COAST-SS is available in an updated AllelicSeries R package (github.com/insitro/Alle...). Congrats to all coauthors @zmccaw.bsky.social, Jianhui Gao, Rounak Dey, Simon Tucker, Yiyan Zhang, Jessica Gronsbell, Xihao Li, Emily Fox, @codushlaine.bsky.social and the Research Team @insitro.bsky.social!

Further, because COAST-SS operates on summary statistics instead of individual-level genotypes, it can complete a genome-wide scan 100x faster than the original COAST.

Of the 108 candidate allelic series identified in our analysis, 80 had prior rare variant support, and 23 of the remaining 28 associations had common variant support. The remaining 5 associations, all with HDL, are reported here for the first time.

In a combined analysis of UKB and MVP with a sample size of up to 840K, we identify 108 Bonferroni-significant genes associated with lipid traits, including 41 for HDL (below), more than doubling the number of genes identified in our previous analyses.

Though we recommend using LD calculated in a relevant sample, due to the low LD among rare variants, the test remains valid even when variants are assumed to be in linkage equilibrium.

In simulations and real data, we show that COAST-SS is robust to how linkage disequilibrium (LD) is specified.

Here we present COAST-SS, which enables detection of allelic series genes based entirely on per-variant GWAS summary statistics.

The original COAST method (www.cell.com/ajhg/fulltex... enables identification of dose-response relationships between genes and phenotypes (“allelic series genes”), empowering rare variant association testing, but requires access to individual-level genetic and phenotypic data.

@tkaraletsos.bsky.social, Daphne Koller, Hugues Aschard, @mendelrandom.bsky.social, @dgmacarthur.bsky.social, @codushlaine.bsky.social, and additional members of the team at @insitro.bsky.social !

Congrats to the team and our collaborators! @zmccaw.bsky.social, Rounak Dey, Hari Somineni, David Amar, @smukherjee89.bsky.social, Kaitlin Sandor, ...

In summary, associations with ratio traits are not specific to the ratio per se. Practitioners considering GWAS on ratios should examine the motivation for forming the ratio. We provide guidance on how to approach GWAS of related pairs of traits in light of analysis objectives.

In the case where the denominator is heritable and collider bias is a concern, we propose additionally adjusting for a leave-one-chromosome-out polygenic score (LOCO-PGS) of the denominator, which we show corrects for collider bias due to genetic background.

Instead of a ratio model, we propose a model that adjusts for the denominator as a covariate, e.g. to detect the genetic effect on weight while holding height constant. Loci from weight adjusted for height GWAS were enriched for adiposity-related signals more so than loci from the BMI GWAS.

Examining ratio traits commonly used in clinical practice and genetic association testing, such as waist-to-hip ratio and albumin-to-creatinine ratio, we find that loci detected by the ratio model are enriched for denominator signals.

Top “hits” from ratio trait GWASs can be driven entirely by the denominator. This means, for example, that attempts to find genes associated with BMI may return genes associated entirely with height (only), and not weight.

With high-dimensional omics and imaging data, there is a temptation to test all pairs of ratios to increase power. We caution against this approach; GWAS on ratio traits may increase the number of significant associations, but it does so by conflating signals for the numerator and denominator.