NePSTA 队列
4 个德国中心 107 个样本，含空间转录组、H&E 与 EPIC 甲基化等诊断结果；8 项任务的主数据。
论文声明验证队列已在 Zenodo 与 Dryad 公开；训练队列未见公开存放。
数据入口 ↗论文报告 · 以数据为中心
空间转录组与图学习，怎样用于脑肿瘤诊断？
NePSTA · Nature Cancer，2025 · 8 项任务的数据、划分与结果
这项研究把 Visium 空间转录组与图神经网络结合，用于中枢神经系统肿瘤的常规诊断。共 8 项分析。复用前要分清：前 4 项检查数据质量并与常规检测对照，不训练模型；后 4 项训练模型，按患者划分数据。
数据
8 项任务总览
| 任务 | 数据与规模 | 标签或参照 | 划分 | 主要结果 |
|---|---|---|---|---|
| 空间转录组质控 比较不同处理方式、组织类型下的数据质量 | NePSTA 107 个样本 + Ren 等 11 个外部样本 | 无标签，按处理方式、组织类型分组比较 | 不划分 | 各组每 spot UMI 数无显著差异；spot 数主要取决于活检还是整块切除 |
| 虚拟免疫组化 由空间表达推断 Ki67、GFAP、NeuN 的蛋白丰度 | NePSTA 中 12 名有相邻切片 IHC 的参与者 | 相邻切片的 IHC 信号强度 | 不划分，与参照对比 | 相关系数：Ki67 0.47，GFAP 0.32，NeuN 0.57 |
| 拷贝数变异推断 由空间表达推断染色体增益和缺失 | NePSTA 中有甲基化芯片拷贝数结果的参与者 | 甲基化芯片的拷贝数判定 | 不划分，与参照对比 | 一致率 81.2%；三分类 AUC 80.37% |
| 细胞组成与微环境 估计每个 spot 的细胞类型，按甲基化亚类比较微环境 | NePSTA 全部样本 + GBMap（16 个数据集、110 名患者） | GBMap 细胞类型特征；甲基化亚类用于分组 | 不划分 | 间充质亚型免疫细胞丰富；RTK 亚型缺少 T 细胞和巨噬细胞浸润 |
| 组织区域分割 把切片分为主瘤、坏死、浸润区、白质和皮层 | NePSTA：队列 1 的 41 人、队列 2 的 27 人 | 神经病理专家的 H&E 区域标注 | 按患者：41 人训练（内部十折），27 人验证 | 验证集准确率 87.43% |
| 甲基化亚类分类 由空间数据预测肿瘤区域的 DNA 甲基化亚类 | NePSTA 中有 EPIC 甲基化结果的参与者；9.7 万个肿瘤子图 | EPIC 甲基化芯片的亚类 | 按患者五折交叉验证 | 患者级准确率 0.893；3 跳邻域子图准确率 0.999，无邻域 0.409 |
| MGMT 启动子甲基化 由空间数据预测 MGMT 启动子甲基化状态 | NePSTA 中有 MGMT 结果的 53 人 | EPIC 甲基化芯片的 MGMT 状态 | 一个队列 30 人训练，另一队列 23 人评估 | 准确率 99%；F1、精确率、召回率均为 1.0 |
| CDKN2A/B 缺失 预测 CDKN2A/B 纯合或杂合缺失，并比较缺失区域的细胞组成 | NePSTA 中有 CDKN2A/B 结果的 53 人；GBMap 用于邻域分析 | 甲基化芯片的 CDKN2A/B 状态 | 一个队列 30 人训练，另一队列 23 人评估 | 子图准确率 85.4%；缺失区域富集单核细胞与血管内皮细胞 |
逐项分析
01 空间转录组质控
- 做什么
- 检查 Visium 空间转录组的数据质量受哪些因素影响：石蜡或冰冻处理、组织类型、样本大小。临床使用要求质量稳定，这是后续任务的前提。
- 输入 → 输出
- 每个 spot 的表达计数、坐标、H&E 图像与样本处理信息 → 各组的 UMI 数、每 spot 细胞数和组织覆盖的 spot 数。
- 方法
- Cell Ranger 处理，SPATA 导入；每个 spot 归一化到 10,000 计数并取对数，回归批次和核糖体、线粒体比例。不训练模型。
- 数据与规模
- NePSTA 4 个德国中心 107 个样本（海德堡 50、弗莱堡 41、曼海姆 15、梅明根 1）：队列 1 为 3 个中心的 66 个，队列 2 为弗莱堡的 41 个。另用 Ren 等已发表的 11 个样本作外部对照。
- 结果
- 各队列、各样本类型的每 spot UMI 数无显著差异；UMI 数与细胞密度强相关，IDH 野生型胶质母细胞瘤最高；非肿瘤组织每 spot 细胞数明显更低；组织覆盖的 spot 数主要取决于活检还是整块切除。
- 复用时注意
- 论文总数 130 包含 11 个外部对照和 12 个健康皮层样本，NePSTA 本身是 107 个。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 11
“For data analysis and quality control, we used the Cell Ranger pipeline from 10X Genomics.”
