Pedigree versus population analysis

When genealogical information is available, PopPAnTe evaluates the relatedness matrix from the known pedigree relationships using a recursive procedure and assuming pedigree founders as unrelated [9]. This results in a variance-covariance matrix that is usually both symmetric and semi-positive definite. Therefore, the maximum likelihood estimates of the variance components can be assessed through efficient Cholesky decomposition.

When the genealogical information is not available, a GSM can be estimated from genome-wide genetic data with any of several well-established tools, such as PLINK [10], GCTA [11], or LDAK [12], and given as input to PopPAnTe. The property of positive-definiteness does not always hold for GSMs. A bending procedure [13] is used by default to transform the matrix when it is not positive semi-definite –but the user has the option to use a LU decomposition instead. Additionally, PopPAnTe implements the QR decomposition to solve the rare cases where the variance-covariance matrix is not invertible and neither the Cholesky nor the LU decompositions can be used.

To speed-up the evaluation of the variance components, PopPAnTe allows the user to set an arbitrary threshold below which individuals can be considered as unrelated. Otherwise, the user has the option of using the value of expected kinship between second or third cousins [14, 15].

Region-based testing

When predictors can be ordered in space, as in the case of epigenetic markers, PopPAnTe allows the computation of region-based association tests by gathering information from flanking predictors included in a sliding window of user-defined size, whose values are replaced by their first principal component. By definition the first principal component accounts for as much of the variability in the data as possible, and can thus be used to summarise the joint distribution of all variables included in a given region for gene- or region-based association studies (e.g., [16, 17]).

Data pre- and post-processing

Quantile normalisation [18] can be automatically applied to improve normality of both response variables and predictors. Moreover, PopPAnTe implements two approaches to correct the association test for unwanted biological and technical variability (e.g., batch effects). When the source of the confounders is known, it can be directly included in the association model. To deal with unknown sources of biological and technical co-variation, PopPAnTe can integrate into the association model the principal components that are required to explain a user-specified percentage of variation.

PopPAnTe implements the Benjamini-Hochberg procedure (BH step-up procedure) to control the false discovery rate [19], and, to aid in results interpretation and further analyses, it generates basic Quantile-Quantile and Manhattan plots – the latter only when genomic data that can be ordered in space (e.g., CpG loci) are used as predictors.

Finally, when the genealogical information is available, to determine whether an association has been generated by a uniform contribution of all the families within the sample, or by a strong contribution of a small number of families, PopPAnTe reports, for each test, the percentage of families showing a positive contribution and the Gini coefficient [20] assessed on family contribution to the χ ² statistics.

Simulation study

We simulated quantitative response and predictor variables in three-generation families (maternal and paternal grandparents, parents, and two offspring).

In the first simulated scenario, we aimed to test the relationship between running time and number of subjects included in the analysis. Therefore, we simulated 11 independent datasets including an increasing number of three-generation families (from 10 to 1000, thus comprising 80 to 8000 individuals) and one response and one predictor variable for all simulated subjects.

In the second simulated scenario, we aimed to test the relationship between running time and number of variables included in the analysis. Consequently, we fixed the number of families included in each dataset (125 families, corresponding to 1000 individuals) and generated 7 independent datasets with one response and an increasing number of predictors ranging from one to 10,000.

Both scenarios were simulated 100 times and the median time necessary for the testing step recorded. Simulations were performed on a Mac BookPro 2.3 GHz, Intel Core i7, 16 GB RAM; Java version 1.7. Default parameters were used for the Java Virtual Machine (1 GB of memory, 1 thread).

Case study 1: Epigenome-wide association study in a Qatari family study

We carried out an epigenome-wide association study of body mass index (BMI) using extended families from Qatar. The Qatari population is an isolated inbred population characterised by a large number of consanguineous families [21]. A detailed description of the subjects and methylation data included in this study has been previously reported in Zaghlool et al. [22]. Briefly, we used genome-wide methylation and SNP data generated from whole blood on the Infinium HumanMethylation450 Bead-Chip (Illumina Inc, San Diego, CA) and the Illumina HumanOmni2.5-8M BeadChip, respectively. DNA methylation Beta-values were measured for 123 individuals, 88 with both genotype and BMI data in 13 multigenerational families. We used GCTA to calculate a GSM between pairs of individuals using all autosomal SNP markers with minor allele frequency >0.01. We compared heritability estimates of the methylation values at CpG loci and their association with BMI in the Qatari family study using either the family information or the inferred GSM. Age, sex, and cell-type proportions as estimated using the Houseman method [23] were included in the model as fixed effects. We observed a very high concordance correlation coefficient [24] of the effect size estimates for the association between CpG methylation states and BMI (r β=0.99; Fig. 1, left), as well as of the CpG-specific component of genetic and environmental variances (\(r\sigma _{a}^{2}=0.99\) and \(r\sigma _{e}^{2}=0.90\), respectively; Fig. 1, right).

Case study 2: transcriptome-wide association study in UK twins

In the second case study, we carried out a transcriptome-wide association study with BMI in a cohort of healthy female Caucasians twins. The TwinsUK adult twin registry includes about 12,000 subjects, predominately females [25]. Genotyping of the TwinsUK dataset was performed with a combination of Illumina HumanHap300, HumanHap610Q, 1M-Duo and 1.2MDuo 1M chips and imputation was performed using the IMPUTE software package (v2), as previously described [26]. Expression profiling in subcutaneous adipose tissue was measured using Illumina Human HT-12 V3 BeadChips for 825 female individuals [27], 778 of whom had both genotype data and BMI information. We used LDAK to calculate a GSM based on allelic correlation across autosomes. Before calculation, we excluded SNPs with minor allele frequency <0.01. We compared effect size estimates of the gene expression profiles versus BMI, using either the family information or the inferred GSM. Age was included in the model as a fixed effect. We observed a very high concordance correlation coefficient of the effect size estimates for the association between gene expression levels and BMI (r β=0.99; Fig. 2, left), as well as of the gene-specific genetic component of genetic and environmental variances (\(r\sigma _{a}^{2}=0.99\) and \(r\sigma _{e}^{2}=0.99\), respectively; Fig. 2, right).