Integrated bioinformatics and machine learning analysis identify ACADL as a potent biomarker of reactive mesothelial cells.

Mesothelial cells with reactive hyperplasia are difficult to distinguish from malignant mesothelioma cells on cell morphology. This study aimed to identify and validate potential biomarkers that distinguish mesothelial cells from mesothelioma cells through machine learning combined with immunohistochemistry (IHC) experiments. We integrated the gene expression matrix from three GEO datasets (GSE2549, GSE12345, GSE51024) to analyze the differently expressed gene (DEGs) between normal and mesothelioma tissues. Then three machine learning algorithms, least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and random forest (RF) were used to screen and obtain four shared candidate markers, including ACADL, EMP2, GPD1L, HMMR. The receiver operating characteristic curve (ROC) analysis showed that the area under the curve (AUC) for distinguishing normal from mesothelioma was 0.976, 0.943, 0.962, and 0.956, respectively. The expression and diagnostic performance of these candidate genes were validated in another two independent datasets (GSE42977 and GSE112154), indicating that the performances of ACADL, GPD1L, and HMMR were consistent between the training and validation datasets. Finally, the optimal candidate marker ACADL was verified by IHC assay. ACADL was strongly stained in mesothelial cells, especially for reactive hyperplasic mesothelial cells, but was negative in malignant mesothelioma cells. Therefore, ACADL has the potential to be used as a specific marker of reactive hyperplasic mesothelial cells in the differential diagnosis of mesothelioma.