We believe that three aspects of PRS construction merit particular attention and need to be explored further in future studies. First, the method employed to construct the PRS. Conventional machine learning approaches, where the model is trained on raw genotype data, have been reported to outperform the GWAS-based approach used here37. However, such approach was not feasible because raw genotype data in large-scale studies were not available. In a GWAS-based approach, summary statistics data of GWAS are used to estimate risk score coefficients of genotype dosage. After initial use in schizophrenia21, this approach has proven successful in capturing and predicting the genetic influence on multiple complex polygenic traits38,39. Here we showed a PRS constructed in a GWAS-based approach successfully stratified patients into risk groups with distinct PTSD risk and severity levels in a cohort that is independent of the discovery GWAS samples. We expect uncertainties in the likelihoods and estimates will become lower as more data are amassed. The expected rate of this improvement is estimated from a theoretical analysis. Furthermore, advances in novel methodological approaches may accelerate this pace.
