A hybrid machine learning framework with two-step feature selection for identifying key biomarkers and drug targets in monkeypox

Monkeypox (Mpox) is a viral disease that has garnered global attention due to its human-to-human transmissibility and cross-species transmission. Recent outbreaks in both endemic and non-endemic regions highlight the urgent need to elucidate its molecular mechanisms and evolutionary dynamics. As of now, there are no approved antiviral therapies available for the treatment of Mpox. In this work, we proposed a robust machine learning (ML) and bioinformatics approach to uncover candidate biomarkers associated with Mpox. Two microarray datasets and one RNA-seq dataset were analyzed to screen an initial panel of genes according to p-values. A two-stage feature selection pipeline based on Fisher Score (FS) and Recursive Feature Elimination (RFE) was implemented on the selected gene set. The resulting optimized gene group was subsequently input into our suggested Multilayer Perceptron (MLP) model, which demonstrated superior classification accuracy compared to other classifiers. A total of 33 shared genes were discovered between the differentially expressed genes (DEGs) and the final selected group. Gene Ontology (GO) and KEGG signaling pathway analysis revealed critical biological functions and signaling pathways linked to the common genes. A protein–protein interaction (PPI) network analysis revealed 10 key genes. Regulatory network analysis, incorporating TF-gene and TF-miRNA interactions, highlighted five hub genes with strong associations with regulatory mechanisms. Additionally, molecular docking revealed that ATG3 , TRIM14 , and DUOX1 exhibited strong interactions with therapeutic compounds. These results indicate that the discovered hub genes may serve as potential candidates for prompt detection, prognosis, and therapeutic targeting in Mpox infection.