“Standard import procedures include normalizing gene expression, which is achieved using the Seurat version 4.0 package.”
展开论文原文与位置
Source location: Paper | PDF p. 2
“Our study involved samples from four German medical centers (Heidelberg, n = 50; Freiburg, n = 41; Mannheim, n = 15; Memmingen, n = 1), along with external controls published recently9 (n = 11) and 12 healthy cortex samples, totaling 130 samples.”
“The first cohort (n = 66) comprised samples from three centers, encompassing a range of CNS pathologies from highly malignant glioblastomas (GBs) to epilepsy-associated (glio)neuronal tumors.”
“These samples underwent comprehensive spatial transcriptomics profiling, alongside state-of-the-art diagnostic workups, including morphology inspection, EPIC methylation arrays, classifier predictions, IHC and NGS.”
展开论文原文与位置
Source location: Paper | PDF p. 2
“In cohort 1, the majorit of samples (n = 42, 63.6%) were processed from paraffin-embedded tis sue, with only 24 samples (36.6%) processed from freshly frozen tissue Conversely, cohort 2 included approximately half of the samples (n = 18 43.9%) processed from paraffin-embedded tissue.”
“First, we aimed to explore and quantify potential confounders of the technology, as clinical applications demand robust and consistent quality readouts.”
展开论文原文与位置
Source location: Paper | PDF p. 2; Paper | PDF p. 24
“Our study involved samples from four German medical centers (Heidelberg, n = 50; Freiburg, n = 41; Mannheim, n = 15; Memmingen, n = 1), along with external controls published recently9 (n = 11) and 12 healthy cortex samples, totaling 130 samples.”
“The first cohort (n = 66) comprised samples from three centers, encompassing a range of CNS pathologies from highly malignant glioblastomas (GBs) to epilepsy-associated (glio)neuronal tumors.”
“The second cohort (n = 41) was a single-center cohort from Freiburg, similarly profiled (Fig. 1b).”
“In this study we report a total of 130 samples”
02 虚拟免疫组化
- 做什么
- 由空间表达推断诊断用蛋白标记的丰度，生成可以代替免疫组化染色的“虚拟染色”。
- 输入 → 输出
- 每个 spot 的表达谱，经贝叶斯超分辨率细化 → 每个 spot 的蛋白丰度图。
- 数据与规模
- 只有 12 名参与者同时有空间转录组和相邻切片的免疫组化，评估只能用这 12 人。
- 评估
- 免疫组化图像取蓝色通道转灰度、缩放到 0–1，逐 spot 提取信号强度，与推断丰度求相关。
- 结果
- Ki67 0.47，GFAP 0.32，NeuN 0.57，P 值均极小。
- 复用时注意
- 参照来自相邻切片而非同一切片，逐 spot 对应依赖两张切片的配准。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 2; Paper | PDF p. 12
“We start by converting the IHC image to grayscale by extracting the blue-channel (img[,,3]) and transform into data.frame format using the reshape2::melt() function. The pixel intensity values were extracted and rescaled to a range of 0–1.”
