By GlioState Editorial Team作者：GlioState 编辑团队 · Published发布 · Last updated最后更新

AI Virtual Cell — Baseline Knowledge Base · First-Round Review

Generated 2026-10-12 (PT) | 139 items (current count), see the Library; all papers verified against Europe PMC / Crossref, web items verified against official pages. English version faithfully translated from the Chinese original.

1. Definition and milestones

  • Definition: Bunne et al. in Cell (2024) proposed the "AI Virtual Cell (AIVC)" — built on multimodal, multiscale universal representations, able to simulate in silico how cell states change under arbitrary perturbations, and to serve as a "virtual instrument" that guides experiments VC0001. Follow-ups include an oncology-oriented commentary VC0003, a review of preclinical translation VC0004, and an operational definition of a "virtual cell world model" proposed at ICML 2026 VC0118.
  • Milestones: scGen (2019) VC0015 → genome-wide Perturb-seq (2022) VC0040 → Geneformer / GEARS / CPA (2023) VC0009, VC0013, VC0014 → scGPT / scFoundation (2024) VC0008, VC0011 → Arc Virtual Cell Challenge (2025) VC0002, VC0105 → State and Tahoe-100M formally published in Cell in 2026 VC0019, VC0042. A Nature news feature summarizes the competition VC0007.

2. Key teams and models

  • CZI: the CELLxGENE data platform VC0043, TranscriptFormer (12 species, 112 million cells) VC0021, VC0102, the reasoning model rBio VC0101, and the Virtual Cells Platform VC0100.
  • Arc Institute: State VC0019, VC0104, scBaseCount VC0044, and the Virtual Cell Challenge 2025/2026 VC0105, VC0107, VC0058.
  • Companies: Xaira's X-Cell (4.9-billion-parameter diffusion model) and the X-Atlas dataset VC0033, VC0045, VC0108; Tahoe (formerly Vevo) with Tahoe-100M VC0041; Recursion/Valence with TxPert and the "predict–explain–discover" framework VC0018, VC0115; GenBio's AIDO.Cell VC0022, VC0113; Google/Yale's C2S-Scale VC0032; Noetik's tumor spatial models (licensed to GSK) VC0111; Bioptimus VC0114.
  • Academic models: UCE VC0012, GeneCompass VC0029, spatial foundation models such as Nicheformer / Novae / CellPLM VC0024, VC0025, VC0023, and pathology–spatial bridging models VC0026, VC0028. A perturbation-proteomics-based virtual cell model was published in Nature in 2026 VC0036.
  • Teams in China: BioMap (an improved scFoundation) took first place in the 2025 Challenge and Sichuan University took second VC0105.

3. Progress in brain tumors

  • State spaces and atlases (the "foundation" for modeling): Neftel proposed four GBM cell states VC0059; the GBmap reference atlas VC0060; spatial organization studies VC0061, VC0062, VC0092, VC0094; IDH-mutant glioma VC0063, VC0090; H3K27M diffuse midline glioma VC0064, VC0065; pediatric high-grade glioma (pHGG) VC0089; medulloblastoma VC0079, VC0080, VC0081; meningioma VC0082, VC0083, VC0084; brain metastases VC0085, VC0086, VC0087, VC0088.
  • Plasticity and perturbation data: immune cells drive transitions to MES-like states VC0066; spatial epigenomic niches VC0093; BRD2 VC0095; in vivo Perturb-seq in GBM (including radiotherapy) VC0070 is currently the resource closest to "virtual-cell training data for brain tumors".
  • Digital twins / mechanistic models: a mechanistic digital twin of the GBM astrocyte (preprint) VC0071; a GBM attractor model VC0072; a multiscale model plus reinforcement learning to optimize treatment scheduling VC0073; agent-based models and the classic "virtual brain tumor" VC0074, VC0075, VC0076; generative MRI-based growth prediction for DMG and a methodological critique of it VC0077, VC0078.
  • Organoids: the GBO organoid biobank VC0067, Core2Edge VC0096, automated drug screening VC0097.
  • Conclusion: as of this search round, no peer-reviewed paper was found that systematically applies single-cell foundation models such as scGPT, Geneformer or State to perturbation prediction in glioma (several Europe PMC queries returned no results). "AI foundation models" in brain tumors currently concentrate on MRI and pathology VC0098, VC0099, VC0039.

