CIT-Lasso: a scalable approach beyond guilty by association for identifying causal variants from genome-wide summary statistics

Abstract

We present CIT-Lasso, a framework that uses only summary statistics to identify, genome-wide, sets of variants carrying non-redundant information on a phenotype, distinguishing likely causal variants from correlated variants that are merely associated. The open-source implementation completes genome-wide analysis in under 15 min on one CPU. In simulations, it outperforms existing methods in false discovery rate control, power, and fine-mapping resolution. Applied to an Alzheimer's disease meta-analysis, it identified 82 loci, 37 beyond conventional GWAS; prior MPRA and CRISPR-Cas9 studies corroborate prioritized variants. Results on other 67 large-scale GWAS reveal the method's generalizability to make discoveries beyond conventional GWAS pipeline.