Analyzing admixture in phased v.unphased dataset

Determination of haplotype phase is becoming increasingly important as we enter the era of large-scale sequencing because many of its applications, such as imputing low-frequency variants and characterizing the relationship between genetic variations in different populations. Haplotype phase can be generated through laboratory-based experimental methods, or it can be estimated using computational approaches. We assess the haplotype phasing method that is aravailable in BEAGLE software, focusing in particular on using its output in ADMIXTURE analysis.

For simplicity's sake we have selected individuals from "Balto-Slavic"  cluster (the cluster attribution of individuals were inferred from Dienekes's Mclust using 11 MDS dimension), which is the major cluster of our project. Here is an all inclusive list of IDs for selected participants of our project:

V158
V157
V160
V202
V169
V170
V171
V174
V176
V177
V180
V181
V188
V189
V196
V205
V208
V211
V215
V218
V220
V221
V222
V228
V225
V232
V236
V237
V235
V231
V244
V246
V238






We had  thinned the genotype data of selected individuals to c.100 000 SNPs, removing SNPs in strong LD and  low quality SNps.  After that we used  GERMLINE pipeline for phasing PLINK format data with BEAGLE and processing in GERMLINE (phasing was performed in a homologous populations), Then, in order to assess possible discrepancies between phased and unphased data, we performed ADMIXTURE analysis (with 4 assumed clusters K=4) separately for original unphased dataset and BEAGLE-phased dataset.

To our surprise, we haven't be able to find expected signficant differences between phased and unphased multi -SNP markers genotypes (the range of difference is c.1-5%).

Unphased data:

Phased data:

Spreadsheet with ADMIXTURE results can be found here.

Simulated SNP-populations of MDLP

Yesterday I had set out to repeat  "simulation" experiments with a SNP dataset of my project's dataset, using PLINK's simulation techniques first described (in terms of population genetics) by Dienekes (the analogous experiments were performed by Harappa DNA BGA project and Eurogenes BGA project).
Synthetic "ancestral" populations (Altaic, Anatolian-Balkanian, Balto-Slavic, North-Atlantic, Scandinavian, Volga-Uralic and Celto-Germanic) were simulated using standard PLINK's simulation routine, with each ""synthetic" population including 5 generated synthetic individuals:

plink --simulate wgas1.sim --make-bed --out sim11

plink --simulate wgas2.sim --make-bed --out sim12

etc. ..

In data simulation, we assumed that  each of 7 clusters defined by specific combination of  allele frequencies of c.100000 Snps (obtained from ADMIXTURE K=7 run under unsupervised model) represents one ancestral pupulation.

Since I was interested in PCA loadings of "ancestral" populations, i used Eigensoft for explicit modeling  differences between different components" along continuous axes of variation. The calculated PCA loadings were then visualized as interactive biplot in R-package BiplotGUI  using the following  Biplot's command:

> Biplots(Data = PCA[, -1], groups = [, 1])

Afterwards  i performed three statistical tests on imported PCA loadings: linear regression, circular regression and procrusted analysis.

MDLP modification of DIYDodecad calculator: additional instructions/ideas

I have come up with another idea of how to use the estimated frequencies of my project for inferring the origin of shared HIR segments. Suppose, for example, that a Lithuanian/Belorussian and a Norwegian share some HIR segment. This could be:

1)Balto-Slavic-like ancestry in the Norwegian individual
2)Scandinavian-like ancestry in the Lithuanian individual
3)third party ancestry in both individuals

Using byseg 500 50 mode in DIYDodecad, AncestryFinder file, and MDLP allele frequencies, i was able to predict the origin of the HIR segment by picking one of three scenarios:

1) if the Lithuanian sees an excess of Scandinavian, then he should pick the second scenario
2) if he sees nothing unremarkable, the first one
3) if an excess of some component (relatively) low in both Lithuanian and Norwegian (e.g., Caucassian-Anatolian), the third.


After analyzing my "Scandinavian" "matches" from AF's file, i can conclude that DIYDodecad analysis (in byseg mode) reveals the presence of all three possible scenarios:

Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
X Sweden Sweden United States United States 1 156.4 160 3.6 5.6
60%-Scenario nr.1 (Scandinavian-like ancestry in the Balto-Slavic individual)
X Denmark Denmark Denmark Denmark 1 87.9 94.3 6.4 6.3
Scenario nr.1 (Scandinavian-like ancestry in the Balto-Slavic individual)
X Norway Norway Norway Norway 1 62.1 65.4 3.3 5
Scenario nr.1 (Scandinavian-like ancestry in the Balto-Slavic individual)
Anonymous0454 Finland Finland Finland Finland 1 104.8 110.3 5.5 6.4
Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
X Sweden Not Provided Not Provided Not Provided 2 82.2 88.1 5.9 5
60%-Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
X Sweden Sweden United States United States 3 174.9 178.1 3.2 5
80%-Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
Anonymous0353 Denmark Denmark Denmark Denmark 3 68.4 73.3 4.9 8.1
50%-Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
Anonymous0245 Denmark Denmark Denmark Denmark 4 70.5 77.3 6.8 5.3


