We propose a novel method for the task of protein subfamily identification, that is, finding subgroups of functionally closely related sequences within a protein family. In line with phylogenomic analysis, the method first builds a hierarchical tree using as input a multiple alignment of the protein sequences, then uses a post-pruning procedure to extract clusters from the tree. Differently from existing methods, it constructs the hierarchical tree top-down, rather than bottom-up, and associates particular mutations with each division into subclusters. The motivating hypothesis for this method is that it may yield a better tree topology, with more accurate subfamily identification as a result, and additionally indicates functionally important sites and allows for easy classification of new proteins. A thorough experimental evaluation confirms the hypothesis. The novel method yields more accurate clusters and a better tree topology than the state-of-the-art method SCI-PHY, identifies known functional sites, and identifies mutations that, alone, allow for classifying new sequences with an accuracy approaching that of hidden Markov models.