4. Critiques and benchmarks

Deep models do not outperform linear baselines VC0048; foundation models underperform simple baselines on perturbation prediction VC0049; zero-shot performance is limited VC0050; PertEval-scFM VC0051; Systema shows that systematic variation leads to overestimation VC0052; evaluation principles VC0053; the gap between models and baselines persists VC0054; attention weights do not provide gene-regulatory-network information VC0056. The 2025 Challenge exposed that its metrics could be "gamed", and in 2026 the format changed to zero-shot VC0105, VC0058. On the other hand, X-Cell reports scaling laws for perturbation prediction and claims to surpass linear baselines (not yet peer reviewed) VC0033, VC0110.

5. Conference calendar (2026–2027)

See the Conferences page for the full, data-driven list. Past meetings: ICLR LMRL (2026-04) VC0116, ICML FM4LS (2026-07) VC0117, ISMB 2026 VC0128, EANO 2026 (Rome) VC0123. Dates for RECOMB / ISMB 2027 are not yet confirmed.

6. Research outlook and gaps (brain tumors)

  1. Lack of brain-tumor perturbation data: Perturb-seq has mainly been run in cell lines such as K562 and RPE1; there are currently no large-scale Perturb-seq or drug-perturbation data in GSC, organoid or brain-slice systems (Tahoe-100M includes only a few glioma cell lines — to be re-checked).
  2. "State transitions" are a natural modeling task: whether transitions among NPC/OPC/AC/MES states, and state drift after radiochemotherapy or TTFields, can be predicted could be tested with GBmap plus longitudinal samples VC0069.
  3. Spatial and TME dimensions: the glioma infiltrative margin and neuron–tumor synapses have not yet entered virtual cell models.
  4. Cross-modal bridging: use paired samples of MRI/pathology foundation models and spatial transcriptomics to build practical "patient digital twins".
  5. Evaluation: linear baselines and zero-shot controls are essential (see VC0048, VC0052).
  6. Gap area: spinal cord tumors (intramedullary ependymoma, astrocytoma) have almost no single-cell atlases, let alone virtual cell studies — a distinctive niche for this hub.

7. Site plan

Sections: ① weekly digest (auto-updated, de-duplicated by seen IDs); ② model library (cards: task, data, open-source or not, benchmark performance); ③ glioma state-atlas navigator (Neftel states, GBmap, CELLxGENE entry); ④ benchmark & critique tracker; ⑤ conference calendar; ⑥ clinician's perspective column (bilingual).

First article topics:

  1. What is an AI virtual cell? A close reading of Bunne 2024 VC0001;
  2. Why do linear baselines keep beating deep models? VC0048, VC0049;
  3. The four GBM states and plasticity: a natural target for virtual cells VC0059, VC0066;
  4. In vivo Perturb-seq comes to glioma VC0070;
  5. Two years of the Arc Challenge and how teams from China performed VC0105, VC0058;
  6. From the "virtual brain tumor" to digital twins: twenty years of mechanistic models VC0075, VC0073, VC0071;
  7. Xaira vs. Noetik: cell-first or tissue-first? VC0108, VC0111;
  8. A survey of single-cell atlases of brain metastases and meningioma;
  9. What data can neurosurgeons contribute? (intraoperative margin sampling, longitudinal recurrence samples);
  10. A round-up of virtual-cell and AI abstracts at SNO 2026.

AI 虚拟细胞（AI Virtual Cell）领域基线知识库 · 首轮综述

生成日期：2026-10-12（PT）｜条目共 139 条（当前数量），见文献库；所有论文均经 Europe PMC / Crossref 核验，网页条目经官网核验。

一、领域定义与里程碑

  • 定义：Bunne 等在 Cell（2024）提出“AI 虚拟细胞（AIVC）”——以多模态、多尺度的通用表征为基础，可在计算机中模拟细胞在任意扰动下的状态变化，并作为“虚拟实验仪器”指导实验 VC0001。后续有面向肿瘤的解读 VC0003、临床前转化综述 VC0004，以及 ICML 2026 提出的“虚拟细胞世界模型”操作性定义 VC0118。
  • 里程碑：scGen（2019）VC0015 → 全基因组 Perturb-seq（2022）VC0040 → Geneformer / GEARS / CPA（2023）VC0009、VC0013、VC0014 → scGPT / scFoundation（2024）VC0008、VC0011 → Arc 虚拟细胞挑战赛（2025）VC0002、VC0105 → State、Tahoe-100M 于 2026 年正式发表在 Cell VC0019、VC0042。Nature 新闻特稿概述了这场竞赛 VC0007。

