A polygenic score is commonly derived using genome-wide genotype data to summarize the genetic contribution to a particular disease at the individual level. Usually it is constructed as a linear combination of SNP genotype weighted by the SNP-wise regression coefficient of the SNP to the phenotype using SNPs with p values smaller than a particular threshold. Commonly a range of thresholds are used which can pose problems with multiple comparisons as well as over-fitting. Here, an alternative weighting scheme is proposed, making use of the local true discovery rate, estimated from summary statistics. Two methods of estimation are proposed—maximum likelihood and kernel density estimation. Simulation studies using real and artificial data suggest this new weighting scheme is highly comparable with standard polygenic scores using the best possible p value threshold in prediction, even though this threshold is not normally known in practice.