单核转录组揭示脑出血后神经炎症与髓鞘修复的时序动态
下划线 = 中文容易“变味”或丢信息的词句：红线带编号 = 20 条注释卡（悬停看、点击跳转）。译文按原文顺序全文翻译；英文原文附于文末
作者：Zhan Chen、Qinglin Wang、Rong Xiang、Ruoqi Ding、Jin Tao、Qinfeng Peng、Shaoshuai Wang、Nannan Cheng、Mengke Zhao、Jiaxin Li、Qidi Xue、Chuanyu Liu、Xuemei Chen、Longqi Liu、Junmin Wang、Jian Wang（王建，通讯作者）、Mingyue Wang（通讯作者）
单位（原文署明）：郑州大学基础医学院／郑州大学附属医院神经外科相关团队；华大生命科学研究院（BGI Research，深圳）；卡罗林斯卡学院（Karolinska Institutet，瑞典）等
出处：Clinical and Translational Medicine（临床与转化医学）2025;15(10):e70486
DOI：10.1002/ctm2.70486 PMID：41017072 PMCID：PMC12477064
译自英文原题：Single-nucleus transcriptomic profiling reveals temporal dynamics of neuroinflammation and myelin repair after intracerebral haemorrhage
本页为学习用途的非官方全文中译本，开放获取（CC BY）原文见文末附录。
摘要
背景 脑出血（intracerebral haemorrhage, ICH）进展迅速、病理过程复杂、治疗手段有限，是卒中最严重的亚型之一。血液外溢进入脑实质会触发一连串炎症反应，构成继发性损伤。单核 RNA 测序（snRNA-seq）数据使我们对出血脑内细胞异质性与动态互作有了更深入的认识。免疫细胞在塑造神经炎症中起关键作用；然而，缺乏全面的纵向研究限制了我们对抗炎／促炎过程随时间演变的理解，也给靶向治疗策略的开发带来困难。
方法 我们在胶原酶诱导的 ICH 小鼠模型中，于损伤后第 1、3、7、14、28 天，并设naive 对照，用 snRNA-seq 描绘基因表达随时间的变化。
结果 我们获得了 281 577 个高质量转录谱，代表 21 种不同的细胞类型。共表达网络分析揭示出一个显著的“炎症模块”，在 ICH 全程保持活跃。单细胞转录组与免疫荧光染色的整合分析提示，多种表达 Mif 的细胞可能参与局部炎症，并可能通过 Cd44、Cd74 等受体–配体对与巨噬细胞发生作用。随时间推移，小胶质细胞似乎越来越成为促炎信号的主要接收者。在消退期，少突胶质细胞呈现与成熟度提升和髓鞘再生一致的转录特征，而这一过程可能由 T 细胞介导的相互作用所促进。
结论 这些发现从系统层面呈现了 ICH 进展与消退过程中细胞类型特异的反应以及免疫介导的相互作用。
要点（原文 Highlights）
- 在多个时间点（第 1、3、7、14、28 天）建立 ICH 小鼠模型，构建高质量的单核 RNA 测序图谱。
- 计算分析提示，ICH 早期的巨噬细胞募集可能涉及迁移抑制因子（MIF）信号通路。
- 在消退期，T 细胞可能与形成髓鞘的少突胶质细胞相互作用，从而可能促进 ICH 后的髓鞘再生。
引言
ICH 是卒中的严重亚型，2019 年占全球全部卒中的 27.9%。ICH 治疗策略的进展有限，过去几十年患者预后仅有小幅改善。ICH 的病理生理涉及扩张中的血肿造成的直接压迫，以及由血肿和血液来源代谢产物触发的继发性生理与细胞反应。这些挑战凸显了持续研究、开发创新治疗策略的紧迫性。ICH 时血液漏入脑内，诱发一系列有害过程，包括脑水肿、凝血酶毒性和神经炎症。
由小胶质细胞、星形胶质细胞、巨噬细胞（MAC）和 T 细胞等介导的神经炎症，是继发性脑损伤的关键驱动因素，并与不良临床结局相关。它加重脑损伤、阻碍组织修复，并促发卒中后的运动与认知障碍。小胶质细胞与 MAC 是炎症的核心介导者，传统上依据实验模型被分为 M1（促炎）与 M2（抗炎）两个亚型；但越来越多的证据表明，这些细胞呈现的是功能谱系，而非离散状态。Zhang 等人在 ICH 后的血肿周围水肿（PHE）组织中鉴定出 12 种小胶质细胞亚型，揭示出包括 DNA 修复、退变、疾病相关和过渡性炎症状态在内的多种表型。这些发现凸显了 ICH 微环境中小胶质细胞表型的复杂性。
除神经炎症外，脱髓鞘也是 ICH 的关键病理特征。髓鞘破坏会扰乱神经信号传导，导致功能障碍。脱髓鞘过程可被少突胶质前体细胞（OPC）逆转，后者对恢复髓鞘完整性与改善恢复至关重要：OPC 增殖并分化为成熟少突胶质细胞（MOL），以修复受损髓鞘。新出现的证据还提示，调节性 T 细胞（Treg）通过调节免疫反应促进髓鞘再生与神经修复。
snRNA-seq 使人们能够更细致地理解细胞对 ICH 的反应。该技术提供脑细胞的高分辨率转录谱，便于鉴定不同的细胞群体及其分子通路。这类见解对于解开免疫细胞相互作用的复杂性及其对炎症的贡献至关重要。
本研究采用 snRNA-seq 研究胶原酶诱导的 ICH 小鼠模型中神经炎症与髓鞘再生的细胞与分子动力学。我们鉴定了不同的细胞簇及其分子特征，揭示了参与炎症与髓鞘再生的关键细胞互作与调控机制。我们的发现提出了促进 ICH 后功能恢复的具体治疗靶点，为开发治疗策略提供了基础。
结果 1 小鼠血肿周围组织的转录组全景
为考察 ICH 后基因表达的时间动态，我们采用了胶原酶诱导的 ICH 模型——这一成熟方法通过降解血管基底膜引发 ICH。我们在 ICH 后的关键时间点（第 1、3、7、14、28 天）监测基因表达，并取病灶周围脑组织消化后进行 snRNA-seq 分析（图 1、表 1）。细胞依据线粒体基因比例、基因数和双细胞（doublet）评分进行过滤，以剔除低质量细胞（图 S1）。共鉴定出 281 577 个细胞，平均每个细胞检出 2010 个基因。
图 1 ICH 模型病灶周围脑组织与 naive 对照的单细胞全景。（A）使用 snRNA-seq 对 ICH 模型进行转录组图谱构建的流程示意。（B）点图显示 21 个显著细胞群体标志基因的平均表达水平；颜色代表平均标化基因表达量（z 分数），点大小代表该群体中检出该标志基因的细胞百分比。（C）ICH 小鼠病灶脑组织 snRNA-seq 数据的聚类 UMAP 可视化；外圈半径代表各细胞类型细胞数的 log10 变换值。共鉴定 21 个细胞群体：DGGRC（齿状回颗粒神经元）、MEGLU（间脑与中脑兴奋性神经元）、DEINH（间脑与中脑抑制性神经元）、GNBL（谷氨酸能神经母细胞）、NGNBL（非谷氨酸能神经母细胞）、PEP（肽能神经元）、TEINH（端脑抑制性中间神经元）、TEGLU（端脑谷氨酸能神经元）、MSN（端脑投射抑制性神经元）、MAC（巨噬细胞）、T（T 细胞）、AC（星形胶质细胞）、MGL（小胶质细胞）、OPC（少突胶质前体细胞）、OLG（少突胶质细胞）、EPEN（室管膜细胞）、CHOR（脉络丛上皮细胞）、EN（内皮细胞）、PER（周细胞）、FB（成纤维细胞）、VLM（血管软脑膜细胞）。（D）折线图（左）显示主要细胞类型在六个时间点（naive、第 1、3、7、14、28 天）的相对比例变化；柱状图（右）为方差分析（ANOVA）的 F 统计量，按细胞类型着色，并归为神经元、免疫细胞、胶质细胞、血管细胞四大类。（E）散点图显示 TEGLU_0 的差异表达基因，前 100 位以浅绿高亮。（其余分图说明见文末英文原图注）
我们首先用 scVI 消除不同实验条件来源的批次效应（见“材料与方法”），再用 scVI 生成的降维矩阵进行 UMAP 与 Leiden 聚类。我们人工鉴定出 67 个细胞簇，并依据文献中的标志基因将其注释为 21 个主要细胞类型（图 S2、S3，表 S1）。这些细胞类型包括神经元（9 个）、免疫（2 个）、胶质（4 个）与血管（6 个）细胞簇（图 S4、S5）。各细胞类型在各时间点与各取样部位均有出现，但存在明显的时间偏好：抑制性神经元在纹状体中更丰富；外周血免疫细胞在 naive 组相对稀少，这很可能是因为血脑屏障（BBB）完整、限制了外周血免疫细胞浸润。我们还观察到端脑谷氨酸能神经元（TEGLU）簇内部显著的多样性（图 S6），这与大脑皮层复杂的构筑与生物学功能相符。
为考察 ICH 后病灶内细胞类型构成的变化，我们用 propeller 分析其比例（图 1D、图 S7）。神经元细胞比例，尤其是 TEGLU 与 TEINH，在 ICH 后下降，这与周围神经元的死亡有关。TEGLU_0 的线粒体基因百分比升高（图 S8、S9）；既往研究提示，线粒体过度分裂会加重神经元焦亡与神经炎症。我们用成熟的基因集评估程序性细胞死亡过程，发现 TEGLU_0 在铁死亡、坏死性凋亡与焦亡上的评分均高于其他亚型（图 S10）。此外，BBB 破坏后我们观察到大量外周免疫细胞浸润脑实质，且各具不同的时间偏好：MAC 在早期更突出，T 细胞在后期更多见。这一观察提示，外周免疫细胞在 ICH 进展中承担着不同的生物学角色。
结果 2 炎症共表达模块持续激活的细胞来源
为解释哪些细胞类型参与了神经炎症、并考察 ICH 后的分子动态，我们用 hdWGCNA 进行共识共表达网络分析。首先依据无标度拓扑拟合曲线开始进入平台区的拐点选择软阈值功率（图 S11），以既良好近似无标度网络、又避免功率过高而降低网络连通性。我们确定的最优阈值为：naive（5）、第 1 天（6）、第 3 天（6）、第 7 天（5）、第 14 天（5）、第 28 天（7）。随后我们为各时间点的基因表达矩阵分别构建网络，再整合网络并鉴定基因模块，在六个网络中共鉴定出 10 个不同的模块（图 2A）。每个模块由高共表达的基因构成，很可能反映了协调的调控机制与共同的生物学过程。
整合 hdWGCNA 与 Harmony 后，我们得到协调化模块特征基因（hME），作为概括各共表达模块基因表达谱的度量。为定位各模块中的枢纽基因，我们计算每个基因基于特征基因的连接度（kEM）并鉴定枢纽基因——它们是生物学功能的关键指示物（图 S12，表 S2）。随后我们以枢纽基因为特征对共表达矩阵作可视化，标注出排名前三的基因（图 2A）。模块 6 富集促炎基因，模块 2 则呈现髓鞘形成相关基因（图 2D）。边（edge）可视化显示各枢纽基因之间共表达关系强烈，凸显了其功能重要性。
为考察模块之间可能的相互作用，我们用 hME 计算模块间的 Pearson 相关系数。模块 2 与模块 9 显著强相关，且两者的表达区域与少突胶质谱系高度吻合（图 2B）。有趣的是，模块 9 也表达于星形胶质细胞与免疫细胞，提示这些细胞可能支持神经元的髓鞘包被。
我们进行差异模块特征基因（DME）分析，通过在各时间点比较 ICH 组与 naive 对照，以鉴定呈时间特异性变化的模块（图 2E）。模块 4 主要富集于星形胶质细胞（AC）与成纤维细胞（FB），在第 7 天之后显著升高（图 2F、图 S13）。差异表达分析显示，模块 4 前 50 位枢纽基因中有若干在第 14 天的 AC 中较 naive 组上调，包括与星形胶质细胞成熟相关的基因（Ndrg2、Gpr37l1）以及细胞外基质重塑相关基因（Timp3、Sdc4、Clu；图 S14）。此外，空间转录组分析显示星形胶质细胞与成纤维细胞在病灶边界处处于相邻区域（图 S15）。这些发现提示，星形胶质细胞与成纤维细胞协同介导胶质瘢痕形成，从而填补出血后留下的脑内空腔。更重要的是，模块 6 在 ICH 后持续上调（相较 naive 组），在第 3 天和第 7 天表达升高（图 2E）。免疫相关细胞——包括 MAC、小胶质细胞（MGL）与 T 细胞——是该模块的主要贡献者（图 2C）。该模块富集于免疫效应活性、白细胞活化与抗原加工呈递相关通路（图 2D）。模块 6 的持续激活说明炎症在 ICH 后发挥关键作用。
结果 3 神经炎症亚型的免疫图景动态与功能转向
