The complexity of genetic architecture of traits increases the challenges in polygenic risk score (PRS) construction and reduces accuracy of prediction, given that a single PRS method might not summarize the genomic susceptibility of traits comprehensively. Recent work has developed two-phase designs in re-sequencing studies where only informative subsamples are selected for cost-effective data collection. Here, we propose an optimization approach integrating multiple PRS methods in a two-phase design for re-sequencing studies. Set in linear regression framework, our model utilizes a constrained residual dependent sampling (RDS) design to accommodate a mixture of two PRS methods: LassoSum and LD-pred-inf, which posit contrasting assumptions with respect to the trait underlying genetic architecture. The method is evaluated against the traditional RDS designs with single or both PRS methods via simulation and applied to data from Northern Finland Birth Cohort of 1966 study.