MFGE8 protein, human

All linkage and distance calculations were performed after $Lo{g}_2\left(x+1\right)$ transformation.
The starting point of the dendrogram construction was the 265 clusters. For each gene, we computed average expression, trinarization with f = 0.2, trinarization with f = 0.05 and enrichment score. For each cluster we also know the number of cells, annotations, tissue distribution and samples of origin.
We defined major classes of cell types based on prior knowledge: neurons, astroependymal, oligodendrocytes, vascular (without VLMC), immune cells and neural crest-like. For each class, we defined pan-enriched genes based on the trinarization 5% score. Each class (except neurons) was tested against neurons, to find all the genes where the fraction of clusters with trinarization score = 1 in the class was greater than the fraction of clusters with trinarization score > 0.9 among neurons.
In order to suppress batch effects (mainly due to ambient oligodenderocyte RNA in hindbrain and spinal cord samples), we collected the unique set of genes pan-enriched in the non-neuronal clusters, as well as a set of non-neuronal genes that we believe to have tendency to appear in floating RNA (Trf, Plp1, Mog, Mobp, Mfge8, Mbp, Hbb-bs, H2-DMb2) and a set of immediate early genes (Fos, Jun, Junb, Egr1). These genes were set to zero within the neuronal clusters to avoid any batch effect when clustering the neuronal clusters. We further removed sex specific genes (Xist, Tsix, Eif2s3y, Ddx3y, Uty, and Kdm5d) and immediate early genes Egr1 and Jun from all clusters.
We bounded the number of detected genes in each cluster to the top 5000 genes expressed, followed by scaling the total sum of each cluster profile to 10,000.
Next, we selected genes for linkage analysis: from each cluster select the top N = 28 enriched genes (based on pre-calculated enrichment score), perform initial clustering using linkage (Euclidean distance, Ward in MATLAB), and cut the tree based on distance criterion 50. This clustering aimed to capture the coarse structure of the hierarchy. For each of the resulting clusters, we calculated the enrichment score as the mean over the cluster divided by the total sum and selected the 1.5N top genes. These were added to the previously selected genes.
Finally, we built the dendrogram using linkage (correlation distance and Ward method).

For removing milk fat globules (MFGs), somatic cells, and cell debris, raw milk samples were centrifuged at 2000× g for 20 min at 4 °C in an A508-C rotor (Kubota, Tokyo, Japan) using model 7000 centrifuge (Kubota), as described previously with slight modifications [22 (link),23 (link)]. Defatted milk was pre-warmed at 37 °C for 10 min, acetic acid (AA) was mixed with the milk [milk/AA = 100 (volume)], and the resulting milk was stirred for 5 min at room temperature, followed by centrifugation at 5000× g at 25 °C for 20 min in an R14A rotor (Hitachi Koki, Tokyo, Japan) using Himac CR20GII centrifuge (Hitachi Koki). Casein was pelleted and supernatant (whey) was filtered sequentially through 1.0, 0.45, and 0.2-μm filters (GA-100, C045A047A, and C020A047A, Advantec, Tokyo, Japan). EVs were isolated from 50 mL of whey using two successive ultracentrifugation steps (UC): at 100,000× g for 1 h at 4 °C in a P42A angle rotor (Hitachi Koki), and at 100,000× g for 1 h at 4 °C in a P42ST swing rotor (Hitachi Koki) using Himac CP60E ultracentrifuge (Hitachi Koki). After the first UC, the supernatant was discarded and pellet was suspended in 8 mL of phosphate buffered saline to transfer it to 13PET tube (Hitachi Koki). The second UC was performed next, and supernatant was discarded. Pelleted EVs were stored at −80 °C for further use. Isolation of EVs was confirmed by detecting EV-surface-marker proteins CD9, CD63, CD81, and MFG-E8 by western blot analysis, as described previously [23 (link)].