神经炎症是 ICH 后继发性损伤的关键驱动因素，但认识仍不充分。共表达网络分析显示，免疫通路在 ICH 全程持续上调与激活。小胶质细胞、星形胶质细胞、MAC 与 T 细胞是神经炎症反应的主要贡献者（图 2F）。这些细胞的炎症评分升高，在第 7 天达到峰值，其中 MAC 与小胶质细胞是主要来源（图 3D）。我们进行亚聚类分析，鉴定出 4 个星形胶质细胞亚型、3 个小胶质细胞亚型和 3 个 MAC 亚型（图 3A、图 S16）。我们采用 milo——一种基于K 近邻（KNN）图的差异丰度（DA）分析方法——量化 ICH 进展中的动态变化（图 3B）。免疫细胞中，MAC 主要出现在早期，而 T 细胞在后期富集。胶质细胞中，AC_4 与 MGL_2 随时间呈现动态变化（图 3B）。BBB 破坏后免疫细胞浸润并诱导炎症，而与 ICH 相关的 AC 与 MGL 亚型则反映了胶质细胞对出血的应答。这些显著改变的亚型可能对 ICH 病理产生重要影响。
图 3 ICH 模型内神经炎症亚型的细胞异质性。（A）星形胶质细胞（AC）、小胶质细胞（MGL）、巨噬细胞（MAC）与 T 细胞在 snRNA-seq 数据中的亚型 UMAP 可视化。（B）蜂群图显示 ICH 后各时间点神经炎症亚型的差异丰度，显著变化以红、蓝高亮。（C）点图显示星形胶质细胞的差异表达基因，按亚型着色。（D）箱线图显示各星形胶质细胞亚型的神经炎症评分。（E）火山图展示以小胶质细胞亚型 MGL_1 为对照的差异上调基因，按亚型着色。（F）柱状图（下）显示六个时间点各亚型相对比例变化；柱状图（上）为 ANOVA 的 F 统计量，按亚型着色。（G）点图显示巨噬细胞的差异表达基因，按亚型着色。（H）柱状图显示 MAC 中上调差异基因的 GO 富集。（I）代表性免疫荧光图像显示 ICH 后第 7 天与第 14 天病灶核心内 CD3⁺ T 细胞（红），DAPI（蓝）标记细胞核。（J）箱线图显示 ICH 后第 7 天与第 14 天的 T 细胞计数，p 值由 t 检验得出。（K）折线图显示经 TCellSI 评估的 T 细胞功能评分（左：细胞毒性；右：调节性）在 ICH 后随时间的变化。
为进一步刻画不同细胞亚型的分子特征，我们用伪批（pseudo-bulk）方法分析差异表达基因（DEG）及其相关功能通路（表 S3）。AC_2 与 AC_3 的基因表达谱提示区域特化的星形胶质细胞状态：AC_2 高表达的 Thbs4 与白质星形胶质细胞相关，AC_3 表达的 Agt 与 Slc6a11 则与丘脑星形胶质细胞的标志物相关（图 3C）。有意思的是，AC_4 的丰度在第 7 天达到峰值，与炎症反应的高峰相吻合（图 3B、3D）。该簇高表达 Serpina3n——已知该基因可通过激活 NF-κB 信号通路促进神经炎症，这与既往描述的反应性星形胶质细胞谱相符。MGL_1 为稳态小胶质细胞，以 P2ry12 与 Tmem119 高表达为标志，与维持小胶质细胞状态相关（图 3E）。MGL_3 是对 ICH 反应最早的小胶质细胞亚型，激活细胞周期通路以促进增殖（图 3E、3F）。MGL_2 主要富集于白细胞迁移与伤口愈合通路（图 3H）。
ICH 后外周免疫细胞的浸润是对脑内微环境变化的应答。MAC_3 主要见于第 3 天，清除血肿中的铁离子以维持铁稳态（图 3G）。第 7 天时 MAC_1 保持较高比例，通过激活补体通路与免疫反应促进细胞碎片清除（图 3G）。MAC_2 高表达 Cd74 并富集于免疫反应相关通路；此外，MAC_2 表达 Ms4a4c——一种新的 M2 表面标志物。MAC_2 还富集于“T 细胞活化的正向调控”通路，进一步支持其潜在的保护功能（图 3H）。这些结果提示，MAC 在 ICH 进展中经历了从清铁到清碎片、再到抗炎的功能转变，说明分阶段调控可优化其治疗潜力。
T 细胞数量不足以进一步亚聚类；但它们主要出现在修复期，提示其可能参与 ICH 后的恢复（图 3F）。为提供空间信息，我们在第 7 天与第 14 天的脑切片上对 CD3 进行免疫荧光染色。两个时间点均可在血肿周围区域见到 CD3⁺ T 细胞（图 3I），与单细胞数据提示的亚急性期存在在性质上一致。为进一步探索其潜在功能，我们采用 TCellSI 评分，发现 T 细胞状态呈动态转变：细胞毒性评分在第 7 天达峰，而调节性评分逐渐升高并持续至第 14 天，提示恢复期可能发生功能转型（图 3K）。
结果 4 亚型间细胞通讯由促炎向抗炎信号的动态转折
为更好理解不同细胞亚型如何参与神经炎症，我们考察其在细胞通讯网络中的相互作用，重点看促炎与抗炎信号通路。值得注意的是，相互作用数量先随天数增加、至第 7 天达到峰值，其趋势与炎症模块和炎症评分相似（图 4A、图 S17）。
图 4 ICH 模型内神经炎症亚型间的细胞–细胞互作。（A）ICH 神经炎症亚型中预测的配体–受体（L–R）互作数量，按亚型着色。（B）ICH 模型中各互作组的互作概率差之和（相对信息流），分为促炎（上）与抗炎（下）。（C）层级图显示第 1 天（左）、第 3 天（中）、第 7 天（右）的迁移抑制因子（MIF）信号网络；节点为细胞类型，边为互作，宽度与互作概率成正比，按亚型着色。（D）点图显示第 1 天与第 3 天 MIF 信号通路 L–R 的平均表达水平；颜色为平均标化表达量（z 分数），点大小为该群体中检出该基因的细胞百分比。（E）代表性免疫荧光染色显示 MIF 信号：MIF 为配体，CD44 与 CD74 为受体。（F）圆形图显示第 14 天（左）与第 28 天（右）的 MIF 信号网络。（G）点图显示第 14 天与第 28 天 CD226 信号通路配体–受体的平均表达水平。
所鉴定的促炎通路呈现不同的时间变化模式（图 4B）。某些通路在第 1 天活性显著升高，其中 CCL 与迁移抑制因子（MIF）通路对解释早期炎症反应尤具意义。Ccl2、Ccl6、Ccl7 与 Ccl12 等趋化因子（图 S18）由脑实质内多种细胞分泌，促成免疫细胞向损伤部位募集。此外，表达 Cd44 与 Cd74 的 MAC 与多种表达 Mif 的细胞互作（图 4D），提示 MIF 刺激的 CD74–CD44 受体复合物可能通过激活 ERK 通路促进细胞存活与增殖。空间转录组显示，ICH 后第 1 天 MAC 在血肿周围明显聚集，而 Mif 由多种细胞广泛表达，与我们的数据一致（图 S19、S20）。免疫荧光进一步显示 Mif 信号与 Cd74/CD44 信号在空间上邻近（图 4E），支持二者可能发生互作。这些互补的观察表明，表达 Cd74 的 MAC 可能对周围细胞来源的 Mif 作出应答，从而参与急性期的局部炎症反应。
在 ICH 后第 7 与第 14 天，我们观察到免疫细胞活化增强，表现为 CD45、MHC-I 与补体介导的吞噬水平升高（图 4B）。CD45 是活化白细胞的标志物；MHC-I 向 CD8⁺ T 细胞呈递抗原，促进免疫反应；补体介导的吞噬使 MAC 得以清除碎片与病原体。在较晚阶段，ITGAL–ITGB2、MHC II 类与 LCK 等通路更为突出。ITGAL–ITGB2 通路对白细胞黏附与迁移至关重要，有助于免疫细胞募集至损伤部位；MHC II 类分子向 CD4⁺ T 细胞呈递胞外抗原，增强适应性免疫应答。我们比较了三个关键时间点（naive 组、ICH 后第 1 天与第 7 天）细胞对炎症通路的信号强度（图 S21）。结果显示主导性的受体信号发生时间转移：ICH 后 24 小时以 MAC_2 为主要信号接收细胞，至第 7 天则完全转为 MGL_2 主导。这一转变提示，ICH 后期的炎症管理需要更多关注小胶质细胞的激活。
整体上存在由急性期以促炎为主向后期更偏抗炎状态的转向（图 4B）。VISTA 通路——一条抑制性免疫检查点通路——在 ICH 后下调。小胶质细胞表达的配体 Vsir 与星形胶质细胞分泌的蛋白 Igsf11 之间的互作，凸显了星形胶质细胞作为炎症反应调节中介的潜在作用（图 4B）。我们观察到第 7 天来自 AC_2 的 TGFβ 通路（Tgfb2 −（Tgfbr1 + Tgfbr2））上调，以及来自 T 细胞的 GDF 通路（Gdf15 − Tgfbr2）上调（图 S22、S23）。两条通路均通过 Tgfbr2 传递信号，而据报道 Tgfbr2 可抑制小胶质细胞激活。此外，GDF 已被证明可调控甘油三酯代谢，并有助于提高对炎症损伤的耐受。值得注意的是，T 细胞表面高表达的 Cd226 与炎症细胞上的 Pvr 直接互作，可防止局部炎症过度（图 4G）。这一互作是维持局部免疫稳态、通过在免疫检查点处平衡正负信号来保护组织的重要调控机制。我们在炎症环境中鉴定出抗炎通路，这可能为调控 ICH 后炎症提供策略。随着对这一复杂细胞间互动认识的深入，我们对促炎与抗炎信号之间精细平衡的理解也更加清晰。
结果 5 少突胶质谱系动态与 T 细胞–少突胶质细胞互作的可能
随着 ICH 后期神经炎症消退，免疫微环境的转变可能为髓鞘修复创造有利条件。在共表达网络分析中，免疫细胞对模块 9（髓鞘维持）有显著贡献（图 2C）。
为考察少突胶质谱系细胞的变化及其与神经炎症的关系，我们对 OPC 与 OLG 进行了亚聚类，依据经典标志物鉴定出六个不同的亚簇（图 5A）。我们进一步将疾病相关少突胶质细胞（DOL）与 MOL 作比较（图 5B，表 S4）。DOL 在 ICH 后增多，表达经典损伤相关标志物如 Serpina3n，并上调激肽释放酶相关肽酶（Klk8、Klk9）与成纤维细胞生长因子受体（Fgfr4）等基因。此外，表达增殖标志的少突胶质前体细胞（POPC）簇高表达 Serpina3n，在第 1 天与第 3 天增多，细胞周期评分高（图 5C、5E）。新形成的少突胶质细胞（NFOL）簇在第 7 天增多（图 5C、5F）。这些结果表明，OPC 在应激下加速细胞周期、促进增殖以支持髓鞘修复。
图 5 T 细胞促进少突胶质谱系的髓鞘再生。（A）OPC 与 OLG 的 snRNA-seq 转录组数据中少突胶质谱系的 UMAP 可视化。DOL，疾病相关少突胶质细胞；MFOL，形成髓鞘的少突胶质细胞；MOLL，成熟少突胶质细胞；NFOL，新形成的少突胶质细胞；OPCs，少突胶质前体细胞；POPCs，增殖性少突胶质前体细胞。（B）堆叠小提琴图显示少突胶质谱系各亚型标志基因的平均表达水平，按表达中位数着色。（C）折线图（左）显示各亚型相对比例在六个时间点的变化；柱状图（右）为 ANOVA 的 F 统计量。（D）流线图显示投影于 UMAP 空间的 RNA 速率流场。（E）ICH 后少突胶质谱系各转变的加权 RNA 速率评分的时间分辨定量。（F）小提琴图显示按相对 CytoTRACE 评分的少突胶质谱系分化潜能差异（评分越低代表分化潜能越高）。（G）热图显示 T 细胞的配体活性，按 AUPR（精确率–召回率曲线下面积）着色。（H）热图显示 T 细胞与 MFOL 之间的调控潜能。（I）韦恩图展示“髓鞘再生轨迹基因”与“少突胶质谱系细胞 DEG”在预测靶基因上的重叠。（J）散点图显示 Vcan、Chn2、Eml1 与 Larp6 在少突胶质谱系细胞中沿伪时间的表达动态。（K）代表性免疫荧光染色显示 ICH 后第 14 天病灶核心（右）与对侧半球（左）的 CD3⁺ T 细胞（绿）与 MBP⁺ 髓鞘（红）。（L）箱线图显示同侧与对侧纹状体中髓鞘碱性蛋白（MBP）信号的定量，p 值由配对检验得出。
为进一步深入理解修复过程，我们采用 Dynamo——一套根植于动力系统理论与微分几何的工具——分析少突胶质谱系的剪接 mRNA（图 5D）。时间特异的 RNA 速率分析显示，ICH 后少突胶质谱系细胞发生位移（图 5E）。naive 组反映出活跃的少突胶质发生（oligodendrogenesis）过程；第 1 天时该轨迹被 ICH 打断，提示髓鞘形成的早期受损。在神经炎症高峰、尤其是第 3 天与第 7 天，NFOL 的成熟开始恢复。此外我们观察到两条轨迹：一条由 MOL 指向 DOL，另一条由 OPC 指向 POPC。这些路径提示少突胶质谱系细胞对 ICH 的初始应答。尽管在 ICH 后期分化轨迹趋于重新对齐 naive 组，但总体上仍更紊乱，凸显了 ICH 对少突胶质细胞动态的长期影响。为量化该轨迹，我们计算了各阶段的 RNA 速率分布（图 5E），呈现相似趋势。我们采用 CytoTRACE2 评估各亚簇的分化潜能，验证了轨迹的可靠性（图 5F）。
既往研究显示，T 细胞在多发性硬化（MS）中通过调控 OPC 增殖与分化促进髓鞘修复；但其在 ICH 中的具体作用仍基本未知。因此我们用 NicheNetR 探索第 7 天与第 14 天 T 细胞与 MFOL 之间的互作。通过配体–受体分析，我们发现神经细胞黏附分子（Ncam1）——在轴突接触后促进少突胶质细胞存活与突起生长——与 Gfra1、Fgfr1 和 Fgfr2 互作（图 5H）。
