Serum Fusion Transcripts to Assess the Risk of Hepatocellular Carcinoma and the Impact of Cancer Treatment through Machine Learning

Abstract

Hepatocellular carcinoma (HCC) is one of the most lethal malignancies for human. Early diagnosis of HCC is crucial in reducing the mortality of the disease. In this study, a panel of 9 fusion transcripts in the serum samples from 136 individuals using TaqMan qRT-PCR was analyzed. Seven fusion genes were frequently detected in the serum samples of HCC patients, including MAN2A1-FER (100%), SLC45A2-AMACR (62.3%), ZMPSTE24-ZMYM4 (62.3%), PTEN-NOLC1 (57.4%), CCNH-C5orf30 (55.7%), STAMBPL1-FAS (26.2%) and PCMTD1-SNTG1 (16.4%). Machine learning models were constructed based on serum fusion gene levels to predict HCC occurrence in the training cohort using leave-one-out-cross-validation approach. One of the machine learning models called 4-fusion genes logistic regression model (MAN2A1-FER<40, CCNH-C5orf30<38, SLC45A2-AMACR<41, PTEN-NOLC1<40) produced a 91.5% accuracy in the training cohort. The same model generated an accuracy of 83.3% in the testing cohort. When serum α-fetal protein (AFP) level was incorporated into the machine learning model, a 2-fusion gene+AFP logistic regression model (MAN2A1-FER<40, CCNH-C5orf30<38, AFP) was found to generate an accuracy of 94.8% in the training cohort. The same model generated 95% accuracy in both the testing cohort and the combined cohorts. Cancer treatment reduced most of the serum fusion transcript levels. Serum fusion gene machine learning models may serve as important tools in screening HCC and monitoring the impact of HCC treatment.