Comparison of the evaluated methods on synthetic data sets with various sample sizes and numbers of genomic regions harboring interacting variants. The values shown are the differences in prediction accuracy compared to AMB (top), and the number of region-kernels selected by each method (bottom). Larger samples allow using more expressive models with a larger number of kernels. The advantage of MKLMM methods over AMB, and of MKLMM-Adapt over MKLMM-Poly2, increases with sample size and decreases with the number of regions. The median lines in the bottom row are sometimes not shown because they intersect with the other quartiles, owing to the discrete nature of these data.
