First of all, this is a really cool finding for any of you who are into longevity research.
Secondly, this is a really cool example of a project that incorporates some machine learning, but also some hand-derived rules using domain expertise.
Basically, the researchers took a set of stem cells, then studied the properties of those cells. At each stage, they grouped the cells based on whether or not those cells would be capable of regenerating the flatworm. Finally, they ended up with the relevant cell.
(You might think of that as a decision tree)
At one stage, they use t-SNE to visualize clusters of different types of cells: