← Reviews← 综述 · Draft初稿
Glioblastoma cell states and AI virtual cells胶质母细胞瘤细胞状态与虚拟细胞
Author: GlioState Editorial Team | First draft 2026-10-13 (PT) | Every citation is a verified item in this site's library (VC id) linked to its DOI or official page; preprints are marked and are not peer reviewed.
1. What are glioblastoma (GBM) "cell states"?
- The Neftel four-state framework: Using single-cell RNA-seq, Neftel et al. (Cell 2019) described four malignant cell states in GBM — NPC-like (neural-progenitor-like), OPC-like (oligodendrocyte-progenitor-like), AC-like (astrocyte-like) and MES-like (mesenchymal-like). Most tumors contain several states; their proportions are associated with genetic alterations in CDK4, PDGFRA, EGFR and NF1, and cells can transition between states (plasticity) VC0059.
- Drivers of plasticity: interactions between immune and cancer cells drive transitions to MES-like states VC0066; BRD2 bromodomain regulation of state plasticity modulates therapy response VC0095; spatial epigenomic niches are linked to plasticity (preprint) VC0093; hypoxia coordinates the spatial landscape of myeloid cells and affects survival VC0094.
Other frameworks and complementary views:
- Neurodevelopmental hierarchy: GBM recapitulates a normal neurodevelopmental hierarchy VC0068;
- Spatial organization: multi-layered spatial organization VC0061, bidirectional tumor–host interdependence VC0062, niche-specific states from core to edge (preprint) VC0092;
- Evolution under therapy: a single-cell atlas of GBM under treatment VC0069;
- Reference atlas: GBmap, a single-cell and spatial reference atlas of >1M cells VC0060;
- Other gliomas: IDH-mutant glioma VC0063, VC0090; H3K27M diffuse midline glioma VC0064, VC0065; a longitudinal atlas of pediatric high-grade glioma VC0089.
- Take-home: cell states are continuous, transitional and shaped by the microenvironment and by therapy, not fixed "subtype labels" — which is exactly what a virtual cell aims to model.
2. How do AI virtual cell models represent and predict states and perturbations?
- Definition: Bunne et al. (Cell 2024) define the AI Virtual Cell (AIVC) as built on multimodal, multiscale universal cell representations that simulate in silico how cell states change under arbitrary perturbations VC0001; a "virtual cell world model" has since been proposed VC0131.
- Representation (cell embeddings): single-cell foundation models such as scGPT VC0008, Geneformer VC0009 and scFoundation VC0011 map cells into an embedding space in which states such as NPC/OPC/AC/MES could, in principle, be located.
- Predicting perturbation responses: GEARS VC0013, CPA VC0014, neural optimal transport (CellOT) VC0016, Arc's State VC0019; generative approaches such as AlphaCell (preprint) VC0137 and PerturbLDM (preprint) VC0138; unified genetic/chemical perturbation embeddings in UniPert-G2CP VC0130.
Attempts in GBM so far:
- scFOCAL/ISOSCELES links GBM cell states to L1000 drug signatures to predict state-specific sensitivity and combinations VC0132;
- DELPHAI (preprint) predicts temozolomide perturbation responses in GBM organoids VC0129;
- In vivo Perturb-seq in GBM and its microenvironment dissects drivers and radiotherapy responses — exactly the paired "perturbation → state" data virtual cells need VC0070;
- Boolean-network attractor analysis from scRNA-seq treats states as stable attractors of a regulatory network VC0072.
- Evidence to keep in view: benchmarks show deep models do not yet reliably beat simple linear baselines for perturbation prediction VC0048, VC0049; zero-shot performance of foundation models is limited VC0050; the Virtual Cell Challenge focuses on zero-shot generalization across contexts VC0058.
- As of this search round: no peer-reviewed study was found that systematically uses general models such as scGPT / Geneformer / State to predict GBM cell-state transitions (a search finding, not proof of absence).
3. GBM-specific digital twins and simulation
- Classic tissue/imaging-scale models: the proliferation–invasion "virtual and real brain tumors" model VC0075; a multiscale DTI-based model of spread along white-matter tracts VC0076.
- Patient-specific xenograft simulation (PLoS Comput Biol 2026): simulates growth and invasion of patient-derived GBM xenografts on mouse brain anatomy, fits parameters with approximate Bayesian computation (ABC), and models treatment via parameter perturbation VC0133.
- TumorTwin (BMC Med Inform Decis Mak 2026): an open, modular Python framework for patient-specific tumor digital twins, demonstrated on high-grade glioma growth and radiotherapy response VC0134.
- Cell/mechanistic scale: a proteome-calibrated mechanistic digital twin of the GBM astrocyte (preprint) VC0071; an agent-based hybrid biophysical model VC0074; a multiscale mechanistic model combined with reinforcement learning to optimize glioma combination-treatment scheduling VC0073.
- Generative AI for imaging growth: a proof-of-principle MRI growth-forecasting study in pediatric DMG VC0077, and a critique raising domain-shift and statistical-nesting concerns VC0078.
- Experimental "twin" platforms: a GBM organoid biobank VC0067, the Core2Edge organoid–brain slice infiltration model VC0096, and automated personalized organoid drug screening VC0097.
- Gap: imaging/tissue-scale twins and single-cell-state virtual cells are still largely separate tracks; no mature coupling of the two was found.
4. Evidence status at a glance
| Study | Data type | Model | Independently validated? | Clinical distance |
|---|---|---|---|---|
| Neftel four states VC0059 | scRNA-seq + model experiments | State scoring / lineage analysis | Reproduced in later atlases (e.g. VC0060) | Research tool; not used for clinical decisions |
| GBmap VC0060 | Single-cell + spatial | Reference atlas | A resource, not a predictive model | Far |
| In vivo Perturb-seq VC0070 | In vivo CRISPR perturbation + scRNA-seq | Experimental data | Data awaiting models | Far |
| scFOCAL/ISOSCELES VC0132 | scRNA-seq + L1000 | State–drug signature integration | Experimental validation in paper; no independent replication found | Mid-far (preclinical) |
| DELPHAI VC0129 | GBM organoids + TMZ | Perturbation prediction + LLM agent | Preprint, not peer reviewed | Far |
| GBM attractors VC0072 | scRNA-seq | Boolean network | Computational | Far |
| Xenograft simulation VC0133 | Mouse xenografts | Anatomy-aware growth model + ABC | Fitted to mouse data | Far (preclinical) |
| TumorTwin VC0134 | Imaging (HGG demo) | Patient-specific digital twin framework | Framework/demo; no prospective validation found | Mid (usable framework, not clinically validated) |
| RL scheduling VC0073 | Multiscale mechanistic model | Model + RL | Computational | Far |
| Pediatric DMG generative growth VC0077 | MRI | Generative AI | Proof of principle; critiqued VC0078 | Mid-far |
Note: "independently validated" and "clinical distance" are the author's qualitative judgments based on study type, not results of a systematic review.
5. Open problems
- Scarce GBM perturbation data: large perturbation datasets come mostly from cell lines such as K562 and RPE1; paired perturbation data in GSCs, organoids and in vivo remain rare (VC0070 is an exception).
- Are state transitions predictable? Whether a virtual cell can predict shifts in NPC/OPC/AC/MES proportions after radiochemotherapy or other interventions could be tested against pre/post-treatment atlases (VC0069).
- Baselines: any GBM virtual-cell claim should be compared with simple linear baselines (VC0048, VC0049).
- Coupling scales: how to connect single-cell state models with TumorTwin-style imaging twins (VC0134) and anatomy-aware simulation (VC0133).
- Space and microenvironment: states are shaped by immune cells, hypoxia and niches (VC0066, VC0094, VC0092); modeling tumor cells alone may not be enough.
- Spinal cord intramedullary tumors: almost no single-cell atlases and no virtual-cell studies; the library has no items on them yet.
6. 临床点评 / Clinical commentary (to be written by the author)
[Placeholder] To be written by the GlioState Editorial Team: a neuro-oncology view of what these models may mean for surgical sampling, multi-region biopsy, recurrence assessment and trial design. Not yet written — do not cite.
