Genomic prediction with kinship-based multiple kernel learning produces hypothesis on the underlying inheritance mechanisms of phenotypic traits

Abstract

Background: Genomic prediction encompasses the techniques used in agricultural technology to predict the genetic merit of individuals towards valuable phenotypic traits. It is related to Genome Interpretation in humans, which models the individual risk of developing disease traits. Genomic prediction is dominated by linear mixed models, such as the Genomic Best Linear Unbiased Prediction (GBLUP), which computes kinship matrices from SNP array data, while Genome Interpretation applications to clinical genetics rely mainly on Polygenic Risk Scores.

Results: In this article, we exploit the positive semidefinite characteristics of the kinship matrices that are conventionally used in GBLUP to propose a novel Genomic Multiple Kernel Learning method (GMKL), in which the multiple kinship matrices corresponding to Additive, Dominant, and Epistatic Inheritance Mechanisms are used as kernels in support vector machines, and we apply it to both worlds. We benchmark GMKL on simulated cattle phenotypes, showing that it outperforms the classical GBLUP predictors for genomic prediction. Moreover, we show that GMKL ranks the kinship kernels representing different inheritance mechanisms according to their compatibility with the observed data, allowing it to produce hypotheses on the normally unknown inheritance mechanisms generating the target phenotypes. We then apply GMKL to the prediction of two inflammatory bowel disease cohorts with more than 6500 samples in total, consistently obtaining results suggesting that epistasis might have a relevant, although underestimated role in inflammatory bowel disease (IBD).

Conclusions: We show that GMKL performs similarly to GBLUP, but it can formulate biological hypotheses about inheritance mechanisms, such as suggesting that epistasis influences IBD.

Fig. 1

Figure showing the comparison between the 9 methods we benchmarked on the CATTLE dataset. The MEAN, FH, and CKA GMKL approaches using 1 to 5 kernel matrices are respectively shown in shades of green, red and blue. The gray bars show the additive GBLUP model (light) and the optimal GBLUP model (dark), which always use the optimal set of kinship matrices. The phenotypes ranging from zero to four involve only A and D effects, with Pheno:0 being 100% additive, Pheno:2 being 50/50%, and Pheno:4 being purely Dominant. Phenotypes 5–9 include also epistatic effects. They are composed by a base 33% A and D components, plus a 34% epistatic component that is additive-additive (Pheno:5), additive-dominant (Pheno:6), and dominant-dominant (Pheno:7). Pheno:8 contains a mixture of all effects

Fig. 2

Figure showing the balanced accuracy (BAC), area under the ROC curve (AUC), and area under the precision-recall curve (AUPRC) for the benchmark of our GMKL approaches with GBLUP and other ML methods on the IBDSNP dataset

Fig. 3

Figure showing the comparison between the nine methods we benchmarked on the IBDWES dataset. The MEAN, FH, and CKA GMKL approaches using one to five kernel matrices are respectively shown in shades of green, red, and blue. The black to gray bars show the additive GBLUP models, while the bottom bars show the performance obtained by the Neural Networks methods proposed in [50]