Noninvasive diagnosis of hepatocellular carcinoma by integrating the genetic, epigenetic and fragmentation features of cell-free DNAs

Cell-free DNAs (cfDNAs) released from the apoptosis and necrosis cells provide abundant information for screening cancers. Recent studies have revealed different types of features in cfDNAs which can be applied to auxiliary diagnosis of cancers. In this work, we proposed a staked ensemble model to effectively integrate three classes of features extracted from the bisulfite sequencing of cfDNA samples, including four types of fragmentation features (fragment size coverage, fragment size distribution, end motif, breakpoint motif), one type of genetic feature (Copy Number variation), and one type of epigenetic feature (methylation fragment level). For each type of features, a base model is constructed which ensembles six machine learning algorithms to predict cancer scores for every sample. Then, the cancer scores from six base models are ensembled to generate the final stacked model. To demonstrate the effectiveness of the staked ensemble model, it was applied to discriminate samples with hepatocellular carcinoma (HCC) and health controls based on the whole genome bisulfite sequencing data of cfDNAs. The results show that it outperforms four popular methods with single or multi types of features, achieving a sensitivity of 93.33%, a specificity of 93.40%, a accuracy of 95.67% and an AUC of 0.9815 on the validation dataset. Conclusively, the staked model can effectively integrate six types of features of cfDNAs and has great potential for screening cancers.