mStruct: A New Admixture Model for Inference of Population Structure in Light of Both Genetic Admixing and Allele Mutations

1
mStructA New Admixture Model for Inference of
Population Structure in Light of Both Genetic
Admixing and Allele Mutations

• Suyash Shringarpure and Eric Xing
• School of Computer Science
• Carnegie Mellon University
• ICML 2008

Presented by Haojun Chen
2
Outline

• Background
• Structure Model
• mStruct Model
• Experiment Results
• Summary

3
Background

• Allele one member of a pair or series of
  different forms of a gene
• Population structure analysis aim to shed light
  on evolutionary history of modern human
  population
• Microsatellites and single nucleotide
  polymorphisms (SNP) data base of population
  structure analysis
• State-of-the-art method Structure

4
Structure Model
x Microsatellite alleles
unique set of population-specific
multinomial distributions
vector of multinomial parameters,
a.k.a., allele frequency profile (AP), of the
allele distribution at locus i in ancestral
population k total number of observed
marker alleles at locus I total
number of marker loci total number of
individuals individual-specific admixing
coefficient vector
5
Pitfall of Structure

• There is no mutation model for modern individual
  alleles with respect to common prototypes in the
  modern populations
• Every unique allele in the modern population is
  assumed to have a distinct ancestral frequency,
  rather than allowing the possibility of it just
  being a descendent of some common ancestral allele

6
mStruct Model
set of
ancestral alleles mutation parameter
associated with locus frequencies of the
ancestral alleles total number of ancestral
alleles
Microsatellite mutation model SNP mutation model
7
Generative Process

• Generative process for Structure
• where
• Generative process for mStruct
• step 2.2 above is replaced by

8
mStruct Model Inference

• MCMC slow
• Variational inference for hidden variable
• variational EM for hyperparameter

9
Synthetic Data
Twenty microsatellite genotype datasets with 100
individuals from 3 ancestral populations at 50
genotype loci
10
HGDP Microsatellite Data

• Model selection by BIC (Bayesian Information
  Criterion) score

11
HGDP Microsatellite Data
1056 individuals from 52 populations at 377
autosomal microsatellite loci
am-spectrum spectrums of different ancestral
populations gm-spectrum spectrums of
different geographical populations
12
Contour of Mutation Rates
13
Summary

• mStruct takes into account genetic admixture and
  allele mutation effects
• mStruct extended LDA which allows noisy
  observations
• Variational inference algorithm that allows
  tractable inference developed for mStruct
• Other application images, text and so on