进一步考察 T 细胞配体对 MFOL 靶基因的影响，我们鉴定出 Tgfb1（图 5G）；据报道它可在成人中枢神经系统中促进 MS 的髓鞘再生。我们的数据提示，Tgfb1 通过 Tgfbr1、Itgb8、Itgav 与 App 等受体调控 MFOL 中多个靶基因的表达（图 5H）。将“髓鞘再生轨迹关键基因”与“预测靶基因”作交集分析，得到若干共享基因，其中 Chn2、Eml1、Larp6、Ndrg1 与 Vcan 在少突胶质谱系中高表达（图 5I、图 S24）。我们进一步考察沿伪时间轨迹呈现动态表达的代表性基因（图 5J、图 S25），发现 Vcan 与 Chn2 优先表达于早期 OPC 阶段；晚期少突胶质分化基因 Eml1 在中间分化阶段达峰；少突胶质细胞特异的 RNA 结合蛋白（RBP）Larp6 则在 MFOL 阶段显著升高。ICH 后第 14 天的免疫荧光分析显示，病灶核心内有密集的细胞群体，但髓鞘极少（图 5K）。MBP 信号定量显示，同侧纹状体的总髓鞘面积较对侧显著减少（图 5L）。对单个髓鞘面积的累积分布分析显示，ICH 后第 14 天病灶区偏向更大的髓鞘尺寸，这可能归因于髓鞘破坏后的结构松散。在病灶边界，T 细胞常紧邻这些发生改变的髓鞘结构（图 5K），提示 T 细胞与少突胶质细胞之间可能存在细胞层面的相互作用。尽管我们的发现指向 ICH 后期 T 细胞–少突胶质细胞互作的可能，但 T 细胞介导髓鞘修复的直接证据仍然缺乏，尚需进一步研究（详见讨论）。
讨论
我们用 snRNA-seq 描绘了胶原酶诱导的 ICH 小鼠模型中跨越时间的转录组变化，聚焦损伤后第 1、3、7、14、28 天这五个关键时间点。与以往主要关注急性期、且集中于血肿或血肿周围区域的单细胞转录组研究相比，我们的方法提供了更宽的时间视野。例如，Zhang 等人对人类 PHE 组织进行了 scRNA-seq，揭示了相当程度的免疫异质性；另一项研究中，Ye 等人用啮齿类白质出血模型展示了髓系细胞与 OPC 群体的变化，突出与炎症相关的脂质代谢重编程和铁死亡信号。不过，本研究的一项局限在于使用 naive 动物而非假手术（sham）动物作为对照。这种做法提供了正常生理状态的基线，却无法扣除手术操作本身的影响；未来纳入 sham 对照的研究将能更精确地评估 ICH 特异的效应。
我们在数据集中鉴定出 67 个不同细胞簇，对应 21 个主要细胞类型；随后系统描绘了 ICH 后病灶周围细胞类型的变化，突出其随时间的动态位移。由于血肿及 ICH 造成的继发性损伤，神经元比例显著下降；相反，胶质细胞与免疫细胞的比例在特定时间点出现不同程度的升高。值得注意的是，MAC 与 MGL 在 ICH 早期显著增多，而 AC、OLG 与 T 细胞主要在中期至后期增多。
神经炎症持续时间很长，即使血肿已被清除，其机制仍错综复杂，且尚缺乏明确的治疗策略。既往研究指出，活化的胶质细胞与免疫细胞可能在神经炎症中发挥重要的调控作用。通过共识共表达网络分析，我们观察到免疫–炎症模块在 ICH 全程强烈激活，变化在第 7 天最为显著。这一现象在炎症通路评估中得到进一步验证。在 ICH 早期，炎症反应促使小胶质细胞快速活化并招募 MAC，以有效清除受损细胞碎片；然而过度的炎症会导致组织损伤迁延、阻碍修复并加重继发性损伤。尽管认识已有进展，我们对炎症进展的确切分子驱动因素与时间线的理解仍有空白。
通过对炎症相关细胞类型作亚簇分析，我们鉴定出一种仅在 ICH 后出现的 Serpina3n⁺ 星形胶质细胞亚型，代表损伤相关星形胶质细胞。既往在癫痫中的研究显示，活化的星形胶质细胞释放的 Serpina3n 可通过 NF-κB 信号通路促进促炎因子释放。ICH 后，MAC 与小胶质细胞——脑内固有免疫细胞——显著增殖，对神经炎症有实质贡献。细胞通讯分析提示，早期 MAC 的募集可能涉及 MIF 相关信号。我们揭示了 ICH 后不同时间点 MAC 的异质性：早期以 Ftl1⁺ MAC 亚型为主，主要清除铁离子、维持铁稳态，有利于神经元存活；Cd74⁺ 与 Ms4a4c⁺ MAC 亚型出现在较晚阶段，说明 MAC 构成随时间发生转变。尽管对 ICH 后 MAC 的多样角色已有一定认识，但驱动 MAC 功能转换的调控机制仍不清楚，未来研究需要阐明这些机制，以开辟新的治疗策略。此外，小胶质细胞对 ICH 的应答是分化为两个可能各司其职的亚型。我们在早期首先鉴定出一个增殖能力很强的小胶质细胞亚型，它对补偿铁死亡等细胞死亡机制造成的小胶质细胞损失至关重要；另一个小胶质细胞亚型则表现出增强的吞噬与抗原呈递功能，有助于清除血肿与胶质瘢痕。ICH 第 7 天之后，该亚型成为炎症信号通路的主要靶标，提示调控小胶质细胞活化对管理 ICH 后期炎症至关重要。这种转变可能源于修复期 MAC 数量的下降，从而让小胶质细胞承担更主导的角色。值得注意的是，小胶质细胞在这一阶段也可能有助于限制炎症：例如 Li 等人报道，新补充的小胶质细胞呈现较低的炎症特征；Spittau 等人则证明小胶质细胞中的 TGFβ 信号起关键抗炎作用，进一步支持小胶质细胞在 ICH 亚急性至慢性阶段成为炎症的关键调节者。调控免疫细胞的活化及其功能状态之间的转换，是管理 ICH 后神经炎症的有前景的治疗靶点；阐明这些转换背后的机制，可为调节免疫反应的策略提供依据，从而有望减轻炎症并促进组织修复。
局灶性炎症病灶可造成大量白质损伤，常导致脱髓鞘，进而破坏神经细胞功能。近期研究提示，髓鞘修复与新分化及成熟少突胶质细胞相关。我们鉴定出完整的少突胶质谱系，并观察到 OPC 与 OLG 对 ICH 后脑损伤微环境的应答。富集于白细胞迁移与免疫反应的 DOL 高表达 Serpina3n——一种与脱髓鞘相关的信号；Serpina3n 在 POPC 中也高表达，提示改变的微环境可能影响 OPC 的细胞周期，刺激其增殖。我们的 RNA 速率分析表明，ICH 导致早期少突胶质发生的紊乱，随后在后期分化增强，这可能反映了一种内在的再生机制，在炎症消退过程中促进髓鞘修复。
Treg 在促进髓鞘形成中的作用是一个新兴的学术关注点。我们在数据集中评估了 T 细胞，发现其在 ICH 后期由细胞毒性功能转向调节性功能。T 细胞与 MFOL 之间的互作鉴定出 NACM1 与 FGFR2 等支持细胞存活与髓鞘形成的通路；TGF-β1 被鉴定为髓鞘维持的关键配体，其潜在靶基因为少突胶质细胞特异的 RNA 结合蛋白 Larp6。然而，T 细胞数量稀少，使我们无法更精确地区分不同亚型的作用。此外，虽然 Larp6 在结直肠癌中抑制鞘磷脂合成，但要在未来确认其在髓鞘再生中的作用，仍必须开展进一步的通路实验。
总体而言，本研究推进了对 ICH 单细胞转录组图景的认识，并为 ICH 后潜在的分子机制提供了见解。
材料与方法
ICH 小鼠模型
基于我们既有的 ICH 建模方案，本研究使用雄性 C57BL/6 小鼠，以 70% N₂O 与 30% O₂ 混合气体中的异氟烷麻醉（诱导 4%、维持 2%）。小鼠置于脑立体定位仪上，沿矢状缝作 1–1.5 cm 切口；切口区域填以石蜡后，用连接微量注射泵的玻璃微电极，将胶原酶 VII-S（0.075 U 溶于 0.5 µL 无菌生理盐水；Sigma）注入左侧纹状体，坐标：前囟前 0.8 mm、中线侧方 2.0 mm、颅骨表面下 3.5 mm，使用双臂脑立体定位仪（Stoelting, 51503D）。胶原酶以 0.1 µL/min 恒速输注。术中通过体温维持装置（RWD）将小鼠体温维持在 37.0 ± 0.5 °C。naive 组为未经处理的雄性 C57BL/6 小鼠；术后 24 小时内死亡的小鼠被排除在分析之外。
伦理批准
本研究的伦理批准来自郑州大学生命科学伦理审查委员会（批准号 ZZUIRB 2022–148）与华大基因机构审查委员会（BGI ETHICAL CLEARANCE，批准号 BGI-IRBE23025）。
组织取材与单核分离
于 ICH 后第 1、3、7、14、28 天采集模型小鼠脑组织，每个时间点一只小鼠。脑切片取自血肿周围区域：具体而言，将 ICH 模型小鼠的脑置于脑模具中，按预定坐标切取皮层与纹状体——切片 1（Paxinos 与 Franklin 小鼠脑图谱第 29–38 图）与切片 2（第 21–29 图），以确保相邻血肿取材区域的解剖一致性与可重复性。所采组织立即速冻，备作 snRNA-seq。单核分离按既往方案并稍作修改：简要而言，解冻组织剪碎后与裂解缓冲液（10 mM Tris–HCl，pH 7.4；10 mM NaCl；3 mM MgCl₂；0.1% IGEPAL CA-630；1% BSA；1× 蛋白酶抑制剂；1 U/µL RNase 抑制剂）混合，转入 1 mL 杜恩斯（Dounce）匀浆器，先用松配的 A 杵匀浆 25–50 次，再以紧配的 B 杵匀浆 25 次。匀浆液经 40 µm 滤网过滤，4 °C、500 × g 离心 5 min 沉淀细胞核。细胞核以 DAPI 染色，用 Countstar 自动细胞计数仪计数以确定核悬液浓度；依计数结果取适当体积，在细胞重悬缓冲液中调整终浓度至 1000 核/µL，用于建库。
按厂商方案，使用 DNBelab C 系列单细胞建库试剂盒（MGI Tech）制备 snRNA-seq 文库。简要而言，单核悬液用于液滴生成、破乳、磁珠收集、逆转录与 cDNA 扩增，形成带条形码的文库；随后按厂商方案构建带接头的文库。浓度以 Qubit ssDNA 检测试剂盒（Thermo Fisher Scientific, Q10212）测定；文库定量后，在中国国家基因库（深圳）的 DNBSEQ 平台上以双端测序方式测序，策略为：读段 1 长 41 bp、读段 2 长 100 bp。
定量与统计分析
原始数据处理。DNBSEQ-T10 的原始测序读段用 PISA（v1.1）过滤并拆分。比对前，用 Drop-seq（v1.13）去除 poly(A) 尾、为读段附加细胞与 UMI 条形码，并滤除条形码碱基质量低的读段。读段以 STAR（v2.7.4a）比对至 GRCm38 基因组，并用 Sambamba（v0.7.0）排序。对 snRNA-seq 数据，合并外显子与内含子读段以统计每个基因的转录本数；最后用 PISA 生成“核 × 基因”的 UMI 计数矩阵。
去除双细胞。各文库以 Scrublet（v0.2.3）去除双细胞。Scrublet 通过在数据集内模拟人工双细胞、构建最近邻分类器来检测双细胞，无需先验知识或预聚类。按 Scrublet 默认参数执行，剔除与模拟伪双细胞相似度最高的前 5% 细胞。
去除低质量细胞。我们对每个文库采用中位数绝对偏差（MAD）法，动态识别并剔除离群细胞：基于 UMI 计数、基因数、线粒体百分比与核糖体百分比计算 MAD，以 5 倍 MAD 为阈值决定保留哪些细胞。此外设定基本质控标准：UMI 计数 >1000、基因数 >500、线粒体百分比 <10%。
去除环境 RNA。用 celda（v1.18.1）的 DecontX 函数以默认设置降低环境 RNA（ambient RNA）噪声，用于估计并去除单细胞基因组数据中的污染。
整合、聚类与细胞类型注释
获得过滤后的基因 UMI 计数矩阵后，我们采用半监督流程作初步注释：用 Scanpy（v1.10.2）合并全部样本数据并提取高变基因子矩阵；随后用深度学习工具 scVI 去除批次效应。基于批次校正后的降维矩阵，以 Leiden 算法聚类，分辨率参数设为 5。我们以 Shi 等人的数据集构建模型并据此注释本数据集，随后人工校订注释。表中所引全部文献均已独立核实，可通过所附 PMID 溯源。
细胞簇比例分析
不同组间的细胞簇比例以单因素方差分析（ANOVA）后接 Dunnett 多重比较检验进行比较。全部统计分析在 R 中使用 speckle 包（v1.0.0）完成。
共识共表达网络分析
为在基因表达水平上考察全局转录变化，我们按 https://smorabit.github.io/hdWGCNA/index.html 的教程，用 hdWGCNA（v0.3.03）进行共识共表达网络分析；此外进行 DME 分析，以鉴定 ICH 后上调或下调的模块。
基于邻域的差异丰度分析
我们使用 milopy（v0.1.0）框架检测精细细胞群体随时间的动态变化：由 KNN 图构建邻域，用线性模型进行 DA 检验。为捕捉时间进程，将时间点（D1、D3、D7、D14、D28）按序编码为整数，作为连续变量纳入设计公式。
