I’m doing a per site ML optimization for an alignment of 485 sequences using JTT model . I observed that for some positions the log likelihood is  -1e+06. What is the meaning of this result?
Thanks in advance


gustavo

Dear Gustavo,

This is indicative of numerical underflow (e.g. the likelihood of some sites falls below MIN_DOUBLE). This 'hack' effectively excludes sites which underflow from the analysis.

You can change a #define (USE_SCALING_TO_FIX_UNDERFLOW to 1) in in calcnode.h and recompile HyPhy. This will turn on a procedure to correct for underflow but slow down likelihood calculations quite a bit - about 3x depending on how many sites underflow - which is why it is turned off by default. Also, currently this fix only works for nucleotide and codon data (but not for a.a.). Let me add the appropriate fix for this case and I'll post a note when this is done.

Cheers,
Sergei