“When juxtaposed with consecutive participant sections, our virtual stainings exhibited robust alignment with IHC-derived results (Fig. 2b).”
展开论文原文与位置
Source location: Paper | PDF p. 2; Paper | PDF p. 5
“To this end, we devised an innovative computational module named ‘inferred IHC’. This module harnesses super-resolution spatial transcriptomics through the Bayesian inference11 to forecast protein abundance, rendering it a viable diagnostic surrogate for traditional IHC (Fig. 2a,b).”
“Paired IHC and spatia transcriptomics (for analysis of a–d) was available for 12 participants.”
展开论文原文与位置
Source location: Paper | PDF p. 2; Paper | PDF p. 5
“Paired IHC and spatia transcriptomics (for analysis of a–d) was available for 12 participants.”
“When juxtaposed with consecutive participant sections, our virtual stainings exhibited robust alignment with IHC-derived results (Fig. 2b).”
展开论文原文与位置
Source location: Paper | PDF p. 5
“Paired IHC and spatia transcriptomics (for analysis of a–d) was available for 12 participants.”
03 拷贝数变异推断
- 做什么
- 由空间表达推断染色体增益和缺失，与甲基化芯片的拷贝数结果比对，判断能否替代。
- 输入 → 输出
- 表达谱按 1 Mbp 染色体分箱、10 kbp 插值、loess 平滑 → 每个区域判为增益、缺失或无变化。
- 方法
- SPATA2 拷贝数估计与参照判定；三分类 AUC 按类别两两比较后取平均。
- 数据与规模
- NePSTA 中所有有甲基化芯片拷贝数结果的参与者，按染色体臂分层分析。
- 结果
- 一致 81.2%（都判无变化 70.3%、都判增益 5.5%、都判缺失 5.4%）；判定相反：增益 0.05%、缺失 0.02%。三分类 AUC 80.37%。
- 复用时注意
- 81.2% 中有 70.3% 是两种方法都判为无变化；一致检出增益或缺失的只占 10.9%。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 11
“Chromosomal bins were created using the SPATAwrapper::Create.ref.bins() function, with a bin size of 1 Mbp used for this study. Data were then rescaled and interpolated over a 10-kbp window, with normalization achieved using a loess regression model”
展开论文原文与位置
Source location: Paper | PDF p. 4; Paper | PDF p. 4
“Direct comparison of methylation-based CNV against our inferred CNV detection revealed a consensus of CNV alterations in 81.2% cases (consensus no alterations, 70.3%; consensus gains, 5.5%; consensus loss, 5.4%), only 0.05% divergent gains (gains detected by Visium and loss detected in the methylation-based CNV analysis) and 0.02% divergent losses (Fig. 3d).”
“We performed multiclass receiver operating characteristic (ROC) analysis to validate the accuracy of detecting gains and losses or diploid chromosome sets, demonstrating an overall area under the curve (AUC) of 80.37% (Fig. 3f).”
展开论文原文与位置
Source location: Paper | PDF p. 4
“To more precisely examine the potential variability in accuracy of CNV detection across different chromosomal regions”
展开论文原文与位置
Source location: Paper | PDF p. 4
“Direct comparison of methylation-based CNV against our inferred CNV detection revealed a consensus of CNV alterations in 81.2% cases (consensus no alterations, 70.3%; consensus gains, 5.5%; consensus loss, 5.4%), only 0.05% divergent gains (gains detected by Visium and loss detected in the methylation-based CNV analysis) and 0.02% divergent losses (Fig. 3d).”
