Genetic Analysis of Complex Traits

Jing Hua Zhao

July 20, 2010, NIST

2010

About the title

Complex traits refer to common diseases or traits with no clear modes of Mendelian inheritance Reduced penetrance, heterogeneity, phenocopy, pleiotrophy, (Lander & Schork 1994), environmental factors, examples include diabetes, heart diseases, mental disorders, height, body-mass index (BMI) Methods include the assessment of familial aggregation for heritability, identification of major gene effect, study of cosegregation of genetic marker with putative disease-predisposing loci in the so-called linkage studies and association studies in search of frequency differences between cases and controls and/or correlation between genotype and phenotype as a quantitative trait. Morton et al. (1983), Khoury et al. (1993), Thomas (2004)

A sketch
We provide an overview of genetic analysis of complex traits in humans in the context of large volume of genetic data. While there are many analytical issues, our focus is more on the practical side. We provide specific examples of genetic association study. We are not limited to R, and would provide examples using systems other than R whenever appropriate. Our hope remains to be that this will serve as a forum for a range of issues and a contact point for future researches. Questions are welcome during the sessions.

Contents
The presentation consists of four parts: I. Overview

II. Analytic tools and association testing III. Miscellaneous topics IV. OpenMx and NCBI2R V. Conclusion You may find materials from useR!2008 and useR!2009 tutorials relevant. They are both available from my personal home page.

What have changed?

We are quite far with topics in both useR!2008 and useR!2009, esp. genome-wide association studies (GWAS) of directly genotyped and imputed SNPs and interaction analysis. It is routine with Stata function which automating analysis by SNPTEST. We have updated results regarding genetic predisposition score from the EPIC-Norfolk study. We have a better understanding of the SNP annotation, via UCSC/galaxy and in particular NCBI2R. There are other changes, e.g., functions MiMa has been replaced with metafor package. I am more proficient with R and have consolidated R/gap functions fbsize, pbsize, ccsize, and added ab and masize. We have explored raw storage mode in R which is central to snpMatrix and GenABEL. We provide further examples. We add examples for chromosome X data, and obtained imputed genotypes for analysis. We also add materials regarding OpenMx. In the future, more materials on Bioconductor can be added.

Monographs
Morton NE. Rao DC, Lalouel JM. Methods in Genetic Epidemiology. Karger, 1983. Khoury MJ, Beaty TH, Cohen BH. Fundamentals of Genetic Epidemiology. Oxford University Press, 1993. Falconer DS, Mackay TFC. Introduction to Quantitative Genetics, 4e. Longman, 1996 Hartl D, Clark AG. Principles of Population Genetics, 3e. Sinauer Associates, Inc. 1997 Lange K. Mathematical and Statistical Methods for Genetics Analysis. 2e, Springer 2002 Sorensen D, Gianola D. Likelihood, Bayesian, and MCMC Methods in Quantitative Genetics. Springer 2002 Thomas DC. Statistical Methods in Genetic Epidemiology, Oxford University Press, 2004 Armitage P, Colton T (Eds). Encyclopedia of Biostistics, 2e, Wiley, 2005

Monographs
Elston RC, Johnson WD. Basic Biostatistics for Geneticists and Epidemiologists, A Practical Approach. Wiley, 2005 Ahrens W, Pigeot I (Eds). Handbook of Epidemiology. Springer, 2005 Balding DJ, Bishop M, Cannings C (Eds). Handbook of Statistical Genetics, 3e, Wiley, 2007 Siegmund D, Yakir B. The Statistics of Gene Mapping. Springer 2007 Wu R, Ma C-X, Casella G. Statistical Genetics of Quantitative Traits-Linkage, Maps and QTL. Springer, 2007 Speicher MR, Antonarakis SE, Motulsky AG. Vogel and Motulskys Human Genetics: Problems and Approaches, 4e, Springer, 2010 Lin S, H Zhao (Eds). Handbook on Analyzing Human Genetic Data: Computational Approaches and Software. Springer, 2010

I Overview

Terminology
Genes, Chromosome, markers Alleles, genotypes, haplotypes Phenotypes, mode of inheritance, penetrance Mendelian laws of inheritance, Hardy-Weinberg equilibrium, linkage disequilibrium Association tests for single or multiple SNPs Population stratification Multiple testing Gene-environment interaction (GEI)

Topics
Organization
Genetic epidemiology

Linkage studies Association studies GWAS

The landscape Study design

Genetic epidemiology
It is the study of the role of genetic factors in determining health and disease in families and in populations, and the interplay of such genetic factors with environmental factors, or a science which deals with the aetiology, distribution, and control of diseases in groups of relatives and with inherited causes of disease in populations (http://en.wikipedia.org). It customarily includes study of familial aggregation, segregation, linkage and association. It is closely associated with the development of statistical methods for human genetics which deals with these four questions. The last two questions can only be answered if appropriate genetic markers available (Elston & Ann Spence. Stat Med 2006;25:3049-80).

Linkage studies
It is the study of cosegregation between genetic markers and putative disease loci, and has been very successful in localizing rare, Mendelian disorders but since has difficulty for traits which do not strictly follow Mendelian mode of inheritance, considerable linkage heterogeneity and it has limited resolution. It typically involves parametric (model-based) and nonparametric (model-free) methods, the latter most commonly refers to allele-sharing methods. The underlying concepts are nevertheless very important. It can still be useful in providing candidates for fine-mapping and association studies. With availability of whole genome data, it is possible to infer relationship or correlation between any individuals in a population.

Association studies
They focus on association between particular allele and trait; it is only feasible with availability of dense markers. It has traditionally applied to both relatives in families and population sample. For the latter there has been serious concern over spurious association due to difference in allele frequencies between hidden subpopulations in a sample. A range of considerations has been made (Balding. Nat Rev Genet 2006;7:781-91) but the availability of whole genome data refreshes our understanding and perspectives.

GWAS
Any study of genetic variation across the entire human genome designed to identify genetic association with observable traits or the presence or absence of a disease, usually referring to studies with genetic marker density of 100,000 or more to represent a large proportion of variation in the human genome (Pearson & Manolio. JAMA 2008;299:1335-44), or simply look for associations between DNA sequence variants and phenotypes of interest (Donnelly. Nature 2008; 456:728-31). It is associated with the common disease common variant hypothesis (CD-CV). Common polymorphisms (MAF>1%) might contribute to susceptibility to common diseases, so that GWAS of common variants might be used to map loci contributing to common diseases. It therefore helps to catalog millions of common variants in the human population, massive genotypes to large number of individuals, and appropriate analytical framework (Altshuler et al. Science 2008; 322:881-888).

The landscape
A catalog of published GWASs is maintained by Office of Population Genomics at the National Human Genome Research Institute (NHGRI) and available from http://www.genome.gov/GWAStudies As of 3/2010, there were 779 published genome-wide associations at p<5x10-8 for 148 traits. As of 06/2010, the table included 587 publications. For instance, for body mass index, it includes the major publications from GWASs with 100000 SNPs, namely, Thorleifsson et al. Nat Genet 2009;41:18-24, Willer et al. Nat Genet 2009; 41:25-34; Loos et al. Nat Genet 2008;40:76875, Fox et al. BMC Med Genet 2007;8:S18, Frayling et al. Science 2007;316:889-94. Furthermore, there were Benzinou et al. Nat Genet 2008;40:943-5, Meyre et al. Nat Genet 2009;41:157-9.

Published Genome-Wide Associations through 3/2010, 779 published GWA at p<5x10-8 for 148 traits
NHGRI GWA Catalog

Context
The population under study must be characterized to allow the selection of patients likely to share a genetic cause of disease Thousands of cases and controls may be needed if a study is to have sufficient statistical power to identify the alleles of interest it creates bioinformatics challenges and raises questions about how to identify true positive signals

Christensen & Murray. New Eng J Med 2007;356:1094-7

International collaborative projects

The HapMap project (http://hapmap.ncbi.nlm.nih.gov/) was a study of 270 people from the Yoruba in Nigeria (30 trios), Japanese (45 unrelated individuals), Han Chinese (45 unrelated individuals) and CEPH (30 trios). The 1000 genome project (http://www.1000genomes.org) aims to sequence at least one thousand anonymous participants. It still undergoes revision. The database of genotypes and phenotypes (dbGaP) (http://www.ncbi.nlm.nih.gov/sites/entrez?db=gap) was developed to archive and distribute the results of studies that have investigated the interaction of genotype and phenotype. The genetic analysis workshops (GAWs) (http://www.gaworkshop.org/) are a collaborative effort among genetic epidemiologists to evaluate and compare statistical genetic methods. For each GAW, topics are chosen that are relevant to current analytical problems through simulated or real data.

A conceptual picture based on a test of H0: =0 vs H1: = 1 >0 from a normal distribution

Sample size calculation for normal distribution

Study designs

Three common genetic association designs involving unrelated individuals (left), nuclear families with affected singletons (middle) and affected sib-pairs (right). Males and females are denoted by squares and circles with affected individuals filled with black colors and unaffected individuals being empty Risch & Merikangas. Science 1996;273:1516-7, Zhao. J Stat Soft 2007;23(8):1-18

Sample sizes required for association detection using population data

Power of linkage versus association

Power calculation under matched design

GEI of type-2 diabetes

Legends in the project manual were perhaps confusing so it is worthwhile to re-present here. Matched case-control study Type I error rate () = 0.00001 (two-sided) Continuous environmental factors with standard deviation 1, and interaction odds ratio (Rge) = 1.2 ~ 4 K = 0.05 (done for 0.1 ~ 0.15) Sample size (N) = 500 ~ 12,000 Additive model Allele frequency (p) = 0.05, 0.1, 0.2, 0.3 We supplied these to Quanto 1.0 (http://hydra.usc.edu/gxe, now available on the Epidemiology Unit machines) Gauderman WJ. Stat Med 21:35-50, 2002

One more example of EDNAR application

The calculation is as a linear function of proportion of variance explained, significant level and sample size. proc power; ods output output=op; multreg model = fixed alpha = 0.00001 0.000001 0.0000005 nfullpred = 1 ntestpred = 1 rsqfull = 0.001 to 0.005 by 0.001 rsqdiff = 0.001 to 0.005 by 0.001 ntotal = 10000 to 25000 by 1000 power = .; run;

Power by %variance explained (R2)

R2 Sample size =10-5 10000 15000 20000 25000 =5x10-7 10000 15000 20000 25000 0.031 0.124 0.290 0.489 0.29 0.67 0.90 0.98 0.67 0.95 1.00 1.00 0.9 1.0 1.0 1.0 0.98 1.00 1.00 1.00 0.10 0.29 0.52 0.72 0.52 0.86 0.97 1.00 0.86 0.99 1.00 1.00 0.97 1.00 1.00 1.00 1 1 1 1 0.1 0.2 0.3 0.4 0.5

Power by=1e-5,1e-6,5e-7
10000 20000 10000 20000

1 e -0 5 0 .0 0 1
1.0 0.8 0.6 0.4 0.2 0.0

1 e -0 5 0 .0 0 2

1 e -0 5 0 .0 0 3

1 e -0 5 0 .0 0 4

1 e -0 5 0 .0 0 5

1 e -0 6 0 .0 0 1

1 e -0 6 0 .0 0 2

1 e -0 6 0 .0 0 3

1 e -0 6 0 .0 0 4

1 e -0 6 0 .0 0 5
1.0 0.8

power

0.6 0.4 0.2 0.0

5 e -0 7 0 .0 0 1
1.0 0.8 0.6 0.4 0.2 0.0 10000 20000

5 e -0 7 0 .0 0 2

5 e -0 7 0 .0 0 3

5 e -0 7 0 .0 0 4

5 e -0 7 0 .0 0 5

10000

20000

10000

20000

S a m p le s ize

Two-stage design on main effect

The goal is to reduce cost without compromising efficiency. Given our study sample and SNPs of interest are defined, a staged design furnishes collection of all information in several steps. In the simplest and well-studied two-staged design of genetic case-controls studies, a proportion of individuals is genotyped at all of the SNPs and a proportion of the most significant ones is selected and to be carried over as replication study at the second stage. Skol et al. Nat Genet 2006, 38(2):209-13 (check the associate website for a program called CaTS). It is implemented in the function tscc within R/gap.

FTO-BMI-T2D Mendelian randomisation

FTO-T2D association is gone once BMI is included in the model. This has been used in the so-called Mendelian randomisation study disentangling the causal association of BMI-T2D (Freathy et al. Diabetes 2008, 57:1419-26). There is association between FTO and BMI (a). There is epidemiological association between BMI and metabolic traits (b). The association between FTO genotype and metabolic traits is mediated by BMI (c=axb).

Power calculation
We can of course perform simulations to obtain power estimate but it would be somewhat involved. Instead, we calculate standard error of FTO-BMI-T2D can be calculated which can form the basis of power calculation (Kline RB. Principles and Practice of Structural Equation Modeling, 2nd Edition, The Guilford Press 2005). We implement this in ab function in R/gap. We have for EPIC-Norfolk 25,000, SNP-BMI regression coefficient (SE) of 0.15 (0.01), and BMI-T2D log(1.19) (0.01). We consider =0.05. Criticism arised from this posthoc power calculation could be alleviated when we allow for a range of sample sizes to be considered in the next slide.

Two-stage GEI
A case-only design is used as the first stage. This is to be followed by a second stage involving both cases and controls. Kass & Gold. Handbook of Epidemiology 2004; I.7 Murcray et al. Am J Epidemiol 2008; 169:219-26 Li & Conti. Am J Epidemiol 2008; 169:497-504 Thomas D. Nat Rev Genet 2010 (Epub)

References
Armitage P, Colton T. Encyclopedia of Biostatistics, Second Edition, Wiley 2005 Balding DJ, Bishop M, Cannings C. Handbook of Statistical Genetics, Third Edition, Wiley 2007 Elston RC, Johnson W. Basic Biostatistics for Geneticists and Epidemiologists: A Practical Approach. Wiley 2008 Haines JL, Pericak-Vance M. Genetic Analysis of Complex Diseases, Second Edition. Wiley 2006 Rao DC, Gu CC (Eds). Genetic Dissection of Complex Traits, Volume 60, Second Edition (Advances in Genetics). Academic Press 2008 Thomas DC. Statistical Methods for Genetic Epidemiology. Oxford University Press 2004

A summary
We have restricted our focus and leave out a lot of details to cover a rapid moving field with limited time. It seems that the practice of study designs and data analysis cannot be changed in a short run, but we have already seen steady increase in use of R.

Case study: GWAS of obesity-related traits

Background Study design Statistical analysis On-going research

EPIC study
The European Prospective Investigation into Cancer and Nutrition (EPIC) is coordinated by Dr Elio Riboli, Head of the Division of Epidemiology, Public Health and Primary Care at the Imperial College London. EPIC was designed to investigate the relationships between diet, nutritional status, lifestyle and environmental factors and the incidence of cancer and other chronic diseases. EPIC is the largest study of diet and health ever undertaken, having recruited over half a million (520,000) people in ten European countries: Denmark, France, Germany, Greece, Italy, The Netherlands, Norway, Spain, Sweden and the United Kingdom.

EPIC-Norfolk study
EPIC-Norfolk participants are men and women (based on over 30,000 people) who were aged between 45 and 74 when they joined the study, who lived in Norwich and the surrounding towns and rural areas. They have been contributing information about their diet, lifestyle and health through questionnaires, and through health checks carried out by EPIC nurses.

Case-cohort design for EPIC-Norfolk study

It originally followed case-control design (e.g., WTCCC with seven cases and common controls) with 3425 cases and 3400 controls. It is potentially more powerful. Controls are selected. It has then been changed into case-cohort design, in which cases are defined to be individuals whose BMI above 30 and controls are a random sample (subcohort) of the EPIC-Norfolk cohort which includes obese individuals. The subcohort is representative of the whole population and allows for a range of traits to be examined. The analysis is potentially more involved but established.

Case-cohort design
The distribution of body mass index (BMI) is the casecohort design of the EPIC-Norfolk study of obesity is a combination of the sub-cohort sample and case sample which is truncated from the whole cohort at BMI=30 Zhao. J Stat Soft 2007;23(8):1-18
Histogram of bmi

Frequency

0 15

200

400

600

800

20

25

30 bmi

35

40

45

Power/sample size
It started with assessment of how the power is compromised relative to the original case-control design. This was followed by power/sample size calculation using methods established by Cai and Zeng (2004) as implemented in an R function, noting a number of assumptions. More practically, it was also envisaged that a proper representative sample of a total of 25,000 individuals would be 10%; the subcohort is then approximately 2,500. The total sample was split between two stages.

GeneChips
Affymetrix 500K Data were available for 3850 individuals Illumina 317K It came at a later time Data quality appears to be poor? The focus has therefore been Affy500K, but with a possible comeback.

Analysis
An incremental approach was adopted since the storage and computing power were somewhat uncertain. This was predated with controls from the breast cancer study, involving about 400 individuals with Perlegen 250K GeneChips. QC including call rates and HWE was feasible with SAS/Genetics (~30GB) which provides a good estimate of the storage for all individuals (~380GB). The Linux platform seemed favourable.

EPIC400 analysis

The analysis for GWAS

QC including visualisation of clustering, outliers, was largely done by colleagues at Sanger (as for WTCCC) The overall strategy was data partition, i.e., by chromosome and further by region (30) in each chromosome, largely on a long, skinny data format A major advantage is that the analysis can be resumed whenever the system experiences problems We stuck to SAS to allow for reliability and flexibility with or without SAS/Genetics, for BMI/obesity as continuous and binary outcomes are readily tackled with REG/LOGISTIC procedures most outputs are available from the output delivery system (ODS) The picture was eventually changed with a revised coding algorithm and the use of imputed data

Additional analysis
Population stratification via EIGENSTRAT SAS is very handy since a single put statement is sufficient to generate the output. Collaborative (e.g. height) and consortium work (GIANT) On the UK side, this is mainly involved with IMPUTE/SNPTEST, with inputs on strand, standard error, quantitative traits, outputs. This facilitates meta-analysis considerably.

The first report

Meta-analysis for odds of obesity

LDL

Height

BMI/obesity

Further on BMI

Reflection on the study design

Current practice
Linux clusters are now ready for comprehensive analyses and greatly facilitated by Linux/awk script which is light. awk proves very useful and can be transformed to Perl. In fact, any statistical package which processes data elements would be less efficient. An example is the transformation of long, wide, transposed format noted earlier. They call C/C++ programs such as IMPUTE/SNPTEST. We use Stata package to automate SNPTEST, and in some instances involved C/C++ code. SAS is still useful for data preparation, and in a sense less professional than DBMS such as Oracle but enjoys a large user community and has facility for data analysis. SAS 9.2 PROTO procedure allows for C/C++ to be called.

FTO/physical activity--BMI/WC association

FTO variant, rs1121980, was genotyped in 20,374 participants (39-79 years) from the EPIC-Norfolk Study. Physical activity (PA) was assessed by a validated selfreported questionnaire. The interaction between rs1121980 and PA on BMI and waist circumference (WC) was examined by including the interaction term in mixed effect models. Our results show that PA attenuates the effect of FTO rs1121980 genotype on BMI and WC.

Main effect of FTO

FTO--physical activity interaction

References
Bodmer W, Bonilla C. Nat Genet 2008;40:695-701 EPIC: http://epic.iarc.fr/ EPIC-Norfolk: http://www.srl.cam.ac.uk/epic Long AD et al.. Science 1997; 275:1328 Loos R et al. Nat Genet 2008; 40:468-75 Prentice RL. Biometrika 1986;73:1-11 Risch N, Merkangas K (1996) Science 1997;273:1516-7 Sandhu MS et al. Lancet 2008; 371:483-91 Thomas DC. Net Rev Genet 2010; 11:259-72 Vimaleswaran KS. Am J Clin Nutr 2009; 90:425-428 Weedon MN et al. Nat Genet 2008;40:575-83 Willer et al. Nat Genet 2009; 41:25-34 Zhao JH. J Stat Soft 2007;23(8):1-18 Zhao JH et al. CCIS 2007;2:781-90

II Association analysis

Topics
Elements of association analysis Analytic tools R packages Examples Appendix

Elements of association analysis

Quality control: call rates, Hardy-Weinberg equilibrium and minor allele frequencies and others such as clustering of genotypes, relatedness and population stratification. Test of associations often through linear regression for continuous trait, and through logistic regression for binary, the proportion of variance explained for LR is measured through R2 while the score statistic under additive model is equivalent to the Armitage trend test. Genotype imputation: mostly often through HapMap CEU sample, involving ~2.5 million SNPs Graphical presentation Interpretation, replication Report of findings

GRAMMAR
It refers to genome-wide rapid association using mixed model and regression, and implemented in R/GenABEL. The method first obtains residuals adjusted for family effects and subsequently analyzes the association between these residuals and genetic polymorphisms using least-squares methods. It can also involves selected polymorphism to be followed up with the full measured genotype analysis (Aulchenko et al. Genetics 2007; 177:577-85). yi = + j j X ij + Gi + ei Initial model: We have the residuals ei = yi ( + j j X ij + Gi ) yi * ei = + i g i + i Linear regression: Measured genotype model: yi = + i g i + j j X ij + Gi + ei The method adjusts for familial relationship, computationally fast, and ready to incorporate methods developed for unrelated individuals in the second stage.

Graphics and association plots

Plot of summary statistics Pedigree-drawing LD plot Q-Q plot -- contrasting observed versus expected log-p values Manhattan plot -- distribution of genome-wide p values Regional association plot -- including recombination, contribution from imputed SNPs and top hits from consortium meta-analysis

Analytical tools
There are several reviews on Human Genomics, and an active list is maintained at http://linkage.rockefeller.edu LINKAGE, GENEHUNTER, Merlin, PAP, SAGE, SOLAR ETDT, EHPLUS, FBAT, QTDT, UNPHASED, SAS/Genetics R (genetics, gap, haplo.stats, haplin, kinship) For GWAS HaploView PLINK, SNPGWA IMPUTE, MACH, BinBam EIGENSTRAT METAL SAS, Stata, R (snpMatrix, GenABEL, SNPassoc)

Connections with R
Occasionally, these software will be cross-referenced. Analyses with specialized programs such as IMPUTE/SNPTEST and PLINK are illustrated in the useR!2008 tutorial. snpMatrix provide connect with PLINK file, e.g., narac <- read.plink("narac.bed","narac.bim","narac.fam")

Basic R packages
genetics haplo.stats gap Rassoc, HardyWeinberg, kinship, multic, pedigree, identity

See CRAN task view for Genetics (http://cran.rproject.org/web/views/Genetics.html), as with an earlier review on the motivation for analysis with R statistical and computational environment (Zhao & Tan. Hum Genomics 2006;2:258-65) It also refers to Rgenetics projects whose packages are now available from http://www.bioconductor.org.

Hardy-Weinberg equilibrium tests

Suppose g is a data frame containing genotype counts for a list of SNPs. We can obtain exact HWE p value as follows. library(gap) head(g) comhom het rarehom 1 12879 6699 961 2 13463 6214 799 for(i in 1:2) print(snp.HWE(as.numeric(g[i,]))) [1] 0.01843766 [1] 0.01542034

Gene-counting method: ABO blood type

library(VGAM) abodat <- data.frame(A = 186, B = 38, AB = 13, O = 284) fit <- vglm(cbind(A, B, AB, O) ~ 1, ABO, abodat) fit coef(fit) Coefficients: (Intercept):1 (Intercept):2 -1.303414 -2.941384 Degrees of Freedom: 2 Total; 0 Residual Residual Deviance: 0.3917573 Log-likelihood: -8.372631 > Coef(fit) pA pB 0.21359094 0.05014533

Transmission/disequilibrium test (TDT)

library(gap) x <- matrix(c(0,0, 0, 2, 0,0, 0, 0, 0, 0, 0, 0, 0,0, 1, 3, 0,0, 0, 2, 3, 0, 0, 0, 2,3,26,35, 7,0, 2,10,11, 3, 4, 1, 2,3,22,26, 6,2, 4, 4,10, 2, 2, 0, 0,1, 7,10, 2,0, 0, 2, 2, 1, 1, 0, 0,0, 1, 4, 0,1, 0, 1, 0, 0, 0, 0, 0,2, 5, 4, 1,1, 0, 0, 0, 2, 0, 0, 0,0, 2, 6, 1,0, 2, 0, 2, 0, 0, 0, 0,3, 6,19, 6,0, 0, 2, 5, 3, 0, 0, 0,0, 3, 1, 1,0, 0, 0, 1, 0, 0, 0, 0,0, 0, 2, 0,0, 0, 0, 0, 0, 0, 0, 0,0, 1, 0, 0,0, 0, 0, 0, 0, 0, 0),nrow=12) xx <- mtdt2(x,refcat="12")

We obtain results similar to ETDT (Sham PC, Curtis D (1995) An extended transmission/disequilibrium test (TDT) for multi-allelic marker loci. Ann. Hum. Genet. 59:323-336).

Haplotype analysis
library(haplo.stats) mc4r.map <- read.table("mc4r.map",as.is=TRUE) snps <- mc4r.map[,2] M <- length(snps) a1 <- sprintf("%s%s",snps,rep(".a1",M)) a2 <- sprintf("%s%s",snps,rep(".a2",M)) a1a2 <- c(a1,a2) for(i in 1:M) {a1a2[2*i-1] <- a1[i];a1a2[2*i] <- a2[i]} mc4r <read.table("mc4r.ped",col.names=c(paste("v",1:6,sep=""),a1a2)) pheno <- read.csv("mc4r.csv",sep="\t",skip=11) cohort <- subset(pheno,cohort==1) attach(cohort) mc4r.12 <haplo.score(bmi,mc4r[id,7:30],x.adj=sex+age,locus.label=snps[1:12 ])

Generalized linear models

mc4r.geno <setupGeno(mc4r[id,7:dim(mc4r)[2]],locus.label=snps) attr(mc4r.geno,"unique.alleles")[1:12] mc4r.12 <haplo.score(bmi,mc4r.geno[,1:24],x.adj=sex+age,locus.label =snps[1:12]) mc4r.data <- data.frame(geno=mc4r.geno,cohort) mc4r.gauss <- haplo.glm(bmi ~ sex + age + geno, family = gaussian, na.action="na.geno.keep", allele.lev=attributes(geno)$unique.alleles, data=mc4r.data, locus.label=snps, control = haplo.glm.control(haplo.freq.min=0.02)) mc4r.gauss detach(cohort)

Gene-gene, gene-environment interaction

haplo.glm is considerably slower but it is among the few facilities for GEI analysis A recent analysis of SNCA-LRRK2 interaction with Parkinsons disease snca_assign <- read.dta("snca_post.dta") snca_lrrk2_int1 <haplo.glm(formula=y~ssex+snca_assign$hap12*lrrk 2,family="binomial",data=clean,locus.label=s2) snca_lrrk2_int1 Object snca_assign contains effective haplotype assignment based on SNCA and used as covariate for SNCA-LRRK2 interaction analysis. We can also use haplo.interaction from SNPassoc.