亚型分析、差异表达分析与 GO 富集
为进一步考察目标细胞类型的生物学功能，我们分别对 AC、MGL、MAC、T、OLG 与 OPC 的转录组数据用 Scanpy 作亚组分析；依据各组细胞数与生物学功能，经多轮测试确定适当的分辨率。随后用 Libra R 包（v1.7）作差异表达分析——其伪批方法可在单细胞 RNA-seq 数据中控制生物学重复。先将同一样本内同一簇细胞的原始基因计数汇总，构建伪批表达矩阵；再用 run_de() 函数在既定细胞亚型与其各自参照群体之间作比较，内部以 DESeq2 建模与统计检验。p 值 <0.05 且 |log2 倍数变化| >1 的基因视为显著差异表达。此外，用 clusterProfiler（v4.12.6）的 compareCluster 函数对多组上调基因作富集分析。
细胞–细胞通讯
我们用 CellChat 包（v1.5.0）分析神经炎症亚型内的通讯网络，以完整的 cellChatDB.mouse 作为受体–配体数据库识别潜在互作。核心函数 computeCommunProb 采用 truncatedMean 方法计算各细胞亚型在关键互作分析中的平均基因表达。按 https://github.com/sqjin/CellChat 的教程，为每个时间点构建细胞–细胞通讯网络，并以 netP 数据槽识别通路；最后聚焦与神经炎症相关的信号通路。
多评分分析
为描述不同细胞类型之间的连续谱变化，我们对核心数据作了多种评分分析。首先，基于炎症反应相关基因（GO:0006954），用 Scanpy 的 score_genes 函数分析免疫细胞的炎症反应水平，以 naive 组作为标定基线。其次，我们评估少突胶质谱系的分化潜能并对细胞周期分期评分：分化潜能用 CytoTRACE2（v1.0.0）以默认参数处理原始计数矩阵，得到每个细胞的分化潜能评分；细胞周期评分用 Scanpy 的 score_genes_cell_cycle 函数评估 S 期与 G2M 期。最后，我们用 TCellSI（v0.1.0）评估 T 细胞的功能评分：基于标准化表达矩阵，用 TCSS_scRNAseqCalculate 函数完成；该过程利用 HomoloGene（v1.4.68.19.3.27）完成人–鼠同源基因转换。
RNA 速率与向量场分析
我们按官方教程（https://dynamo-release.readthedocs.io/en/latest/）用 dynamo（v1.3.2）重建表达动态向量场。简要而言，使用少突胶质谱系 snRNA-seq 数据的剪接与非剪接计数矩阵，计算脱髓鞘与髓鞘再生过程中的 RNA 速率、加速度与曲率；通过将特定基因表达水平设定为 25，模拟扰动这些基因后对髓鞘再生 RNA 速率的影响。
少突胶质谱系转变速率的量化
为量化沿少突胶质谱系的转变速率，我们使用 Pearson 转变矩阵。对处于给定状态的每个细胞，计算其向下一个状态细胞（如 OPC → NFOL）的转变权重之和，并设最小权重阈值以滤除伪连接；该评分代表向目标状态转变的潜力。我们在既定谱系阶段（OPC → NFOL → MFOL → MOLL → DOL）与各发育时间点上重复该流程；转变速率评分按时间点与谱系转变汇总并作图。
NicheNet 分析
为探索 T 细胞如何驱动髓鞘形成，我们用 NicheNetR（v1.0.0）识别介导 T 细胞与 NFOL、MFOL 相互作用的潜在配体。首先将 NFOL 与 MFOL 设为“接收细胞”，鉴定在至少 5% 细胞中表达的基因，聚焦第 7 天与第 14 天相较 naive 阶段的差异表达基因；随后基于背景基因预测活性配体，推断相应受体与靶基因并找出排名靠前的配体；最后锁定关键靶基因，供后续扰动分析使用。
免疫荧光
多重免疫荧光染色在石蜡包埋脑切片上进行，使用基于酪胺信号放大（TSA）的试剂盒（#AFIHC027，艾方生物）。脱蜡并以柠檬酸缓冲液（pH 6.0）作抗原修复后，切片与一抗 4 °C 孵育过夜；洗涤后加 HRP 标记二抗，继以 TSA 荧光团放大。每轮染色之间用洗脱缓冲液（#G1266，艾方）作抗体剥离，从而依次检测多个靶标。细胞核以 DAPI 复染，用抗荧光淬灭封片剂封片。
所用抗体组合：MIF 信号通路（图 4E）：五色组合，含 CD44（1:1500，#AF14754）、CD74（1:1500，#AF06806）、MIF（1:1500，#AF05273）、TMEM119（1:1500，ab209085，abcam）与 F4/80（1:1500，#SAD002）；另作三色组合：MIF（1:1000，BS-1044R，Bioss）、CD74（1:8000，GB115427，Servicebio）、CD44（1:2000，GB112054，Servicebio）。T 细胞浸润（图 3I）：CD3（1:2000，#GB12014，Servicebio）。T 细胞–OLG 交互（图 5K）：双标组合 CD3（1:2000，#GB12014）与 MBP（1:5000，#GB12226，Servicebio）。图像以配备相应滤光片的多光谱荧光扫描仪采集。
免疫荧光分析
提取对应髓鞘染色的红色通道，施加适当阈值转换为二值掩膜；用形态学运算（开、闭、填洞）精修掩膜，并在 ROI Manager 中手工校正感兴趣区。以 Analyse Particles 功能定量，最小尺寸阈值设为 100 µm² 以排除小噪声信号；记录髓鞘阳性总面积及其占视野的比例。T 细胞计数在同一软件中手工完成。
作者贡献与声明
作者贡献 王建（Jian Wang）、王明跃（Mingyue Wang）与王军民（Junmin Wang）构思研究；Mingyue Wang、Qinfeng Peng、Shaoshuai Wang、Nannan Cheng、Mengke Zhao、Jiaxin Li 与 Qinglin Wang 完成实验；Zhan Chen、Qinglin Wang、Rong Xiang 与 Jin Tao 分析与解释数据；Zhan Chen、Junmin Wang、Jian Wang 与 Mingyue Wang 起草与修订稿件；Ruoqi Ding 完成验证实验；Longqi Liu、Chuantyu Liu 与 Xuemei Chen 提供建议。全体作者审阅并批准稿件。
利益冲突 作者声明无利益冲突。
伦理声明 全部动物实验经郑州大学实验动物管理与使用委员会（批准号 ZZUIRB 2022–148）与华大研究院（批准号 BGI-IRB E23025）批准，并按国家实验动物护理与使用指南及 ARRIVE 指南执行。
▲ 本部分翻译难点（对照学习用）
附录 · 英文原文
Single‐nucleus transcriptomic profiling reveals temporal dynamics of neuroinflammation and myelin repair after intracerebral haemorrhage
Abstract
Intracerebral haemorrhage (ICH) progresses rapidly with complex pathology and limited treatment options, making it a severe subtype of stroke. The extravasation of blood into the brain parenchyma triggers a cascade of inflammatory responses, contributing to secondary injury. Single‐nucleus RNA sequencing (snRNA‐seq) data have enabled more profound insights into the cellular heterogeneity and dynamic interactions within the haemorrhagic brain. Immune cells play a crucial role in shaping neuroinflammation. However, the lack of comprehensive longitudinal studies limits our understanding of the temporal evolution of these inflammatory processes, posing a challenge to the development of targeted therapeutic strategies.
We used snRNA‐seq in collagenase‐induced ICH mouse models at Days 1, 3, 7, 14 and 28 post‐injury, alongside naive controls, to profile the dynamics of gene expression over time.
We obtained 281 577 high‐quality transcriptional profiles representing 21 distinct cell types. Co‐expression network analysis revealed a prominent ‘inflammation module’ that remained active throughout ICH. Integrative single‐cell transcriptomic and immunofluorescence staining suggested that the various Mif‐expressing cells may contribute to local inflammation, potentially engaging macrophages via receptor–ligand pairs such as Cd44 and Cd74. Over time, microglia appeared to serve as key recipients of pro‐inflammatory signals increasingly. During the resolution phase, oligodendrocytes exhibited transcriptional signatures consistent with enhanced maturation and remyelination, which T cell‐mediated interactions may have facilitated.
These findings offer a systems‐level perspective on cell‐type–specific responses and immune‐mediated interactions during ICH progression and resolution.
Establish intracerebral haemorrhage (ICH) mouse models at various time points (Days 1, 3, 7, 14, 28) and construct a high‐quality single‐nucleus RNA sequencing (snRNA‐seq) atlas.Computational analyses suggest that macrophage recruitment in the early stage of ICH potentially involves migration inhibitory factor (MIF) signalling pathways.T cells may interact with myelin‐forming oligodendrocytes during the resolution phase, potentially contributing to remyelination after ICH.
