Leveraging machine learning and single-cell RNA sequencing strategies to develop a risk prognosis scoring based on liquid-liquid phase separation feature genes in pediatric hepatoblastoma.
AI interpretation is pending for this paper.
Open original publication →What the AI sees
Not AI summarized yet.
Research significance
Pending deeper interpretation.
Source abstract
BACKGROUND: Considerable evidence highlights the intricate association between liquid-liquid phase separation (LLPS) and tumorigenesis, progression, and therapy resistance. However, there has been limited exploration of the role of LLPS in hepatoblastoma (HB). This study integrates machine learning techniques with single-cell RNA sequencing (scRNA-seq) to systematically analyze the molecular features of LLPS-associated genes in HB and establish the first LLPS-based prognostic prediction model for HB. METHODS AND MATERIALS: RNA-seq data from the Gene Expression Omnibus (GEO) database were utilized, integrating GSE133039 and GSE75271 as the training cohort and GSE81928 and GSE132037 as the validation cohort. Differential expression analysis was performed on the training cohort, identifying 124 HB-specific differentially expressed genes. Weighted gene co-expression network analysis (WGCNA) was then applied to identify four functional modules containing 11,757 genes. A set of 3612 known LLPS-related genes was used to filter 11 hub genes. Hub genes were further selected through Random Forest and Support Vector Machine-Recursive Feature Elimination (SVM-RFE), with ZCCHC12, CDH13, and CDKN2A identified as the optimal candidates. A LLPS-based risk score (LlpsHBScore) was constructed using Lasso regression, and its performance was evaluated in the testing cohort. The correlation between the immune microenvironment and LlpsHBScore was analyzed using CIBERSORT, EPIC, MCPcounter, and quantiSeq algorithms, and the results were validated using the scRNA-seq dataset GSE186975. RESULTS: In both the training and testing cohort, as well as single-cell dataset, expression levels of ZCCHC12, CDH13, and CDKN2A were significantly upregulated in HB. The LlpsHBScore was positively correlated with the expression of these genes. The nomogram based on LlpsHBScore effectively diagnosed HB and accurately predicted the 2-year and 5-year survival rates of patients. Further analysis revealed that the LlpsHBScore was significantly correlated with the expression of drug resistance gene clusters and the infiltration level of CD8+ T cells. CONCLUSIONS: This study identifies elevated expression levels of ZCCHC12, CDH13, and CDKN2A in HB, contributing to a better understanding of their potential involvement in the disease. We successfully developed a machine learning-based risk score (LlpsHBScore), and a visual diagnostic tool (nomogram) with high clinical application potential. These findings offer new insights into the molecular subtyping, drug resistance mechanisms, and optimization of targeted therapies for HB.