Improving genetic risk prediction by leveraging pleiotropy

Abstract

An important task of human genetics studies is to predict accurately disease risks in individuals based on genetic markers, which allows for identifying individuals at high disease risks, and facilitating their disease treatment and prevention. Although hundreds of genome-wide association studies (GWAS) have been conducted on many complex human traits in recent years, there has been only limited success in translating these GWAS data into clinically useful risk prediction models. The predictive capability of GWAS data is largely bottlenecked by the available training sample size due to the presence of numerous variants carrying only small to modest effects. Recent studies have shown that different human traits may share common genetic bases. Therefore, an attractive strategy to increase the training sample size and hence improve the prediction accuracy is to integrate data from genetically correlated phenotypes. Yet, the utility of genetic correlation in risk prediction has not been explored in the literature. In this paper, we analyzed GWAS data for bipolar and related disorders and schizophrenia with a bivariate ridge regression method, and found that jointly predicting the two phenotypes could substantially increase prediction accuracy as measured by the area under the receiver operating characteristic curve. We also found similar prediction accuracy improvements when we jointly analyzed GWAS data for Crohn's disease and ulcerative colitis. The empirical observations were substantiated through our comprehensive simulation studies, suggesting that a gain in prediction accuracy can be obtained by combining phenotypes with relatively high genetic correlations. Through both real data and simulation studies, we demonstrated pleiotropy can be leveraged as a valuable asset that opens up a new opportunity to improve genetic risk prediction in the future.

Fig. 1

Prediction accuracy of different methods on the BARD-SZ data. “BVR”: bivariate ridge regression; “UVR”: univariate ridge regression. The numbers in the brackets are the mean AUCs achieved by each method in the 50 repeats.

Fig. 2

Prediction accuracy of the bivariate ridge regression after shuffling the SNP identities and of ridge regression. Red plots represent the results of bivariate ridge regression and blue plots represent those of ridge regression. The percentage below each red plot represents the fraction of the SNPs that were shuffled.

Fig. 3

The prediction accuracy of bivariate and univariate ridge regression for BARD and SZ with subsamples.

Fig. 4

Prediction accuracy of different methods on the CD-UC data. “BVR” and “UVR” are defined as in Figure 1. The numbers in the brackets are the mean AUCs achieved by each method in the 50 repeats.

Fig. 5

Prediction accuracy of the bivariate ridge regression after shuffling the SNP identities and of ridge regression. Red plots represent the results of bivariate ridge regression and blue plots represent those of the ridge regression. The percentage below each red plot represents the fraction of the SNPs that were shuffled.

Fig. 6

The prediction accuracy of bivariate and univariate ridge regression for CD and UC with subsamples.

Fig. 7

Simulation results for the case when the two diseases have equal sample sizes and h² levels and m = 1000. “BVR” and “UVR” are defined as in Figure 1. Two h² levels (0.3 and 0.6) and two sample sizes (1000 and 2000) were simulated. The proportion of shared causal SNPs, γ was varied from 0 to 1 with an increment of 0.25. The numbers below the UVR box plots are the sample sizes. Following the UVR box plots are the box plots representing the results of BVR with the same sample sizes at different γ values (below the BVR box plots).

Fig. 8

Simulation results for the case when the two diseases have unequal sample sizes and equal h² levels and m = 1000. “BVR” and “UVR” are defined as in Figure 1. One of the diseases has 2000 samples and the other has 1000 samples. Two h² levels (0.6 and 0.3) were simulated. The proportion of shared causal SNPs, γ was varied from 0 to 1 with an increment of 0.25. The numbers below the UVR box plots are the h² levels. Following the UVR box plots are the results of BVR with the same h² levels at different γ values (below the BVR box plots).

Fig. 9

Simulation results for the case when the two diseases have equal sample sizes and unequal h² levels and m = 1000. “BVR” and “UVR” are defined as in Figure 1. One of the diseases has h² = 0.6 and the other has h² = 0.3. Two sample sizes (2000 and 1000) were simulated. The proportion of shared causal SNPs, γ was varied from 0 to 1 with an increment of 0.25. The numbers below the UVR box plots are the sample sizes. Following the UVR box plots are the results of BVR with the same sample sizes at different γ values (below the BVR box plots).