Late-stage dedifferentiation and epigenetic memory of cancer stem cells in hepatocellular carcinoma

Cancer stem cells (CSCs) drive recurrence and drug resistance in hepatocellular carcinoma (HCC), but their origin remains controversial: are they tumour-initiating cells or late-stage dedifferentiation products? Direct human single-cell evidence linking bipotent progenitors (BPs) and CSCs has been lacking.

We integrated single-cell RNA sequencing (scRNA-seq) data from 109 samples (44 patients; 410,608 cells) across five public cohorts and generated EpCAM-enriched scRNA-seq from two additional HCC patients. Single-cell somatic mutations were inferred from the transcriptomic data, yielding 384,867 high-confidence variants across 31,908 cells from 20 patients. Clonal evolution was reconstructed through copy number variation (CNV) phylogenies and transcription-coupled-repair-based cell-of-origin inference. CSC and non-CSC subpopulations from Huh7 cells were flow-sorted before and after two weeks of culture and profiled by targeted bisulfite sequencing. A core-imprint risk score was evaluated in multiple cohorts and validated on a 97-case tissue microarray by multiplex immunofluorescence.

Unexpectedly, BPs harboured higher mutation burdens than other non-malignant parenchymal cells, and CSCs harboured higher mutation burdens than most other tumour cells, challenging their role as genomically quiescent ancestors. CNV phylogenies and evolutionary distances placed CSCs at the most distal branches of the tumour tree, while cell-of-origin analysis identified BPs as a pre-malignant precursor arising from hepatocyte dedifferentiation. RNA velocity, pseudotime and SNP-integrated lineage reconstruction converged on this directionality, with CSCs arising at the terminus of tumour evolution, reproduced at single-patient resolution in the EpCAM-enriched samples. Mechanistically, CSCs upregulated DNA methyltransferases (DNMTs), and ~78% of CSC-specific methylation changes were stably retained after CSC differentiation but not reproduced during de novo stemness acquisition, indicating locked-in epigenetic memory. A 16-gene core-imprint risk score specifically predicted early recurrence (≤2 years), and CD13⁺ CD133⁺ CSCs showed elevated 5-hydroxymethylcytosine correlating with poor prognosis.

We propose a framework in which hepatocytes dedifferentiate into BPs as a pre-malignant state, undergo malignant transformation, and a subset acquires stemness through DNMT-mediated reprogramming stabilized by epigenetic memory. These findings challenge the classical stem cell origin hypothesis, showing that CSCs in established HCC are late-stage dedifferentiation products, provide a rationale for targeting CSC epigenetic stability, and offer a biomarker for early recurrence.