nickbrazeau / hmmertime Goto Github PK

Appears to be tied to genotype with very little genotyping error? Possibly due to lack of het calls? See MCMC.cpp for comments and debugging

Bob's comments to revisit
◦ I've gone through the model assumptions again and am pretty happy with what we've got. My main reservation is how realistic our assumptions about all strains having somewhat equal IBD probabilities really is. It seems likely to me that in real data we might have:
▪︎ The majority of samples completely independent
▪︎ Some samples that have IBD between two strains with recent ancestry
▪︎ Some more samples that have more distance IBD, meaning faster switching and a lower overall relatedness
▪︎ A handful of more complex samples that have a combination of the above. i.e. two strains recent and another that is more distant
◦ This is not really what we're assuming right now, as currently if two samples are related then it increases the chance that more samples are related, which I think is not quite right. I think I might have a play with whether we can infer separate k and f between different strains. This seems a more reasonable assumption to me.
◦ I think what we have in emission probabilities is all correct, as is the method of getting transition matrix from Eigenvalues, as is the forward-algorithm method of integrating over space, as is updating the COI. So essentially all the main workings of the method I still think are good
◦ There is a chance that we may need some more MCMC mixing machinery, especially if I get the multiple-ancestry thing above going, as there are likely to be multiple plausible solutions in some cases. I have this pretty well defined now from drjacoby/Tapestry, so can copy over.

bimodal -- one time short, one time long?

option for user input of PLAF