Dear colleagues,
I am investigating the population genetic structure of a forest-dwelling bird species using the STRUCTURE program on 9 STR loci.
As my dataset is unbalanced (11 sampling sites, ranging from 4 samples to 56 samples per site, on average 17 samples per site), I want to use the alternative ancestry prior described by Wang (2017).
My question: when entering the alternative ancestry prior "initial ALPHA=0.33" (as the assumed number of populations is three) and checking "seperate Alpha for each population" in the menu "ancestry model>Advanced>Configure", I am unsure on how to deal with ALPHAPRIORA and ALPHAPRIORB. Do these stay in default (0.05 and 0.001 respectively) or do I have to adapt them accordingly?
Other parameters: admixtured model, correlated freq.
Thank you for your help!
Wang J (2017) The computer program structure for assigning individuals to populations: easy to use but easier to misuse. Mol Ecol Resour. 2017 Sep;17(5):981-990
Hi Florian,
Alpha is a prior used by STRUCTURE for running the MCMC. When using the Wang (2017) recommendation, you indeed need to ensure
"#define POPALPHAS (B) Individual alpha for each population" is set to 1. The difference now is that alpha will move independently for each population. What prior value you use for this alpha is depending on your data. The only way to find out what this should be is run a smaller subset of your data and check how your priors are behaving. Additionally, you need to ensure your chains are converging (as for any Bayesian analysis). That being said, as far as I know, not many people actually do this and use the default priors given by STRUCTURE.
Hope this helps!
Lotte
Florian,
my impression from checking the Structure manual and menu is that there are actually four options for setting/inferring ALPHA:
(menu of the frontend of 2.3.4 on Linux: New Parameter Set -> Ancestry Model -> Use Admixture Model: Advanced -> Infer ALPHA check: Configure -> we are here)
Initial ALPHA: default=1, Jinliang Wang suggests to use 1/K .
This is the initial ALPHA for the MCMC chain used by all four options. Over the run of the MCMC chain this value should change and converge to a new value within the chain (it should not jump around), and also converge to the same value between runs.
Option1:
"Separate Alpha for each population": (default: unchecked)
If unchecked, all populations/clusters share the same degree of admixture. The parameter ALPHA is the same for all populations and this one parameter is inferred using either a uniform prior or a Gamma distributed prior.
Suboption 1.1:
"Use a uniform prior for ALPHA": box is checked (it's the default)
The next two rows of settings are applicable, setting ALPHAMAX (default=10.0) and ALPHAPROPSD (default=0.025).
If you run the default settings, it's this option and these settings.
Suboption 1.2:
"Use a uniform prior for ALPHA": uncheck the box
Now the prior for ALPHA is assumed to be a Gamma distribution with its parameters defined in rows three and four, setting ALPHAPRIORA and ALPHAPRIORB.
You are still inferring only a single ALPHA for all populations. Just a different prior is used for this one parameter.
Option 2:
"Separate Alpha for each population": check the box
My impression is that this is called the "alternative prior" by Jinliang Wang. Here a different degree of admixture is inferred for each population. There is a discussion of/warning (identifiability) about this option in the manual under 6.3 "Estimating admixture proportions when most individuals are admixed". My note would be that you are adding K more parameters, thus your dataset needs to be sufficiently informative.
Suboption 2.1:
"Use a uniform prior for ALPHA": box is checked (its the default)
If I understand it right, each of the population-specific ALPHAs is inferred using a uniform prior (same settings for the priors of all separate ALPHAs).
Suboption 2.2:
"Use a uniform prior for ALPHA": uncheck the box
Now the priors for the population-specific ALPHAs are assumed to be each a Gamma distribution characterized by the same parameters.
****
How to decide which option and parameter values to use for your concrete dataset? You could run the different options and compare their overall likelihoods, as well as check if the cluster memberships of individuals make sense. For parameter values of the priors, you could choose extreme values (maybe in the manual) and check if there is an impact on the overall likelihood and again if the clustering results are reasonable.
If you are only expecting to find three clusters, you might not be in the parameter space that Wang 2017 considers. Eg. if several of your small population samples really belong to one large population and you expect them to cluster together, sample sizes might not differ that drastically in the end.
Also, you could visualize the genetic structure in you dataset using a PCA to get an overview over what to expect (eg. using adegenet in R). While, PCAs are also sensitive to sample sizes, they are a completely different approach, might give you a different visual idea and provide an independent result to which to compare your Structure solution.
Jutta
Thank you for your answers! I agree with you both.
When running a STRUCTURE analysis, besides implementing a hierarchical approach!, one should also consider to use Wang's alternative ancestry prior, especially in case of unbalanced sampling (Wang 2017 Mol Ecol Res). Referring to a personal communication with the author, this means checking the box "seperate alpha for each population" (initial alpha=1/K is suggested, but not mandatory).
Also take care in estimation of K, using LnP(D) and Evannos delatK.
Concerning ALPHAPRIORA and ALPHAPRIORB however I am still not sure on how to deal with those, but I will do test runs with them modified from the default as well.
As you wrote, Jutta, verifying your STRUCTURE results by different (such as multivariate) analyses is a must-do. Besides PCA, I also like the senstive DAPC (which manages to display the differences between populations while at the same time minimizing influence of differences within populations) (implemented in adegenet as well).
Cheers
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Sampling Design
- Sample Size