04 细胞组成与微环境
- 做什么
- 用单细胞参考图谱估计每个 spot 的细胞类型组成，比较不同甲基化亚类的肿瘤微环境。
- 输入 → 输出
- spot 表达谱与 GBMap 细胞类型特征 → 每个 spot 的细胞类型丰度，按亚类汇总。
- 方法
- Cell2location 反卷积（500 次迭代，后验采样 1,000 次）；另用空间互相关指数衡量基因对的空间关联。
- 数据与规模
- NePSTA 所有有甲基化分类的样本。GBMap 整合 16 个数据集、110 名患者，在这里是细胞类型字典，不是测试集。
- 结果
- 间充质亚型免疫细胞丰富，髓系细胞呈免疫抑制特征；受体酪氨酸激酶（RTK）亚型明显缺少 T 细胞和肿瘤相关巨噬细胞浸润。
- 复用时注意
- 论文以图和文字描述比较结果，没有给出可直接复用的数值表。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 12
“We trained the model on a graphics processing unit for 500 iterations to ensure computational efficiency. After training, we used the export\_posterior function to extract the posterior distribution of cell type proportions”
展开论文原文与位置
Source location: Paper | PDF p. 4
“Leveraging the robust and well-validated Cell2lo cation algorithm, combined with an extensive GB reference dataset (GBMap), we predicted the abundance of myeloid, T cell and stromal subpopulations across all methylation classes (Fig. 4a).”
展开论文原文与位置
Source location: Paper | PDF p. 4
“The meth ylation Mes subtype exhibited immune-rich microenvironments with an immunosuppressive myeloid profile, while others (RTKI and RTKII) displayed a notable absence of T cell and tumor-associated macrophage infiltration (Extended Data Fig. 3).”
展开论文原文与位置
Source location: Paper | PDF p. 4
“Leveraging the robust and well-validated Cell2lo cation algorithm, combined with an extensive GB reference dataset (GBMap), we predicted the abundance of myeloid, T cell and stromal subpopulations across all methylation classes (Fig. 4a).”
05 组织区域分割
- 做什么
- 把切片自动分成主瘤、坏死、浸润区、白质和皮层（非肿瘤切片另有健康皮层和蛛网膜），供后续分子预测使用。
- 输入 → 输出
- 以每个 spot 为中心的 3 跳子图 → 该子图的组织类别。
- 模型
- 3 层图同构网络（GIN），全局平均池化；交叉熵损失，Adam 优化。
- 数据与划分
- 只用组织标注足够的切片：队列 1 的 41 人训练，训练集内十折交叉验证；队列 2 的 27 人验证。
- 结果
- 以神经病理专家的分割为参照，验证集准确率 87.43%。
- 复用时注意
- 报告记录的节点特征含组织标注编码，而分割目标也是组织类别；复用前需对照论文代码，确认训练时是否使用。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 13
“Expression profiles (top 5,000 genes), CNAs, histological annotations (one-hot encoded) and encoded H&E image vectors were included”
“Edge features were defined by spatial proximity with up to six neighbors per node”
展开论文原文与位置
Source location: Paper | PDF p. 6; Paper | PDF p. 11
“A pivotal step before subgroup prediction involved automated spatial transcriptomics data segmentation, aiming to classify subgraphs histologically.”
“Manual segmentation of the histological regions was performed by a neuropathologist according to the morphological features of the H&E scan.”
