altered average_pairwise_rf_distance to avoid divide-by-zero in the c…

[marybarker]

…ase of a tree shaped DAG, and added option to apply Sankoff only to a subset of DAG nodes

[willdumm]

Looks good! Sorry I didn't think of that divide-by-zero thing.

The sankoff stuff looks great! But I was thinking, what if we wanted to add an extra character to the already-computed sequence (for phylogeography) and only do sankoff on that character? For that it might be useful to be able to leave alone unambiguous bases in the existing node labels, and only modify those that are ambiguous (we could relabel by adding an N to the end of all internal sequences, then add the appropriate region character to the end of each leaf node sequence, then do sankoff). This would be like what use_internal_node_sequences does for ete trees.

Alternatively, it might be useful to just specify which label attribute we do Sankoff on, so by default it could be the sequence attribute, but it could also be the region attribute or something. I guess this would be a keyword argument passed to sankoff_downward.

I'm not suggesting that these things need to make it into this PR but I'm just thinking about how we can reuse this idea for something like adding an extra character for phytogeography...

[marybarker]

Great idea! Yeah, it would be good to do Sankoff on a subset of the sites, leaving the remainder fixed. This will take a bit of extra thought(I've made an issue #55), since sites are not independent and hence it will require either a method to compute transition cost for fully resolved sequences or else a the full-length cost vector expansion, even when only a subset of the sites are being recomputed.