Adjustment for multiple testing

Bonferroni.sig(wga, model="log-add",alpha=0.05) library(qvalue) q <- qvalue(p) plot(q) library(multtest) adj <c("Bonferroni","Holm","Hochberg","SidakSS","SidakSD", "BH","BY") mt <- mt.rawp2adjp(p,adj) mt.reject(cbind(mt$rawp,mt$adjp),seq(0,0.1,0.001))$r

Manhattan and regional association plots

library(gap) # for the Framingham data analysis png("figures.pdf") par(mfrow=c(2,1),mai=c(1,1,0.2,0.8),ps=7) qqunif(test$np,bg="blue",bty="n",xlim=c(0,6),cex=0.02) par(las=2) mhtplot(test,usepos=TRUE,pch=21,colors=rep(c("blue","green"), 11),cutoffs=c(4,5,6),cex=0.02) dev.off() # DGI example for asplot asplot("rs10811661", "CDKN2A/CDKN2B region", "9", CDKNlocus, CDKNmap, CDKNgenes, 5.4e-8, c(3,6))

C D K N2 A/C D K N2 B r e g io n
L D (r^2 ) 8 0 .8 0 .5 0 .2 0 .0 im p . 60 P =5 .4 e -0 8

-log10(Observed p)

6 rs1 0 8 1 1 6 6 1 4 40 Recombination rate (cM/Mb)

20

0 0
CDKN2B CDKN2A

21900

22100 C hro m o so m e 9 p o sitio n (kb )

22300

Association plot

R packages for GWAS

SNPassoc GenABEL P2BAT snpMatrix

While the first three are available from CRAN, snpMatrix is available from BioConductor. Other packages include, multtest, meta, rmeta, CAMAN, qvalue, ROCR.

Setup for GWAS

CRAN http://cran.r-project.org/web/packages/index.html > setRepositories() > install.packages(c(SNPassoc, GenABEL)) > library(GenABEL) BioConductor > source("http://bioconductor.org/biocLite.R") > biocLite("snpMatrix") > library(snpMatrix)

Notes on S4 class
We illustrate with two classes > library(snpMatrix) > showClass(snpMatrix) > library(GenABEL) > showClass(scan.gwaa) It is more informative with the following commands > class?snpMatrix > class?scan.gwaa Later we will omit the command prompt (>). We will also give examples of creating object with new() function.

Example MC4R SNPs and BMI

To make a smooth exposition we use our study of SNPs near MC4R and body mass index as reported by Loos et al. Nat Genet 2008;40:768-775. The MC4R gene is located on chromosome 18 and we will focus on SNPs rs17782313 and rs17700633 at positions 56002077 and 57000671 according to NCBI build 35, all genotypes being on forward strand. These were based on 3850 population-based individuals at stage 1 of the case-cohort study from which 3552 individuals remained after quality controls. We will run through SNPassoc, snpMatrix and GenABEL packages on data as contained in files mc4r.ped, mc4r.map and mc4r.csv

SNPassoc
library(SNPassoc) map <- read.table("mc4r.map",sep="\t",as.is=TRUE) info <- data.frame(snp=map[2],chr=map[2],pos=map[4]) ped <- read.table("mc4r.ped",sep="\t",as.is=TRUE) names(ped) <- c(paste("v",1:6,sep=""),map[,2]) pheno <read.csv("mc4r.csv",sep="\t",skip=11,header=TRUE,as.is=TR UE) is.cohort <- pheno$cohort==1 cohort <- subset(pheno,is.cohort) snp <- ped[,-c(1:6)][is.cohort,] snps <- dim(snp)[2] for(i in 1:snps) { substr(snp[,i],2,2) <- "/"; empty <- (snp[,i]=="0/0"); snp[empty,i] <- NA }

Analysis
mc4r <- setupSNP(snp,1:snps,sep="/",sort=TRUE,info=info) summary(mc4r) plot(mc4r$rs17782313) plot(mc4r$rs17700633,type=pie) hwe <- tableHWE(mc4r) mc4r.ld <- LD(mc4r) summary(mc4r.ld) mc4r.ld$"R^2" attach(cohort) association(bmi ~ sex+age+rs17782313,data=mc4r) wga <- WGassociation(bmi ~ sex+age+1,model="logadd",data=mc4r) png("mc4r.png") qqpval(wga$"log-additive") dev.off()

Summary statistics for two SNPs

Q-Q plot

Comments
SNPassoc is essentially designed for dealing with unrelated individuals but with considerable enhancements from genetics and haplo.stats. It implements permutation tests for binary traits through scanWGassociation(,nperm=) and permTest() It is possible to conduct gene-gene interaction: mc4r.ip <interactionPval(bmi~sex+age,data=mc4r,model="lo g-add") plot(mc4r.ip) We got a very good feel of the kind of analysis it may involve and this is a very simple example.

snpMatrix
library(snpMatrix) mc4r <- read.snps.pedfile("mc4r.ped") summary(mc4r) mc4rsnps <- row.names(mc4r$snp.support) head(mc4r$snp.support) head(mc4r$subject.support) # quality controls mc4r.qc <- summary(mc4r$snp.data) head(mc4r.qc) mc4r.ld <- ld.snp(mc4r$snp.data) plot.snp.dprime(mc4r.ld,"mc4r.eps",scheme="rsq") # ps2pdf mc4r.eps # xpdf mc4r.pdf # LD(rs17782313, rs17700633) mc4r$snp.support[c(1,12),] pair.result.ld.snp(mc4r$snp.data,1,12)

PCA and identity-by-state analysis

# PCA mc4r.xxt <- xxt(mc4r$snp.data,correct.for.missing=TRUE) mc4r.pc <- eigen(mc4r.xxt, symmetric=TRUE) loadings <- snp.cor(mc4r$snp.data, mc4r.pc$vectors[,1:10]) # identity-by-state analysis mc4r.ibs <- ibs.stats(mc4r$snp.data) mc4r.count <- ibsCount(mc4r$snp.data) mc4r.dist <- ibsDist(mc4r.count) mc4r.clust <- hclust(mc4r.dist) plot(mc4r.clust) Note this is based on XXT, in an order of N2, where X and N are the genotype data matrix and number of individuals (see vignette for details).

Phenotype data and case-control analysis

# Phenotype data pheno <- read.csv("mc4r.csv",skip=11,sep="\t") pheno$cc<-ifelse(pheno$bmi>=30,1,0) attach(pheno) # Case-control analysis of all individuals cc.test <- single.snp.tests(cc,snp.data=mc4r$snp.data) summary(cc.test) class(cc.test) showClass("snp.tests.single") chi.squared(cc.test,1) #qq.chisq(cc.test@chisq[,1])

Meta-analysis
# a meta-analysis cc.test <single.snp.tests(cc,snp.data=mc4r$snp.data,score=TRU E) cc.test2 <- pool(cc.test,cc.test) summary(cc.test2) cc.test.sign <- effect.sign(cc.test) table(cc.test.sign) cc.test.sign[1:12] cc.test.switch <- switch.alleles(cc.test,c(1,12)) effect.sign(cc.test.switch)[1:12]

OLS estimation and retrospective analysis

# ordinary least squares estimates reg.rhs <snp.rhs.tests(bmi~sex+age,family="gaussian",subset=(cohort==1),s np.data=mc4r$snp.data) class(reg.rhs) showClass("snp.tests.glm") reg.rhs@df qq.chisq(reg.rhs@chisq) print(reg.rhs) # retrospective models reg.lhs <snp.lhs.tests(mc4r$snp.data,~bmi,~sex+age,subset=(cohort==1)) class(reg.lhs) showClass("snp.tests.glm") reg.rhs@df qq.chisq(reg.lhs@chisq,df=2)

Genotype imputation
It is customarily to impute genotypes in a large study based on a small sample of fully-genotyped individuals, e.g., hapmap, so as to conduct association tests for large number of SNPs. It is also useful for meta-analysis of SNPs from different platforms such as Affymetrix 500K and Illumina 550K. As it is snpMatrix implements genotype imputation between sets of markers based on same individuals; more generally this involves genotypes from HapMap.

Hapmap and imputation

# ideally we would use 60/90 founders and a combination of hapmap CEU and our study sample url.p1 <- "ftp://ftp.hapmap.org/hapmap/genotypes" url.p2 <- "/latest_ncbi_build35/fwd_strand/non-redundant/" url.p3 <- "genotypes_chr18_CEU_r21a_nr_fwd.txt.gz" hapmap <- paste(url.p1,url.p2,url.p3,sep="") chr18 <- read.HapMap.data(hapmap) chr18snps <- row.names(chr18$snp.support) summary(chr18) sel <- chr18snps%in%mc4rsnps impute.from <- chr18$snp.data[,!sel] impute.to <- chr18$snp.data[,sel] pos.from <- chr18$snp.support$Position[!sel] pos.to <- chr18$snp.support$Position[sel] mc4r.imp <- snp.imputation(impute.from, impute.to, pos.from, pos.to) summary(mc4r.imp) plot(mc4r.imp)

QC for chromosome X
# we omit the X.ped data/map here owing to their size X <- read.snps.pedfile("X.ped",X=TRUE) X.qc <- summary(X$snp.data) X.col <- col.summary(X$snp.data) SNPs <- subset(X.col, Call.rate>=0.90&MAF>=0.01&z.HWE>=1e-6) write.csv(row.names(SNPs), "X.snps", quote=FALSE, row.names=FALSE) library(foreign) write.dta(X.col,"Xqc.dta")

By default, X.map is called which contains lines as follows

X X X SNP_A-1787762 SNP_A-1788139 SNP_A-1789223 0 0 0 148021903 135986846 5694766

snp.imputation
It is notable with the definition of snp.imputation that given two set of SNPs typed in the same subjects, this function calculates regression equations which can be used to impute one set from the other in a subsequent sample. We customarily use external data (e.g., available from HapMap, 1000 genomes or elsewhere) and our sample jointly, treating non-typed SNPs as missing. CRAN packages such as mice should facilitate this on the phenotype side.

Comments
snpMatrix has explicit treatment of chromosome X. It also provides some facilities for dealing with family data. The retrospective method would be more appropriate with data involving the kind of sample selection here. Please check for snpMatrix vignette for use of hexbin package. It is possible to take advantage of the S4 class facility as implemented in the package when coded genotypes are available from or to other sources, e.g., m1 <- new(snp.matrix,dm1) m2 <- new(snp.matrix,dm2) m <- snp.rbind(m1,m2) write.snp.matrix(m,m.dat)

GenABEL: Flowchart (Q Zhang from WUSTL)

phdata: phenotypic data (data frame) gtdata: genotypic data (snp.data-class) snp.data()

gwaa.data-class

nbytes: number of bytes used to store data on a SNP nids: number of people male: male code idnames: ID names nsnps: number of SNPs nsnpnames: list of SNP names chromosome: list chromosomes corresponding to SNPs coding: list of nucleotide coding for SNP names strand: strands of the SNPs map: list SNPs positions 2-bit storage gtps: genotypes (snp.mx-class) 0 00
1 2 3 Save 01 10 11 75%

load.gwaa.data(phenofile = "pheno.dat", genofile = "geno.raw)

convert.snp.text() from text file (GenABEL default format) convert.snp.ped() from Linkage, Merlin, Mach, and similar files convert.snp.mach() from Mach format convert.snp.tped() from PLINK TPED format convert.snp.illumina() from Illumina/Affymetrix-like format

Data manipulation
snp.subset: subset data by snp names or by QC criteria add.phdata: merge extra phenotypic data to the gwaa.data-class. ztransform: standard normalization of phenotypes rntransform: rank-normalization of phenotypes npsubtreated: non-parametric adjustment of phenotypes for medicated subjects

QC and summary statistics

summary.snp.data: summary of snp data (Number of observed genotypes, call rate, allelic frequency, genotypic distribution, P-value of HWE test check.trait: summary of phenotypic data and outlier check based on a specified p/FDR cut-off check.marker: SNP selection based on call rate, allele frequency and deviation from HWE HWE.show: showing HWE tables, Chi2 and exact HWE Pvalues perid.summary: call rate and heterozygosity per person ibs: matrix of average IBS for a group of people & a given set of SNPs hom: average homozygosity (inbreeding) for a set of people, across multiple markers

SNP association scans

scan.glm performs snp association test, e.g., scan.glm((y~x1+x2++CRSNP", family = gaussian(), data, snpsubset, idsubset). scan.glm.2D:2-snp interaction scan. ccfast: case-control association analysis by computing chisquare test from 2x2 (allelic) or 2x3 (genotypic) tables. emp.ccfast obtains Genome-wide significance (permutation) for ccfast scan. qtscore: association test (GLM) for a trait (quantitative or categorical) emp.qtscore() is genome-wide significance (permutation) for qscore() scan. mmscore: score test for association between a trait and genetic polymorphism, in samples of related individuals (needs stratification variable, scores are computed within strata and then added up). egscore: association test, adjusted for possible stratification by principal components of genomic kinship matrix (snp correlation matrix).

Haplotype association scans

scan.haplo: haplotype association test using GLM in R library scan.haplo.2D: 2-haplotype interaction scan

Results as in scan.gwaa class

from scan.glm, scan.haplo, ccfast, qtscore, emp.ccfast,emp.qtscore Names: snpnames list of names of SNPs tested P1df: p-values of 1-d.f. (additive or allelic) test for association P2df: p-values of 2-d.f. (genotypic) test for association Pc1df: p-values from the 1-d.f. test for association between SNP and trait; the statistics is corrected for possible inflation effB: effect of the B allele in allelic test effAB: effect of the AB genotype in genotypic test effBB: effect of the BB genotype in genotypic test Map: list of map positions of the SNPs Chromosome: list of chromosomes the SNPs belong to Idnames: list of subjects used in analysis Lambda: inflation factor estimate, as computed using lower portion (say, 90%) of the distribution, and standard error of the estimate Formula: formula/function used to compute p-values Family: family of the link function / nature of the test

Table and graphics

descriptives.marker(): descriptives.trait(): descriptives.scan(): plot.scan.gwaa(): plot.check.marker(): table of marker info. table of trait info. table of scan results plot of scan results plot of marker data (QC etc.)