Establish intracerebral haemorrhage (ICH) mouse models at various time points (Days 1, 3, 7, 14, 28) and construct a high‐quality single‐nucleus RNA sequencing (snRNA‐seq) atlas.
Computational analyses suggest that macrophage recruitment in the early stage of ICH potentially involves migration inhibitory factor (MIF) signalling pathways.
T cells may interact with myelin‐forming oligodendrocytes during the resolution phase, potentially contributing to remyelination after ICH.
INTRODUCTION
Intracerebral haemorrhage (ICH) is a severe subtype of stroke, accounting for 27.9% of all strokes worldwide in 2019. Limited progress has been made in therapeutic strategies for ICH, with only modest improvements in patient outcomes over the past few decades. The pathophysiology of ICH involves direct pressure from the expanding haematoma, as well as secondary physiological and cellular responses triggered by the haematoma and blood‐derived metabolites. These challenges underscore the pressing need for ongoing research to develop innovative therapeutic strategies for ICH., Blood leakage into the brain during ICH induces a cascade of harmful processes, including brain oedema, thrombin toxicity and neuroinflammation.
Neuroinflammation, mediated by cell types such as microglia, astrocytes, macrophages (MAC) and T cells, is a key driver of secondary brain injury and is associated with poor clinical outcomes., It exacerbates brain damage, hinders tissue repair and contributes to post‐stroke motor and cognitive impairments. Microglia and MAC, pivotal mediators of inflammation, are traditionally categorised into M1 (pro‐inflammatory) and M2 (anti‐inflammatory) subtypes based on experimental models. However, accumulating evidence suggests these cells exhibit a functional spectrum rather than discrete states. Zhang et al. identified 12 microglial subtypes in peri‐haematomal oedema (PHE) following ICH, revealing diverse phenotypes including DNA repair, degenerative, disease‐associated and transitional inflammatory states. These findings highlight the phenotypic complexity of microglia in the ICH microenvironment.
In addition to neuroinflammation, demyelination is a critical pathological feature of ICH. The breakdown of the myelin sheath disrupts neuronal signal transmission, leading to functional impairments., The process of demyelination can be reversed by oligodendrocyte precursor cells (OPCs), which are vital for restoring myelin integrity and improving recovery. OPCs proliferate and differentiate into mature oligodendrocytes (MOL) to repair damaged myelin. Emerging evidence also highlights the role of regulatory T cells (Tregs) in promoting remyelination and neural repair by modulating immune responses.
Single‐nucleus RNA sequencing (snRNA‐seq) has enabled a more nuanced understanding of cellular responses to ICH. This technique provides high‐resolution transcriptional profiles of brain cells, facilitating the identification of distinct cellular populations and their molecular pathways., Such insights are crucial for unravelling the complexity of immune cell interactions and their contributions to inflammation.
In this study, we employed snRNA‐seq to investigate the cellular and molecular dynamics of neuroinflammation and remyelination in a collagenase‐induced ICH mouse model. We identified distinct cell clusters and their molecular characteristics, revealing key cellular interactions and regulatory mechanisms involved in inflammation and remyelination. Our findings identify specific therapeutic targets to promote functional recovery after ICH, providing a foundation for developing treatment strategies.
RESULTS
Transcriptomic landscape of mouse peri‐haematoma tissue
To examine the temporal dynamics of gene expression following ICH, we utilised the collagenase‐induced ICH model. This well‐established methodology triggers ICH by breaking down the vascular basement membrane. We monitored gene expression at critical time points after ICH (1, 3, 7, 14 and 28 days). We digested brain sections around the lesion for snRNA‐seq analysis (Figure and Table). Cells were filtered based on the mitochondrial gene ratio, gene count and doublet score to remove low‐quality cells (Figure). A total of 281, 577 cells were identified, with an average of 2010 genes detected per cell.
[FIGURE 1] The single‐cell landscape of mouse brain tissue surrounding the lesion in the ICH model and naive controls. (A) Schematic of the workflow for transcriptomic profiling of the ICH model using single‐nucleus RNA sequencing (snRNA‐seq). (B) The dot plot shows the mean expression levels of marker genes for the 21 significant cell populations. The colour represents the average scaled gene expression level (z‐score), and the dot size represents the percentage of cells in which the marker gene was detected for each population. (C) Uniform Manifold Approximation and Projection (UMAP) visualisation of clustering in the snRNA‐seq transcriptomic data from the ICH mouse brain lesion tissue. The outer circles represent the log10‐transformed cell numbers for each cell type. A total of 21 cell populations are identified: DGGRC (dentate gyrus granule neurons), MEGLU (di‐ and mesencephalon excitatory neurons), DEINH (di‐ and mesencephalon inhibitory neurons), GNBL (glutamatergic neuroblasts), NGNBL (non‐glutamatergic neuroblasts), PEP (peptidergic neurons), TEINH (telencephalon inhibitory interneurons), TEGLU (telencephalic glutamatergic neurons), MSN (telencephalon projecting inhibitory neurons), MAC (macrophages), T (T cells), AC (astrocytes), MGL (microglia), OPC (oligodendrocyte precursor cells), OLG (oligodendrocytes), EPEN (ependymal cells), CHOR (choroid plexus epithelial cells), EN (endothelial cells), PER (pericytes), FB (fibroblasts) and VLM (vascular leptomeningeal cells). (D) Line plots (left) show the changes in relative proportions of primary cell types across six time points (naive, Day 1, Day 3, Day 7, Day 14 and Day 28). Bar plots (right) show F‐statistics using analysis of variance (ANOVA), coloured by cell types. They are categorised into four major groups: neurons, immune cells, glial cells and vascular cells. (E) The scatter plot illustrates the DEGs of TEGLU_0, with the top 100 genes highlighted in light green. (F) The violin plot shows the programmed cell death pathway scores in TEGLU subtypes.
Initially, we used scVI to remove batch effects across libraries from different experimental conditions (see Materials and Methods section). The dimensionality reduction matrix generated by scVI was used for subsequent Uniform Manifold Approximation and Projection (UMAP) and Leiden clustering. We manually identified 67 clusters and annotated them to 21 major cell types using curated marker genes from the literature (Figures,, and Table). These cell types included neuronal (9), immune (2), glial (4) and vascular (6) cell clusters (Figures and). Each cell type was present at all time points and locations. However, distinct temporal preferences of cell types were observed (Figure). Inhibitory neurons were more abundant in the striatum, while peripheral blood immune cells were comparatively scarce in the naive group. This scarcity was likely due to the intact blood–brain barrier (BBB), which restricted the infiltration of immune cells from the peripheral blood. We observed pronounced diversity within the telencephalic glutamatergic neurons (TEGLU) cluster (Figure), consistent with the cerebral cortex's complex architecture and biological functions.
To investigate shifts in cell types within the lesion after ICH, we analysed their proportions using propeller analysis (Figures and). The proportion of neuronal cells, particularly TEGLU and TEINH, declined after ICH, which contributes to the death of surrounding neurons. The TEGLU_0 exhibits an increased mitochondrial percentage (Figures and). Previous research indicated that excessive mitochondrial fission exacerbates neuronal pyroptosis and neuroinflammation., We assessed programmed cell death processes using well‐established gene sets. TEGLU_0 had a higher score than other subtypes in ferroptosis, necroptosis and pyroptosis (Figure). Furthermore, we observed the infiltration of numerous peripheral immune cells into the brain parenchyma following the disruption of BBB, each exhibiting distinct temporal preferences. Specifically, the prevalence of MAC was more pronounced in the early stages, while T cells exhibited a greater presence in the later stages. This observation suggests that peripheral immune cells assume varying biological roles in the progression of ICH.
Cellular contributions to continued activation of inflammatory co‐expression modules
To explain which cell types contributed to the neuroinflammation and examine the molecular dynamics following ICH, we performed a consensus co‐expression network analysis using hdWGCNA. First, we selected the soft‐thresholding power based on the point where the scale‐free topology fit curve (Figure) began to plateau, indicating a good approximation to a scale‐free network while avoiding excessively high powers that would reduce network connectivity. We determined the following optimal thresholds: Naive (5), Day 1 (6), Day 3 (6), Day 7 (5), Day 14 (5) and Day 28 (7). We then systematically constructed separate networks for the gene expression matrix at each time point, integrated networks and identified gene modules. We identified 10 distinct modules across six networks (Figure). Each module consisted of genes with high co‐expression, likely reflecting coordinated regulatory mechanisms and involvement in shared biological processes.
By integrating hdWGCNA and Harmony, we generated harmonised module eigengenes (hMEs), which served as metrics for summarising the gene expression profiles of co‐expression modules. To pinpoint the pivotal genes within each module, we calculated the eigengene‐based connectivity (kEM) of each gene and identified the hub genes, which serve as critical indicators of biological functions (Figure and Table). Subsequently, we visualised the co‐expression matrix using hub genes as features, highlighting top three genes (Figure). Module 6 was enriched in pro‐inflammatory genes, while Module 2 exhibited myelination‐related genes (Figure). The edge visualisation indicates that each hub gene exhibits strong co‐expression, underscoring their functional significance.
[FIGURE 2] The co‐expression network illustrates changes in gene expression following ICH. (A) UMAP visualisation of the hub gene co‐expression network for the ICH model and Naive group. Nodes, coloured by module assignment, represent genes, with the top three hub genes annotated for each module. Edges, downsampled for clarity, represent co‐expression relationships. (B) Correlation plots show the correlation between each module based on their harmonised module eigengenes (hMEs), with purple indicating positive and green indicating negative correlations. The areas of the squares show the absolute value of the corresponding correlation coefficients. (C) UMAP plot of all cell populations from single‐nucleus RNA sequencing (snRNA‐seq), coloured by module assignment. (D) Bar plot showing the top three gene ontology (GO) enrichment pathways, with the y‐axis representing log10 p values (values over 20 were truncated for clarity) and bars coloured by module assignment. (E) Heatmap displaying differential module eigengene (DME) results comparing the ICH model to Naive across time points. The Wilcoxon rank‐sum test with Bonferroni correction was applied: not significant (p >.05); ∗p <.05; ∗∗p <.01; ∗∗∗p<.001; ∗∗∗∗p<.0001. (F) Dot plot showing the enrichment of differentially expressed genes (DEGs) of primary cell types in each module, coloured by module assignment. Dot size represents the logarithm of the odds ratio.
To investigate potential interactions between modules, we calculated Pearson correlations between them using hMEs. Module 2 and Module 9 notably exhibited strong correlations, with their expressed regions closely aligned with the oligodendrocyte lineage (Figure). Interestingly, Module 9 was also expressed in astrocytes and immune cells, suggesting that these cells might support neuronal ensheathment.
We performed differential module eigengene (DME) analysis to identify modules that exhibit time‐specific changes by comparing the ICH groups to naive controls across all time points (Figure). Module 4, primarily enriched in astrocytes (AC) and fibroblasts (FB), exhibited a significant increase after Day 7 (Figures and). Differential expression analysis revealed that several of the top 50 hub genes in Module 4 were up‐regulated in AC at Day 14 compared to the naive group, including genes related to astrocyte maturation process (Ndrg2, Gpr37l1) and extracellular matrix remodelling (Timp3, Sdc4, Clu; Figure). Moreover, spatial transcriptomic analysis indicated that astrocytes and fibroblasts were located in adjacent regions at the lesion border (Figure). These findings suggest that astrocytes and fibroblasts cooperatively mediate glial scar formation, thereby contributing to the filling of brain cavities left by the haemorrhage. More importantly, Module 6 showed sustained up‐regulation after ICH compared to the naive group, with elevated expression observed on Days 3 and 7 (Figure). Immune‐related cells, including MAC, microglia (MGL) and T cells, are the primary contributors to this module (Figure). It was enriched in immune effector activity, leukocyte activation and antigen processing and presentation (Figure). The sustained activation of Module 6 suggests that inflammation plays a crucial role after ICH.
Dynamic immune landscape and functional shifts of neuroinflammation cell type during ICH progression
Neuroinflammation, a key driver of secondary injury following ICH, remains inadequately understood., Co‐expression network analyses revealed persistent up‐regulation and activation of immune pathways throughout ICH progression. Microglia, astrocytes, MAC and T cells were the principal contributors to the neuroinflammatory response (Figure). The inflammatory scores of these cells were elevated, reaching a peak at Day 7, with MAC and microglia being the primary sources (Figure). We conducted a sub‐clustering analysis and identified four astrocyte subtypes, three microglial subtypes and three MAC subtypes (Figures and). We employed milo, a differential abundance (DA) analysis approach based on K‐nearest neighbour (KNN) graphs, to quantify the dynamic changes during ICH progression (Figure). Among immune cells, MAC predominantly appeared at earlier stages. In contrast, T cells were enriched at later stages. AC_4 and MGL_2 exhibited dynamic temporal changes in glial cells, as revealed by their shifting patterns over time (Figure). The immune cells infiltrated and induced inflammation after the disruption of the BBB, while the AC and MGL subtypes associated with ICH reflected glial responses to the haemorrhage. These markedly altered subtypes may significantly impact ICH pathology.
