User:R. Eric Collins/MBL/ML

Maximum Likelihood

Paul Lewis

Derrick Zwickl

Large ML inference

• GARLI
• branch length optimization makes up majority of ML inference time because changing one branch affects lengths of every other branch.
  • this problem gets untenable with large trees (e.g. >50 taxa)
• how accurate does a tree likelihood need to be?
• evolutionary (not technically genetic) algorithm
  • start with a large number of individuals (parameter sets)
  • compete them against each other, unfit ones die out (and some fit ones, but never best fit)
  • can give starting tree that you believe, even if it has polytomies

• difficulty due to parameter estimation: nucleotide < amino acid < codon (because of transition probability estimation)

• when to use codons
  1. mixture of divergent and closely related sequences
  • not always better, plus it takes longer
  1. when the models get better

Questions:

1. which branches take longest to calculate? leaves? deep branches?
2. settings contstraints... are there any algorithms that 'fill in' the valleys (level the mountains)
3. when do you need a tree this good? use cases for depth-of-tree'ing would be useful
4. ancestral state reconstruction... indel/gap models