ParallABEL
An R Library for Generalized Parallelization of GenomeWide Association Studies

http://parallabel.r-forge.r-project.org/ http://www.sci.psu.ac.th/units/genome/CGBR/ParallABE L/index.html Sangket et al. BMC Bioinformatics 2010; 11:217

Applied to MC4R data

library(GenABEL) convert.snp.ped("mc4r.ped","mc4r.map2","mc4r.out",strand="+ ") csv <- read.csv("mc4r.csv",skip=11,sep="\t",as.is=TRUE) attach(csv) csv2 <- data.frame(id,sex=2-sex,cohort,age,bmi,zbmi,rbmi) write.table(csv2,"mc4r.csv2",sep=" ",row.names=FALSE) mc4r <- load.gwaa.data(phe = "mc4r.csv2", gen = "mc4r.out", force = TRUE) Note that the map2 file now has three columns: chromosome, SNP names and positions. It also explicitly allow for strand. The addition of phenotypic information is via the load.gwaa.data, which requires specification of id and sex (0=female, 1=male) in a strictly way.

Analysis
HWE.show(mc4r) r2 <- r2fast(mc4r) dp <- dprfast(mc4r) rho <- rhofast(mc4r) descriptives.trait(mc4r) descriptives.marker(mc4r) use <- csv$cohort==1 qt.bmi <- qtscore(bmi~sex+age,data=mc4r,idsubset=use) plot(qt.bmi) However, as is shown here once the gwaa.data is defined a range of analyses can be rather straightforward. Again we only focus on the cohort sample (cohort==1).

Scatter plot of p values

GAW15 Expression quantitative trait

There is substantial individual variation in expression level of genes, which is smaller in monozygotic twins than among individuals of other relationships, suggesting a genetic component (Morley M, Molony CM, Weber TM, Devlin JL, Ewens KG, Spielman RS, Cheung VG: Genetic analysis of genome-wide variation in human gene expression. Nat 2004, 430:743-747). Genetic Analysis Workshop 15 problem 1 provided 14 three-generation families 2554 expression quantitative traits 2882 SNP genotypes Chromosomal positions of these SNPs These information was contained in comma-delimited files each with appropriate header. This simple example serves to illustrate the basic analysis involved.

Getting data into R

We first got data into R,
id <- read.table("LINKAGE.PED",header=T,as.is=T,sep=",") phn <read.table("LINKAGE.PHN",header=T,as.is=T,sep=",") snp <read.table("LINKAGE.SNP",header=T,as.is=T,sep=",",na.st ring="0/0") map <read.table("LINKAGE.MAP",header=T,as.is=T,sep=",") pheno <- merge(id,phn,by=c("FAMID","ID")) ped <- merge(pheno,snp,by=c("FAMID","ID"))

Now the object ped has all the necessary information. We omit details of association testing but pedigree diagrams.

Pedigree diagrams
library(kinship) pdf("pedfile.pdf"); attach(ped) uid <- unique(ped$FAMID) for (j in 1:length(uid)) { selected <- FAMID==uid[j] id <- ID[selected] dadid <- FA[selected] momid <- MO[selected] sex <- SEX[selected] par(xpd=TRUE) ped <- pedigree(id, dadid, momid, sex) plot(ped, id=paste(\n,id,sep=)) title(uid[j]) k <- kinship(id,dadid,momid) print(k) } detach(ped); dev.off()

A typical pedigree diagram

GAW16 Framingham data

Data management through SAS QC and basic association statistics via PLINK Estimation of inflation factor by snpMatrix Cross-check with GRAMMAR procedure from R/GenABEL library(GenABEL) # this is an example of Framingham data for GAW16 convert.snp.tped(tped = "chrall.tped", tfam = "pheno.tfam", out = "chrall.raw", strand = "+") df <- load.gwaa.data(phe = "pheno.dat", gen = "chrall.raw", force = TRUE) Longitudinal data with SAS, Stata and Mplus. Rpackages include gee, nlme and packages which handle family data, e.g., kinship, GWAF, pedigreemm. Graphics via R/gap

References
Elston RC. Introduction and overview. Stat Meth Med Res 9(6, special issue), 2000 Balding DJ. Nat Rev Genet 7:781-791, 2006 Elston RC, Anne Spence M. Stat Med 25:3049-3080, 2006 McCarthy MI, Abecasis GR, Cardon LR, Goldstein DB, Little J, Ioannidis JPA, Hirschhorn JN. Genome-wide association studies for complex traits: Consensus, uncertainty and challenges. Nat Rev Genet 9:356369, 2008 Zheng G, Marchini J, Geller NL. Introduction to the special issue: Genome-wide association studies. Stat Sci 24(4, special issue), 2009

Appendix- GWAS with SAS and Stata

Procedures in SAS/BASE and other modules provide graphics, database support and internet connectivity. SAS/STAT provides standard procedures including linear and logistic regressions or generalized linear (nonlinear, mixed ) model as well as covariance and linear structure model (CALIS), MULTTEST. SAS/Genetics includes procedures for summarizing marker data (ALLELE), inferring and tagging haplotypes (HAPLOTYPE and HTSNP), association testing in population-based (CASECONTROL) and family-based (FAMILY) samples.

The transposed data format

rs17782313 rs8097644 rs9947403 rs639407 rs11665563 rs11663816 rs619662 rs727406 rs8089366 rs11152217 rs9955666 rs17700633 rs9946888 rs9961245 rs17066774 TT CC CC AA CC TT GG GG GG GG GG GG TT CC GG CT CC TT GG CT CT AA GG GT GT AG AG CT CT GG TT AC CC AA CC TT AG GT GG GG GG GG CT CT GG TT CC CC AA CC TT GG GG GG GG GG GG CT CT GG TT CC CC AA CC TT GG 0 GG GG GG GG TT CC GG TT CC CC AA CC TT GG 0 GG GG GG GG CT CT GG TT CC CC AA CC TT GG GG GG GG GG AG CT CT GG CC CC TT GG TT CC AA GG TT GG AA AA TT CC GG

Data preparation
data long (keep=&snpid id &vlist a1a2 add n); set data; fid=open("data"); length id $11. add 3. a1a2 $3.; format add 1.; set map point=_n_; n=0; do col=2 to attrn(fid,"nvars"); iid=col-1; set &trait (keep=&vlist) point=iid; if &inc=1 then do; id=varname(fid,col); a1a2=vvaluex(id); add=.; if a1a2 ne " " then do; a1=substr(a1a2,1,1); a2=substr(a1a2,3,1); add=(a1=b)+(a2=b); n+1; end; output; end; end; rc=close(fid); run;

Analysis
ods select none; proc allele data=long genocol; by rsn notsorted; var a1a2; ods output markersumm=ms allelefreq=out.af; run; proc reg data=long; by rsn notsorted; ods output parameterestimates=bmipm; model bmi = age add / b stb; quit; proc logistic data=long descending; by rsn notsorted; ods output parameterestimates=obpm CLOddsPL=obclpm; model obesity = age add / expb clodds=pl; run;

Stata
It is a general-purpose, modern and easy to use statistical analysis system (e.g., http://en.wikipedia.org/wiki/Stata). Functions for genetic data includes summary statistics, test of Hardy-Weinberg equilibrium, haplotype estimation, tagging and association analysis. It allows for C/C++ routines to be used for computer intensive tasks. My colleague has implemented SNPTEST-based GWA analysis to automate a variety of sample and analyses for imputed genotypes. There is also a good implementation for meta-analysis (metan, etc), as with a set of functions for instrumental variable regressions in our context.

Programs by David Clayton

ginsheet- Read genotype data from text files. gloci - Make a list of loci. greshape - Reshape a file containing genotypes to a file of alleles. gtab - Tabulate allele frequencies within genotypes and generate indicators (performs Hardy-Weinberg Equilibrium testing). gtype - Create a single genotype variable from two allele variables. htype - Create a haplotype variable from allele variables. mltdt - Multiple locus TDT for haplotype tagging SNPs (htSNPs). origin - Analysis of parental origin effect in TDT trios. pseudocc - Create a pseudo-case-control study from case-parent trios. pscc - Experimental version of pseudocc in which there may be several groups of linked loci. pwld - Pairwise linkage disequilibrium measures. rclogit - Conditional logistic regression with robust standard errors. snp2hap - Infer haplotypes of 2-locus SNP markers. tdt - Classical TDT test. trios - Tabulate genotypes of parent-offspring trios.

Programs by Adrian Mander

gipf - Graphical representation of log-linear models. hapipf - Haplotype frequency estimation using EM algorithm and log-linear modelling. pedread - Read pedigree data file (in pre-Makeped LINKAGE format), similar to ginsheet pedsumm - Summarises a pre-Makeped LINKAGE file. pedraw - Draws one pedigree in the graphics window plotmatrix - Produces LD heatmaps displaying graphically the strength of LD between markers. profhap - Calculates profile likelihood confidence intervals for results from hapipf swblock - A step-wise hapipf routine to identify the parsimonious model to describe the Haplotype block pattern. qhapipf - Analysis of quantitative traits using regression and log-linear modelling when phase is unknown. hapblock - attempts to find the edge of areas containing high LD within a set of loci

Other programs
By Mario Cleves gencc - Genetic case-control tests genhw - Hardy-Weinberg Equilibrium tests qtlsnp - A program for testng associations between SNPs an a quantitative trait. By Catherine Saunders co_power - Power calculations for Case-only study designs. gei_matching geipower - Power calculations for Gene-Environment interactions. ggipower - Power calculations for Gene-Gene interactions. tdt_geipower - Power calculations for Gene-Environment interactions via TDT analysis. tdt_ggipower - Power calculations for Gene-Gene interactions via TDT analysis. By Neil Shephard genass - Performs a number of statistical tests on your genotypic data and collates the results into a Stata formatted data set for browsing.

Programs for GWAS

By Chuck Huber phasein/phaseout input/output with PHASE haploviewin/haploviewout input/output with HAPLOVIEW By Jianan Luan qc genomic control using p values gwa genomewide analysis using SNPTEST By Jing Hua Zhao stata_snphwe a Stata plugin for exact test of HardyWeinberg equilibrium using genotype counts

III Miscellaneous topics

Topics

Meta-analysis Risk prediction Instrumental variable method and structural equation modeling Gaussian graphical models and networks Extreme value modeling

Meta-analysis
Some circulations within the GIANT consortium considered two studies with sample sizes 32000 and 8000 both with p values 1e-8, we have a combined twosided p value of 1.49e-14 but also yields p=4.89e-8 with p1=1e-4 and p2=1e-5 (weighted z-score method from metap in gap). In general, it statistically combines data from multiple studies in the consortium to learn about association (level of significance) and factors related to variations in its magnitude (effect size). We have test of significance = size of effect x size of study, e.g., 12=r2N (Kramer & Rosenthal. Comprehensive Clinical Psychology 3-15, Elsevier 1998)

Combining independent tests

Fishers method One can use truncated p values Stouffers method is based on normal approximation. The R implementation is straightforward with sum(-2 * log(pvalues))) and sum(qnorm(1pvalues)) / sqrt(k)).

2 2k

= 2 ln P i = 1,..., k ,
i

z =1

K (1 P )
k 1 i =1 i

Fishers method has limitations in Giving equal weight to studies with different sizes No test of heterogeneity No point estimate to become more precise as K increases However, there is suggestion about bias regarding msSNP.

Regression models for meta-analysis

Fixed effects model is unable to account for heterogeneity since deviations from i and are assumed to be explained by random error. Random effects model. It is assumed that each study has its own effect distribution against a common distribution. The popular DerSimonian-Laird (DL, moment) estimator equates the expectation of the heterogeneity statistic. We can include covariates in the model to make study-specific adjustments, i.e., meta-regression. Simple heterogeneity (SH) model uses GLS with strictly positive variance estimate.

= + , i = 1,..., k ,
i i

~ N (0, )
2 i i

= + b + , i = 1,..., k ,
i i i

b ~ N (0, ), ~ N (0, )
2 2 i i i

( ) ( k 1) =
k 2 2 i =1 k i i 2 k 4 k 2 i =1 i i =1 i i =1 i

= + z + b + , i = 1,..., k ,
i 1 i i i

b ~ N (0, ), ~ N (0, )
2 2 i i i

Var ( ) = + = (r + 1) E ( ) = X , Var ( ) = V
2 2 2 i i 2

Measure of Heterogeneity
Cochrans Q, Q = ik=1 ( i ) 2 / i , can be referred to a chisquared distribution with k-1 degrees of freedom.
2

I2, defined as 100%(Q-df)/Q, which expresses the percentage of between-study variability that is attributable to heterogeneity rather than chance. Thresholds of 20%, 50%, and 75% are suggested to have low, moderate and high heterogeneity (Higgins et al. BMJ 2003; 327:57-60). It has been suggested that cQ~x2(v) with Q being heterogeneity chi-square, has excellent property (Bohning et al. 2008).

