search

veg / hyphy-analyses

HyPhy standalone analyses
MIT License
37 stars 17 forks source link

profile likelihood? #8

Closed jodyhey closed 4 years ago

jodyhey commented 4 years ago

it would be great if the multi-synonymous code generated confidence intervals for rates/ratios thanks

spond commented 4 years ago

Resolved via c7c05eb2a20110a921e0e8556c01ebb8d27f61ec

jodyhey commented 4 years ago

Profile likelihood (confidence intervals) does not seem to be working in MG94. e.g. --ci Yes does not lead to ci's in the output.

spond commented 4 years ago

Dear @jodyhey,

You mean in the FitMG94.bf file?

Best, Sergei

jodyhey commented 4 years ago

I think so. I'm reporting what Shak told me, that FitMG94.bf was not returning confidence intervals.

spond commented 4 years ago

Dear @jodyhey,

That's because that option only applies to the multi-synonymous model. I haven't added it to the FitMG94.bf model:)

Best, Sergei

jodyhey commented 4 years ago

Guess I misread the update re #8. We are using ci’s as crude way is assessing dn/ds >1.

Sent from my iPhone

On Apr 10, 2020, at 11:02 AM, Sergei Pond notifications@github.com wrote:

 Dear @jodyhey,

That's because that option only applies to the multi-synonymous model. I haven't added it to the FitMG94.bf model:)

Best, Sergei

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.

spond commented 4 years ago

Dear @jodyhey,

Do you want me to add a proper LRT for this?

Best, Sergei

jodyhey commented 4 years ago

The MSS model on MG94 simulated data turned up as we hoped, i.e. the MSS estimated dn/ds values are similar to the omega value put into the simulation. This supports the idea that MSS dn/ds values can be the basis of an improved tests for selection under a conventional interpretation, and given the low correlation between MSS estimated dn/ds values and MG94 estimated dn/ds values, there is probably a pretty good story to tell about how MG94 dn rates are way off.
fwiw we saw 4 genes with confidence interval lower bound > 1, in the 7aa MSS model. In MG94 all of these genes had estimated dn/ds << 1 So a short answer would be 'yes', it would be nice to be able to do LLR tests.

spond commented 4 years ago

Dear @jodyhey,

Sure, I'll add LLR; do you want them for each ratio, or just dN/dS?

Best, Sergei

jodyhey commented 4 years ago

If if does not add a ton of work, then all would probably be useful. The way these things usually go is that you add one thing, which opens some new ways to think about things, and then you want it generalized.