Obtain a sub-optimum tree with LH cutoff?

[oushujun]

Hello,

Thanks for developing and maintaining this fascinating method.

I want to search an ML tree in my population data. I know it will be very difficult to converge into a bifurcating tree, but I only want to learn about large branches currently, and not hoping to spend a large chunk of time to "optimize" tips. I notice the likelihood value increased quickly and keep increasing just a slight bit (but taking a couple hours each round). In this case, can I stop the run and get the current tree? Or is that any way to set a likelihood cutoff to stop the run? I notice the "--spr-cutoff" parameter but not quite understand what it means and how to set a proper value.

Below is the command I used.

raxml-ng --msa Chr1.imputed.fa --model GTR+G --threads 50 --seed 12315

Any suggestions are welcome. Thanks ahead!

Shujun

[amkozlov]

Hi Shujun,

the relevant option for this would be --lh-epsilon which defines the likelihood cutoff to terminate SPR cycles.

--spr-cutoff is a different cutoff which is used in the subtree traversal heuristic, described in section 3.2 of this paper here:

https://sco.h-its.org/exelixis/pubs/VLSI2007.pdf

Hope this helps, Alexey

[oushujun]

Hi Alexey,

Thank you for your clarification. I have problems to understand the concept of LH epsilon. I see the cutoff of this parameter is 0.1, is that mean once the LH difference between this round and the previous round is less than 0.1, the ML search will stop?

To implement this parameter, can I just add it to my previous command and run it in the same folder to harvest the checkpoints that raxml-ng stored and run without this parameter?

Thanks, Shujun

[stamatak]

Dear Shujun,

Could we take this discussion to the RAxML google group?

Like this it's likely that more users might also answer your question.

Alexis

On 08.05.2018 20:23, Shujun Ou wrote:

Hi Alexey,

Thank you for your clarification. I have problems to understand the concept of LH epsilon. I see the cutoff of this parameter is 0.1, is that mean once the LH difference between this round and the previous round is less than 0.1, the ML search will stop?

To implement this parameter, can I just add it to my previous command and run it in the same folder to harvest the checkpoints that raxml-ng stored and run without this parameter?

Thanks, Shujun

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/amkozlov/raxml-ng/issues/44#issuecomment-387478016, or mute the thread https://github.com/notifications/unsubscribe-auth/AA1w-gCSWJfrC-DZuyaQ-jd09N65YA3Hks5twdR6gaJpZM4T1yD7.

-- Alexandros (Alexis) Stamatakis

Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology

www.exelixis-lab.org

[oushujun]

Dear Alexis,

Sorry I didn't realize both programs use the same support site. I will move forward to the RAxML Google group.

Thanks for all these wonderful programs!

Shujun

On Tue, May 8, 2018, 2:16 PM Alexis Stamatakis notifications@github.com wrote:

Dear Shujun,

Could we take this discussion to the RAxML google group?

Like this it's likely that more users might also answer your question.

Alexis

On 08.05.2018 20:23, Shujun Ou wrote:

Hi Alexey,

Thank you for your clarification. I have problems to understand the concept of LH epsilon. I see the cutoff of this parameter is 0.1, is that mean once the LH difference between this round and the previous round is less than 0.1, the ML search will stop?

To implement this parameter, can I just add it to my previous command and run it in the same folder to harvest the checkpoints that raxml-ng stored and run without this parameter?

Thanks, Shujun

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/amkozlov/raxml-ng/issues/44#issuecomment-387478016,

or mute the thread < https://github.com/notifications/unsubscribe-auth/AA1w-gCSWJfrC-DZuyaQ-jd09N65YA3Hks5twdR6gaJpZM4T1yD7 .

-- Alexandros (Alexis) Stamatakis

Research Group Leader, Heidelberg Institute for Theoretical Studies Full Professor, Dept. of Informatics, Karlsruhe Institute of Technology

www.exelixis-lab.org

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/amkozlov/raxml-ng/issues/44#issuecomment-387494145, or mute the thread https://github.com/notifications/unsubscribe-auth/AFt-NHPPpj4-I0h0TrMYN8NdwswonNsuks5tweDfgaJpZM4T1yD7 .

[amkozlov]

Thank you for your clarification. I have problems to understand the concept of LH epsilon. I see the cutoff of this parameter is 0.1, is that mean once the LH difference between this round and the previous round is less than 0.1, the ML search will stop?

that's exactly how it works. Please note, that lh-epsilon applies to model and branch length optimization as well. So you might want to run --evaluate command on a resulting topology with a lower lh-epsilon to refine model/brlen estimates.

To implement this parameter, can I just add it to my previous command and run it in the same folder to harvest the checkpoints that raxml-ng stored and run without this parameter?

yes, this should work.

Best, Alexey