Single cell transcriptome analysis of human embryonic stem cell-derived neurons spanning the rostrocaudal and dorsoventral axes

Our inability to derive the neuronal diversity that comprises the posterior central nervous system (pCNS) using human pluripotent stem cells (hPSCs) poses an impediment to understanding human neurodevelopment and disease in the hindbrain and spinal cord. We established a modular, monolayer differentiation paradigm that recapitulates both rostrocaudal (R/C) and dorsoventral (D/V) patterning, enabling derivation of diverse pCNS neurons with discrete regional specificity. Expansive single-cell RNA-sequencing (scRNAseq) analysis coupled with a novel computational pipeline allowed us to detect hundreds of transcriptional markers within region-specific phenotypes, enabling discovery of gene expression patterns across R/C and D/V developmental axes.

Processed data matrix: For each of the 6 samples from direct differentiation (GSE186696) and 14 samples from modular differentiation (GSE186697), We merged the gene expression matrices from each sample into a single matrix while taking the union of the genes from each matrix. The combined matrix is [12,543 cells x 20,598 genes] for the direct differentiation dataset and [49,959 cells x 23,941 genes] for the multiple generation dataset. We transformed the values of these matrices by taking their square root and standardizing each cell’s expression profile by dividing by the mean expression of a gene in each cell for the subsequent clustering analysis.

Clusters - HOX profile clusters: We applied Louvain clustering (k=13) for the visualization of the simultaneous expression of HOX genes in our data set (GSE186697), which revealed inter- and intra-sample HOX profile heterogeneity.

Clusters - primary clusters: We applied sparse non-negative matrix factorization (NMF) (Kim and Park, 2008) based clustering to define cardinal cell groups within our data set (GSE186697) in an unbiased manner and identified 25 primary clusters.

Clusters - subpopulation subclusters: We organized related primary clusters into 17 different groups, and then developed and applied a consensus clustering based approach with the goal of defining robust subclusters representing subtypes of known cardinal populations.

Consensus_graph_matrix: We regrouped our 25 primary cell clusters into 17 subgroups based on similarity of the cell types assigned to each cluster, and created a consensus graph of cell co-clustering relationship per subgroup. For every pair of cells in a subgroup, we counted the proportion of times the two cells were in the same cluster (across multiple clustering approaches), and generated a weighed graph of cells with weights corresponding to this proportion.

我们无法利用人源多能干细胞（hPSCs）推导出构成后脑和脊髓的后中枢神经系统（pCNS）中的神经元多样性，这限制了我们对人类神经发育和疾病的理解。我们建立了一种模块化、单层分化范式，该范式能够重现从头至尾（R/C）和背腹（D/V）的图案，从而允许产生具有离散区域特异性的多种pCNS神经元。通过对单细胞RNA测序（scRNAseq）的广泛分析以及新型计算管道的应用，我们能够在区域特异性表型中检测到数百个转录标记，从而实现了跨越R/C和D/V发育轴的基因表达模式的发现。

处理后的数据矩阵：对于直接分化（GSE186696）的6个样本和模块化分化（GSE186697）的14个样本，我们将每个样本的基因表达矩阵合并为一个单一矩阵，并取每个矩阵中基因的并集。直接分化数据集的合并矩阵为[12,543细胞 x 20,598基因]，多代数据集的合并矩阵为[49,959细胞 x 23,941基因]。我们通过对这些矩阵的值取平方根并进行标准化处理，即将每个细胞的表达谱除以每个细胞中基因的平均表达量，从而对矩阵进行转换。

聚类 - HOX基因表达模式聚类：我们对数据集（GSE186697）中HOX基因的同时表达应用了Louvain聚类（k=13）进行可视化，揭示了样本内和样本间的HOX基因表达模式异质性。

聚类 - 初级聚类：我们应用基于稀疏非负矩阵分解（NMF）的聚类方法（Kim and Park，2008），以无偏的方式定义数据集（GSE186697）中的核心细胞群，并识别出25个初级聚类。

聚类 - 亚群次级聚类：我们将相关的初级聚类组织成17个不同的组，然后开发并应用了一种基于共识聚类的策略，旨在定义代表已知核心人群亚型的稳健次级亚群。

共识图矩阵：我们根据分配给每个聚类的细胞类型相似性，将25个初级细胞聚类重新分组为17个子组，并为每个子组创建了细胞共聚类关系的共识图。对于子组中每对细胞，我们计算了两个细胞在同一聚类（跨多个聚类方法）中的比例，并生成了一个带有相应比例权重的细胞加权图。