PMBCClg=clustergram(PMBC_perc_filtered(I(1:2000),:),'Standardize',0,'DisplayRange',7.5,'Colormap',colormap(jet),'RowPDist','correlation','ColumnPDist','spearman','Symmetric',false,'linkage','complete', 'OptimalLeafOrder',false)
Now you get a clustergram which can be annotated interactively following Matlab Clustergram Manual to make it look like the following
And save (by right clicking) individual clusters into workspace as Cluster1, Cluster2 etc.
If you want to see expression levels of a set of genes, you can put the names of them in text.dat
And then import this list to an array
GenesOfInterest=table2array(readtable('test.dat','ReadVariableNames',false,'Delimiter','\t'))
Then plot a HeatMap matching the cell order of your clustergram
HeatMap(PMBC_perc(GenesOfInterest,get(PMBCClg,'ColumnLabels')),'Standardize',2,'Colormap',colormap(jet),'Symmetric',true,'DisplayRange',2.5)
If you only know the gene name (e.g. HOPX) but not the gene ID, you can look it up using this:
get(PMBCdm(strcontain('HOPX',get(PMBCdm,'RowNames')),:),'RowNames')
It will show
Before we do this, let's first drop down our manually assigned cell clusters from clustering analysis. Based on the coloring in the example image, the clusters are:
ClusterID=repmat([0],size(get(PMBCClg,'ColumnLabels')));
ClusterID(GetOrder(get(Cluster1,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=1;
ClusterID(GetOrder(get(Cluster2,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=2;
ClusterID(GetOrder(get(Cluster3,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=3;
ClusterID(GetOrder(get(Cluster4,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=4;
ClusterID(GetOrder(get(Cluster5,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=5;
ClusterID(GetOrder(get(Cluster6,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=6;
ClusterID(GetOrder(get(Cluster7,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=7;
ClusterID(GetOrder(get(Cluster8,'ColumnLabels'),get(PMBCClg,'ColumnLabels')))=8;
Now inspect top 30 PCs for the normalized dataset (you can use normalized values, if you use the function zscore()).
[COEFF,SCORE,latent]=PCAPlot(double(PMBC_perc_filtered(I(1:2000),get(PMBCClg,'ColumnLabels'))),get(PMBCClg,'ColumnLabels'),get(PMBC_perc_filtered(I(1:2000),:),'RowNames'),UMICount,100,30);
Among the outputs you get an image of PC scores, where each row represents each PC's scores across individual cells, ordered in the same way as the clustergram you generated.
Based on this diagram, we see that only the top 10 PCs show cell-type specificity. So we only use these 10 PCs to plot t-SNE.
Now calculate t-SNE coordinates and plot
mappedX=tsne(double(PMBC_perc_filtered(I(1:2000),get(PMBCClg,'ColumnLabels')))','Algorithm','barneshut','NumPCAComponents',10,'Distance','spearman','Perplexity',30);
scatter(mappedX(:,1),mappedX(:,2),50,[0.8 0.8 0.8],'.')
You should right click mappedX in Workspace window to save it to a file. So that next time you can use mappedX=importdata(); to import it
Now paint the t-SNE plot with cluster colors.
hold on
scatter(mappedX(ClusterID==1,1),mappedX(ClusterID==1,2),50,[0 0.45 0.74],'.')
scatter(mappedX(ClusterID==2,1),mappedX(ClusterID==2,2),50,[0.85 0.33 0.1],'.')
scatter(mappedX(ClusterID==3,1),mappedX(ClusterID==3,2),50,[0.93 0.69 0.13],'.')
scatter(mappedX(ClusterID==4,1),mappedX(ClusterID==4,2),50,[0.49 0.18 0.56],'.')
scatter(mappedX(ClusterID==5,1),mappedX(ClusterID==5,2),50,[0.47 0.67 0.19],'.')
scatter(mappedX(ClusterID==6,1),mappedX(ClusterID==6,2),50,[0.3 0.75 0.93],'.')
scatter(mappedX(ClusterID==7,1),mappedX(ClusterID==7,2),50,[0.64 0.08 0.18],'.')
scatter(mappedX(ClusterID==8,1),mappedX(ClusterID==8,2),50,[1 0 0],'.')
Of course, you can do an unsupervised clustering using the top 10 PCs I mentioned above. kmeans_colors=distinguishable_colors(10); scatter(mappedX(:,1),mappedX(:,2),50,kmeans_colors(kmeans(SCORE(:,1:10),10),:),'.');
figure
scatter(mappedX(:,1),mappedX(:,2),50,PMBC_perc('ENSG00000105369_CD79A',get(PMBCClg,'ColumnLabels')),'.')
colormap(hot)
