Figure 3 is basically just a different way of visualizing figure 2C I think. Pseudotime is a concept from single-cell RNA-seq and would not apply in your case as you know what the timepoints are, so you can order the columns with that knowledge rather than using a machine learning approach. Clustering rows is then just hclust to derive clusters. It would be then on you how you call each cluster, e.g. with the labels you use above depending on how drastic the changes are between timepoints.