GoKawiilHuman tissue
Post-mortem human tissue samples
Eyeballs from individuals diagnosed with MS were obtained from the Netherlands Brain Bank, and those from control donors without recognizable neuropathological changes were obtained from Johns Hopkins University. Detailed information about donors used for snRNA-seq, immunohistochemistry, qPCR and smFiSH are provided in Supplementary Table 1. The macula region of these donors was used for snRNA-seq while smFISH was performed on tissue adjacent to the macula. qPCR and immunohistochemistry were conducted with the peripheral retina. Paraffin-embedded cortex tissue for histopathology was obtained from the UK Multiple Sclerosis Tissue Bank at Imperial College London. Samples were classified as chronic active lesions or normal appearing grey matter according to the histopathological assessment provided by the UK Multiple Sclerosis Tissue Bank’s histopathology reports. Although no statistical methods were used to predetermine sample size, our sample size is comparable to those reported in previous studies12,13,60.
Nucleus isolation and library preparation
Eyeballs were enucleated from deceased healthy controls and people with MS. Only tissue from patients with a post-mortem interval ≤24 h was used. To isolate the macula region, fresh-frozen eyeballs were positioned in a CM3050 S cryostat (Leica Microsystems) at –20 °C, with the lens facing downward and the optic disc oriented toward the examiner. The macula was identified as a distinct yellow spot on the retina, and the tissue was stored at −80 °C until use. All of the subsequent steps were performed on ice according to a previously published protocol with minor adaptations. In brief, on each experimental day, the frozen tissue was transferred to ice-cold NP40 lysis buffer (0.1% NP-40 (Thermo Fisher Scientific, 85124), 10 mM Tris pH 8.0, 1 mM CaCl 2 , 8 mM MgCl 2 , 15 mM NaCl, 0.02 U µl–1 DNase I (Merck Millipore, D4527)). The retina was transferred to a Dounce homogenizer in 1 ml lysis buffer supplemented with 0.2 U µl−1 Ribolock RNase Inhibitor (Thermo Fisher Scientific, EO0382) and homogenized 20 times with both loose and tight pestles. The homogenate was passed through a 100 µm cell strainer and centrifuged at 500g for 5 min. All buffers, except for the washing buffer, were supplemented with 0.16 U µl−1 Ribolock RNase inhibitor. Pelleted nuclei were resuspended in staining buffer (Tris base buffer: 10 mM Tris pH 8.0, 1 mM CaCl 2 , 8 mM MgCl 2 , 15 mM NaCl, 1 U ml−1 DNase I) containing 0.02% Tween-20 and 2% BSA, with primary antibodies against NeuN (1:250, Merck Millipore, ABN91) and RBPMS (1:250, Abcam, ab194213) for 15 min at 4 °C. Nuclei were washed, centrifuged at 500g for 5 min and stained with secondary antibodies (anti-chicken 647, 1:500; Jackson ImmunoResearch, 703-605-155; and anti-rabbit PE, 1:200, BioLegend, Poly4064) for 15 min at 4 °C. After another wash step, nuclei were filtered through a 70 µm strainer, resuspended in sorting buffer (Tris base buffer with 2% BSA), and Hoechst (1:2,000) was added to visualize nuclei. NeuN+RBPMS+ nuclei were sorted using the BD FACS Aria III device running BD FACSDiva v.9.0.1 into Ames medium (Sigma-Aldrich, A1420) with 1.5% BSA. Sorted NeuN+RBPMS+ nuclei were pelleted at 500g for 5 min at 4 °C, resuspended in approximately 20 µl of 1% BSA in PBS, visually inspected, counted in a Neubauer chamber and adjusted to a concentration of around 1,000 nuclei per µl. Nuclei were then loaded onto the 10x Chromium Single Cell Chip G (10x Genomics) with a targeted recovery of about 12,000 nuclei per channel. Libraries were generated according to the manufacturer’s protocol using the Chromium Single Cell 3′ Reagent Kit version 3.1 (dual index), measured using the Agilent Bioanalyzer (TapeStation 4150) and sequenced on the Illumina NovaSeq 6000 platform (paired-end), aiming for a sequencing depth of around 30,000 reads per nucleus.
