Machine learning-based pathomics model predicts ANGPT2 expression and prognosis in hepatocellular carcinoma.

Abstract

Angiopoietin 2 (ANGPT2) is a promising prognostic marker and therapeutic target in hepatocellular carcinoma (HCC). However, assessing ANGPT2 expression and prognosis from histopathological images with naked eye is challenging. In this study, machine learning was employed to develop a pathomics model that analyzed histopathological images to predict ANGPT2 status. 267 cases, obtained from TCGA-HCC were divided into training and testing set. 91 cases from a single center were employed as a validation set. ANGPT2 was demonstrated up-regulated in HCC and patients with high ANGPT2 expression had a significant overall survival (OS) decline in TCGA-HCC cohort. Histopathological features in the training set were extracted, screened, and incorporated into a gradient boosting machine (GBM) model that generated pathomics score (PS), which successfully identified ANGPT2 expression level in three sets and showed remarkable risk stratification for OS in TCGA-HCC cohort (P < 0.0001) and the single center cohort (P = 0.001). Multivariate analysis suggested that PS could serve as a predictor for prognosis (P < 0.001). Bioinformatics analysis illustrated distinction of tumor growth and development related gene enriched pathways, VEGF-related genes expression and immune cell infiltration in different PS value. Our research indicates that histopathological image features can enhance prediction of molecular status and prognosis in HCC. The integration of image features with machine learning has potential for improving prognosis prediction in HCC.