I just experimented with building n-grams for each anagram candidate and calculating the dissimilarity of pairs within each group, then multiplying this score by the length of the word.

The results look pretty decent, but I get "BASIPARACHROMATIN" vs "MARSIPOBRANCHIATA" as number 1.

https://gist.github.com/aclemmensen/725f60a0746382521faefa0d...