A Multivariate Genome-Wide Association Study of Wing Shape in Drosophila melanogaster

Abstract

Due to the complexity of genotype-phenotype relationships, simultaneous analyses of genomic associations with multiple traits will be more powerful and informative than a series of univariate analyses. However, in most cases, studies of genotype-phenotype relationships have been analyzed only one trait at a time. Here, we report the results of a fully integrated multivariate genome-wide association analysis of the shape of the Drosophila melanogaster wing in the Drosophila Genetic Reference Panel. Genotypic effects on wing shape were highly correlated between two different laboratories. We found 2396 significant SNPs using a 5% false discovery rate cutoff in the multivariate analyses, but just four significant SNPs in univariate analyses of scores on the first 20 principal component axes. One quarter of these initially significant SNPs retain their effects in regularized models that take into account population structure and linkage disequilibrium. A key advantage of multivariate analysis is that the direction of the estimated phenotypic effect is much more informative than a univariate one. We exploit this fact to show that the effects of knockdowns of genes implicated in the initial screen were on average more similar than expected under a null model. A subset of SNP effects were replicable in an unrelated panel of inbred lines. Association studies that take a phenomic approach, considering many traits simultaneously, are an important complement to the power of genomics.

Keywords: Drosophila wing; GP map; developmental genetics; genome-wide association analysis; multivariate GWAS; phenomics.

Figure 1

Capturing wing shape with a spline model. Closed circles are landmarks formed by the intersection of splined veins and open circles are semilandmarks used to represent the locations of veins. (A) Splines fitted to a typical D. melanogaster wing. Colored lines are the splines. (B) Blue overlay represents the range of shape variation among Drosophila Genome Reference Panel lines.

Figure 2

Relationship between experiments and analyses. Ellipses represents stocks or results that are the input to this work. Boxes with no fill are experiments or preliminary analyses only described in Materials and Methods. Filled boxes generate the major results presented. DGRP, Drosophila Genome Reference Panel; GO, gene ontology; LASSO, Least Absolute Shrinkage and Selection Operator; LD, linkage disequilibrium; MAC, minor allele count; MANOVA, multivariate analysis of variance analysis; ME-NC, Maine-North Carolina; PCA, principal component analysis.

Figure 3

Interlaboratory repeatability. (A) High repeatability of line effect sizes for shape across laboratories. (B) Low repeatability of size across laboratories, despite high intralaboratory repeatability (Figure S1B).

Figure 4

Manhattan plots of the log₁₀ inverse P-values from multivariate analysis (upper panel) and univariate analysis of PC1 (principal component 1) (lower panel). Solid red line is P = 0.00007, the cutoff for a 5% false discovery rate (FDR) using the Storey and Tibshirani analysis of the multivariate data. The same cutoff is also applied in to the PC1 analysis. Green points are the four SNPs that reach the 5% FDR cutoff from analysis of just the PC1 P-values.

Figure 5

Quantile–quantile plot of observed vs. expected P-values genome-wide in the multivariate analysis. Black: all SNPs; red: SNPs with minor allele frequency (MAF) < 0.15; blue: SNPs with MAF > 0.15.

Figure 6

Mean measures of multivariate and univariate effect size for SNPs categorized by significance of the univariate test on each PC using P = 0.00007 as a cutoff. Gray squares: total multivariate effect size for SNPs significant in the corresponding univariate analysis; red squares: univariate effect size for SNPs significant in the corresponding univariate analysis (also shown in Figure 5); green diamonds: univariate effect size score for SNPs significant in the multivariate analysis; and blue circles: univariate effect size score for all SNPs. Horizontal reference lines show the mean multivariate effect size for all SNPs (solid line) and for all SNPs significant in the multivariate analysis of variance analysis (dashed line). multiv., multivariate; PC, principal component; signif., significant.

Figure 7

Wing shape deformations inferred for gene knockdowns and SNP effects implicating corresponding genes. All deformations shown magnified 2× except for knirps knockdown. Deformation scale is log2, so +0.2 represents a 15% increase in the corresponding area of the wing. (A) Effects of different levels of Egfr knockdown on wing shape. (B) Comparison of knockdown (left) and LASSO SNP vector (right) for 2R:17440366, which is in an intron of Egfr. The Egfr knockdown is the regression of the shape changes shown in (A) on the level of mifepristone applied. The correlation between these vectors is 0.68. (C) Comparison of knockdown of knirps (left) and LASSO SNP vector for 3L:20685772, 6558-bp downstream from knirps. The correlation between these vectors is 0.54. LASSO, Least Absolute Shrinkage and Selection Operator.

Figure 8

Wing-shape deformations inferred for a SNP in the intron of Lar. Above: effect of SNP 2L:19596734 in the DGRP. Below: effect of SNP 2L:19596734 in the ME-NC2 replication population. Vector correlation between the vectors of effects from the DGRP and the ME-NC2 was 0.84. Both effects shown at 3×. DGRP, Drosophila Genome Reference Panel; ME-NC, Maine-North Carolina.