Exploration of Potential Biomarkers and Immune Landscape for Hepatoblastoma: Evidence from Machine Learning Algorithm

This study aimed to investigate the immune landscape in hepatoblastoma (HB) based on deconvolution methods and identify a biomarkers panel for diagnosis based on a machine learning algorithm. Firstly, we identified 277 differentially expressed genes (DEGs) and differentiated and functionally identified the modules in DEGs. The Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis and GO (gene ontology) were used to annotate these DEGs, and the results suggested that the occurrence of HB was related to DNA adducts, bile secretion, and metabolism of xenobiotics by cytochrome P450. We selected the top 10 genes for our final diagnostic panel based on the random forest tree method. Interestingly, TNFRSF19 and TOP2A were significantly down-regulated in normal samples, while other genes (TRIB1, MAT1A, SAA2-SAA4, NAT2, HABP2, CYP2CB, APOF, and CFHR3) were significantly down-regulated in HB samples. Finally, we constructed a neural network model based on the above hub genes for diagnosis. After cross-validation, the area under the ROC curve was close to 1 (AUC = 0.972), and the AUC of the validation set was 0.870. In addition, the results of single-sample gene-set enrichment analysis (ssGSEA) and deconvolution methods revealed a more active immune responses in the HB tissue. In conclusion, we have developed a robust biomarkers panel for HB patients.