二、主要团队与模型

  • CZI：CELLxGENE 数据平台 VC0043、TranscriptFormer（12 个物种、1.12 亿细胞）VC0021、VC0102、推理模型 rBio VC0101、Virtual Cells Platform VC0100。
  • Arc Institute：State VC0019、VC0104、scBaseCount VC0044、虚拟细胞挑战赛 2025/2026 VC0105、VC0107、VC0058。
  • 企业：Xaira 的 X-Cell（49 亿参数扩散模型）及 X-Atlas 数据集 VC0033、VC0045、VC0108；Tahoe（原 Vevo）的 Tahoe-100M VC0041；Recursion/Valence 的 TxPert 和“预测-解释-发现”框架 VC0018、VC0115；GenBio 的 AIDO.Cell VC0022、VC0113；谷歌/耶鲁的 C2S-Scale VC0032；Noetik 的肿瘤空间模型（已授权 GSK）VC0111；Bioptimus VC0114。
  • 学术模型：UCE VC0012、GeneCompass VC0029、Nicheformer / Novae / CellPLM 等空间基础模型 VC0024、VC0025、VC0023，以及病理-空间桥接模型 VC0026、VC0028。基于扰动蛋白质组的虚拟细胞模型已发表于 Nature 2026 VC0036。
  • 中国团队：百图生科（scFoundation 改进版）获 2025 挑战赛第一，四川大学获第二 VC0105。

三、脑肿瘤方向进展

  • 状态空间与图谱（建模的“地基”）：Neftel 提出 GBM 的四种细胞状态 VC0059；GBmap 参考图谱 VC0060；空间组织研究 VC0061、VC0062、VC0092、VC0094；IDH 突变型 VC0063、VC0090；H3K27M 弥漫中线胶质瘤 VC0064、VC0065；儿童高级别胶质瘤（pHGG）VC0089；髓母细胞瘤 VC0079、VC0080、VC0081；脑膜瘤 VC0082、VC0083、VC0084；脑转移 VC0085、VC0086、VC0087、VC0088。
  • 可塑性与扰动数据：免疫细胞驱动 MES 样状态转变 VC0066；空间表观生态位 VC0093；BRD2 VC0095；GBM 体内 Perturb-seq（含放疗）VC0070 是目前最接近“脑肿瘤虚拟细胞训练数据”的资源。
  • 数字孪生 / 机制模型：GBM 星形细胞的机制性数字孪生（预印本）VC0071；GBM 吸引子模型 VC0072；多尺度模型+强化学习优化治疗时序 VC0073；主体模型与经典“虚拟脑肿瘤”VC0074、VC0075、VC0076；DMG 的 MRI 生成式生长预测及其方法学批评 VC0077、VC0078。
  • 类器官：GBO 类器官库 VC0067、Core2Edge VC0096、自动化药筛 VC0097。
  • 结论：截至本轮检索，尚未找到将 scGPT、Geneformer、State 等单细胞基础模型系统性用于胶质瘤扰动预测的同行评审论文（多组 Europe PMC 检索均无结果）。脑肿瘤领域的“AI 基础模型”目前集中在 MRI 和病理方向 VC0098、VC0099、VC0039。

四、批评与基准

深度模型不如线性基线 VC0048；基础模型在扰动预测上不如简单基线 VC0049；零样本表现有局限 VC0050；PertEval-scFM VC0051；Systema 指出系统性变异会造成高估 VC0052；评估原则 VC0053；模型与基线差距仍在 VC0054；注意力权重不能提供调控网络信息 VC0056。2025 挑战赛暴露了指标可被“刷分”的问题，2026 年改为零样本赛制 VC0105、VC0058。另一方面，X-Cell 报告了扰动预测的缩放规律，并称已超越线性基线（尚未经同行评审）VC0033、VC0110。

五、会议日历（2026–2027）

时间会议条目
2026-11-12至15SNO 2026，费城VC0122
2026年11月下旬虚拟细胞挑战赛 2026 结果公布（时间来自二手来源）VC0107
2026-12-11至12NeurIPS 2026 研讨会，悉尼（脑与身体基础模型 / 生物智能体 / SIMBIOCHEM）VC0119、VC0120、VC0121
2027-02-15至18Keystone 单细胞生物学会议，班夫VC0125
2027-04-02至07AACR 2027，奥兰多VC0124
2027-10-04至07Keystone AI 与基因调控会议（日期待复核）VC0126
2027-11-18至21SNO 2027，凤凰城VC0122