Implementations
SAS has no built-in procedure for meta-analysis but can customarily done via PROCs GLM (fixed effects/inverse variance) and more often MIXED as well as macros. Stata has a comprehensive collection of meta-analysis, notably metan. R hosts several package at CRAN (e.g., meta, rmeta) . S-PLUS has user-written packages, e.g., hblm. Others such as HLM, MLwiN, WinBUGS. Customized programs

Useful URLs
CAMAN (Computer Assisted Analysis of Mixtures) http://www.charite.de/biometrie/schlattmann/book/ improved.ci (function for the improved confidence interval using DL method) http://www.statistik.tu-dortmund.de/ma_book.html hblm (Hierarchical Bayes Linear Model Programs) ftp://ftp.research.att.com/dist/bayes-meta/ CAMAP (Computer-Assisted Meta-Analysis with the Profile Likelihood) http://www.personal.reading.ac.uk/~sns05dab/Softwar e.html

Fixed-effects meta-analysis
data test; input studyid lor est; col=_n_; row=_n_; value=est; cards; data for 15 studies run; proc mixed method = ml data=test; class studyid; model lor = / s cl; repeated / group = studyid; parms / parmsdata=test eqcons=1 to 15; run;

Random-effects meta-analysis
proc mixed data=test covtest; class studyid; model lor = / s cl outp=predp outpm=predm; repeated diag / r; random studyid / g gdata = test s v; ods output CovParms=cp G=G R=R V=V SolutionF=SF SolutionR=SR; run; data predp; set predp; pvalue=probnorm(resid/stderrpred); run; data predm; set predm;pvalue=probnorm(resid/stderrpred); run;

Stata
use meta5 list in 1/5 metan b se, by(snp) fixedi nograph

WinBUGS
model { for (i in 1:r) { y[i] ~ dnorm(psi[i],w[i]) psi[i] ~ dnorm(theta,t) } theta ~ dnorm(0,1.0E-4) t ~ dgamma(0.001,0.001) tausq <- 1/t } list(y = c(0.864, 0.646, 0.272, 0.916, 0.867, 0.819, 0.809, 1.212, -0.273), w = c(4.40, 9.89, 16.81, 8.38, 8.15, 10.36, 10.79, 4.40, 15.95), r = 9) list(theta = 0, t = 1, psi = c(0,0,0,0,0,0,0,0,0))

R/meta, R/rmeta, R/CAMAN

library(CAMAN) data(aspirin) aspirin mix <- mixalg(obs="logrr", var.lnOR="var", data=aspirin) library(rmeta) attach(aspirin) annotate <- cbind(name,year) metaplot(logrr,se,labels=annotate) library(meta) mg <- metagen(logrr,se) plot(mg) funnel(mg) metabias(mg, method=linreg)

R/meta and R/metafor with by

library(foreign) setwd(".") meta5 <- read.dta("meta5.dta") attach(meta5) library(meta) s <- by(meta5,snp,function(x) metagen(b,se,data=x)) names(s) names(s$rs998663) library(metafor) ss <- by(meta5,snp,function(x) rma(b,se,data=x)) names(ss$rs998663) # Forest, Funnel, Radial and Residual plots plot(ss$rs998663)

Customized programs
META METAL MetABEL R/snpMatrix

A cautionary note
In a meta-analysis, we compute effect size for each study and combine them but not combine summary data and compute an effects size for the combined data. This allows for a check of consistence regarding effect sizes across studies and minimizes the potential confounders. If we were to pool data across studies and then compute the effect size from the pooled data, we may get the wrong answer, due to Simpons paradox. See Chapter 13 of Borenstein M, Hedges LV, Higgins JPT, Rothstein HR. Introduction to Meta-Analysis. Wiley 2009

Extensions: multivariate Meta-Analysis

Background A gene-based association testing (Neale & Sham) not dissimilar to the usual Fisher p value method Multilocus scan statistics (Hoh & Ott) not taking off Bayesian meta-analysis is more involved and the formulation via summary data as in Verzilli et al. is not not necessarily used. P values adjusted for correlated tests (p_ACT, Conneely & Boehnke) addresses the following question: What is the minimum p value and more importantly given it is obtained what are the significant levels for all others? Problems Covariance of association tests can be poorly estimated given multicollinearity between SNPs at a region/gene.

Statistical models
The data typically involve b, SE from linear regression of nearby SNPs to allow for fixed- and random effects modeling and assessment of statistical significance. It is not obvious how to infer covariance matrix involving these bs. However, we can work around with respect to pair-wise correlations (r). For linear regression, it is known that r and t (=b/SE) is related via a simple expression r2=t2/(n-2+t2). The covariance between pair-wise correlation has the following form.

Covariance between pairs of correlations

Elston RC (1975). On the correlation between correlations. Biometrika 62:133-40

Combination of SNPs via GLS

The results of k independent studies, each with p correlations, can be expressed as the concatenation of the vectors of all available correlations. The large sample variance-covariance matrix is then block diagonal. The estimation of the pooled correlation matrix can then be done via weighting or via a generalized least squares (GLS) framework. A test of homogeneity of correlation matrices among studies can be performed (Becker 1992). We can accommodate the heterogeneity via a random effects model such that population correction for specific study is a result of the population correlation and study specific factor. The implementation (e.g., in R) accounts for variable number of SNPs from each study (Verzilli et al. 2008).

p_ACT and p_ACT_meta

p_ACT is based on multivariate normal (MVN) assumption originally for sample with individual genotypes but recently extended to results from consortium meta-analysis. The basic idea with p_ACT_meta is to find the minimum p value from the collection of correlated SNPs and obtain subsequent p values based on MVN conditional distributions (Holms procedure) using R/mvtnorm. It uses a James-Stein shrinkage estimate as implemented in R/corpcor. A description of mvtnorm appears in The R Journal. However, the omnibus approach noted earlier is appealing.

Summary
It is far from a comprehensive overview but offers some flavour of the kind of thinking and practice. Evidence synthesis with conscious recognition of heterogeneity is in the heart of meta-analysis. Fixed effects analysis is restricted to data of the type found in the studies included, but random effects model generalizes to all studies of the type from which our studies were drawn. Results from both models together with SH model are highly recommended. We have omitted the graphical aspects, e.g., Bax et al. AJE 2009; 169:249-55. An Excel macro is available from http://www.mix-for-meta-analysis.info/index.html

Risk prediction
A set of SNPs can be used in a logistic regression model to predict if an individual is a case or control based on a cut-off probability. An optimal cut-off can be facilitated through receiver operating characteristics (ROC) curve. The ability to classify individuals correctly is measured by area under the ROC curve (AUC, e.g. ~0.5, 0.7-0.8, 0.8-1 for no, acceptable, excellent discrimination). Examples: prostate cancer, obesity, HDL/TG/LDL. A testing example library(verification) obs<- round(runif(100)) pred<- runif(100) A<- verify(obs, pred, frcst.type = "prob", obs.type = "binary") roc.plot(A, main = "Test", binormal = TRUE, plot = "both") roc.plot(A, threshold=seq(0.1,0.9, 0.1), CI=TRUE, alpha=0.1) roc.plot(obs,pred,xlab=1-specificity',ylab='sensitivity',cex=2) AUC <- roc.area(obs,pred)$A

Risk score and BMI in EPIC-Norfolk

=0.15 kg m-2 /allele p=1.54E-22
2500 28

=0.36 cm /allele p=1.10E-18

2500 95

2000 27

2000

Waist (cm)

BMI (kg/m2)

Frequency

1500 26 1000

Frequency

90 1500

1000 85

25 500
500

0 6 7 8 9 10 11 12 13 14 15 16 17

24

0 6 7 8 9 10 11 12 13 14 15 16 17

80

Genetic Risk Score

Genetic Risk Score

Risk score and obesity/overweight

10

Odds ratio

p for trend = 1.03E-12

0.1 6

7 2.6

8 6.6

9 11.1

10 15.6

11

12

13 12.8

14 8.2

15 4.4

16 1.9

17 1.0

%: 1.4

17.9 16.6

Genetic risk score and its proportion in the sample (%)

ROC curve and AUC

1 0.9 0.8 0.7

Model 1 Model 2 Model 3 Null

Sensitivity

0.6 0.5 0.4 0.3 0.2

AUC:
0.1 0 0 0.1

Model 1 (Age + Age2 + Sex): 0.572 (95% CI: 0.560-0.584) Model 2 (SNPs): 0.574 (95% CI: 0.559-0.590) Model 3 (Age + Age2 + Sex + SNPs): 0.597 (95% CI: 0.582-0.612)
0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

1-specificity

Instrumental variable (IV) estimation

It is a method for estimating regression Y=(Z X)b+e parameters b when X are measured with error, W=X+U, and possibly when a second or biased but independent measurement (T) is available. Given cov(T,e) = cov(T,U)=0, cov(T,X)<>0, b=cov(T,Y)/cov(T,W). More formally, 1. T is uncorrelated with X; 2. T is independent of the measurement error U=W-X in the surrogate W; 3. (W,T) is a surrogate for X so that E(Y|Z,X,W,T)=E(Y|Z,X). See Fuller WA. Measurement Error Models. Wiley 1987; Greene WH. Econometric Analysis, 5e. Prentice Hall 2003; Carroll et al. Measurement Error in Nonlinear Models-A Modern Perspective, 2e. CRC 2006; Gelman A, J Hill. Data Analysis Using Regression and Multilevel/Hierarchical Models. Cambridge University Press 2007

IV in simple terms
In an observational study, U represents unmeasured confounders of the XY association. In a randomized trial, U represents variables that affect adherence to treatment assignment and thus influence received treatment X. Z is called an instrumental variable (or instrument) for estimating the effect of X on Y. Rothman KJ, Greenland S, Lash TL. Modern Epidemiology, 3e, Lippincott Williams & Wilkins 2008 a. Z affects X (i.e., Z is an ancestor of X). b. Z affects the outcome Y only through X (i.e., all directed paths from Z to Y pass through X). c. Z and Y share no common causes.

Instrumental-variables regression (IVLS)

We can generalize the model as Ynx1=Xnxppx1+, IVLS or two-stage least squares (2SLS) requires that (i) Xnxp and Znxq with n>qp. (ii) ZX and ZZ have full rank, p and q respectively. (iii) Y=X+. (iv) The i are i.i.d. with mean 0 and variance 2. (v) Z is exogeneous, i.e., Z is independent of . The cases with q>, =, and < p are called over-, just-, and under- identified, respectively. Solution to the system proceeds by multiplying Z on the Y-X model and rescaling by variance such that (ZZ)1/2ZY=(ZZ)1/2ZX + , where = (ZZ)1/2Z Freeman DA. Statistical Models-Theory and Practice, Revised Edition. Cambridge University Press, 2009.

FTO genotype, BMI and metabolic traits

There is epidemiological association between BMI and metabolic traits. There is association between FTO and BMI. The association between FTO genotype and metabolic traits would be mediated by BMI (c=axb).

This is the so-called triangulation approach (Freathy et al. Diabetes 2008; 57:1419-26).

Direct and indirect effects

We can lay out two equations We can plug in the second equation into the first. We proceed with two steps: 1. We first regress TG on SNP. 2. We also regress BMI on SNP. We then have the Wald estimate with = 0 A summary in our setting is Bochud et al. IJE 2008, 37:414-6
2

TG = + BMI + SNP + error BMI = + SNP + error

0 1 2 0 1

TG = + BMI + SNP + error = + ( + SNP ) + error

0 1 2 0 1 0 1

= ( + ) + ( + ) SNP + error
0 1 0 1 1 2

TG = + SNP + error
0 2

BMI = + SNP + error

0 2

= + = ( ) /
2 1 1 2 1 1 2

-1131T>C (rs662799), TG and CHD

1. 1131T>C, a regulatory variant in APOA5, is unrelated to several non-lipid risk factors or LDL cholesterol, and comparatively moderately related to HDL cholesterol and other major lipids. 2. 1131T>C is strongly related to TG concentration in a dose-dependent manner, with every C allele increasing TG by about as much as having type 2 diabetes mellitus. 3. in an analysis of 20 842 cases and 35 206 controls,1131T>C is related to risk of CHD in an analogous dose-dependent manner, with about 18% higher risk per C allele. 4. in an analysis of 302 430 people, risk of CHD with genetically raised TG is concordant with risk of disease with equivalent differences in circulating TG itself. 5.1131T>C is associated with higher VLDL concentration and smaller HDL particle sizepathways through which TG could affect risk of coronary heart disease. (Lancet 375:1634-9, 2010)

SLC2A9, urate levels and metabolic syndrome

This example was reported recently by McKeigue et al. Int J Epidemiol. 2010; 39:907-18 The data contains 583 individuals with sex, age and seven SNPs, one of which is non-synonymous and used as instrumental variable. The R package mediation only accepts data without missing values, so we used 493 individuals. The authors implemented a Bayesian logistic models and have applied JAGS and have argued in favor of this model over probit model.

Parameter with values 1 vs 0 yields lod score of 2.24

Issues with IV
No suitable genetic variant Unreliable gene association Population stratification Linkage disequilibrium Pleiotropy Nonlinear association Weak instrument

See Lawlor et al. (2007). Stat Med 27:1133-67; Didelez & Sheehan (2007). Stat Meth Med Res 16:309-30; Didelez et al. Stat Sci 2010. Pare & Anand (2010) Lancet 375:1584-5

Warnings against categorical data

Three models are involved with binary outcome (y), mediator (M), and predictor (X): y=i1+cX+e1, y=i2+cX+bM+e2, M=i3+aX+e3 such that when (c-c) is is employed, its standard error becomes more complicated than ordinary linear regression. It is often to set the residual variance in logistic regression to be 2/3 and probit regression to 1. The mathematical tractability of multivariate probit distribution makes it appealing in modeling categorical variable with Mplus. A single formula to standardize according to estimates from y=i1+cX+e1. See MacKinnon DP. Introduction to Mediation Analysis. Lawrence Erlbaum Associates, 2008

Mediation analysis
X Z X Y Z X Z X Y Z X
Scenarios of mediation: complete (upper left), partial (lower left), complete (upper right) and partial (lower right) with two mediators

X Y

The SLC2A9 example

library(foreign) snp <- read.dta(mediate.dta) library(mediation) B=lm(x~nsg+sex+age+rs3766404+rs6677604+rs132942 8+rs11582939+CFHR3R1del+rs7517126,data=snp) c=glm(y~x+sex+age+nsg+rs3766404+rs6677604+rs132 9428+rs11582939+CFHR3R1del+rs7517126, family=binomial(link="logit"),data=snp) logitm <- mediate(b, c, sims=10000, treat="nsg", mediator="x") summary(logitm) We obtain comparable results with probit link.

Results
Quasi-Bayesian Confidence Intervals Mediation Effect: -0.006834 95% CI -0.022355 0.002811 Direct Effect: -0.1205 95% CI -0.19597 -0.02652 Total Effect: -0.1273 95% CI -0.20195 -0.03293 Proportion of Total Effect via Mediation: 0.04556 95% CI 0.02904 0.15595

Structural equation modeling

Several examples seen in recent GWAS literature can be modeled via path analysis or put in this framework. It is typically confirmatory based on model-fitting. It has been a rather useful device to study causal relationship. It is natural to study change using longitudinal data. sem package in R is a very good initiative, but it is often necessary to resort to other systems such as EQS, AMOS, Mplus, e.g., the inter-relationship between anthropometric measurements using Mplus. A critique is that SEM relies on conditional independence assumptions with IV being as a special case, so that the assumptions required for causal effects are difficult to satisfy. It is helpful to examine equivalent models.

