Predicting ATC codes at level 3 in python

I am trying to implement the algorithm implemented in this paper on Python. Basically it is a multi labeling algorithm to make out of sample prediction at level 1 of Anatomical Therapeutic Chemical (ATC). Too do that it basically uses as features 3 sets of scores which, being an economist, I don’t know directly. Specifically:

1. interaction score with the drugs in an ATC group;
2. maximum structural similarity score
3. fingerprint score

Now, it is said in the paper that these features can be downloaded from KEGG. I gave a look at the dataset but am actually unable to find them. Hence my questions for you, please, are:

a) Anyone knows a way to retrieve them please?
b) Do you, in general, know algorithms to predict ATC codes at third level of aggregation?

Thank you,

Federico