展开论文原文与位置
Source location: Paper | PDF p. 6
“We used a tenfold cross-validation to train on n = 41”
“samples with sufficient histological segmentation)”
展开论文原文与位置
Source location: Paper | PDF p. 6
“We used a tenfold cross-validation to train on n = 41”
“Model validation on n = 27 participants”
06 甲基化亚类分类
- 做什么
- 由空间数据预测肿瘤区域的 DNA 甲基化亚类，并检验加入周围 spot 的信息能提升多少。
- 输入 → 输出
- 以每个 spot 为中心的 3 跳子图 → 子图的亚类，再汇总为患者级亚类。
- 模型
- NePSTA 图神经网络（3 层 GIN）；对照为无邻域的线性模型，以及 1–3 跳的 GIN 和图注意力网络。
- 数据与划分
- 训练用 97,000 个肿瘤子图和 12,000 个健康对照子图；按患者五折交叉验证（胶质母细胞瘤与 IDH 突变胶质瘤），邻域比较在 84 名患者上评估。
- 结果
- 子图准确率：线性模型 0.409；GIN 1 跳 0.914、2 跳 0.981、3 跳 0.999；图注意力网络 1 跳 0.874、2 跳 0.983、3 跳 0.995。患者级：未见过的患者准确率 0.893、F1 0.870；五折交叉验证 0.913、F1 0.896。
- 复用时注意
- 0.999 是子图级、0.893 是患者级，单位不同，不能互换引用。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 13; Paper | PDF p. 13
“Expression profiles (top 5,000 genes), CNAs, histological annotations (one-hot encoded) and encoded H&E image vectors were included”
“Edge features were defined by spatial proximity with up to six neighbors per node”
“A three-hop neighborhood for a query spot included all spots within three edges, capturing the spatial context and connectivity within the defined distance.”
展开论文原文与位置
Source location: Paper | PDF p. 7; Paper | PDF p. 13
“From 107 EPIC-characterized participants, datasets were divided into training and validation subsets. For multibiopsy samples, datasets were split by biopsy cores; for single-specimen samples, manual segmentation was performed using SPATA2 to create regions.”
“We divided the samples into two distinct spatial transcriptomic datasets. For samples with multiple biopsies, we treated the raw data from each biopsy as separate datasets. In cases where the entire tissue was on a single slide, we segmented the samples as illustrated in Fig. 5a.”
展开论文原文与位置
Source location: Paper | PDF p. 7; Paper | PDF p. 13
“From 107 EPIC-characterized participants, datasets were divided into training and validation subsets. For multibiopsy samples, datasets were split by biopsy cores; for single-specimen samples, manual segmentation was performed using SPATA2 to create regions.”
“To avoid bias from similar normalization and preprocessing in both the training and the validation cohorts, we processed each dataset individually after splitting the data at the count level.”
展开论文原文与位置
Source location: Paper | PDF p. 13 — plus Web | publisher supplementary PDF, Supplementary Table 1
“From 107 EPIC-characterized participants, datasets were divided into training and validation subsets. For multibiopsy samples, datasets were split by biopsy cores; for single-specimen samples, manual segmentation was performed using SPATA2 to create regions.”
“Training data comprised 97,000 subgraphs from tumor datasets and 12,000 subgraphs from healthy controls, created using a three-hop neighborhood approach.”
“Supplementary Table 1: Evaluation of different models to explore the impact of extended neighborhoods for the predictive power of NEPSTA GIN: Graph isomorphic network, GAN: Graph attention network. The model evaluation was performed on 84 patients.”
07 MGMT 启动子甲基化
- 做什么
- 由空间数据预测 MGMT 启动子甲基化状态；这一信息无法从推断的拷贝数中读出。
- 模型
- NePSTA 图神经网络，加一个 MGMT 预测头。
- 数据与划分
- 两个队列中有 MGMT 结果的 53 人：一个队列 30 人训练，另一个队列 23 人评估。
- 结果
- 准确率 99%，F1、精确率、召回率均为 1.0。
- 复用时注意
- 评估集 23 人时，患者级准确率只能是 95.7%、100% 这样的值；99% 应是子图级结果，引用前需核对。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 13
“Expression profiles (top 5,000 genes), CNAs, histological annotations (one-hot encoded) and encoded H&E image vectors were included”
“Edge features were defined by spatial proximity with up to six neighbors per node”
展开论文原文与位置
Source location: Paper | PDF p. 7; Paper | PDF p. 21
“Extended Data Fig. 6 | Heterogeneity of CDKN2A/B alterations. (a) Schematic of the GNN workflow Since the MGMT and CDKN2A/B status is available in both cohorts (n = 53 patients) we train (n = 30 patients) and evaluate (n = 23 patients) on separate cohorts.”