Mplus code
Model: Title: zltg on zlbmi; snp1: rs1121980 from FTO zlbmi on snp1; snp2: rs17782313 from MC4R zltg on snp1; zlbmi : BMI Model indirect: zlwst : waist zltg ind snp1; zltg : Triglycerides Output: zsys : SBP Standardized; zdia : DBP Data: File is effectsize.dat ; Variable: Names are snp1 snp2 zlbmi zlwst zltg zsys zdia; Missing are all (-9999) ; Usevariables are snp1 zlbmi zltg;

Two mediator model

Mplus for two mediator model

TITLE: two mediator example; DATA: NOBS = 400; NGROUPS = 1; FILE IS mediate2.dat VARIABLE: NAMES ARE ID x m1 m2 y; USEVARIABLES ARE x m1 m2 y; ANALYSIS: TYPE IS GENERAL; ESTIMATOR IS ML; ITERATIONS = 1000; CONVERGENCE = 0.000001; MODEL: y ON m1 m2 x; m1 ON x; m2 ON x; m1 with m2; MODEL INDIRECT; y IND x;

More complicated scenarios

X1 X2

Y1 Y2

When Y1 becomes X2 and X2 becomes Y2, the cross-lagged model can be used to study reverse causation, especially with longitudinal data. It becomes clear that we will be most comfortable with the SEM framework, as is also illustrated with the following slide.

Latent mediator model

Bayesian networks
Rule-based systems with certainty factors have serious limitations as a method for knowledge representation and reasoning under uncertainty, and attention towards a probabilistic interpretation of certainty factors leads to Bayesian networks. It can be described briefly as an acyclic directed graph (DAG) which defines a factorization of a joint probability distribution over the variables represented by the nodes of the DAG. The process of construction involves identification of the relevant variables and their causal relations, which leads to DAG specified in terms of a set of conditional probabilities.

Example-GAW15 Problem 1 data

It was a published data (Morley et al Nature 2004, 430: 743-74) on baseline expression levels of 8793 genes in immortalised B cells from 194 individuals in 14 CEPH pedigrees, shown to have linkage and association and evidence of substantial individual variations. In particular, correlation was examined on expression levels of 31 genes and 25 target genes corresponding to two master regulatory regions. We apply Bayesian network analysis to gain further insight into these findings. If the expression level of a given gene is regulated by certain proteins then it should be a function of the active levels of these proteins. Due to biological variability and measurement errors, the function would be stochastic rather than deterministic. Expression levels of genes are proxies for the activity level of the proteins they encode, although there are numerous examples where activation or silencing of a regulator is carried out by post-transcriptional protein modifications

Methods
Gene expression levels as continuous variables were assumed to follow a multivariate normal distribution, and consistent with a Bayesian network with linear Gaussian conditional densities. The prior of this network is characterised by a prior network reflecting our belief in the joint distribution of the variables in question, and equivalent sample size (ESS) effectively behaving as if it was calculated from a prior data set of that size. For instance, without a priori knowledge of the regulatory network, the prior network could be one where all expression levels are independent in order to avoid explicitly biasing the learning procedure to a particular edge. The learning procedure starts with a training set and evaluates networks according to an asymptotically consistent scoring function that is obtained through the Bayesian framework. The so-called causal structure assumes that dependencies between variables are due to causal relationships between variables in the model.

Left. Importance of the dependencies. Right. Solid arc has direct causal influence (direct meaning that causal influence is not mediated by any other variable that is included in the study). Dashed arc indicates there are two possibilities, but we do not know which holds. Dashed line without any arrow heads indicates there is a dependency but we do not know the reciprocal dependence. From Zhao et al. BMC

Proc 1:S52, 2007

Highlights of the analysis

The series of papers on these data stress the importance of Intermediate phenotypes. Without a priori biological hypothesis, it serves as an exploratory tool for subsequent confirmatory analysis. This particular analysis highlights the potential usefulness of pathway analysis. An apparent limitation of this work, though not uncommon in gene-expression studies, is the relatively small sample size used. To fully elucidate the biological pathways involved may be difficult, as for instance CYCS is involved in a number of pathways. Statistical robustness and biological interpretability remain as the two main challenges for Bayesian network analyses, to which replication, bootstrap and benchmarking have been proposed. Our inference of gene networks also exploits the covariance structure of the data, like structural equation modelling, but is exploratory or hypothesis-generating rather than confirmatory or hypothesis-driven. A number of other software systems are of interest.

A Gaussian graphical model

We models measurements in EPIC-Norfolk data. The full, sub and final models give deviances of 0, 86.5, and 3.5, corresponding to df=0,1,1, respectively.
library(ggm) all <- read.dta("ggm.dta") all <- subset(all,!is.na(height+hip)) cor(all) grm <UG(~weight*bmi+weight*waist+weight*hip+waist*hip+waist *bmi+hip*bmi) fit <-fitConGraph(grm,cor(all),n=2413) grm <UG(~weight*bmi+weight*waist+weight*hip+waist*bmi+hip* bmi) fit <-fitConGraph(grm,cor(all),n=2413) grm <- UG(~bmi*waist+bmi*hip) fit <-fitConGraph(grm,cor(all),n=2413)

Extreme value theory

It is concerned with questions related to extreme values in sequences of random variables and in stochastic processes, e.g. Mn=max(X1,,Xn). An established results state that P((Mn-bn)/an)->H(x) which are of three types and can be combined into a single Generalized Extreme Value (GEV) distribution. The distribution of X conditionally on some high threshold often has a limit which follows Generalized Pareto Distribution (GPD). An associate model considers r largest order statistics. See Finkenstdt B, Rootzn H. Extreme Values in Finance, Telecommunications, and the Environment Chapman and Hall/CRC 2003 and also http://www.stat.unc.edu/postscript/rs/semstatrls.pdf

Annual maximal levels of River Nidd

The data can be used as follows, library(evir) qplot(nidd.annual) data(nidd.annual) nidd.gev <- gev(nidd.annual) plot(nidd.gev) meplot(nidd.annual) shape(nidd.annual) pfit <- gpd(nidd.annual, threshold=200) plot(pfit) quant(nidd.annual)

Summary
We have covered a variety of topics ranging from metaanalysis to causal modelling, which is expected to be more familiar with more genetic variants being established. They are general since some topics are also quite familiar to researchers at other fields (e.g., psychology, social science, econometrics) where for instance structural equation modelling are routinely used.

References
Bohning D, Kuhnert R, Rattanasiri S. Meta-Analysis of Binary Data Using Profile Likelihood. CRC Press, 2008 Conneely KN, Boehnke M. AJHG 2007;81:1158-68 Demidenko E. Mixed Models. Wiley, 2004 Harris et al. Stata J 2008; 8:3-28 Hartung J, Guido K, Sinha BK. Statistical Meta-Analysis with Applications. Wiley, 2008 Normand S-L. T. Stat Med 1999;18:321-59 Rao DC, Gu CC. Genetic Dissection of Complex Traits, 2e. Academic Press, 2008 Schlattmann P. Medical Applications for Finite Mixture Models. Wiley, 2009 Sidik & Jonkman. Appl Stat 2005; 54:367-84 Sterne J. Meta-Analysis in Stata. Stata Press, 2009. Sutton AJ, Abrams KR, Jones DR, Sheldon TA, Song F. Methods for Meta-Analysis in Medical Research. Wiley, 2000 Verzilli et al. AJHG 2008; 82:859-72 Whitehead A. Meta-Analysis of Controlled Clinical Trials. Wiley, 2002

References
Krzanowski WJ, Hand DJ. ROC Curves for Continuous Data. CRC 2009 Pepe, M.S. The Statistical Evaluation of Medical Tests for Classification and Prediction. Oxford University Press, 2003 Gonen M. Analyzing Receiver Operating Characteristic Curves with SAS, SAS Institute Inc., 2007 Loehlin JC. Latent Variable Models-An Introduction to Factor, Path, and Structural Equation Analysis. 4e, Lawrence Erlbaum Associates, 2004 Kline RB. Principles and Practice of Structural Equation Modeling. 2e, The Guiford Press, 2005 Bollen KA, PJ Curran. Latent Curve Models-A Structural Equation Perspective. Wiley, 2006 Kjaerulff UB, AL Madsen. Bayesian Networks and Influence DiagramsA Guide to Construction and Analysis. Springer, 2008 Emmert-Streib F, Matthias D. Analysis of Microarray Data-A NetworkBased Approach. Wiley-VCH, 2008 Junker BH, F Schreiber (Ed). Analysis of Biological Networks. Wiley, 2008

IV OpenMx and NCBI2R

Topics
Heritability estimation Background Family data Twin data OpenMx Summary Information retrieval with NCBI2R Further information

Definition of genetic heritability

Genetic heritability is defined for a quantitative trait as the proportion of variation attributable to genetic factors, and extended to categorical traits through reference to a liability model. The value of the genetic heritability varies according to factors taken into account. Let phenotype P has mean and variance 2 from a linear model P=a+d+c+e where a, d, c and e represent additive, dominance, common environment and individual specific environment, then genetic heritability in the narrow sense is a2 /2 in contrast to genetic heritability as g2 /2, which can include epistasis. Hopper JL. Heritability. In Armitage, P. Colton T (Eds). Encyclopedia of Biostatistics, 2e, Wiley 2005.

Some clarifications
For a binary trait, such as whether or not an individual has a disease, heritability is not the proportion of disease in the population attributable to or caused by, genetic factors. For a continuous trait, genetic heritability is not a measure of the proportion of an individuals score attributable to genetic factors. Heritability is not about cause per se, but about the causes of variation in a trait across a particular population. As heritability varies according to which factors are considered, there is no unique value of genetic heritability of a characteristic. It also varies from population to population. A poorly measured trait will apportion to measurement error leading to lower estimate of genetic heritability.

Heritability studies
Family studies Adoption (rearing of a nonbiological child) studies Migrant studies migrants carry a risk reflecting country of origin Twin study differences between monozygotic and dizygotic twins can be attributed to genetic influence

The case of obesity

BMI is often used as surrogate measurement, such that those with BMI >=25 and >=30 are considered as overweight and obesity. The heritability estimate of BMI had a range of 30-70%, and 50-90% from twin studies. Maes et al showed to be ~70% based on meta-analysis. Peterson et al. (twin study) JAMA 256:2958, 1986 Stunkard et al (adoption study) . New Eng J Med 314:193-8, 1986 Maes et al. Behav Genet 27:325-51, 1997 Friedman JM. Nat Med 10:563-9, 2004

Path analysis of nuclear family data

Recall that for the model P=G+C+E, can be reexpressed with path coefficients so that P=hG+cC+E and to allow for intergenerational difference this can be written as P=hzG+cyC+E for parents. We can also allow for correlation between parental phenotypes (homogamy) as with gene-environment correlations. We have the following path analysis model, noting that it uses the notion of environmental indices and sometime transformation of the phenotype for normality.

Mixed homogamy model

Karlin et al. Am J Hum Genet 1983; 35:695-732

h hz c cy p m u fF fM b i iF iM

Effect of child's genotype on child's phenotype Effect of parental genotype on parental phenotype Effect of child's environment on child's phenotype Effect of parental environment on parental phenotype "Primary" correlation between parental phenotypes due to phenotypic homogamy Correlation between parental genotypes due to social homogamy Correlation between parental environments due to social homogamy Effect of father's environment on child's environment Effect of mother's environment on child's environment Effect of common sibship environment on child's environment Effect of child's environment on child's index Effect of father's environment on father's index Effect of mother's environment on mother's index Correlation between parental genotype and parental phenotype Correlation between parental environment and parental phenotype Correlation between adult's environment and spouse's genotype due to social homogamy Total correlation between parental genotype and parental environment

s a

Segregation analysis of family data

When family data is available, we can examine major gene effect(s) together with collective loci with small and individually unmeasurable effects in a so-called mixed model (Morton NE, MacLean CJ. Am J Hum Genet 1974;26:489-503). It can also incorporate parameters for covariates such as age, sex and race.

Segregation analysis of NIDDM

Cook et al. Diabetologia 1994; 37:1231-40

PATHMIX, ATRIBUTE and POINTER

PATHMIX and ATRIBUTE provides path analysis of nuclear family data for both quantitative and binary traits. POINTER (Lalouel JM, Morton NE. Hum Hered 1981; 31:312-21) provides MLE of d - degree of dominance, which ranges between 0 for a recessive gene and 1 for a dominant; t - displacement between the two homozygotes of the major gene; q - gene frequency of allele leading to affection. Some attempt was made to include in R/CGR bundle. In addition, PAP and more recently JPAP has also implemented the mixed model of segregation and linkage.

Obesity in familial NIDDM

The genetic model specified phenotype as the sum of independent effects attributed to the segregation of alleles at major loci, the transmission of polygenes, and random factors specific to the individual. The parameters were the total mean (g), the total standard deviation (a), the frequency of the allele determining high BMI at locus L (qL), the dominance at locus L (dL), the displacement at locus L (tL), polygenic heritability (h2), and parent-to-offspring transmission probabilities (t1, t2, and t3 with values 1, 0.5, and 0 for Mendelian inheritance) for the three genotypes at one locus. Displacement is the difference, in within genotype SDs, between the means of two homozygotes. Dominance is the difference between the mean for heterozygotes and the mean for homozygotes, for low BMI relative to the displacement. The polygenic heritability is the proportion of the variance within major-locus genotypes, owing to polygenic inheritance. Hasstedt et al. Am J Hum Genet 1997; 61:668-77

Parent-offspring model with LISREL

Boomsma et al. Behav Genet 1989; 19:123-41

SOLAR
SOLAR (Sequential Oligogenci Linkage Analysis Routine, Almasy L, Blangero J. Am J Hum Genet 1998; 62:1198-211) uses likelihood ratio tests to evaluate heritability by comparing a purely polygenic model with a sporadic model in the case of testing heritability. In a polygenic model, h2r is the total additive genetic heritability. In a linkage model (with one or more locus specific elements) h2q1 represents the heritability associated with the first locus, and h2r represents the residual genetic variance. In a oligogenic model, there may also be h2q2, h2q3, etc. Sung J, et al. JCEM 2009; 94:4946-52. reported a recent study of twins with adjusted (age, sex, age2, age2 x sex, total calorie intake, smoking and alcohol use) heritabilities for waist circumference (59%), glucose (59%), HDL (77%), TG (46%).