Disclaimer
For research and educational purposes only. This is not medical advice and not a recommendation of any diagnostic or therapeutic approach, product, company or investment. Patients should consult their own physicians.
作者：GlioState 编辑团队｜初稿 2026-10-13（PT）｜所有引用均为本站文献库（VC 编号）中已核验条目，并链接至 DOI 或官方页面；预印本已标注，未经同行评审。
一、胶质母细胞瘤（GBM）的“细胞状态”是什么？
- Neftel 四状态框架：Neftel 等（Cell 2019）基于单细胞 RNA 测序，将 GBM 恶性细胞归纳为四种状态——NPC 样（神经祖细胞样）、OPC 样（少突胶质前体细胞样）、AC 样（星形胶质细胞样）和 MES 样（间充质样）；同一肿瘤内通常并存多种状态，各状态比例与 CDK4、PDGFRA、EGFR、NF1 等遗传改变相关，并且细胞可在状态间转换（可塑性）VC0059。
- 可塑性的驱动因素：免疫细胞与肿瘤细胞的相互作用可推动细胞向 MES 样状态转变 VC0066；BRD2 溴结构域调控状态可塑性并影响治疗反应 VC0095；空间表观基因组微环境与可塑性相关（预印本）VC0093；低氧塑造髓系细胞的空间分布并与生存相关 VC0094。
其他框架与补充视角：
- 神经发育层级：GBM 重现正常神经发育层级 VC0068；
- 空间组织：多层次空间结构 VC0061、肿瘤—宿主双向依赖 VC0062、从核心到边缘的生态位特异状态（预印本）VC0092；
- 治疗下的演化：放化疗下 GBM 的单细胞演化图谱 VC0069；
- 参考图谱：GBmap（>100 万细胞的单细胞与空间参考图谱）VC0060；
- 其他胶质瘤：IDH 突变型胶质瘤 VC0063、VC0090；H3K27M 弥漫性中线胶质瘤 VC0064、VC0065；儿童高级别胶质瘤纵向图谱 VC0089。
- 小结：细胞状态是连续、可转换、受微环境与治疗影响的，而非固定的“亚型标签”——这正是“虚拟细胞”希望建模的对象。
二、AI 虚拟细胞如何表示与预测状态和扰动？
- 定义：Bunne 等（Cell 2024）提出 AI 虚拟细胞（AIVC）：以多模态、多尺度的通用细胞表示为基础，在计算机中模拟细胞状态在任意扰动下的变化 VC0001；近期有“虚拟细胞世界模型”的提法 VC0131。
- 表示（细胞嵌入）：单细胞基础模型如 scGPT VC0008、Geneformer VC0009、scFoundation VC0011，可将细胞映射到嵌入空间，理论上可在其中定位 NPC/OPC/AC/MES 等状态。
- 预测扰动响应：GEARS VC0013、CPA VC0014、神经最优传输 CellOT VC0016、Arc 的 State VC0019；生成式方法如 AlphaCell（预印本）VC0137、PerturbLDM（预印本）VC0138；遗传与化学扰动统一表示 UniPert-G2CP VC0130。
GBM 中已有的尝试：
- scFOCAL/ISOSCELES 将 GBM 细胞状态与 L1000 药物特征关联，预测状态特异的药物敏感性及组合 VC0132；
- DELPHAI（预印本）在 GBM 类器官中预测替莫唑胺扰动响应 VC0129；
- 体内 Perturb-seq 在 GBM 及其微环境中解析驱动基因与放疗反应——这类数据是训练虚拟细胞所需的“扰动—状态”配对数据 VC0070；
- 基于 scRNA-seq 的布尔网络吸引子分析，把状态视为调控网络的稳定吸引子 VC0072。
- 需要清醒看待的证据：多项基准显示，深度模型在扰动预测上尚未稳定超过简单线性基线 VC0048、VC0049；基础模型的零样本表现有限 VC0050；虚拟细胞挑战赛关注跨情境的零样本泛化 VC0058。
- 截至本轮检索：尚未发现在同行评审文献中，系统地用 scGPT / Geneformer / State 等通用模型预测 GBM 细胞状态转换的研究（为检索结论，非穷尽证明）。
三、GBM 的数字孪生与模拟
- 组织/影像尺度的经典模型：增殖—侵袭（“虚拟与真实脑肿瘤”）模型 VC0075；基于 DTI 的沿白质纤维扩散多尺度模型 VC0076。
- 患者特异异种移植模拟（PLoS Comput Biol 2026）：在小鼠脑解剖结构上模拟患者来源 GBM 异种移植的生长与侵袭，用近似贝叶斯计算（ABC）拟合参数，并通过参数扰动模拟治疗 VC0133。