HLA-MHC group -Scenario III - third party (Celto-Germanic) ancestry in both individuals
X Sweden Sweden United States United States 6 25.8 34.2 8.4 5.4
X Sweden Denmark United States United States 6 27.6 34.5 6.9 5.1
Anonymous0207 Sweden Sweden Not Provided Not Provided 6 25.6 34.1 8.5 5.3
X Norway Not Provided Belgium Not Provided 6 29.7 36.1 6.4 5.3
Anonymous0370 Sweden Sweden Not Provided Sweden 6 30.7 36.6 5.9 5.4
X Denmark Denmark Denmark Denmark 6 25.6 35.5 9.9 5.9
Anonymous0439 Denmark Not Provided United States Hungary 6 26.3 36 9.7 6
X Denmark Denmark Denmark Denmark 6 26 34 8 5.1
X Norway Norway Norway Norway 6 25.6 34.2 8.6 5.6

Chr14. group - 70%-Scenario nr.2 (Balto-Slavic-like ancestry in the Scandinavian individual)
Anonymous0001 Norway Norway Norway Norway 14 38 48.8 10.8 6.5
Anonymous0037 Sweden Norway United States United States 14 39.2 48 8.8 5.1

Scenario nr.1 (Scandinavian ancestry in the Balto-Slavic individual)
Anonymous0002 Denmark Denmark United States United States 15 43.9 51.3 7.4 6

I have modified DIYDodecad calculator v2.0 for my BGA project (MDLP). If you want to analyze your genotype using allelic frequencies, estimated from the latest K=7 Admixture run, you should download MDLP.zip file (File -> Download original) and extract its contents to the directory of DIYDodecad. 

Then launch the MDLP add-on by entering the command DIYDodecadWin MDLP.par


You can download MDLP add-on for DIYDodecad calculator here.

Root Means Square Comparison Excel 2007 Macro Enabled XLSM spreadsheet for the Magnus Ducatus Lithuaniae Project data.

Mark Jost has posted a Root Means Square Comparison Excel 2007 Macro Enabled XLSM spreadsheet for the Magnus Ducatus Lithuaniae Project data.

The root-mean-square is used measure of the differences between ID. The smallest RMS number calculated is the closest overall to you.

If used in any other version please check your original column cells and verify your alternate software is handling sorting and reporting correctly.

Here is a link

http://tinyurl.com/MDLP-Project-AdmixtureK7-RMS

And some example screenshots:

MDS and PCA plots: for V157-V247

I have decided to publish MDS and PCA plots made (in PDF formats) for the latest run (see my previous post)

MDS plot
PCA plot

Project's update: results for participants V157-V247

Here is the latest update of MDL project. For the time being, i have decided to run ADMIXTURE and STRUCTURE analysises simultaneously on the same "LD-pruned" dataset (a subset of the whole MDL dataset with markers corrected for some genetic effects (the admixture coefficients for each participant and references can be downloaded in Excel spreadsheet).

When we look at the global apportionment of ancestry  coefficients in STRUCTURE program, we would expect to see similar but perhaps slightly different patterns than with the ADMIXTURE results. However, the difference between the outcome of those 2 programs was more significant than expected. First of all the STRUCTURE program estimated higher West_European ancestry (especially in samples the from Caucassian region) but the North_East_European and South European estimates were within 10% of one another and overall the estimates were reasonably close in the most of cases (check this spreadsheet for technical details - Fst values, genetic distances etc.). Thus we may conclude that the estimates from the two programs are staistically correlated, while the differences are explained by the different approach to estimating the co-ancestry coefficients. STRUCTURE assumes admixture formed by k ancestral subpopulations, which are defined empirically based on genetic distance between samples in a study set provided by the markers used in the analysis. Not specifying the source or parental populations a priori in this way, the clustering of samples using a variety of k values can be screened and the k that gives the greatest likelihood, and/or the most distinct subpopulation clusters or the most clearly defined mosaic of ancestry affiliations can be selected as best. If adequate AIMs are used, when these methods are run on samples representative for the major continental populations, they tend to identify clusters at low values of k that fall along continental lines, and clusters at higher values of k that fall along ethnic subdivisions defined by other methods such as Principle Components Analysis (Rosenberg et al. 2002)

We started the ADMIXTURE analysis from k=3, resolving Caucassian, North-West European from Asian group. Moving to a more complex K=4 and k=5-population model, the Scandinavian component and the Baltic-North Sea components emerged, resolving  from the Altaic and the Anatolia-Balkanic component  to form a North-West European out group,while both the Altaic and Anatolia-Balkanic groups show little intergroup affinity. The Baltic-North Sea component still show substantial fractional affiliation with the Scandinavian group. Although the k=4 model may seem useful since it neatly provide the most obvious anthropological partitions, the affinity between the Balkanic and two North_European components at least with these markers suggests that this k=4 population model is not what we are looking for either. However, at k= 5, the incidence of fractional affiliations dropped dramatically and for the first time in the analysis of k-value models, we can see the simplest of the clearly partitioned models, breaking the populations used in our project into five groups (Anatolian-Balkanic    Scandinavian    Balto-Slavic    Altaic    Celto-Germanic) that comport with specific major Eurasian metapopulation groups. The results are the same as k=4, except that  the Baltic-North Sea combined group splits now into two new subgroups Balto-Slavic and Celto-Germanics, and the overall incidence of outside-group affiliation was much lower.

In our experiments, higher k-values showed equally clean genetic partitioning but k=5 was the simplest of the clearly partitioned models.

Due to some specific requests, i have also decided to calculate IBS matrix between project's particpants and references (the matrix itself has been visualized in a standard form of a dendrogramm). You also have an opportunity to learn  Mclust clustering of each participant  and look at  the average values of homozygocity and inbreeding for all participants.