbered claim is a statement we checked against the sources listed with it. Tap a number in the text to see its sources in place.

    1. [1]

      A 2024 perspective by Charlotte Bunne and co-authors defines AI virtual cells as high-fidelity simulations of cells and cellular systems under different conditions, learned directly from biological data across measurements and scales. confirmedas of 2024-10-14

    2. [2]

      The 2024 perspective envisions virtual cells helping to identify drug targets, predict how cells respond to perturbations and scale up hypothesis exploration. confirmedas of 2024-10-14

    3. [3]

      Geneformer, published in Nature in 2023, is an attention-based deep learning model pretrained on about 30 million single-cell transcriptomes to make predictions in network biology where data are limited. confirmedas of 2023-05-31

    4. [4]

      scGPT, published in Nature Methods in 2024, is a single-cell foundation model built on a generative pretrained transformer across more than 33 million cells, adapted for tasks including cell type annotation and perturbation response prediction. confirmedas of 2024-02-26

    5. [5]

      A 2025 Nature Methods study compared five foundation models and two other deep learning models with deliberately simple baselines for predicting gene-expression changes after single or double genetic perturbations, and none outperformed the baselines. confirmedas of 2025-08-04

    6. [6]

      Arc said the 2025 results showed perturbation prediction models were not yet consistently beating naive baselines across all metrics, though they improved on perturbation discrimination and on identifying differentially expressed genes. confirmedas of 2025-12-06

    7. [7]

      Arc's case for virtual cells rests on drug development, noting that about 90% of drugs fail clinical trials because of poor efficacy or unintended side effects, and saying a highly predictive virtual cell model could help researchers discover new drugs entirely in silico. confirmedas of 2025-12-06

    8. [8]

      The 2024 perspective lists data needs, evaluation strategies and community standards among the requirements for realising virtual cells. confirmedas of 2024-10-14

    9. [9]

      In June 2025 the Arc Institute launched the Virtual Cell Challenge, a public competition with a grand prize worth $100,000 for the model that best predicts how cells respond to genetic perturbations. confirmedas of 2025-06-26

    10. [10]

      Arc framed the challenge around whether a computational model can stand in for an actual Perturb-seq experiment, predicting gene-expression changes from a starting cell state and a genetic perturbation. confirmedas of 2025-12-06

    11. [11]

      For the 2025 challenge Arc generated a dataset of about 300,000 single-cell profiles from H1 human embryonic stem cells with 300 CRISPRi gene perturbations, chosen as a shift from typical training data so that models had to generalise. confirmedas of 2025-12-06

    12. [12]

      Arc reported that more than 5,000 people from 114 countries registered for the 2025 challenge, more than 1,200 teams submitted results and more than 300 made final submissions. confirmedas of 2025-12-06

    13. [13]

      Arc reported that the winning approaches combined deep learning with classical statistical features, and that almost all models did worse than the baseline on mean absolute error. confirmedas of 2025-12-06

    14. [14]

      The 2026 Virtual Cell Challenge asks for zero-shot prediction of gene knockdown responses in a new Arc-generated dataset of cell lines the models have not seen. confirmedas of 2026-08-01

    Revision history (1)
    1. Page created.

    Created Nov 14, 2026. Last reviewed by an editor on Nov 14, 2026. Next scheduled review: Mar 14, 2027.

    Cite this page

    "Virtual cell models: AI that predicts how cells respond." ContentLora, updated Nov 14, 2026. https://contentlora.com/wiki/virtual-cell-models

    Spotted an error? Suggest a correction or emailcorrections@contentlora.com.