Chains wouldn't converge in MrBayes 3.2.7

[gocevia]

Hello, maybe we completely misunderstand something, but we are dealing with the problem at the moment that I would define as a lack of convergence in 3.2.7 version compared to 3.2.6. When we run an mcmc analysis of a nucleotide alignment in 3.2.7, the average standard deviation of the split frequencies gets stuck around certain value quickly and fails to decrease in spite of 20 mln generations. This is observed at either single-processor or mpi version. When we run the same alignment with the same parameters on 3.2.6 on a single processor, we do get a convergence. However, when we tried to install a parallel 3.2.6 version on a cluster, it turned out to be 3.2.7. We took a source code for MrBayes 3.2.6 from here: https://github.com/NBISweden/MrBayes/releases/tag/v3.2.6, but indeed after compilation, it turned out to be 3.2.7 and behaved as described above (i.e. showing no convergence).

Therefore I would be grateful for two comments actually: (1) what may be wrong with 3.2.7 compared to 3.2.6 that the first doesn't converge at all, while the second does? (2) is it possible to get a correct source code for the 3.2.6 version to get it compiled and try as mpi?

Thank you in advance!

[zhangchicool]

I got v3.2.6 compiling the files in mrbayes-3.2.6.tar.gz...
The convergence issue needs more investigation, but we cannot reproduce the behavior without an input data!

[gocevia]

Hello! Thank you for reply. In fact, as this was really long ago, I'll need some time to get back to the data. We finally stopped at recognizing the fact that there was no convergence only when "covarion=yes" option was used. When we turned it off, the problem vanished. The application of a covarion model didn't turn out to be so necessary in this case, so we just stopped using it altogether. But I'd be happy if you could help me investigate it further. What kind of data exactly do you need here? Best regards, Alexander Kudryavtsev пт, 21 апр. 2023 г. в 06:34, Chi Zhang ***@***.***>:

…

I got v3.2.6 compiling the files in mrbayes-3.2.6.tar.gz... The convergence issue needs more investigation, but we cannot reproduce the behavior without an input data! — Reply to this email directly, view it on GitHub <#259 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AW6YXN6NB264KHBGE722DB3XCH53FANCNFSM5KJMOYFQ> . You are receiving this because you authored the thread.Message ID: ***@***.***>

-- Dr. Alexander Kudryavtsev Laboratory of Cellular and Molecular Protistology Zoological Institute, Russian Academy of Sciences Universitetskaya nab., 1 199034 St. Petersburg, Russia 007 961 8001974 Email me ***@***.***> Laboratory of Cellular and Molecular Protistology web page <https://www.zin.ru/departments/protistology/index_en.html> My Researchgate <https://www.researchgate.net/profile/Alexander-Kudryavtsev-2>

[tiago-simoes]

I have encountered the exact same problem of convergence whenever I use combined evidence data in 3.2.7, with several different prior options. They work perfectly well with 3.2.6. I'm happy to share my input data sets to help solve this issue. Regards, *Tiago R. Simões, PhD * Postdoctoral Fellow Dept. Organismic & Evolutionary Biology, Museum of Comparative Zoology, Harvard University tiago-simoes.com; Google Scholar <https://scholar.google.ca/citations?user=WhSUGsMAAAAJ&hl=en>

…

On Fri, Apr 21, 2023 at 9:13 AM gocevia ***@***.***> wrote: Hello! Thank you for reply. In fact, as this was really long ago, I'll need some time to get back to the data. We finally stopped at recognizing the fact that there was no convergence only when "covarion=yes" option was used. When we turned it off, the problem vanished. The application of a covarion model didn't turn out to be so necessary in this case, so we just stopped using it altogether. But I'd be happy if you could help me investigate it further. What kind of data exactly do you need here? Best regards, Alexander Kudryavtsev пт, 21 апр. 2023 г. в 06:34, Chi Zhang ***@***.***>: > I got v3.2.6 compiling the files in mrbayes-3.2.6.tar.gz... > The convergence issue needs more investigation, but we cannot reproduce > the behavior without an input data! > > — > Reply to this email directly, view it on GitHub > <#259 (comment) >, > or unsubscribe > < https://github.com/notifications/unsubscribe-auth/AW6YXN6NB264KHBGE722DB3XCH53FANCNFSM5KJMOYFQ > > . > You are receiving this because you authored the thread.Message ID: > ***@***.***> > -- Dr. Alexander Kudryavtsev Laboratory of Cellular and Molecular Protistology Zoological Institute, Russian Academy of Sciences Universitetskaya nab., 1 199034 St. Petersburg, Russia 007 961 8001974 Email me ***@***.***> Laboratory of Cellular and Molecular Protistology web page <https://www.zin.ru/departments/protistology/index_en.html> My Researchgate <https://www.researchgate.net/profile/Alexander-Kudryavtsev-2> — Reply to this email directly, view it on GitHub <#259 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AUH42UC5GAOF273E4ISJ55TXCKBYXANCNFSM5KJMOYFQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[zhangchicool]

Thanks for the feedback. I am not familiar with the covarian model, I think someone else should look into this. It can be really tricky to debug convergence issues, as the program runs without giving an error.
Perhaps I can look into the issue caused by the combined evidence data, if it is related to the fossilized birth-death model. You can send me the data and commands used (zhangchi@ivpp.ac.cn), I will have a look using my spare time (not anytime soon, I guess).