Identification of the Association between Hepatitis B Virus and Liver Cancer using Machine Learning Approaches based on Amino Acid

Primary liver cancer has been a common reason for death from cancer globally. The most common type of primary liver cancer is the hepatocellular carcinoma (HCC). The major cause of HCC is chronic infections with hepatitis B virus (HBV). In this research, we used next generation sequencing (NGS), which has been very widely used to produce deep, efficient, and high-quality sequence data. NGS was used to sequence the pre-S region of the HBV genome of total 139 patients, which contain 94 HCC patients and 45 chronic HBV (CHB) patients. We generated two types of datasets. Firstly, for the data of amino acid occurrence frequency, we used basic local alignment search tool (BLAST) to map each NGS short read and translated each alignment into amino acid by DNA codon table. The input features are the occurrence frequencies of 20 basic amino acids using Shannon entropy. We picked 40 patients with 27 HCC and 13 CHB as the independent testing set. Then we used machine learning methods including logistic regression, random forest and support vector machine (SVM) to construct the classification models and make the prediction. The AUC values on the independent testing set for those machine learning methods (logistic regression, random forest and SVM) are 0.946, 0.923 and 0.960 respectively. Secondly, for the data of word pattern frequency of amino acids, we calculated word pattern frequencies of amino acids of all individuals and compared them using Euclidean distance. The input features are the frequencies of amino acid word of length 2, which is normalized by dividing the total occurrence number of all words. What's more, word pattern frequencies of amino acids were used to construct the classification models for HCC status using machine learning methods. Principal coordinate analysis (PCoA) was also used to visualize the associations between patient clusters, the HCC disease status of patients, and the fraction of HBV genotypes. We found that word patterns are powerful for the analysis of the HBV sequences from the aspect of amino acids because the AUC values of the classification models for machine learning methods are all above 0.9. Hence, our study showed that word pattern frequencies of amino acids is powerful for revealing the underlying principles of the occurrence of HCC triggered by HBV. Our essential findings consist of three parts. Firstly, all machine learning methods can generate classification models with high AUC values. Then, we can find some certain positions of amino acids or word patterns of amino acids that the mutation occurred on those positions will induce the HCC. Last, PCoA is associated with the disease status (HCC or CHB) and the fraction of genotype B (or C).