Benchmarking PRS methods for risk prediction of 36 complex traits in UK Biobank by establishment of a large-scale PRS computation platform PRS-hub

Polygenic risk score (PRS) provides a personalized measure of genetic contribution to complex traits, spurring the development of numerous methods for its estimation. Currently, there are no widely comparative studies that are sufficiently tailored to externally benchmark the efficacy of various methods, nor are there standardized platforms for efficient, large-scale PRS computation. To bridge these gaps, we develop a web-based PRS computation platform called PRS-hub and present a comprehensive evaluation of 13 cutting-edge single- and multi-ancestry PRS methods in 36 complex traits utilizing large cohort from UK Biobank. Results of single-ancestry methods demonstrate that although LDpred2 has robust superior performance across a broad spectrum of complex traits for accuracy and computational efficiency, other methods remain valuable and outperform for certain traits. In the case of multi-ancestry methods, PRS-CSx and X-Wing have comparable performance yet LDpred2-meta outperforms them both in European and African ancestry. Moreover, we find that increasing the panel size of LD reference significantly elevates PRS performance within the confines of 50 to ~1,000 and reaches the plateau phase when LD sample size exceed 1,000. Based on these results, we recommend employing the best-performing PRS method for each trait as identified in our study. The PRS-hub platform will aid in the development and clinical utility of the PRS field.