Can strict clock fit better than uncorrelated relaxed clock?

[happy little beast]

Dear community,

I got some unexpected output when testing clock models. I ran the same dataset using strict clock or uncorrelated lognormal relaxed clock. I also did MLE using PS/SS following MCMC. Sampling was sufficient. I guess that UCLD is always more realistic than strict clock especially for deep trees, and wanted to see how much better it is. Here H1 is UCLD, and H0 is strict. My (M1-M0) is some -300. I think it means that strict clock fits better than UCLD, which is unexpected. Meanwhile I also tested some other loci and their M1-M0 are quite positive as expected.

My questions are:

1) Are my analysis and intepretation correct? i.e., does it mean that strict clock is better than UCLD?

2) Can strict clock possibly fit better than UCLD on real datasets? What's the biological implication underneath?

Thank you very much!

[Artem B]

Hello!

Yes it can. I advice you to read BEAST 2 book http://www.beast2.org/book/ where authors comprehensively consider the question of clock choosing. You also can check tutorials on BEAST 1 and BEAST 2 websites, there is a lot of information about the clock comparison out there.

You can always verify the coefficient of variation (CV) in Tracer employing relaxed clock. If CV is about 0.1 it indicates that strict clock fit better (there is no significant rate heterogeneity among branches).
Biological implication is that you study closely related organisms. In other words, why your substitution rate should be different from my if we are both sapiens sapiens? :)
Sometimes when the specific lineage has very different rate it can indicate bad sequencing or dating. 
The similar intuitive mistake is about GTR model as the most fit model (since it compute everything that can be computed)

Regards,
Artem

[happy little beast]

Hello Artem,

Thank you very much for answering my question! I quickly read the relevant parts of the book you recommended. Very helpful!

I checked the C.V. It is around 0.15 to 0.2. What does this number suggest? (weak/moderate/strong violation to clock?)

My taxon sampling was very diverse, much higher than intraspecies level. I used hundreds of loci. Maybe the average of a large number of loci smoothed out flucturations in clock rates, making them together behave like one clock? Is this plausible?

Best,

[Artem B]

Unfortunately I can not answer you since I work with haploid viral data only and it has its own specifics regarding evolution rate within taxa.

My taxon sampling was very diverse, much higher than intraspecies level. I used hundreds of loci. Maybe the average of a large number of loci smoothed out flucturations in clock rates, making them together behave like one clock? Is this plausible?

It suggests that rate of evolution of different lineages deviates from mean clock rate value by 15-20%.As practice shows it's 'border' CV values and you are right with an implementation of PS/SS here. If others priors are correct, then a strict clock model is really correct choice.

 I checked the C.V. It is around 0.15 to 0.2. What does this number suggest?

If you study interspecific taxa then you may use StarBeast2 (addition BEAST 2 package) since coalescent models used by BEAST (e.g. constant population size, Bayesian Skyline etc.), assume that considering population is well mixed (has no significant structure), otherwise the inferred results can be/will be biased. 
Analyzing a big set of different taxa is really challenge. Here is the brilliant tutorial about species trees with relaxed clock: https://taming-the-beast.org/tutorials/species-tree-clocks/

Regards,
Artem

[happy little beast]

Dear Artem,

I appreciate your answers to my questions. Certainly very helpful and inspiring!

Best regards,