Yajie Zhao, 赵亚杰

We’re excited to share our latest publication in @NatureComms: “Integrating common and rare variants improves polygenic risk prediction across diverse populations” 🧬🧪🫀 nature.com/articles/s41467-0… Sincere thanks to Jacob Williams (lead), @AndrewHaoyu (senior), @Tony_Chen6, Xing Hua, @wendyWongSW, Kai Yu, @GENES_PK, and all @uk_biobank + @AllofUsResearch participants whose data made this work possible. Most polygenic risk score (PRS) models focus on common variants. But rare variants can have larger effects for the people who carry them, and they may explain part of the genetic risk not captured by common-variant PRSs. To address this, we developed RICE (polygenic Risk predictions Integrating Common and rarE variants), a framework for integrating rare variant signals with common-variant PRSs across ancestries. RICE provides a unified framework to: • Build a robust common-variant PRS (RICE-CV) by ensembling existing PRS methods • Identify rare variant sets independent of RICE-CV using STAARpipeline + functional annotations • Collapse significant rare variant sets into burden scores • Train rare-variant PRSs (RICE-RV) with penalized regression • Combine RICE-CV and RICE-RV in an independent validation set for integrated risk prediction We evaluated RICE using extensive simulations and sequencing data from @uk_biobank and @AllofUsResearch, up to 740M genetic variants from 361,939 individuals across diverse ancestries and 11 complex traits. A few key findings: • RICE-CV consistently improved common-variant PRS performance, with average R² gains of 9.4% over the best alternative method for continuous traits across trait–ancestry combinations. • Adding rare variants through RICE-RV produced further gains, especially for lipid traits and height. • For lipid traits, incorporating rare variants increased R² by up to ~11.2% in Europeans and ~60.7% in African ancestry compared with common-variant PRS alone. • Rare variants also helped identify individuals with extreme lipid profiles that common-variant PRS alone may miss. For HDL cholesterol, 6.3% of individuals in the top phenotype decile had high RICE-RV but low RICE-CV. • The rare-variant signal was not limited to established high-penetrance lipid genes. RICE-RV captured predictive signal beyond LDLR, APOB, and PCSK9, supporting a broader polygenic rare-variant architecture. Key takeaway: Integrating common and rare variants can significantly improve genetic risk prediction and risk stratification across populations, while rare variants do not improve every trait equally - their value depends on genetic architecture and available sample size. The software and tutorial are open source: RICE package: github.com/jwilliams10/RICE Tutorial: jwilliams10.github.io/RICE_V… Manuscript code: github.com/jwilliams10/RareV…