Multi-tissue transcriptome-wide association studies

Abstract

A transcriptome-wide association study (TWAS) attempts to identify disease associated genes by imputing gene expression into a genome-wide association study (GWAS) using an expression quantitative trait loci (eQTL) data set and then testing for associations with a trait of interest. Regulatory processes may be shared across related tissues and one natural extension of TWAS is harnessing cross-tissue correlation in gene expression to improve prediction accuracy. Here, we studied multi-tissue extensions of lasso regression and random forests (RF), joint lasso and RF-MTL (multi-task learning RF), respectively. We found that, on our chosen eQTL data set, multi-tissue methods were generally more accurate than their single-tissue counterparts, with RF-MTL performing the best. Simulations showed that these benefits generally translated into more associated genes identified, although highlighted that joint lasso had a tendency to erroneously identify genes in one tissue if there existed an eQTL signal for that gene in another. Applying the four methods to a type 1 diabetes GWAS, we found that multi-tissue methods found more unique associated genes for most of the tissues considered. We conclude that multi-tissue methods are competitive and, for some cell types, superior to single-tissue approaches and hold much promise for TWAS studies.

Keywords: complex traits; gene expression; multi-task learning; transcriptome-wide association studies.

Figure 1

Pairwise comparison of performance of the MTL and STL expression prediction methods—$R^2$ on a test set. Each point represents a probe–cell pair. Points above the blue line show increased performance for the method to the left of each plot, while points below the blue line show increased performance for the method underneath the plot. The three numbers represent, clockwise: points with positive $R^2$ above $x=y$ line for the $x$‐axis method, points with positive $R^2$ below the line for the $y$‐axis method, points with negative $R^2$ for both methods. Numbers in brackets represent the corresponding advantage of one method over the other, in terms of $R^2$ (for this calculation negative $R^2$ are taken to be 0). For example, comparing lasso and RF, lasso outperformed RF in 2148 regressions with an advantage of 3.5%, while RF outperformed lasso in 9667 with an advantage of 5.9%, and for 9625 probe–cell pairs neither method achieved a positive $R^2$. MTL, multi‐task learning; RF, random forest; STL, single‐task learning

Figure 2

Power of different methods to detect TWAS association. In the top row, the GWAS and test eQTL traits share causal variant A, while the causal variant for the four background eQTL traits varies (left‐right) from none, to B to A. The bottom row is the same, except the GWAS and eQTL‐test causal variants are different. The total shaded column height is the proportion of TWAS tests that pass p <.05, with lighter shading used to indicate the proportion of tests which would be filtered out proportionality testing at p <.05. The horizontal dotted line is at $y=0.05$, the proportion of false‐positives expected in a well controlled testing procedure in the bottom row. eQTL, expression quantitative trait loci; GWAS, genome‐wide association study; TWAS, transcriptome‐wide association study

Figure 3

Volcano plots for testing association between the predicted gene expression and the T1D status. Grey points are not TWAS‐significant, blue points are TWAS—but not passing proportionality test, and orange points are both TWAS—and proportionality‐significant (SP‐hits). RF, random forest; TWAS, transcriptome‐wide association study

Figure 4

Unique TWAS‐significant hits passing proportionality filtering, by method: lasso (13), RF (21), joint lasso (36), and RF‐MTL (18). RF, random forest; MTL, multi‐task learning; TWAS, transcriptome‐wide association study

Figure 5

A heatmap of genes identified by the four methods after proportionality filtering (top), integrated with a Manhattan plot of type 1 diabetes GWAS. Arrows point to GWAS peaks (red stars) in the vicinity of which (1 Mbp either way) a gene (or several genes, grouped by a bracket) lies. Vertical dotted lines indicate positions of genes; horizontal dotted line is at $-{\log }_{10}p=5$, corresponding to a GWAS significant level of $10^{-5}$; green and purple colors in the Manhattan plot designate alternating chromosomes. Note that the genes in the heatmap are ordered according to their positions, so for any two genes (or groups of genes) an arrow from a leftmost one would point to a peak left of the peak pointed at by the rightmost gene. Any intersection between the arrows is due to the fact that they might point to peaks of vastly different heights. GWAS, genome‐wide association study