Abstract
Purpose: This study aims to develop a liquid biopsy assay to identify HCC and differentially diagnose hepatocellular carcinoma (HCC) from colorectal carcinoma (CRC) liver metastasis. Methods: Thirty-two microRNAs ("HallMark-32" panel) were designed to target the ten cancer hallmarks in HCC. Quantitative PCR and supervised machine learning models were applied to develop an HCC-specific diagnostic model. One hundred thirty-three plasma samples from intermediate-stage HCC patients, colorectal cancer (CRC) patients with liver metastasis, and healthy individuals were examined. Results: Six differentially expressed microRNAs ("Signature-Six" panel) were identified after comparing HCC and healthy individuals. The microRNA miR-221-3p, miR-223-3p, miR-26a-5p, and miR-30c-5p were significantly down-regulated in the plasma of HCC samples, while miR-365a-3p and miR-423-3p were significantly up-regulated. Machine learning models combined with HallMark-32 and Signature-Six panels demonstrated promising performance with an AUC of 0.85-0.96 (p ≤ 0.018) and 0.84-0.93 (p ≤ 0.021), respectively. Further modeling improvement by adjusting sample quality variation in the HallMark-32 panel boosted the accuracy to 95% ± 0.01 and AUC to 0.991 (95% CI 0.96-1, p = 0.001), respectively. Even in alpha fetoprotein (AFP)-negative (< 20ng/mL) HCC samples, HallMark-32 still achieved 100% sensitivity in identifying HCC. The Cancer Genome Atlas (TCGA, n=372) analysis demonstrated a significant association between HallMark-32 and HCC patient survival. Conclusion: To the best of our knowledge, this is the first report to utilize circulating miRNAs and machine learning to differentiate HCC from CRC liver metastasis. In this setting, HallMark-32 and Signature-Six are promising non-invasive tests for HCC differential diagnosis and distinguishing HCC from healthy individuals.