[FIGURE 3] Cellular heterogeneity of neuroinflammation subtypes within the ICH model. (A) UMAP visualisation of subtypes in the snRNA‐seq transcriptomic data from astrocytes (AC), microglia (MGL), macrophages (MAC) and T cells. (B) Beeswarm plot showing differential abundance of neuroinflammation subtypes at each time point after ICH. Significant changes are highlighted in red and blue. (C) Dot plot showing DEGs of astrocytes, coloured by subtype. (D) Box plots showing the neuroinflammation scores of astrocyte subtypes. (E) The volcano plot illustrates the up‐regulated differentially expressed genes of the microglial subtype comparing MGL_1, coloured by subtype. (F) Bar plots (bottom) show the changes in relative proportions of subtypes across six time points (naive, Day 1, Day 3, Day 7, Day 14 and Day 28). Bar plots (up) show F‐statistics using analysis of variance (ANOVA), coloured by subtypes. (G) Dot plot showing DEGs of macrophages, coloured by subtype. (H) Bar plot showing gene ontology (GO) enrichment of up‐regulated DEGs in MAC. (I) Representative immunofluorescence images showing CD3⁺ T cells (red) in the lesion core at Day 7 and Day 14 post‐ICH. DAPI (blue) marks nuclei. (J) Box plots showing T cell counts at Day 7 and Day 14 after ICH. p value was determined by t‐test. (K) Line plots showing inferred T cell functional scores over time after ICH, including cytotoxicity (left) and regulatory (right) programs, as assessed by TCellSI.
To further delineate the molecular signatures of distinct cell subtypes, we analysed differentially expressed genes (DEGs) by pseudo‐bulk and their associated functional pathways (Table). AC_2 and AC_3 exhibited gene expression profiles indicative of regionally enriched astrocyte states. Specifically, Thbs4, highly expressed in AC_2, has been associated with white matter astrocytes, while AC_3 expressed Agt and Slc6a11, markers linked to thalamic astrocytes (Figure). Interestingly, the abundance of AC_4 peaked at Day 7, coinciding with the maximal inflammatory response (Figures and). This cluster exhibited high expression of Serpina3n, a gene known to contribute to neuroinflammation by activating the NF‐κB signalling pathway. It is consistent with previously described reactive astrocyte profiles. MGL_1, homeostatic microglia, marked by high expression of P2ry12 and Tmem119, which were associated with microglial maintenance (Figure). MGL_3, the earliest microglial subtype to respond to ICH, activated cell cycle pathways to promote cell proliferation (Figures and). MGL_2 was primarily enriched in leukocyte migration and wound healing pathways (Figure).
Peripheral immune cell infiltration after ICH responds to microenvironmental changes within the brain. MAC_3, predominantly on Day 3, cleared iron ions from the haematoma to maintain iron homeostasis (Figure). On Day 7, MAC_1 retained a high proportion of cell populations, facilitating the removal of cellular debris by activating complement pathways and immune responses (Figure). MAC_2 showed high expression of Cd74 and was enriched in immune response. Additionally, MAC_2 expressed Ms4a4c, a novel M2 surface marker. Furthermore, MAC_2 was enriched in positive regulation of T cell activation pathways, further supporting its potential protective function (Figure). The results indicated a functional shift in MAC from iron clearance to debris removal and anti‐inflammation during ICH progression, suggesting that stage‐specific regulation can optimise their therapeutic potential.
Additionally, the number of T cells was insufficient for further sub‐clustering. However, they predominantly appeared during the repair phase, suggesting potential involvement in post‐ICH recovery (Figure). To provide spatial context, we performed immunofluorescence staining for CD3 on brain sections at Days 7 and 14. CD3+ T cells were visible in peri‐haematomal regions at both timepoints (Figure), qualitatively consistent with their presence during the subacute phase as suggested by single‐cell data. To further explore their potential functions, we applied TCellSI scoring, which revealed a dynamic shift in T cell states: cytotoxicity scores peaked at Day 7, while regulatory scores gradually increased and remained elevated until Day 14, indicating a possible functional transition during recovery (Figure).
Dynamic shift from pro‐inflammatory to anti‐inflammatory signalling by the cell communication of the subtypes
To better understand how different cellular subtypes contribute to neuroinflammation, we examined their interactions within the cell communication network, with a focus on pro‐inflammatory and anti‐inflammatory signalling pathways. Notably, the interaction number initially increased until Day 7, following a trend similar to the inflammatory module and scores (Figures and).
[FIGURE 4] Cell–cell interactions in neuroinflammation subtypes within the ICH model. (A) The number of predicted ligand–receptor (L–R) interactions in neuroinflammation subtypes of ICH, coloured by subtype. (B) The sum of interaction probability differences (relative information flow) for each interaction group in the ICH model, categorised as pro‐inflammatory (top) and anti‐inflammatory (bottom). (C) Hierarchy plot showing the migration inhibitory factor (MIF) signalling network on Day 1 (left), Day 3 (middle) and Day 7 (right). Nodes represent cell types, and edges denote interactions, with width proportional to interaction probability and colour by subtype. (D) The dot plot showing the mean expression levels of L–R of MIF signalling pathway in Days 1 and 3. The colour represents the average scaled gene expression level (z‐score), and the dot size represents the percentage of cells in which the marker gene was detected for each population. (E) Representative immunofluorescence staining shows MIF signalling. MIF is the ligand, and CD44 and CD74 are the receptors. (F) Circle plot showing the MIF signalling network on Day 14 (left) and Day 28 (right). Nodes represent cell types, and edges denote interactions, with width proportional to interaction probability and colour by subtype. (G) The dot plot showing the mean expression levels of ligand–receptor of CD226 signalling pathway in Days 14 and 28. The colour represents the average scaled gene expression level (z‐score), and the dot size represents the percentage of cells in which the marker gene was detected for each population.
The identified pro‐inflammatory pathways exhibited distinct temporal variation patterns (Figure). Specific signalling pathways, particularly those with heightened activity on Day 1, such as the CCL and migration inhibitory factor (MIF) pathways, were of particular interest in elucidating the early inflammatory response. Chemoattractant factors such as Ccl2, Ccl6, Ccl7 and Ccl12 (Figure), secreted by diverse cell types within the brain parenchyma, facilitate the recruitment of immune cells to the injury site. Furthermore, MAC expressing Cd44 and Cd74 interact with various Mif‐expressing cells (Figure), suggesting that the MIF‐stimulated CD74–CD44 receptor complex may promote cell survival and proliferation via activation of the ERK pathway., Spatial transcriptomics revealed a marked accumulation of MAC around the haematoma on the first day after ICH, and Mif was broadly expressed by multiple cell types, consistent with our data (Figures and). Immunofluorescence staining further showed spatial proximity between Mif signals and Cd74/Cd44 signals (Figure), supporting their potential interaction. These complementary observations indicate that Cd74‐expressing MAC may respond to Mif derived from surrounding cells, which may contribute to local inflammatory responses in the acute phase.
During Days 7 and 14 after ICH, we observed enhanced activation of immune cells, as indicated by increased levels of CD45, MHC‐I and complement‐mediated phagocytosis (Figure). CD45 is a marker for activated leukocytes, while MHC‐I presents antigens to CD8+ T cells, promoting immune responses. Complement‐mediated phagocytosis enables MAC to clearing debris and pathogens. In the later stages, pathways, including ITGAL‐ITGB2, MHC class II and LCK, became more prominent. The ITGAL‐ITGB2 pathway was crucial for leukocyte adhesion and migration, facilitating the recruitment of immune cells to the injury site. MHC class II molecules present extracellular antigens to CD4+ T cells, enhancing the adaptive immune response. We compared the signalling intensity of cellular responses to inflammatory pathways at three key time points: the naive group and Days 1 and 7 after ICH (Figure). Our investigation revealed a temporal shift in predominant receptor signalling, with MAC_2 emerging as the primary signalling receptor at the 24‐h post‐ICH time point, followed by a complete transition to MGL‐2 dominance by Day 7 post‐ICH. This shift suggested that managing inflammation in the later stages of ICH necessitated a greater focus on microglial activation.
There was a broader shift from a predominantly pro‐inflammatory response in the acute phase towards a more anti‐inflammatory state in the later stages (Figure). The VISTA pathway, an inhibitory immune checkpoint pathway,, was found to be down‐regulated after ICH. The interaction between the microglia‐expressed ligand Vsir and the astrocyte‐secreted protein Igsf11 highlighted the potential role of astrocytes as mediators in modulating inflammatory responses (Figure). We observed an up‐regulation of the TGFβ pathway (Tgfb2 − (Tgfbr1 + Tgfbr2)) from AC_2, along with the GDF pathway (Gdf15 − Tgfbr2) from T cells on Day 7 (Figures and). Both pathways signal through the Tgfbr2, which is reported to lead to the suppression of microglial activation. Additionally, GDF has been shown to regulate triglyceride metabolism and contribute to increased tolerance to inflammatory damage. Notably, the high expression of Cd226 on the surface of T cells directly interacts with Pvr on inflammatory cells, preventing excessive local inflammation (Figure). This interaction is a key regulatory mechanism that maintains local immune homeostasis and protects tissues by balancing positive and negative signals at immune checkpoints. Anti‐inflammatory pathways were identified within the inflammatory environment, potentially presenting strategies for modulating inflammation following ICH. As we explored this intricate cellular interplay, we gained a clearer understanding of the delicate balance between pro‐inflammatory and anti‐inflammatory signalling.
Oligodendrocyte lineage dynamics and potential T cell–oligodendrocyte interactions after ICH
As neuroinflammation subsides in the later stages of ICH, the resulting shift in the immune microenvironment may create favourable conditions for myelin repair. In the co‐expression network analysis, immune cells contribute significantly to myelin maintenance in Module 9 (Figure).
To investigate changes in oligodendrocyte lineage cells and the relationship between neuroinflammation. OPCs and OLGs were subdivided into sub‐clusters, identifying six distinct sub‐clusters based on classical markers (Figure). We further compared disease‐associated oligodendrocytes (DOL) with MOL (Figure and Table). DOL, which increased after ICH, expressed classical injury‐related markers, such as Serpina3n and up‐regulated genes, including kallikrein‐related peptidase (Klk8 and Klk9) and fibroblast growth factor receptor (Fgfr4). In addition, proliferative oligodendrocyte precursor cells (POPCs) cluster, highly expressed Serpina3n, increased on Days 1 and 3 and had a high cell cycle score (Figures and). The newly formed oligodendrocytes (NFOL) cluster increased on Day 7 (Figures and). The results indicated that OPC accelerated cell cycle under stress and promoted proliferation for myelin repair.
[FIGURE 5] T cells promote the remyelination of oligodendrocyte lineages. (A) UMAP visualisation of oligodendrocyte lineages in the single‐nucleus RNA sequencing (snRNA‐seq) transcriptomic data from OPC and OLG. DOL, disease‐associated oligodendrocytes; MFOL, myelin‐forming oligodendrocytes; MOLL, mature oligodendrocytes; NFOL, newly formed oligodendrocytes; OPCs oligodendrocyte precursor cells; POPCs, proliferative oligodendrocyte precursor cells. (B) Stacked‐violin plot showing the mean expression levels of marker genes for the oligodendrocyte lineages subtypes, coloured by the median expression. (C) Line plots (left) show the changes in relative proportions of subtypes across six time points (naive, Day 1, Day 3, Day 7, Day 14 and Day 28). Bar plots (right) show F‐statistics using analysis of variance (ANOVA), coloured by subtypes. (D) Streamline plot showing RNA velocity flow projected in the UMAP space. (E) Time‐resolved quantification of weighted RNA velocity scores across oligodendrocyte lineage transitions after ICH. (F) Violin plot showing the difference potential of oligodendrocyte lineages by the relative CytoTRACE score. The lower score represents the higher potential difference. (G) Heatmap showing ligand activity of T cells, coloured by AUPR (area under the precision–recall curve). (H) Heatmap plot showing the regulatory potential between T and MFOL. (I) Venn plot illustrating the overlap between remyelination trajectories genes and oligodendrocyte lineage cells DEGs in predicting target genes. (J) Scatterplot showing pseudotime dynamics of the expression of Vcan, Chn2, Eml1 and Larp6 in oligodendrocyte lineage cells. (K) Representative immunofluorescence staining showing CD3+ T cells (green) and MBP+ myelin (red) in the lesion core (right) versus contralateral hemisphere (left) at Day 14 post‐ICH. (L) Box plots showing quantification of myelin basic protein (MBP) signal in the ipsilateral and contralateral striatum. p value was determined by paired t‐test. (K) Cumulative frequency curves showing the proportion of myelinated fibres across myelin cross‐sectional areas in the lesion (L‐1, L‐2, L‐3) and contralateral (C‐1, C‐2, C‐3) hemispheres.
