AI-Driven Multi-Omics Fusion Framework for Early HCC Prediction and Precision Prognosis

Hepatocellular Carcinoma (HCC) remains one of the primary leading causes of cancer-related mortality globally, chiefly
driven by late-stage clinical detection and a highly heterogeneous tumor microenvironment. Early-stage diagnosis
provides the highest window for curative interventions such as surgical resection or liver transplantation. However,
conventional single-modality screening techniques—including serum alpha-fetoprotein (AFP) measurements and
ultrasonography—frequently demonstrate suboptimal sensitivity and specificity. In this paper, we propose a novel, end-toend AI-Driven Multi-Omics Fusion Framework (Moff-HCC) that integrates transcriptomic (RNA-Seq), epigenomic
(DNA methylation), and proteomic profiles gathered from peripheral blood mononuclear cells and liquid biopsies. Our
framework utilizes an adaptive Graph Convolutional Network (GCN) integrated with a Multi-Head Cross-Attention
mechanism to discover inter-omic regulatory structures while preserving modality-specific biological signals. To resolve
high-dimensional feature spaces and low sample sizes, we introduce a Contrastive Regularized Autoencoder (CRAE) for
robust unsupervised representation learning before downstream classification. Experimental evaluation on a multiinstitutional dataset comprising 1,240 patients demonstrates that Moff-HCC achieves an Area Under the Receiver
Operating Characteristic curve (AUROC) of 0.942 for early HCC detection (BCLC Stage 0/A), significantly
outperforming state-of-the-art single-modality models and traditional late-fusion algorithms. Furthermore, by
incorporating a Cox Proportional Hazards attention block, the model generates a continuous Risk Prognostic Score (RPS)
capable of stratifying patients into low-risk and high-risk survival cohorts (p < 0.001). Explainable AI (XAI) analysis via
integrated gradients identified key multi-omics driver signatures, including aberrant methylation of the CDKN2A
promoter coupled with upregulated GPC3 transcript expressions, providing clinically actionable therapeutic targets. This
holistic computational architecture offers a robust pathway toward non-invasive, early-stage hepatocellular surveillance
and precision patient management