已结束的会议：ICLR LMRL（2026-04）VC0116、ICML FM4LS（2026-07）VC0117、ISMB 2026 VC0128、EANO 2026（罗马）VC0123。RECOMB / ISMB 2027 的日期尚未确认。

六、研究前景与空白（脑肿瘤）

  1. 缺少脑肿瘤扰动数据：Perturb-seq 主要在 K562、RPE1 等细胞系中开展；目前没有 GSC、类器官或脑片体系的大规模 Perturb-seq 或药物扰动数据（Tahoe-100M 只纳入少量胶质瘤细胞系，需复核）。
  2. “状态转换”是天然的建模任务：能否预测 NPC/OPC/AC/MES 之间的转换，以及放化疗或 TTFields 后的状态漂移，可用 GBmap 加纵向样本 VC0069 来检验。
  3. 空间与 TME 维度：胶质瘤浸润边缘和神经-肿瘤突触尚未进入虚拟细胞模型。
  4. 跨模态桥接：利用 MRI/病理基础模型与空间转录组的配对样本，构建可落地的“患者数字孪生”。
  5. 评估：必须设置线性基线和零样本对照（参考 VC0048、VC0052）。
  6. 空白领域：脊髓肿瘤（髓内室管膜瘤、星形细胞瘤）几乎没有单细胞图谱，更没有虚拟细胞研究——这是您的独特切入点。

七、对网站（gliostate.com）的建议

栏目设计：①每周速递（按 seen_ids 去重自动更新）；②模型库（卡片形式：任务、数据、是否开源、基准表现）；③胶质瘤状态图谱导航（Neftel 状态、GBmap、CELLxGENE 入口）；④基准与批评追踪；⑤会议日历；⑥临床医生视角专栏（中英双语）。

首批选题：

  1. 什么是 AI 虚拟细胞？精读 Bunne 2024 VC0001；
  2. 线性基线为何屡胜深度模型？VC0048、VC0049；
  3. GBM 四种状态与可塑性：虚拟细胞的天然靶标 VC0059、VC0066；
  4. 体内 Perturb-seq 进入胶质瘤 VC0070；
  5. Arc 挑战赛两年复盘与中国团队的表现 VC0105、VC0058；
  6. 从“虚拟脑肿瘤”到数字孪生：二十年机制模型史 VC0075、VC0073、VC0071；
  7. Xaira 与 Noetik：细胞优先还是组织优先？VC0108、VC0111；
  8. 脑转移与脑膜瘤的单细胞图谱一览；
  9. 神经外科医生能贡献什么数据？（术中边缘取样、纵向复发样本）；
  10. SNO 2026 中的虚拟细胞与 AI 摘要盘点。

Frequently asked questions常见问题

Which models predict cellular responses to perturbations?哪些模型可预测细胞对扰动的反应？

Examples include GEARS, CPA and Arc Institute’s State model, often built on cell embeddings from single-cell foundation models such as scGPT and Geneformer. Reported performance varies by dataset and context. Independent benchmarks should be checked before relying on any single model.代表性模型包括 GEARS、CPA 和 Arc 研究所的 State，常以 scGPT、Geneformer 等单细胞基础模型的细胞表征为基础。其表现随数据集与情境而异，采用任何单一模型前都应参考独立基准评测。 VC0013 VC0014 VC0019 VC0008 VC0009

What is the Virtual Cell Challenge?什么是虚拟细胞挑战赛？

It is a public benchmark competition launched by the Arc Institute, described as a “Turing test” for virtual cells. The 2026 edition focuses on zero-shot generalization of perturbation prediction across cellular contexts.这是 Arc 研究所发起的公开基准竞赛，被称为虚拟细胞的“图灵测试”。2026 年度赛事聚焦扰动预测在不同细胞情境间的零样本泛化。 VC0002 VC0058

Do deep learning models beat simple baselines for perturbation prediction?深度学习模型在扰动预测上能超过简单基线吗？

Not reliably, according to published benchmarks. A Nature Methods study reported that deep-learning gene perturbation prediction does not yet outperform simple linear baselines, and a benchmark of foundation cell models reached related conclusions. Claims of progress should therefore be compared against such baselines.依据已发表的基准研究，尚不能稳定超过。Nature Methods 的一项研究报告深度学习基因扰动预测尚未超越简单线性基线，另一项基础细胞模型基准研究也得出相近结论。因此任何进展声明都应与此类基线比较。 VC0048 VC0049