Twin ACE model MZ

rMZ = a 2 + c 2

DZ

rDZ = 0.5 a 2 + c 2

Recall that
Expectations of sample correlation E(r)=-(1-2)/(2n)[1-(1-92)/(4n)+] V(r)=(1-2)2/n[1+112/(2n)+] Keeping ES. Introduction to Statistical Inference. Van Nostrand 1962; Dover 1995.

Simple estimation
There we have mean/variance a 2 = 2( r r ) MZ DZ
2 4 1 rMZ

[(

2 nMZ + 1 rDZ

nDZ

Similarly,

c 2 = 2rDZ rMZ
2 4 1 rMZ

2 nMZ + 1 rDZ

nDZ

e 2 = 1 rMZ

(1 r )

2 2 MZ

nMZ

These can be simpler when there are equal numbers of DZ and MZ twins.

Maximum likelihood method

A notable implementation was twinan90 (Williams CJ, Christian JC, Norton JA Jr. (1992). Comp Meth Prog Biomed 38:(2-3):167-176)
library(gap) fs <- file.path(.path.package("gap"),"tests/mzdz.dat") mzdz <- matrix(scan(fs,skip=1),ncol=2,byrow=T) mzdat <- mzdz[1:131,] dzdat <- mzdz[132:206,] twinan90(mzdat,dzdat,xlamb=2) file.show(mzdz.out) file.show(mzdz.log)

The estimation may be unstable.

A simulated twin data

library(mvtnorm) mzm <- as.data.frame(rmvnorm(195, c(22.75,22.75), matrix(2.66^2*c(1, 0.67, 0.67, 1), 2))) dzm <- as.data.frame(rmvnorm(130, c(23.44,23.44), matrix(2.75^2*c(1, 0.32, 0.32, 1), 2))) names(mzm) <- names(dzm) <- names(mzw) <names(dzw) <- c("bmi1","bmi2")

Summary statistics
apply(mzm,2,mean) bmi1 bmi2 22.68876 22.86700 cov(mzm) bmi1 bmi2 bmi1 7.240612 4.698384 bmi2 4.698384 6.921260 cor(mzm) bmi1 bmi2 bmi1 1.0000000 0.6636946 bmi2 0.6636946 1.0000000 apply(dzm,2,mean) bmi1 bmi2 23.32167 23.16760 cov(dzm) bmi1 bmi2 bmi1 9.208189 1.574196 bmi2 1.574196 5.799376 cor(dzm) bmi1 bmi2 bmi1 1.0000000 0.2154175 bmi2 0.2154175 1.0000000

Scatter plots of male and female twins

jpeg("ACE.jpg") par(mfrow=c(2,2)) plot(mzm) plot(dzm) plot(mzw) plot(dzw) dev.off() BMIs in MZ twins are seen to be more correlated than those in DZ twins.

Structural equation modelling

Multiple and multivariate regression Confirmatory factor analysis Latent growth curves Latent differential equations Moderated parameter models Multigroup models Multilevel multivariate models with moderated parameters

Mx, MxGUI, OpenMx

Mx is a well-established software for structural equation modeling and in particular widely used in twin modeling. MxGUI is Windows-based program which greatly facilitate the modeling process. OpenMx is a recent initiative to take advantage of the R environment. It carries over a variety of features from Mx/MxGUI. As it implies, the software is freely available from http://openmx.psyc.virginia.edu As it is written in R, it is considerably simple in our context to explore functionality it provides. We have adapted some examples from OpenMx website and for extended description, it is recommended to visit there.

Mx specification for twin data I

G1: model parameters Data Calc NGroups=4 Begin Matrices; X Lower 1 1 Free Y Lower 1 1 Free Z Lower 1 1 Free W Lower 1 1 Fixed Begin Algebra; A= X*X' ; C= Y*Y' ; E= Z*Z' ; D= W*W' ; End Algebra; End G2: MZ Data NInput-vars=2 NObservations=522 Labels N_t1 N_t2 CMatrix .73865671 .68574888 1.3722835 Matrices= Group 1 Covariances A+C+D+E | A+C+D _ A+C+D | A+C+D+E / Options RSidual End

Mx specification for twin data II

G3: Dizygotic twin pairs Data NInput_vars=2 NObservations=272 Labels N_t1 N_t2 CMatrix 1.0942882 .3712542 .93089623 Matrices= Group 1 H Full 1 1 Q Full 1 1 Covariances A+C+D+E | H@A+C+Q@D _ H@A+C+Q@D | A+C+D+E / Matrix H .5 Matrix Q .25 Start .6 All Options Multiple RSidual End G4: beta-test data calc matrices= Group 1 compute (A|C|D|E) @ (A+C+D+E)~/ options rs nd=3 options multiple end

OpenMx installation and a first session

source('http://openmx.psyc.virginia.edu/getOpenMx.R') library(OpenMx) ?OpenMx data(demoOneFactor) head(demoOneFactor)
x1 x2 x3 x4 x5 1 -0.1086832 -0.4669377 -0.177839881 -0.08093113 -0.07065026 2 -0.1464765 -0.2782619 -0.273882553 -0.15412007 0.09271729

vars <- names(demoOneFactor)

Model specification and fitting

Model <- mxModel("One Factor", type="RAM", manifestVars=vars, latentVars="G", mxPath(from=G, to=manifests), mxPath(from=vars, arrows=2), mxPath(from=G, arrows=2, free=FALSE, values=1.0), mxData(observed=cov(demoOneFactor), type="cov", numObs=500) ) Fit <- mxRun(Model) summary(Fit)

Parameter estimation
name matrix 1 <NA> 2 <NA> 3 <NA> 4 <NA> 5 <NA> 6 <NA> 7 <NA> 8 <NA> 9 <NA> 10 <NA> row col A x1 G A x2 G A x3 G A x4 G A x5 G S x1 x1 S x2 x2 S x3 x3 S x4 x4 S x5 x5 Estimate 0.39715212 0.50366111 0.57724141 0.70277369 0.79624998 0.04081419 0.03801999 0.04082718 0.03938706 0.03628712 Std.Error 0.015549769 0.018232514 0.020448402 0.024011418 0.026669452 0.002812717 0.002805794 0.003152308 0.003408875 0.003678561

Model-fitting statistics
observed statistics: 15 estimated parameters: 10 degrees of freedom: 5 -2 log likelihood: -3648.281 saturated -2 log likelihood: -3655.665 number of observations: 500 chi-square: 7.384002 p: 0.1936117 AIC (Mx): -2.615998 BIC (Mx): -11.84452 adjusted BIC: RMSEA: 0.03088043

Elementary statements
We can obtain a list of commands as usual, i.e., library(help=OpenMx) e.g., mxAlgebra mxMatrix mxData mxEval mxAlgebraObjective

mxFIMLObjective mxRun

Examples
A <- mxMatrix("Full", nrow = 3, ncol = 3, values=2, name = "A") A FullMatrix 'A' @labels: No labels assigned. @values [,1] [,2] [,3] [1,] 2 2 2 [2,] 2 2 2 [3,] 2 2 2 @free: No free parameters. @lbound: No lower bounds assigned. @ubound: No upper bounds assigned.

ACE model
ACE<-function(mzDat=mzData,dzDat=dzData,type="raw",selV=selVars){ twinACEModel <- mxModel("ACE", mxMatrix("Full", 1, 1, TRUE, .6, "a", name="X"), mxMatrix("Full", 1, 1, TRUE, .6, "c", name="Y"), mxMatrix("Full", 1, 1, TRUE, .6, "e", name="Z"), mxAlgebra(X %*% t(X), "A"), mxAlgebra(Y %*% t(Y), "C"), mxAlgebra(Z %*% t(Z), "E"), mxAlgebra(A+C+E, name="V"), mxMatrix("Full", 1, 2, TRUE, 20, "mean", name="expMean"), mxAlgebra(rbind(cbind(A+C+E, A+C), cbind(A+C, A+C+E)), "expCovMZ"), mxAlgebra(rbind(cbind(A+C+E, 0.5%x%A+C), cbind(0.5%x%A+C, A+C+E)), "expCovDZ"), mxModel("MZ", mxData(mzDat, type), mxFIMLObjective("ACE.expCovMZ", "ACE.expMean", selV)), mxModel("DZ", mxData(dzDat, type), mxFIMLObjective("ACE.expCovDZ", "ACE.expMean", selV)), mxAlgebra(MZ.objective + DZ.objective, name="twin"), mxAlgebraObjective("twin"))

Fitting ACE model

twinACEFit <- mxRun(twinACEModel, silent=TRUE) exp_ACE <mxEval(rbind(expCovMZ,expCovDZ,expMean), twinACEFit) est_ACE <- mxEval(cbind(A,C,E,A/V,C/V,E/V), twinACEFit) LL_ACE <- mxEval(objective, twinACEFit) rownames(exp_ACE) <c('CovMZT1','CovMZT2','CovDZT1','CovDZT2','Mean') colnames(exp_ACE) <- c('T1','T2') rownames(est_ACE) <- 'ACE' colnames(est_ACE) <- c('a','c','e','a^2','c^2','e^2')

Fitting AE model
twinAEModel <- mxModel(twinACEModel, mxMatrix("Full", 1, 1, FALSE, 0, "c", name="Y")) twinAEFit <- mxRun(twinAEModel, silent=TRUE) exp_AE <- mxEval(rbind(expCovMZ,expCovDZ,expMean), twinAEFit) est_AE <- mxEval(cbind(A,C,E,A/V,C/V,E/V), twinAEFit) rownames(est_AE) <- 'AE' LL_AE <- mxEval(objective, twinAEFit) LRT_ACE_AE <- LL_AE - LL_ACE

Fitting CE model
twinCEModel <- mxModel(twinACEModel, mxMatrix("Full", 1, 1, FALSE, 0, "a", name="X")) twinCEFit <- mxRun(twinCEModel, silent=TRUE) exp_CE <- mxEval(rbind(expCovMZ,expCovDZ,expMean), twinCEFit) est_CE <- mxEval(cbind(A,C,E,A/V,C/V,E/V), twinCEFit) rownames(est_CE) <- 'CE' LL_CE <- mxEval(objective, twinCEFit) LRT_ACE_CE <- LL_CE - LL_ACE

Fitting E model and summary statistics

twinEModel <- mxModel(twinAEModel, mxMatrix("Full", 1, 1, FALSE, 0, "a", name="X")) twinEFit <- mxRun(twinEModel, silent=TRUE) exp_E <- mxEval(rbind(expCovMZ,expCovDZ,expMean), twinEFit) est_E <- mxEval(cbind(A,C,E,A/V,C/V,E/V), twinEFit) rownames(est_E) <- 'E' LL_E <- mxEval(objective, twinEFit) LRT_ACE_E <- LL_E - LL_ACE exp <- cbind(exp_ACE,exp_AE,exp_CE,exp_E) est <- rbind(est_ACE,est_AE,est_CE,est_E) lls <rbind(cbind(LL_ACE,0),cbind(LL_AE,LRT_ACE_AE),cbind(LL_CE,LRT_A CE_CE),cbind(LL_E,LRT_ACE_E)) df <- c(NA,1,1,2) lls <- cbind(lls,pchisq(2*lls[,2],df,lower.tail=FALSE)) rownames(lls) <c("l(ACE)","l(AE),lrt(AE,ACE)","l(CE),lrt(CE,ACE)","l(E),lrt(ACE,E)") invisible(list(exp=exp,est=est,lls=lls)) }

Heritability estimates
The heritability and 95% boostrap CI estimates by models are as follows, mean 0.6558460 0.6579248 0.0000000 0.0000000 sd 0.04264874 0.03891981 0.00000000 0.00000000 lcl 0.5722545 0.5816420 0.0000000 0.0000000 ucl 0.7394376 0.7342076 0.0000000 0.0000000

ACE AE CE E

This provides a simple estimation, although we could obtain analytical approximation in a more elaborate way.

Summary
The pursuit of precise estimation of genetic vs environmental contributions to complex traits have a long history and currently an indispensible part of genetic epidemiology or statistical genetics. The literature we focused here is largely from 1970s onwards. However, we have gone quite far with our understanding and implementation of procedures. The former includes simple estimation, maximum likelihood methods, path analysis and structural equation modeling while the latter evolves from Fortran, LISREL/Mx/MxGUI to R. Our focus is on the practical side. The new practice with OpenMx rests on the flexible and powerful R computing environment, which makes collaborative work truly possible.

Information retrieval with NCBI2R

Whenever there is a routine task, it calls for a formal programming. GWAS produces many p-values without full context and some annotation is needed. NCBI (http://www.ncbi.nlm.nih.gov/) is a good source with up-to-date information. PubMed All Databases Books OMIM SNP Taxomony A direct access to the URL is convenient but tedious. A really simple solution is to use applications such as NCBI2R.

Setup
As from v1.3, NCBI2R is available from CRAN and the projects homepage has more information: http://drop.io/NCBI2R_package. As usual the package is loaded into R as follows. library(NCBI2R) library(help=NCBI2R) help.start() The package is still under development but most functions should work under the Windows environment.

