Integrating supervised and unsupervised machine learning with mendelian randomization to discover and validate early predictive biomarkers for phase separation-related renal fibrosis

• Gene data from renal fibrosis patients and normal controls were obtained from the GEO database. • WGCNA, GSEA, and three machine learning methods (SVM-RFE, RF, LASSO) were used to analyze and refine the gene set. • A diagnostic model was built and evaluated using a nomogram and gene set enrichment analysis. • Five core genes (CELF6, HSPA1A, MKNK1, NSUN5, PAN2) most linked to renal fibrosis were identified. • Mendelian randomization revealed a causal link between high expression of the viral myocarditis pathway (influenced by these genes) and renal fibrosis development, explaining the mechanism. Kidney fibrosis is a prevalent chronic renal disorder that can result in scarring, organ failure, and systemic complications. Current diagnostic approaches, predominantly core needle biopsy, are limited, especially for early detection. This study seeks to identify biomarkers for early diagnosis by analyzing gene expression datasets and employing advanced algorithms, focusing on genes and pathways involved in phase separation and immune cell infiltration. Gene expression data for kidney fibrosis were obtained from the Gene Expression Omnibus (GEO) database. Differential expression and functional enrichment analyses were performed on samples, including normal controls. Three supervised machine learning algorithms, including Support Vector Machine-Recursive Feature Elimination (SVM-RFE), were used to identify candidate biomarkers. Model performance was assessed through nomograms, batch Receiver Operating Characteristic (ROC) analysis, and boxplots. Single-sample gene set enrichment analysis (ssGSEA) was applied to evaluate gene set activity variations, while Mendelian randomization (MR) inferred causal relationships between enriched pathways and disease progression. The MR model's validity was confirmed using six unsupervised machine learning algorithms. The GSE66494 dataset from the GEO database included 61 samples. Gene set enrichment analysis (GSEA) identified 30 upregulated genes, including TTLL3. Machine learning algorithms highlighted HSPA1A and NSUN5 as key genes in kidney fibrosis. The nomogram (area under the curve, AUC = 1.000) and boxplot (p < 0.05) validated model accuracy. Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis revealed 18 metabolic pathways linked to kidney fibrosis. MR analysis identified significant associations with acute kidney failure (p = 0.01) and craniosynostosis (p = 0.0006), both associated with HSPA1A. NSUN5 and HSPA1A genes, along with the acute kidney failure pathway, are identified as reliable biomarkers for kidney fibrosis, with potential to advance diagnostics and therapeutic development. This study leverages machine learning and Mendelian randomization to identify NSUN5 and HSPA1A as significant biomarkers for early renal fibrosis diagnosis. By integrating advanced algorithms and bioinformatics, the research highlights the importance of phase separation and immune cell infiltration pathways. These findings offer promising directions for enhancing clinical diagnostics, treatment strategies, and pharmaceutical development for chronic kidney disease.