Our polygenic prediction models explained up to 17% and, on average 13% of variance in European cohorts. Given that SNP-heritability estimates are the theoretical upper limit for polygenic prediction, the LDSC estimates are clearly underestimates.

Using SBayesRC, which models a mixture of SNP effect sizes and can better account for large-effect variants, we estimated SNP-heritability at ~19% (vs. 6% from LDSC). Estimates were similar across African and East Asian ancestries.

We found enrichment for:
Upregulated genes in microglia, and downregulated genes in three neuronal subtypes (Sncg, Sst, and L6 IT Car3).

We identified 118 loci in a multi-ancestry GWAS and 9 more in a European-only GWAS (total = 127 loci).
Of these, 48 were novel, including 8 potential drug targets: QPCT, EGFR, KEAP1, SYK, AXL, RRM2B, CACNA1S, and IL23A.

Summary: We analyzed ~180K cases & 2.6M controls, identified 127 loci (48 new), improved heritability estimates (19% in Europeans) & PGS prediction (mean 13%), found potential drug targets, and enrichment in microglia and three neuronal cell types.
More details below ⬇️

We show in simulations and empirical data that this simple way of estimating R2 works surprisingly well, outperforming another published approach.

In a population reference sample (e.g., 1000 Genomes), where the sample disorder prevalence is the same as in the population, the variance of a PGS on the liability scale will be equal to its R2. That is, no phenotype data is required.

We also compared the calibration of BPC to other methods using tuning samples (with geno- and phenotype data) and show that it performs similarly at smaller tuning sample sizes, but worse at larger tuning sample sizes. Because tuning samples are difficult to obtain, BPC may often be preferred.

It is also well calibrated in empirical analyses, where we analyzed 9 disorders of varying genetic architectures

We show in simulations, across different parameter settings, that the BPC approach is very well calibrated, outperforming a published method.

This is achieved by transforming a Bayesian PGS (computed using an existing method, e.g., PRS-CS or SBayesR) to its underlying liability scale, estimating the variances of the PGS in cases and controls based on theory, and applying Bayes’ Theorem to compute the probability.

E.g., when we applied this test to gene boundaries, we found APOE showed substantial differences for LDL, with larger effects and h² in females (i.e., 6% vs. 3%)

6/7

Some traits (e.g., BMI) had global correlations close to one but harbored loci with local correlations as low as –0.12.
Testosterone was the other extreme. It had a global correlation of zero, yet mixed strongly positive and negative local correlations.

4/7

118/157 traits had at least one local genetic correlation significantly different from one. 205 of these loci were negative. Most global genetic correlations are close to one and can mask these differences.

3/7

146 1Mb-loci across 47 traits showed significant h² differences between the sexes. Blood biomarkers (e.g., testosterone, urate) stood out, for example, with a locus on chromosome 4 showing a heritability of 10% in females but ~3% in males for urate.

2/7