Fig. 4: Impact of changing the heritability model when using summary statistics.

a For each of the 225 phenotypes, we compare PRS constructed using LDAK-BayesR-SS assuming either the LDAK-Thin or BLD-LDAK Model, to PRS constructed using our implementations of the existing tools lassosum, sBLUP, LDpred and SBayesR. We measure the accuracy of PRS via R², the squared correlation between observed and predicted phenotypes. The x-axis reports highest R² across the four existing tools, while the y-axis reports the percentage increase in R² if instead of using the existing tool with highest R², we use LDAK-BayesR-SS assuming either the LDAK-Thin or BLD-LDAK Model (improvements above 50% are truncated). b The same as a, except phenotypes are grouped based on highest R² across the four existing tools: 0.01–0.05 (106 phenotypes), 0.05–0.10 (51 phenotypes) or 0.10–0.33 (68 phenotypes). Boxes mark the median increase in R² and the inter-quartile range. Source data are provided within the Source Data file.