TEDML: a new machine learning (ML) approach for predicting thyroid eye disease and identifying key biomarkers.

Abstract

Thyroid eye disease (TED) features immune infiltration and metabolic dysregulation. Understanding these processes and identifying potential biomarkers are crucial for improving diagnosis and treatment. To this end, immune cell infiltration was analyzed and gene set variation analysis (GSVA) was conducted on the GSE58331 dataset to identify differences between TED and normal tissues. Differentially expressed genes were identified using GSE58331 and GSE105149. Subsequently, a prediction model (TEDML) was developed by combining 113 machine learning algorithms to identify key biomarkers. In addition, enrichment analyses were performed to understand biological functions and pathways involved in TED, and drug sensitivity analyses were conducted to identify potential therapeutic agents. Immune infiltration analysis revealed higher levels of CD4+ Tem, CD4+ Tcm, NKT, NK cells and neutrophils in TED patients compared to controls, with lower levels of macrophages M1 and M2. GSVA indicated significant enrichment in immune-related processes and metabolic pathways. The TEDML model, constructed from the Stepglm[forward] algorithm, demonstrated high accuracy (area under curve of 1 on the training set, 0.893 in validation set), identifying six key genes (CSF3R, ALDH1A1, MXRA5, VSIG4, DPP4 and MDH1). Drug sensitivity analysis suggested that azathioprine and methylprednisolone might be effective at different stages of TED, with CSF3R as a potential therapeutic target. Overall, the TEDML model is accurate and reliable, and the identification of CSF3R as a key biomarker and its correlation with drug sensitivity offers new insights into targeted therapy for TED.