To gain deeper insights into the repair process, we applied Dynamo, a tool rooted in dynamical systems theory and differential geometry, to analyse the spliced mRNA of the oligodendrocyte lineage (Figure). Time‐specific RNA velocity analyses indicate a shift of the oligodendrocyte lineage cells in ICH (Figure). The naive group reflects an active process of oligodendrogenesis. On Day 1, this trajectory is disrupted by ICH, suggesting an early impairment of myelin formation. During the peak of neuroinflammation, particularly at Days 3 and 7, the maturation of NFOL begins to recover. Additionally, we observed two trajectories: one from MOL to DOL, and another from OPC to POPC. These pathways indicate an initial response of oligodendrocyte lineage cells to ICH. Although the differentiation trajectories tend to realign with the naive group during the later stages post‐ICH, they remain more disordered overall, highlighting the long‐term impact of ICH on oligodendrocyte dynamics. To quantify this trajectory, the distribution of RNA velocities was computed at each stage (Figure) and showed a similar trend. We employed CytoTRACE2 to assess the differentiation potential across various sub‐clusters, validating the reliability of the trajectory (Figure).
Previous studies have shown that T cells facilitate myelin repair in multiple sclerosis (MS) by regulating OPC proliferation and differentiation. However, their specific role in ICH remains largely unexplored. Therefore, we used NichenetR to explore the interactions between T cells and MFOL on Days 7 and 14. Through ligand–receptor analysis, we uncovered that neural cell adhesion molecule (Ncam1), which promotes oligodendrocyte survival and process outgrowth following axonal contact, interacts with Gfra1, Fgfr1 and Fgfr2 (Figure).
By further investigating the impact of T cell ligands on MFOL target genes, we identified Tgfb1 (Figure), which has been reported to promote remyelination in the adult CNS in MS. Our data suggested that Tgfb1 regulated the expression of multiple target genes in MFOL via receptors like Tgfbr1, Itgb8, Itgav and App (Figure). An intersection analysis of the key genes involved in the myelin regeneration trajectory with the target genes identified shared genes. Among these genes, such as, Chn2, Eml1, Larp6, Ndrg1 and Vcan, were highly expressed in oligodendrocyte lineage (Figures and). We further examined representative genes exhibiting dynamic expression patterns along the pseudotime trajectory (Figures and), and found that Vcan and Chn2 were preferentially expressed in early OPC stage. Eml1, late oligodendrocyte differentiation genes, displayed peak expression in the intermediate differentiation stages. Larp6, an oligodendrocyte‐specific RNA‐binding protein (RBP), exhibited a marked increase in expression specifically at the MFOL stage. Immunofluorescence analysis at Day 14 post‐ICH revealed a dense cellular population within the lesion core, with little myelin present (Figure). Quantification of myelin basic protein (MBP) signal demonstrated a pronounced reduction in total myelin area in the ipsilateral striatum relative to the contralateral side (Figure). Cumulative distribution analysis of individual myelin sheath areas revealed lesion shift towards larger sheath sizes at Day 14 after ICH, which may be attributable to structural loosening secondary to myelin disruption (Figure). At the lesion border, T cells were frequently localised in close apposition to these altered myelin structures (Figure), suggesting a potential cellular interplay between T cells and oligodendrocytes. Although our findings point towards possible T cell–oligodendrocyte interactions during the later stages of ICH, direct evidence for T cell‐mediated myelin repair is still lacking and will require further experimental validation.
DISCUSSION
We performed snRNA‐seq to profile transcriptomic changes over time in a collagenase‐induced ICH mouse model, focusing on five key time points: 1, 3, 7, 14 and 28 days post‐injury. Compared to previous single‐cell transcriptomic studies, which have primarily focused on the acute phase and concentrated on the haematoma or peri‐haematomal regions, our approach offers a broader temporal perspective. For example, Zhang et al. conducted scRNA‐seq on human PHE tissue, uncovering considerable immune heterogeneity. In another study, Ye et al. used a rodent model of white matter haemorrhage to demonstrate shifts in myeloid and OPC populations, highlighting lipid metabolic reprogramming and ferroptosis signalling in association with inflammation. However, one limitation of our study is the use of naive animals as controls rather than sham‐operated animals. While this approach provides a baseline for normal physiological conditions, it does not account for the effects of the surgical procedure itself. Future studies including sham‐operated controls would allow a more precise assessment of ICH‐specific effects.
We identified 67 distinct clusters within the dataset, corresponding to 21 major cell types. Then, we systematically mapped the cell type changes surrounding the lesion site post‐ICH, highlighting their dynamic shifts over time. As a result of the haematoma and secondary injury caused by ICH, the proportion of neurons decreased significantly. Conversely, the proportions of glial and immune cells exhibited distinct increases at specific time points. Notably, MAC and MGL exhibited prominent increases in the early stages of ICH, whereas AC, OLG and T cells predominantly increased during the mid‐to‐late stages.
Neuroinflammation persists for an extended duration, involving intricate mechanisms even after haematoma clearance, and lacks a clearly defined treatment strategy., Previous studies have indicated that activated glial and immune cells may play vital regulatory roles in neuroinflammation., Through consensus co‐expression network analysis, we observed robust activation of the immune‐inflammatory module throughout the progression of ICH, with the most pronounced changes observed on Day 7. This phenomenon was further validated by evaluating inflammatory pathways. In the initial stages of ICH, the inflammatory response promotes the rapid activation of microglia and the recruitment of MAC to clear damaged cellular debris effectively. However, excessive inflammation leads to prolonged tissue damage, hinders healing and exacerbates secondary injuries. Despite advancements in our understanding, there are still gaps in our understanding of the precise molecular drivers and the timeline of inflammation's progression.
By sub‐cluster analysis of inflammation‐related cell types, we identified a Serpina3n+ astrocyte subtype that only emerged after ICH, representing injury‐associated astrocytes. Previous studies in epilepsy have shown that Serpina3n, released by activated astrocytes, can promote the release of pro‐inflammatory factors via the NF‐κB signalling pathway. After ICH, MAC and microglia—the brain's intrinsic immune cells—proliferate significantly, contributing substantially to neuroinflammation., Cell communication analyses suggest that MAC recruitment in the early stage may involve MIF‐associated signalling. We revealed the heterogeneity of MAC at various time points after ICH. In the early stage, Ftl1 + MAC subtypes primarily clear iron ions and maintain iron homeostasis, which benefits neuron survival. The Cd74+ and Ms4a4c+ MAC subtypes appeared at later stages, indicating a shift in MAC composition over time. While some progress has been made in understanding the diverse roles of MAC following ICH, the regulatory mechanisms driving shifts in MAC function remain poorly understood. Future research is needed to clarify these mechanisms and to pave the way for novel therapeutic strategies. In addition, microglia respond to ICH by differentiating into two distinct subtypes that may play specialised roles in mediating the inflammatory response. We initially identified a microglial subtype with high proliferative capacity during the early stage. It plays a vital role in compensating for microglial loss resulting from ferroptosis and other cell death mechanisms. The other microglial subtype exhibits enhanced phagocytic and antigen‐presenting functions, aiding in the clearance of haematomas and glial scars. After Day 7 of ICH, this subtype becomes the primary target of inflammatory signalling pathways, suggesting that modulating microglial activation is critical for managing inflammation in the later stages of ICH. This shift may be due to a decline in MAC numbers during the repair phase, allowing microglia to assume a more dominant role. Notably, microglia may also contribute to limiting inflammation at this stage. For example, Li et al. reported that newly repopulated microglia exhibit a reduced inflammatory profile. Similarly, Spittau et al. demonstrated that TGFβ signalling in microglia plays a crucial anti‐inflammatory role, further supporting the idea that microglia become key regulators of inflammation in the subacute to chronic phases after ICH. Regulating immune cell activation and their transitions between functional states presents promising therapeutic targets for managing neuroinflammation after ICH. Elucidating the mechanisms behind these transitions can inform strategies to modulate immune responses, potentially mitigating inflammation and facilitating tissue repair.
Focal inflammatory lesions can lead to substantial white matter damage, often resulting in demyelination, which disrupts the function of neural cells. Recent studies suggest that myelin repair is associated with newly differentiated and MOL. We identified the complete oligodendrocyte lineage and observed the responses of OPC and OLG to the brain injury microenvironment after ICH. DOL, enriched in leukocyte migration and immune responses, exhibited high expression of Serpina3n, a signal associated with demyelination., Serpina3n was also highly expressed in POPC, suggesting that the altered microenvironment may influence the OPC cell cycle, stimulating their proliferation. Our RNA velocity analyses indicate that ICH leads to early disruptions in oligodendrogenesis, followed by enhanced differentiation at later stages, which may reflect an intrinsic regenerative mechanism contributing to myelin repair during inflammation resolution.
The role of Tregs in promoting myelin formation is an emerging area of academic interest.,, We evaluated T cells in our dataset and found they shifted from cytotoxic to regulatory functions in the later stages of ICH. Interactions between T cells and MFOL identified pathways such as NACM1 and FGFR2, which support cell survival and myelin formation. TGF‐β1 was identified as an essential ligand for myelin maintenance. Its potential target gene is Larp6, an OLG‐specific RBP. However, the scarcity of T cells makes it impossible to distinguish the effects of different subtypes more precisely. Additionally, while Larp6 suppressed sphingomyelin synthesis in colorectal cancer, further pathway experiments are essential to confirm its role in myelin regeneration in future.
Overall, our study advances the understanding of the single‐cell transcriptomic landscape of ICH and provides insights into the underlying molecular mechanisms following ICH.
MATERIALS AND METHODS
ICH mouse model
Based on our established protocol for ICH modelling, in this study, male C57BL/6 mice were anaesthetised with isoflurane in a mixture of 70% N2O and 30% O2 (4% induction and 2% maintenance). The mice were placed in a stereotaxic frame, and a 1–1.5 cm incision was made along the sagittal suture. After filling the incision area with paraffin wax, a glass micropipette connected to a micro‐infusion pump was used to inject collagenase VII‐S (.075 U in.5 µL sterile saline; Sigma) into the left striatum at coordinates.8 mm anterior to bregma, 2.0 mm lateral to the midline and 3.5 mm deep from the skull surface, using a dual‐arm brain stereo locator (Stoelting, 51503D; Stoelting Co.). The collagenase was infused at a constant rate of.1 µL/min using the micro‐infusion pump (Stoelting). The mice's body temperature was maintained at 37.0 ±.5°C during the procedure using a body temperature maintenance device (RWD). The naive group consisted of untreated male C57BL/6 mice, and any mice that died within 24 h post‐surgery were excluded from the analysis.
The ethical approvals for this study were obtained from the Life Science Ethics Review Committee of Zhengzhou University (Approval No. ZZUIRB 2022–148) and the Institutional Review Board of BGI ETHICAL CLEARANCE (Approval No. BGI‐IRBE23025).
Tissue collection and single‐nucleus isolation
Brain tissues were collected from ICH model mice at Days 1, 3, 7, 14 and 28, with one mouse assigned to each time point. Brain sections were obtained from regions surrounding the haematoma. Specifically, brains from the ICH model mice were placed in a brain matrix, and the cortex and striatum were dissected at predefined coordinates—Section 1 (Figures 29–38 of the Paxinos and Franklin Mouse Brain Atlas) and Section 2 (Figures 21–29 of the Paxinos and Franklin Mouse Brain Atlas). This approach ensured anatomical consistency and reproducibility in sampling regions adjacent to the haematoma. The collected tissues were immediately snap‐frozen for subsequent snRNA‐seq. Single‐nucleus isolation was performed following a previously described protocol with slight modifications. Briefly, the thawed tissues were minced and mixed with a lysis buffer (10 mM Tris–HCl, pH 7.4; 10 mM NaCl; 3 mM MgCl2;.1% IGEPAL CA‐630; 1% BSA; 1× protease inhibitor; and 1 U/µL RNase inhibitor). The mixture was transferred to a 1‐mL Dounce homogeniser and homogenised with 25–50 strokes using a loose pestle (pestle A), followed by 25 strokes with a tight pestle (pestle B). The homogenate was filtered through a 40‐µm strainer and centrifuged at 500 × g for 5 min at 4°C to pellet the nuclei. The nuclei were stained with 4′,6‐diamidino‐2‐phenylindole (DAPI) and counted using a Countstar automated cell counter to determine the concentration of the nuclear suspension. Based on the counting results, an appropriate volume was taken to adjust the final concentration to 1000 nuclei/µL in a cell resuspension buffer for library preparation.