Data preprocessing and quality control
Count matrices for each sample were generated by aligning each library to the human reference mRNA transcriptome GRCh38-2020-A using Cell Ranger (v.7.0.1), including both exonic and intronic reads61. The Cell Ranger output was processed using CellBender v.0.3.0 with the default settings (epochs = 150, fpr = 0.01, learning rate = 10−4) to remove ambient RNA and other background noise62. Downstream analysis was performed with Seurat (v.5)63 in R Studio (R v.4.4.1). For each RNA count matrix, the following steps were carried out: cell counts were normalized to a total library size of 10,000 and log transformed (Seurat, NormalizeData). The top 2,000 variable features were identified (Seurat, FindVariableFeatures), followed by data scaling (Seurat, ScaleData) and dimensionality reduction (Seurat, RunPCA, npcs = 50). RPCA integration was performed to integrate the principal components (Seurat, IntegrateLayers) and used as input for k nearest-neighbour graph construction (Seurat, FindNeighbors, dims = 50) and Leiden clustering with a resolution of 2.5 (Seurat, FindClusters), initially deliberately overclustering the dataset. An initial dataset of 351,737 nuclei was subjected to quality control. Potential doublets were identified using the scDblFinder64 package. Given the defined cluster structure in our dataset, we used a cluster-based approach for doublet identification, estimating the standard 10x doublet rate of 0.8% per 1,000 nuclei. To remove non-RGC nuclei, we filtered the dataset for clusters with high expression of RBPMS, the main RGC marker gene. After subsampling, nuclei with abnormally high (greater than mean + 3 s.d.) gene counts, fewer than 2,300 gene counts, a doublet score of >0.5, as well as mitochondrial counts of >5% were removed.
Clustering and cell type annotation
After removing low-quality and non-RGC nuclei from the dataset, we repeated the normalization and clustering pipeline at a resolution of 0.5 (RunPCA, npcs = 30; FindNeighbors, dims = 30). For each cluster, we calculated differentially expressed marker genes compared with every other cluster (Seurat, FindMarkers). Clusters were merged if ≤5 differentially expressed genes were found between them, with an average log 2 -transformed fold change of >2 or <–2 and a P value < 0.05. Moreover, if a cluster contained fewer than 200 nuclei in the control condition or was absent in two or more samples, it was fused with its nearest neighbour based on a Euclidean distance matrix constructed in gene expression space (Seurat, BuildClusterTree; Supplementary Fig. 1e). The remaining 27 clusters were annotated manually based on their respective marker gene expression, using three published healthy human RGC datasets as reference8,9,10. To avoid disease-related transcriptional changes influencing cell type annotation, marker genes were identified exclusively from healthy control samples using Seurat’s FindAllMarkers function. Of the top 30 markers of each cluster, 2–3 are shown in Extended Data Fig. 1e. For each reference dataset, the top 50 marker genes per cell type were selected and used for gene set enrichment analysis (GSEA) through the ClusterProfiler65 package, applying the ranked cluster markers from the healthy control samples of our dataset. For the enrichment analysis in Extended Data Fig. 1c and the cell type marker identification in Extended Data Fig. 1e, the three midget OFF RGC subtypes (MG-OFF1, MG-OFF2, MG-OFF3) and the four midget ON RGC subtypes (MG-ON1, MG-ON2, MG-ON3, MG-ON4) were combined into single midget OFF and midget ON types. This consolidation streamlined marker gene identification and cell type annotation. The merged midget ON and midget OFF RGC clusters were also used for the UMAP representation in Fig. 1c.
Differential gene expression analysis
... continue reading