Abstract
A recent study proposing a new LDAK method (Speed et al. 2017 Nat Genet) reported that functional enrichments (e.g. coding, conserved, regulatory) estimated by LDAK (largest significant enrichment: 2.51x) were much lower than previous estimates obtained using stratified LD score regression (S-LDSC). To investigate this, we developed a method (S-LDSC+LDAK) that combines our S-LDSC method with annotations constructed from LDAK model weights, and determined that this method produced unbiased estimates both in simulations under the S-LDSC model and in simulations under the LDAK model, unlike existing methods. We applied S-LDSC+LDAK to 16 independent UK Biobank traits, and determined that S-LDSC+LDAK enrichment estimates (largest enrichment: 7.51x) were nearly identical to S-LDSC estimates across 28 main annotations. On the other hand, LDAK enrichment estimates (largest enrichment: 3.96x) were substantially lower than S-LDSC estimates (although the discrepancy was smaller than reported by Speed et al., who did not compare the two methods on the same data set). Our results advocate for using S-LDSC in preference to LDAK to infer functional enrichment and confirm the existence of functional annotations that are highly enriched (≫2.51x) for complex trait heritability, providing strong caveats to the LDAK results reported by Speed et al.