Here there is no good answer, classically the “bootstrap replicate” is the best option, but other are now widely used and accepted, aLRT statistics and SH-like branch support are good alternatives and faster than classical bootstrap.

We could perfectly estimate all of those parameters, but each estimation increases the number of parameter and the error associated to the final result.

In order to know the best combination of option David Posada, implemented jModelTest and ProtTest that will test all combination and all evolutionary model, and according to Akaike relations or likelihood ratio tests will return the best option for your alignment (see Evolutionary tests section for details and references).

A starting tree is the tree from which PhyML will start to search for best solutions. If “Tree topology optimization” is not checked, the starting tree will also be your final tree, as only rates and branch length will be optimized.

“User-defined tree” can be used in case you want to know the likelihood of a specific tree, you can upload your tree file and un-check “Tree topology optimization” option.