The Relationship Between Polygenic Risk Scores and Clinical Phenotype in Patients with Phenylketonuria: Genetic Prediction with the Random Forest Model

Summary: Background /Aims: Pathogenic variations in the PAH gene cause phenylketonuria (PKU), a monogenic metabolic disorder. Individuals with the same mutation often exhibit phenotypic variability despite the monogenic nature of the condition. The aim of this study is to create a prediction model using the Random Forest (RF) machine learning algorithm and to examine the relationship between polygenic risk scores (PRS) and phenotypic severity in PKU patients. Methods: In this study, clinical exome sequencing data obtained from 174 PKU patients with molecular validation were retrospectively examined. Approximately 18,000 common variants were retained after being filtered by population frequency and quality for individual-level analysis. All eligible variants (excluding PAH mutations) were used to calculate PRS, and RF (1000 trees, maximum depth = 5) was used for modeling. International criteria were used to classify patients into mild, moderate, and severe phenotypes. Pearson correlation and ROC analysis were used to evaluate the model's performance. Findings: The RF-based PRS model had a high accuracy rate in predicting phenotypic severity (AUC = 0.91, overall accuracy = 84.3%). There was a significant correlation between PRS values and the severity of the phenotype (r = 0.68, p < 0.001). Severe clinical phenotypes were more common in patients with higher PRS. Variants in genes associated with phenylalanine metabolism (e.g., GCH1, QDPR, PTS) were the most significant contributors to risk prediction according to feature importance analysis results. Results: The results indicate that PRS modeling combined with machine learning could be a useful method for predicting the severity of phenotypes in monogenic disorders such as PKU. This integrative approach highlights the regulatory effect of a polygenic background and suggests that PRS could support clinical risk assessment and personalized treatment plans. However, before clinical application, it is very important to validate in various populations.