- TumorTwin（BMC Med Inform Decis Mak 2026）：开源、模块化的 Python 框架，用于构建患者特异的肿瘤数字孪生，并以高级别胶质瘤的生长与放疗反应为示例 VC0134。
- 细胞/机制尺度：GBM 星形细胞的蛋白组校准机制数字孪生（预印本）VC0071；基于主体的混合生物物理模型 VC0074；多尺度机制模型结合强化学习优化胶质瘤联合治疗时序 VC0073。
- 生成式 AI 预测影像生长：儿童 DMG 的 MRI 空间生长预测原理验证研究 VC0077，以及对其领域偏移与统计嵌套问题的批评 VC0078。
- 实验“孪生”平台：GBM 类器官生物库 VC0067、Core2Edge 类器官—脑片侵袭模型 VC0096、自动化个体化类器官药筛 VC0097。
- 缺口：目前影像/组织尺度的孪生与单细胞状态尺度的虚拟细胞基本是两条独立路线，尚未见二者的成熟耦合。
四、证据现状一览
| 研究 | 数据类型 | 模型 | 是否经独立验证？ | 距临床 |
|---|---|---|---|---|
| Neftel 四状态 VC0059 | scRNA-seq + 模型实验 | 状态评分/谱系分析 | 多个后续图谱复现（如 VC0060） | 研究工具，未用于临床决策 |
| GBmap VC0060 | 单细胞 + 空间 | 参考图谱 | 资源本身，非预测模型 | 远 |
| 体内 Perturb-seq VC0070 | 体内 CRISPR 扰动 + scRNA-seq | 实验数据 | 数据，待模型使用 | 远 |
| scFOCAL/ISOSCELES VC0132 | scRNA-seq + L1000 | 状态—药物特征整合 | 文中有实验验证；未见独立复现 | 中远（临床前） |
| DELPHAI VC0129 | GBM 类器官 + TMZ | 扰动预测 + LLM 代理 | 预印本，未经同行评审 | 远 |
| GBM 吸引子 VC0072 | scRNA-seq | 布尔网络 | 计算分析 | 远 |
| 异种移植模拟 VC0133 | 小鼠异种移植 | 解剖感知生长模型 + ABC | 小鼠数据拟合 | 远（临床前） |
| TumorTwin VC0134 | 影像（HGG 示例） | 患者特异数字孪生框架 | 框架/示例，未见前瞻验证 | 中（框架可用，未临床验证） |
| 强化学习排程 VC0073 | 多尺度机制模型 | 模型+RL | 计算研究 | 远 |
| 儿童 DMG 生成式生长 VC0077 | MRI | 生成式 AI | 原理验证；已有批评 VC0078 | 中远 |
注：“是否经独立验证”“距临床”为作者基于文献类型的定性判断，非系统评价结论。
五、开放问题
- GBM 扰动数据稀缺：大规模扰动数据主要来自 K562、RPE1 等细胞系；GSC、类器官和体内的配对扰动数据仍很少（VC0070 是少数例外）。
- 状态转换是否可预测：能否用虚拟细胞预测放化疗或其他干预后 NPC/OPC/AC/MES 比例的漂移，可用治疗前后图谱（VC0069）检验。
- 基线问题：任何 GBM 虚拟细胞声明都应与简单线性基线比较（VC0048、VC0049）。
- 尺度耦合：如何把单细胞状态模型与 TumorTwin 式影像孪生（VC0134）、解剖感知模拟（VC0133）连接。
- 空间与微环境：状态受免疫、低氧与生态位影响（VC0066、VC0094、VC0092），仅建模肿瘤细胞本身可能不够。
- 脊髓髓内肿瘤：几乎没有单细胞图谱，更没有虚拟细胞研究；本库中尚无相关条目。
六、临床点评 / Clinical commentary（待作者撰写）
【占位】此部分由 GlioState 编辑团队撰写：从神经肿瘤学视角讨论这些模型对手术取样、多区域活检、复发评估和临床试验设计可能的意义。当前尚未撰写，请勿引用。
声明
本文仅供科研与教育用途，不构成医疗建议，也不构成对任何诊疗方法、产品、公司或投资的推荐。患者请咨询主治医生。
Frequently asked questions常见问题
What are glioblastoma cell states?什么是胶质母细胞瘤细胞状态？
