A Data-Driven Approach to Polygenic Risk Scoring: Combining Support Vector Machine and Ridge Regression in Diseases Lacking GWAS Data Utilizing Lipedema as a Model

Multifactorial genetic diseases (as currently applicable for incompletely understood genetic conditions such as lipedema) present a polygenic architecture in which numerous common Single Nucleotide Polymorphisms (SNPs) contribute to overall risk with small summed effects. Traditional Polygenic Risk Scores (PRS) rely on GWAS-derived effect sizes, which are often unavailable or poorly transferable for understudied or rare diseases. This study proposes a summary-statistics-independent (external-GWAS-independent) pipeline for PRS construction and illustrates its application in lipedema. Our pipeline uses individual-level data (n=314; 208 cases, 106 controls) and preselected variants chosen based on biological relevance and physio-logical pathways involved. A standardized workflow includes quality control, alignment, variant calling, annotation, and genotype filtering. A linear Support Vector Machine (SVM) ranks variants by absolute weight to identify the most informative features. Ridge regression with L2 regularization is then applied to derive stable PRS weights. Model performance is evaluated using AUC, accuracy, F1 score, and Nagelkerke's pseudo-R². Analyses of PRS distributions and predictor importance emphasize interpretability and capture weak but cumulative signals. The model achieved an AUC of 0.87 and a Nagelkerke's pseudo-R² of 0.13, indicating meaningful capture of aggregated genetic risk. The proposed summary-statistics-independent pipeline offers a transparent and clinically interpretable strategy for PRS development in GWAS-scarce multifactorial diseases. The combined SVM-ridge strategy appears robust and 140 adaptable, with encouraging performance in lipedema and potential for translational risk stratification. These results should be considered proof-of-concept, pending external validation.