Identifying Psoriasis-Associated Genes Using Multi-Objective Optimization and Machine Learning

Psoriasis is a chronic autoimmune skin disease with a complex molecular basis and no definitive cure, making the identification of reliable gene biomarkers essential for early diagnosis and targeted therapy. This study presents a multi‑objective optimization framework that integrates the NSGA‑III algorithm with machine learning to select a minimal yet highly predictive gene signature from psoriasis gene expression data. A comprehensive dataset was compiled from seven GEO repositories, comprising 1,327 samples and 53,952 common features, and preprocessed through rigorous quality control, batch effect correction using ComBat, quantile normalization, and robust scaling. The NSGA‑III algorithm was configured to simultaneously optimize three conflicting objectives: maximizing the F1‑score, minimizing the Log Loss, and minimizing the number of selected features, with candidate gene subsets evaluated using a Support Vector Machine classifier. The optimal model achieved an F1‑score of 91.54% and a Log Loss of 0.28, selecting only 30 features from which a final panel of 20 unique genes was derived, including LAPTM5, VCAN, IGKC, MAPK7, OVOL2, MAP3K9, and PRLR, which are implicated in immune regulation, keratinocyte proliferation, and inflammatory signaling pathways. The proposed approach effectively reduces data dimensionality while preserving high classification accuracy, offering a robust and interpretable framework for biomarker discovery, and the identified gene panel provides promising candidates for future experimental validation and therapeutic targeting in psoriasis.