Potential of Artificial Intelligence Algorithms for Identification of Relevant Diagnostic and Prognostic Biomarkers of Early-Stage Liver Cancer
| Source: arXiv AI
Tags: biomarker-discovery, cancer-AI, XAI, SHAP, hepatocellular-carcinoma, deep-learning, genomics
A deep learning model with SHAP-based XAI identifies DNAJB14 as the most influential gene for early-stage HCC diagnosis, achieving 90.74% accuracy with 15 selected genes — and functional validation confirms DNAJB14 inhibition reverses tumor cell migration and invasion.
Details
Early detection of hepatocellular carcinoma (HCC) depends on reliable molecular biomarkers, but the transcriptomic landscape is complex and noisy. This paper applies deep learning and explainable AI to a semi-supervised transcriptomic HCC dataset across five disease stages to identify effective diagnostic biomarkers.\n\nBest model: using 15 genes selected via SelectKBest, accuracy reaches 90.74%. Lowest loss (0.3187) was achieved with 20 genes. SHAP-based XAI consistently identifies DNAJB14 as the most influential feature across all experiments.\n\nA key strength of this work is functional validation: the authors demonstrate that DNAJB14 inhibition effectively reverses tumor cell migration, invasion, colony formation, and sphere formation — providing biological evidence supporting the model's attribution, not just statistical correlation.\n\nLimitations acknowledged: dataset class imbalance (mitigated via weighted training) and the need for larger datasets across diverse populations for generalizability. Under second revision at a biomedical signal processing journal.