Skip to content

MIXPRS

Summary

MIXPRS is a multi-population polygenic risk score (PRS) integration framework based on the data fission paradigm that combines multiple multi-population PRS methods using only GWAS summary statistics. By utilizing single nucleotide polymorphism (SNP) pruning to mitigate linkage disequilibrium (LD) mismatch and non-negative least squares (NNLS) regression to estimate combination weights, MIXPRS eliminates the need for individual-level tuning data. An extension, MIXPRS+, further integrates clinical PRSs and functional annotations to improve prediction accuracy in underrepresented populations.

Methodology

The MIXPRS framework operates in three key steps: 1. GWAS Subsampling via Data Fission: Partitions target population GWAS summary statistics into independent subsampled training and tuning datasets [1]. To mitigate LD mismatch, variants are pruned using a pairwise correlation threshold below 0.5 within sliding windows of 250 kb (via PLINK) using a matched reference panel like the 1000 Genomes Project, and the LD matrix is approximated using an identity matrix $I$. 2. Weight Estimation: Applies component multi-population PRS methods (specifically JointPRS and SDPRX) on the subsampled training data and estimates combining weights using Lawson–Hanson non-negative least squares (NNLS) regression [2] on the subsampled tuning data. 3. Derivation of MIXPRS: Recalculates PRS effect sizes using the original full GWAS summary statistics from all populations, and combines them using the estimated weights to produce the final score.

MIXPRS+ Extension

MIXPRS+ extends the framework to incorporate: - ClinicalPRS: Clinically oriented PRSs derived from large meta-analysis GWAS. - Annotation-Informed PRS: Incorporates functional annotations via SBayesRC. - Implementation: Uses the full HapMap3 SNP set with 1000 Genomes Project LD reference panels instead of SNP pruning and identity covariance approximation.

Performance and Validation

Across simulations and real-data analyses (using UK Biobank and All of Us datasets) spanning up to 26 traits across non-European populations (EAS, AFR, SAS, and AMR), MIXPRS achieved: - Comparable prediction accuracy to individual-level tuning (IndPRS) and idealized summary-based (IdealPRS) methods. - Consistent outperformance of existing multi-population PRS methods (including JointPRS, XPASS, SDPRX, PRS-CSx, MUSSEL, PROSPER, and BridgePRS). - An average improvement in prediction accuracy in simulations of 3.2% over JointPRS, 213.1% over XPASS, 8.7% over SDPRX, 0.4% over PRS-CSx, 19.4% over MUSSEL, 19.5% over PROSPER, and 47.5% over BridgePRS across four non-European populations.

See Also

Citations

  • [1] Leqi Xu, Yikai Dong, Xiaowei Zeng, Zeyu Bian, Geyu Zhou, Leying Guan & Hongyu Zhao. (2026). MIXPRS enables multi-population and multi-method polygenic risk scores using summary statistics. Nature Genetics, 58(7), 1583–1594. DOI: 10.1038/s41588-026-02637-4. Source paper: s41588-026-02637-4.pdf
  • [2] Lawson, C. L. & Hanson, R. J. (1995). Solving Least Squares Problems. SIAM.