We transformed our results to the liability scale for comparison with narrow-sense heritability estimates (Fig. 7C) (21). Prior studies have estimated that common SNPs explain 25.6, 20.5, and 19% of the genetic variance for SCZ, BPD, and ASD, respectively (51). These may be taken as theoretical upper bounds for additive models, given unlimited common-variant data. By contrast, nonlinear predictors can exceed these limits. Our best liability scores (from just the genotype at QTL-associated variants) are substantially below these bounds, implying that additional data would be beneficial. By contrast, the variance explained by the full DSPN model exceeds that explained by common SNPs in SCZ and BPD, possibly reflecting the influence of rare variants and epistatic interactions (32.8 and 37.4% respectively—the variance of 11.3% for ASD is slightly lower). However, these estimates may be confounded by trait-associated variation that is environmental in origin (fig. S47).