“To harness this insight, we augmented our GNN framework, incorporating multiple MLPs specifically designed to predict MGMT promoter methylation and the loss of CDKN2A/B”
展开论文原文与位置
Source location: Paper | PDF p. 21
“Extended Data Fig. 6 | Heterogeneity of CDKN2A/B alterations. (a) Schematic of the GNN workflow Since the MGMT and CDKN2A/B status is available in both cohorts (n = 53 patients) we train (n = 30 patients) and evaluate (n = 23 patients) on separate cohorts.”
展开论文原文与位置
Source location: Paper | PDF p. 21
“Extended Data Fig. 6 | Heterogeneity of CDKN2A/B alterations. (a) Schematic of the GNN workflow Since the MGMT and CDKN2A/B status is available in both cohorts (n = 53 patients) we train (n = 30 patients) and evaluate (n = 23 patients) on separate cohorts.”
08 CDKN2A/B 缺失
- 做什么
- 预测 CDKN2A/B 缺失（区分纯合与杂合），并比较缺失区域周围的细胞组成。
- 模型
- NePSTA 图神经网络，加一个 CDKN2A/B 预测头；邻域细胞组成用 GBMap 反卷积。
- 数据与划分
- 有 CDKN2A/B 结果的 53 人：一个队列 30 人训练，另一个队列 23 人评估。区域分析使用 IDH 突变肿瘤的立体定向活检。
- 结果
- 区分未缺失与缺失子图的准确率 85.4%。缺失与未缺失区域的肿瘤细胞比例无显著差异，但缺失区域富集单核细胞和血管内皮细胞。
- 复用时注意
- 85.4% 是子图级准确率，不是患者级。
查看本任务的来源依据
展开论文原文与位置
Source location: Paper | PDF p. 13
“Expression profiles (top 5,000 genes), CNAs, histological annotations (one-hot encoded) and encoded H&E image vectors were included”
“Edge features were defined by spatial proximity with up to six neighbors per node”
展开论文原文与位置
Source location: Paper | PDF p. 8; Paper | PDF p. 21
“and further differentiation between homozygous and heterozygous deletion were found to be impossible in cases when the loss was not associated with chr9p loss (Fig. 6a,b).”
“Extended Data Fig. 6 | Heterogeneity of CDKN2A/B alterations. (a) Schematic of the GNN workflow Since the MGMT and CDKN2A/B status is available in both cohorts (n = 53 patients) we train (n = 30 patients) and evaluate (n = 23 patients) on separate cohorts.”
展开论文原文与位置
Source location: Paper | PDF p. 10; Paper | PDF p. 21
“Contrary to our expectations, our findings revealed that the majority of samples with heterozygous deletions (two of three) displayed a mixed genotype, encompassing both partial and complete loss of CDKN2A/B across different biopsy specimens.”
“Extended Data Fig. 6 | Heterogeneity of CDKN2A/B alterations. (a) Schematic of the GNN workflow Since the MGMT and CDKN2A/B status is available in both cohorts (n = 53 patients) we train (n = 30 patients) and evaluate (n = 23 patients) on separate cohorts.”
展开论文原文与位置
Source location: Paper | PDF p. 10
“Contrary to our expectations, our findings revealed that the majority of samples with heterozygous deletions (two of three) displayed a mixed genotype, encompassing both partial and complete loss of CDKN2A/B across different biopsy specimens.”
开始复用前，确认这三件事
- 获取范围：验证队列的公开声明不代表训练数据全部公开，先核对目标任务需要的文件。
- 对应关系：检查表达、图像、分子标签与患者标识能否对应，按任务确认有效交集。
- 实验设计：分别记录每项任务的数据划分、参照标准和评价层级，不把论文整体规模当作每项任务的可用规模。
论文与来源
Ritter 等，Nature Cancer，2025。
Spatially resolved transcriptomics and graph-based deep learning improve accuracy of routine CNS tumor diagnostics