Glioblastoma cell states are recurring transcriptional programs of malignant cells. Neftel et al. (Cell, 2019) described four — NPC-like, OPC-like, AC-like and MES-like — that coexist within most tumors and can interconvert (plasticity). Later work links transitions, for example toward MES-like states, to immune interactions and spatial niches, so states are best seen as dynamic rather than fixed subtypes.GBM 细胞状态是恶性细胞反复出现的转录程序。Neftel 等（Cell，2019）描述了四种状态——NPC 样、OPC 样、AC 样与 MES 样——它们在多数肿瘤中共存并可相互转换（可塑性）。后续研究将状态转换（如向 MES 样转变）与免疫相互作用及空间微环境相关联，因此细胞状态更应被视为动态的，而非固定亚型。 VC0059 VC0066 VC0093 (preprint)（预印本）
Can virtual cells predict GBM treatment response?虚拟细胞能预测 GBM 的治疗反应吗？
Not yet in a clinically validated way. Early studies link GBM cell states to drug signatures (scFOCAL/ISOSCELES) or predict temozolomide response in organoids (DELPHAI, a preprint), and in vivo Perturb-seq has measured radiotherapy responses. General benchmarks show that perturbation models still struggle to beat simple baselines, so these results should be read as preclinical research.目前尚无经临床验证的能力。早期研究将 GBM 细胞状态与药物特征谱关联（scFOCAL/ISOSCELES），或预测类器官的替莫唑胺反应（DELPHAI，预印本）；体内 Perturb-seq 测量了放疗反应。通用基准显示扰动模型仍难以超越简单基线，因此上述结果应视为临床前研究。 VC0132 VC0129 (preprint)（预印本） VC0070 VC0048
What are GBM digital twins?什么是 GBM 数字孪生？
GBM digital twins are computational models calibrated to an individual tumor or model system to simulate its growth and response to treatment. Examples include TumorTwin, an open Python framework demonstrated on high-grade glioma radiotherapy response, anatomy-aware simulation of patient-derived GBM xenografts, and a proteome-calibrated cell-level twin (preprint). None of these has been prospectively validated for clinical use.GBM 数字孪生是针对个体肿瘤或模型系统校准、用于模拟其生长与治疗反应的计算模型。例如：在高级别胶质瘤放疗反应上演示的开源 Python 框架 TumorTwin、患者来源 GBM 异种移植瘤的解剖感知模拟，以及蛋白组校准的细胞层面孪生（预印本）。这些均未经过面向临床的前瞻性验证。 VC0134 VC0133 VC0071 (preprint)（预印本）
Are virtual cell tools used in GBM patient care today?虚拟细胞工具目前是否用于 GBM 临床诊疗？
No, based on the evidence reviewed here. The items in this review are research tools, atlases, preclinical models or frameworks without prospective clinical validation. GlioState is not medical advice; patients should consult their own physicians.依据本文所综述的证据，没有。本文涉及的均为研究工具、图谱、临床前模型或尚无前瞻性临床验证的框架。GlioState 不构成医疗建议，患者请咨询主治医生。 VC0134 VC0059