Following the manufacturer's protocol, the DNBelab C Series Single‐Cell Library Prep Set (MGI Tech Co.) was used for preparing snRNA‐seq libraries. In brief, single‐nucleus suspensions were used for droplet generation, emulsion breakage, bead collection, reverse transcription and cDNA amplification to create barcoded libraries. Indexed libraries were then constructed according to the manufacturer's protocol. Concentrations were measured using a Qubit ssDNA Assay Kit (Thermo Fisher Scientific, Q10212). The libraries were quantified using a Qubit single‐strand DNA Assay Kit (Invitrogen) and sequenced with paired‐end reads on the DNBSEQ platform at China National GeneBank (Shenzhen, China), employing the following sequencing strategy: 41‐bp read length for read 1 and 100‐bp read length for read 2.
Quantification and statistical analysis
Raw data processing
Raw sequencing reads from DNBSEQ‐T10 were filtered and demultiplexed using PISA (Version 1.1). Before alignment, Drop‐seq (Version 1.13) was used to trim poly(A) stretches, append cell and UMI barcodes to the reads and filter out reads with low‐quality barcode bases. Reads were aligned to the GRCm38 genome using STAR (Version 2.7.4a) and sorted with Sambamba (Version 0.7.0). For snRNA‐seq data, both exon and intron reads were combined to count each gene's transcripts. Finally, a nucleus versus gene UMI count matrix was generated using PISA.
Doublet removal
Doublet removal for each library was performed using Scrublet (Version 0.2.3). Scrublet detects doublets by simulating artificial doublets within the dataset and constructing a nearest‐neighbour classifier without requiring prior knowledge or pre‐clustering. Doublet removal was conducted using Scrublet's default parameters, excluding the top 5% of cells that showed the highest similarity to simulated pseudo‐doublets.
Lower‐quality cell removal
We applied the median absolute deviation (MAD) method for each library to dynamically identify and remove outlier cells from the dataset. Specifically, MAD was calculated based on UMI counts, gene numbers, mitochondrial percentage and ribosomal percentage, using a threshold of five times the MAD to determine which cells to retain. Additionally, we imposed basic quality control criteria, requiring UMI counts above 1000, gene numbers exceeding 500 and mitochondrial percentages below 10% to ensure data quality.
Ambient RNA removal
Ambient RNA noise was reduced using the ‘DecontX’ function of celda (Version 1.18.1) with default settings. This function was applied to estimate and remove contamination from single‐cell genomic data.
Integration, clustering and cell‐type annotation
After obtaining the filtered gene UMI count matrices, we employed a semi‐supervised pipeline for preliminary annotation. We utilised Scanpy (Version 1.10.2) to merge all sample data and subset the matrix of highly variable genes. Next, we employed scVI, a deep‐learning approach, to remove batch effects from the data. Based on the batch‐corrected dimensionality reduction matrix, we performed clustering using the Leiden algorithm, setting the resolution parameter to 5. Shi et al.’s datasets were used to construct our model and annotate our datasets accordingly, followed by manual curation of the annotations. All literature cited in Table has been independently verified and can be traced via the provided PMID.
Cell cluster proportion analysis
Cell cluster proportions among different groups were compared using one‐way analysis of variance (ANOVA) followed by Dunnett's multiple comparison test. All statistical analyses were performed in R using the speckle package (version 1.0.0).
Consensus co‐expression network analysis
To investigate global transcriptional changes at the gene expression level, we applied the hdWGCNA (Version 0.3.03) to perform Consensus Co‐expression Network Analysis, following tutorials from https://smorabit.github.io/hdWGCNA/index.html. Additionally, DME analysis was performed to identify modules that are up‐regulated or down‐regulated after ICH.
Neighbourhood‐based differential abundance analysis
We used the milopy (Version 0.1.0) framework to detect dynamic changes in fine‐grained cell populations over time. Neighbourhoods were constructed from the KNN graph and DA testing was performed using a linear model. To capture temporal progression, time points (D1, D3, D7, D14, D28) were ordinally encoded as integers and used as a continuous variable in the design formula.
Subtype analysis, DGE analysis and Gene Ontology (GO) analysis
To further investigate the biological functions of the cell types of interest, we conducted a subgroup analysis using Scanpy for transcriptomic data for the AC, MGL, MAC, T, OLG and OPC, respectively. Based on the number of cells in each group and their biological functions, we performed multiple rounds of testing to determine an appropriate resolution. Subsequently, differential expression analysis was performed using the Libra R package (version 1.7), which implements a pseudo‐bulk approach to control for biological replicates in single‐cell RNA‐seq data. Raw gene counts were first aggregated across cells from the same cluster within each sample to construct a pseudo‐bulk expression matrix. Comparisons were then made between defined cell subtypes and their respective reference populations using the run_de() function in Libra, which internally utilises DESeq2 for modelling and statistical testing. Genes with a p value <.05 and absolute log2 fold change >1 were considered significantly differentially expressed. Additionally, we conducted enrichment analysis for up‐regulated genes across multiple groups using the ‘compareCluster’ function from the clusterProfiler (Version 4.12.6).
Cell–cell communication
We utilised the CellChat package (Version 1.5.0) to analyse communication networks within the neuroinflammation subtypes. We employed the entire cellChatDB.mouse as the receptor–ligand database to identify potential interactions. We used the key function ‘computeCommunProb’ with the ‘truncatedMean’ method to calculate the average gene expression for each cell subtype during the critical interaction analysis. We followed tutorials from https://github.com/sqjin/CellChat and built the cell–cell communication network for each time point, identifying the pathways using the ‘netP’ data slot. Finally, we focused on the signalling pathways related to neuroinflammation.
Multi‐scoring analysis
To describe the continuous spectral variation between different cell types, we conducted various scoring analyses on our essential data. First, we analysed the inflammatory response levels of immune cells using the ‘score_genes’ function in Scanpy, based on the genes associated with the inflammatory response (GO:0006954), with the Naive group serving as a calibration baseline.
Additionally, we assessed the differentiation potential of oligodendrocyte lineages and scored the cell cycle phases. For differentiation potential, we utilised the CytoTRACE2 (Version 1.0.0) with default parameters to process the raw count matrix, yielding differentiation potential scores for each cell. The cell cycle scoring was performed using the ‘score_genes_cell_cycle’ function in Scanpy to assess S phase and G2M phase scores.
Finally, we evaluated the functional scoring of T cells by employing the TCellSI (Version 0.1.0). This scoring was based on the standardised expression matrix using the ‘TCSS_scRNAseqCalculate’ function. This process utilised HomoloGene (Version 1.4.68.19.3.27) to convert homologous genes between humans and mice.
RNA velocity and vector field analyses
We reconstructed the expression dynamics vector fields using dynamo (Version 1.3.2), following guidelines provided by the official tutorials at https://dynamo‐release.readthedocs.io/en/latest/. Briefly, we used the spliced and unspliced counts matrix from the snRNA‐seq data of the oligodendrocyte lineage. RNA velocity, acceleration and curvature were then calculated for demyelination and remyelination. The effects of RNA velocity on remyelination following the perturbation of specific genes were simulated by setting the expression levels of these genes to 25.
Transition velocity quantification across the oligodendrocyte lineage
To quantify transition velocities along the oligodendrocyte lineage, we used the Pearson transition matrix. For each cell at a given state, we computed the summed transition weights towards cells in the subsequent state (e.g., from OPC to NFOL), using a minimum weight threshold to filter spurious connections. This score represents the potential for transition towards the target state. We repeated this procedure across defined lineage stages (OPC → NFOL → MFOL → Moll → DOL) and across developmental time points. Transition velocity scores were aggregated and plotted by time point and lineage transition.
NicheNet analysis
To explore T cells driving myelination, we utilised Nichenetr (Version 1.0.0) to identify potential ligands mediating interactions between T cells and both NFOL and MFOL. First, we designated NFOL and MFOL as the receivers and identified genes expressed in at least 5% of the cells, focusing on DEGs compared both Day 7 and Day 14 with the Naive stage. We then predicted active ligands based on background genes, inferred corresponding receptors and target genes and identified the top‐ranked ligands. Finally, we pinpointed key target genes for subsequent perturbation analysis.
Immunofluorescence
Multiplex immunofluorescence staining was performed on paraffin‐embedded brain sections using a tyramide signal amplification (TSA)‐based kit (#AFIHC027, AiFang Biological Inc.). Following deparaffinisation and antigen retrieval in citrate buffer (pH 6.0), sections were incubated with primary antibodies overnight at 4°C. After washing, HRP‐conjugated secondary antibodies were applied, followed by TSA fluorophore amplification. Between each staining round, antibody stripping was performed using elution buffer (#G1266, AiFang), allowing sequential detection of multiple targets. Nuclei were counterstained with DAPI, and slides were mounted with antifade mounting medium.
The following panels were used: MIF signalling pathway (Figure): A five‐colour panel including CD44 (1:1500, # AF14754, AiFang Biological Inc.), CD74(1:1500, #AF06806, AiFang Biological Inc.), MIF (1:1500, #AF05273, AiFang Biological Inc.), TMEM119 (1:1500, ab209085, abcam) and F4/80 (1:1500, #SAD002, AiFang Biological Inc.).MIF signalling pathway (Figure): A three‐colour panel targeting MIF (1:1000, BS‐1044R, Bioss), CD74 (1:8000, GB115427, Servicebio) and CD44 (1:2000, GB112054, Servicebio).T cell infiltration (Figure): CD3 (1:2000, #GB12014, Servicebio).T cell–OLG crosstalk (Figure): A dual‐labelling panel targeting CD3 (1:2000, #GB12014, Servicebio) and MBP (1:5000, #GB12226, Servicebio).
Images were acquired using a multispectral fluorescence scanner equipped with appropriate filters.
Immunofluorescence analysis
The red channel corresponding to myelin staining was extracted and converted into a binary mask by applying an appropriate threshold. The resulting mask was refined by morphological operations (opening, closing and hole filling), and regions of interest were further corrected manually in the ROI Manager. Quantification was performed with the ‘Analyse Particles’ function, with a minimum size threshold of 100 µm2 to exclude small noise signals. The total myelin‐positive area and its fraction relative to the image field were recorded. T cell numbers were counted manually in the same software.
AUTHOR CONTRIBUTIONS
Jian Wang, Mingyue Wang and Junmin Wang conceived the study. Mingyue Wang, Qinfeng Peng, Shaoshuai Wang, Nannan Cheng, Mengke Zhao, Jiaxin Li and Qinglin Wang performed experiments. Zhan Chen, Qinglin Wang, Rong Xiang and Jin Tao analysed and interpreted the data. Zhan Chen, Junmin Wang, Jian Wang and Mingyue Wang drafted and revised the manuscript. Ruoqi Ding performed the verification experiments. Longqi Liu, Chuanyu Liu and Xuemei Chen provided advice. All the authors reviewed and approved the manuscript.
CONFLICT OF INTEREST STATEMENT
The authors declare no conflicts of interest.
ETHICS STATEMENT
All animal experiments were approved by the Institutional Animal Care and Use Committees of Zhengzhou University (approval No. ZZUIRB 2022‐148) and BGI Research Institute (approval No. BGI‐IRB E23025), and were performed in accordance with national guidelines for the care and use of laboratory animals and the ARRIVE guidelines.
Supporting information
Acknowledgements
ACKNOWLEDGEMENTSThis work was supported by the China National GeneBank (CNGB). We thank the National Supercomputing Center in Zhengzhou for its computational resources. Mingyue Wang is supported by grants from the National Key R&D Program of China (2022YFC3400300). Longqi Liu is supported by the Shenzhen Basic Research Project for Excellent Young Scholars (RCYX20200714114644191) and the National Key R&D Program of China (2022YFC3400405). Jian Wang is supported by the National Natural Science Foundation of China (82371339).