OpenURL and GetPubMed

OpenURL("http://www.ncbi.nlm.nih.gov/") refs <- GetPubMed(MC4R,MC4R.tab) "Number of papers found in PubMed was: 594" names(refs)
[1] "PMID" [7] "IS" [13] "TA" [19] "AID" [25] "VI" [31] "SB" [37] "LR" [43] "TT" [49] "GN" PrintFilters() "TI" "DP" "JT" "PST" "IP" "PMC" "RN" "CN" "OTO" "AB" "AU" "JID" "SO" "PG" "OID" "MH" "EIN" "OT" "OWN" "STAT" "DA" "LA" "PT" "DEP" "EDAT" "MHDA" "CRDT" "CI" "AD" "PHST" "FAU" "GR" "PL" "MID" "PMCR" "DCOM" "CIN" "RF" "SI" "IR" "FIR" "CON" "IRAD" "localcopy" "link

OpenPMID and OpenPDF

We could use the following commands to save our query refs <- GetPubMed(MC4R) MakeExcel(refs,MC4R.tab) We can browse the content fix(refs) We can examine papers in PDF format OpenPDF(refs$PMID[1]) We can examine summary information as in PubMed

OpenPmid(refs$PMID[1])

Annotation
This is achieved with functions available, e.g.,
ScanForGenes, ScanforSNPs GetSNPInfo, GetGeneInfo GetGeneInfo(#) AnnotateDataframe(mydata, selections=c(marker,p,beta)) AnnotateSNPList, AnnotateSNPFile GetIDs() GetGeneTable(#) GetGOs(#) GetInteractions(#) GetPathways(#) GetRegion(snp,4,start,end), GetRegion(gene,X,start,end) GetPhenotypes(#) GetSNPsInGene(#)

Example: obesity-related SNPs

snps <- c("rs6548238", "rs7566605","rs745229","rs1106683","rs1121980","rs9 939609","rs17782313","rs17700633") snps_info <- GetSNPInfo(snps) snps_split <- SplitGenes(snps_info) snps_list <- AnnotateSNPList(snps,"snps.html") snps_file <- AnnotateSNPFile("snps.txt","snps.html") MakeExcel(snps_file,"snps.tab") MakeHTML(snps_file,"snps.html")

GetSNPInfo, GetGeneInfo, GetNeighGenes

marker genesymbol locusID chr chrpos fxn_class species rs6548238 2 624906 Homo sapiens rs7566605 2 118552496 Homo sapiens rs745229 FAM71F1 84691 7 128146128 coding-synonymous, reference Homo sapiens rs1106683 7 131104066 Homo sapiens rs1121980 FTO 79068 16 52366749 intron Homo sapiens rs9939609 FTO 79068 16 52378029 intron Homo sapiens rs17782313 18 56002078 Homo sapiens rs17700633 18 56080413 Homo sapiens chr LowPoint HighPoint locusID 1 18 55852078 56152078 115701,23327

snps_info <- GetSNPInfo(snps) snps_info

GetNeighGenes("18",56002078,150000) # +/- 150Kbp

names(GetGeneInfo(23327)) [1] "locusID" "org_ref_taxname" "org_ref_commonname" [4] "OMIM" "synonyms" "genesummary" [7] "genename" "phenotypes" "pathways" [10] "GeneLowPoint" "GeneHighPoint" "ori" [13] "chr" "genesymbol" "build" [16] "cyto" "approx"

Melanocortin 4 receptor (MC4R)

gid <- GetIDs("MC4R") gnames <- GetGeneNames(gid) ginfo <- GetGeneInfo(gid) ggo <- GetGOs(gid) gint <- GetInteractions(gid) gpheno <- GetPhenotypes(gid) gpath <- GetPathways(gid) gsts <- GetUniSTSFromName("MC4R") gstsinfo <- GetUniSTSInfo(gsts[1])

GetIDs and GetGeneNames

gid "4160" "342784" "79068" "100270981" "9709 "2646 "400652" "1071" "4023" 181" "5443" "26033 "4157" "132789" "9607" "89866" "9317" "23017 "5566""4864" "4852" "4159" "627" "4094"434" "129787" "156" names(gnames) "genename" "genesymbol" "NewlocusID" "CurrentRecord" "LastUpdate" "locusID" "species" gnames$genesymbol "MC4R" "LOC342784" "FTO" "RPL30P9" "HERPUD1" "GCKR "RPS3AP49" "CETP" "LPL" "AGRP" "POMC" "ATRNL1 "MC1R" "GNPDA2" "CARTPT" "SEC16B" "PTER" "FAIM2 "PRKACA" "NPC1" "NPY" "MC3R" "BDNF" "MAF"ASIP" "TMEM18" "ADRBK1"

GetGeneInfo and GetGeneTable

names(ginfo)

GetGeneTable(4160) # positions of exons, DNA/protein Acc # $ExonInfo Where Start Stop Size Set 1 Exon 1 1438 1438 1 2 CodExon 420 1418 999 1 $ACC.DNA Identifier Length Exons 1 NM_005912.2 1438 1 $ACC.Prot Identifier Length Exons 1 NP_005903.2 332 1

"locusID" "org_ref_taxname" "org_ref_commonname "OMIM" "synonyms" "genesummary "genename" "phenotypes" "pathways "GeneLowPoint "GeneHighPoint" "ori "chr" "genesymbol" "build "cyto" "approx"

ginfo[c("locusID","OMIM","chr","GeneLowPoint"," GeneHighPoint","ori","genesymbol, cyto)]

4160 155541 18 342784 18 79068 610966 16 100270981 8 9709 608070 16 2646 600842 2 400652 18 1071 118470 16 4023 609708 8 181 602311 16 5443 176830 2 26033 612869 10 4157 155555 16 132789 613222 4 9607 602606 5 89866 612855 1 9317 604446 10 23017 604306 12 5566 601639 19 4864 607623 18 58038564 57863787 53737875 19970847 56965748 27719706 57816776 56995835 19796582 67516474 25383722 116853124 89984287 44704168 71014994 177898242 16478967 50260680 14202500 21111463 58040001 MC4R 18q22 57865424 + 18q21.32 54148381 + FTO 16q12.2 19971168 - RPL30P9 8p22 56977793 + HERPUD1 16q12.2-q13 27746551 + GCKR 2p23 57817639 + RPS3AP49 18q21.32 57017756 + CETP 16q21 19824770 + LPL 8p22 67517716 AGRP 16q22 25391559 POMC 2p23.3 117708496 + ATRNL1 10q26 89987385 + MC1R 16q24.3 44728612 GNPDA2 4p12 71016872 + CARTPT 5q13.2 177939050 SEC16B 1q25.2 16555736 + PTER 10p12 50297720 FAIM2 12q13 14228559 PRKACA 19p13.1 21166470 NPC1 18q11-q12

GetGOs
ggo
category Function Function Function Function Process Process Process Process Process Process Process Process Component Component Component name evidence pubmed db db_id G-protein coupled receptor activity IEA GO 4930 melanocortin receptor activity TAS 8794897 GO 4977 protein binding IEA GO 5515 receptor activity IEA GO 4872 G-protein coupled receptor IEA GO 7186 protein signaling pathway G-protein signaling, coupled TAS 8794897 GO 7188 to cAMP nucleotide second messenger feeding behavior TAS 9771698 GO 7631 insulin secretion IEA GO 30073 regulation of bone resorption IMP 16614075 GO 45780 regulation of metabolic process IEA GO 19222 response to insulin stimulus IEA GO 32868 signal transduction IEA GO 7165 cytoplasm IDA 18029348 GO 5737 integral to membrane TAS 8392067 GO 16021 plasma membrane TAS 10585465 GO 5886

GetRegion, GetGeneInfo, GetSNPsInGene

"rs79783591" "rs62097821" "rs34114122" "rs13447336" "rs13447331" "rs13447326" "rs17848587"

GetRegion("snp","18",58038564,58040001)
"rs79390404" "rs61741819" "rs13447340" "rs13447335" "rs13447330" "rs13447325" "rs52834737"

"rs78877161" "rs76500026" "rs52820871" "rs52804924" "rs13447339" "rs13447338" "rs13447334" "rs13447333" "rs13447329" "rs13447328" "rs13447324" "rs13447323" "rs2229616" "rs1016862

"rs74679969" "rs35351438" "rs13447337" "rs13447332" "rs13447327" "rs2282556"

"4160"

GetRegion("gene","18",58038564,58040001)

GetGeneInfo(4160) GetSNPsInGene(4160)

GetPathways
gpath name 1 KEGG pathway: Neuroactive ligand-receptor interaction 2 Reactome Event:Signaling by GPCR web 1 http://www.genome.jp/dbgetbin/show_pathway?hsa04080+4160 2 http://www.reactome.org/cgibin/eventbrowser_st_id?ST_ID=REACT_14797

Obesity and MC4R

oid <- GetIDs("obesity[MC4R]") oinfo <- GetGeneInfo(oid) names(oinfo)
[1] "locusID" "Org_ref_taxname" "Org_ref_commonname" "OMIM" "synonyms" [6] "genesummary" "genename" "phenotypes" "phenotypes.HTML" "pathways" [11] "pathways.HTML" "GeneLowPoint" "GeneHighPoint" "Ori" "Chromosome" [16] "genesymbol" "build" "cyto" "approx"

dim(oinfo) 302 19

A tutorial example
favouriteSNP <- "rs4294787" favouriteSNPInfo <- GetSNPInfo(favouriteSNP) pathway <- GetPathways(favouriteSNPInfo$locusID) genes_in_pathway <- GetIDs(pathway$name) #a loop to enable GetSNPsInGene work with multiple genes for (i in 1:length(genes_in_pathway)) { if(!(exists("biglist"))) { biglist <- GetSNPsInGene(genes_in_pathway[i]) } else { biglist<-c(biglist,GetSNPsInGene(genes_in_pathway[i])) } } length(biglist)

165212

Side interest
# An example showing the principle as implemented in the package can be useful for obtaining other information. keywords <- c("Professor","England") nj <- NatureJobs(keywords,"nj",days=7) dim(nj) 120 11 names(nj) names(nj) [1] "JobTitle" "Employer" "Location" "Posted" "Desc" "DaysAgo" "IDnumber" "BigDescription" [9] "WebLink" "LocalLink" "ExpDate"

A summary
NCBI, especially PubMed, has been a major source of biomedical information retrieval in daily research. The process can considerably be facilitated by R/NCBI2R package. A minor issue is that it only retrieves the latest information but earlier information is very useful (e.g., build 35). Hence more experiences need to be gathered. NCBI2R annotates lists of SNPs and/or genes, with current information from NCBI designed to allow those performing the genome analysis to produce output that could easily be understood by a person not familiar with R. It is easy to anticipate that more functionality can be added from the same principle. It is helpful to keep an eye on the package development even if implementation may not necessarily be a priority in our research.

Further information
We often use NCSC genome browser (http://genome.ucsc.edu) coupled with the galaxy system (http://g2.bx.psu.edu), but the facility in R will be complementary. Annotation databases (Lesk AM. Database Annotation in Molecular Biology-Principles and Practice. Wiley 2005) Other packages such as gene2pathway from CRAN (Prediction of KEGG pathway membership for individual genes based on InterPro domain signatures), SubpathwayMiner (Annotation and identification of the KEGG pathways) Packages from BioConductor (http://www.bioconductor.org), e.g., KEGGSOAP (interface to the KEGG SOAP server) library(annotate) ab <- pubmed[18454148] buildPubMedAbst(xmlRoot(ab)[[1]]) pubmed(18454148, disp=browser)

RNCBI
library(RNCBI) ncbi <- NCBI() einfo <- EInfo(ncbi) einfo <- setRequestParameter(einfo, "db", "pubmed")

V Conclusion

General comments
As has been driven by technological advances in genotyping and computational technology, the genetic analysis of complex trait is a dynamic topic in a fastmoving field. The R environment is now indispensible with a great deal of recognition and stability. Nevertheless, there are areas which can be further advanced, e.g., graphics. A range of models available from R remains to be explored and have been supplementary to the main analysis.

Scientific aspects of GWAS

What are the uses? Discovery of new susceptibility loci Elucidate biologic pathways Identify links between these loci and covariates Risk prediction Where would they go? Expanding well-characterized study populations Expanding the range of genetic variation including structural variants and lower-frequency common variants Documenting functional mechanisms responsible for the association signals. Chanock (Personal Communication) and Altshuler et al. Science 2008, 322:881-8

Limitations and practical issues

Limitations It requires large sample sizes It only identifies loci, not genes It detects only common alleles in a population It usually does not go into the expression level Practical issues For individual studies, the issues as of epidemiological studies in general remain and there is uncertainty to declare statistical significance For consortium meta-analysis, there may be difference in quality control, data sharing and variation in complexity of analysis Technological advances, e.g., sequencing, remain to have profound influences.

A great expectation
Ashley et al. Lancet 2010, 375, 1525-35 The authors assessed a patient with a family history of vascular disease and early sudden death. The analysis involved 2.6M SNPs and 752 CNVs showing increased genetic risk for MI, T2D and some cancers.

Summary
The need from various analyses seeds the development in R and shares much in common with many other problems involving large data, such as interactive graphics in combination with publicly available databases, the use of statistical and computational facilities available from the R system. Applications in substantive areas are the constant source of motivation in package development. The implementation is likely to be patchy but with a great prospect, e.g., advanced models and causal pathways. Alternative computing environments are complementary.

References
Murrell P. R Graphics. Chapman & Hall/CRC, 2005
Gentleman R, Carey V, Huber W, Irizarry R, Dudoit S. Bioinformatics and Computational Biology Solutions Using R and Bioconductor. Springer, 2005

Hahne F, Huber W, Gentleman R, Falcon F. Bioconductor Case Studies. Springer, 2008

Spector P. Data Manipulation with R. Springer, 2008 Foulkes AS. Applied Statistical Genetics with R for Populationbased Association Studies. Springer, 2009 Broman KW, Sen S. A Guide to QTL Mapping with R/qtl. Springer, 2009 Gentleman R. R Programming for Bioinformatics. Chapman & Hall/CR, 2009

Robert C, Casella G. Introducing Monte Carlo Methods with R. Springer, 2010