RNA-centered mechanisms sustaining self-renewal in hepatocellular carcinoma stem cells: Effector dosage, protein output and regulatory deployment

Abstract

Hepatocellular carcinoma (HCC) stem cells sustain tumor propagation, relapse and treatment tolerance. However, the RNA-dependent mechanisms preserving their self-renewal remain scattered across studies of miRNAs, lncRNAs, circRNAs, snoRNAs, RNA-binding proteins and RNA-processing enzymes. Here, we examine how RNA molecules and RNA-processing machineries maintain functional self-renewal by controlling self-renewal effectors. Rather than cataloging RNA classes, we organize the evidence by molecular operation. RNA mechanisms first tune effector dosage and availability through transcript repression, stabilization, export, translational access and protein turnover. They also specify protein output through m6A regulation, RNA editing, ribosome biogenesis, tRNA modification and transcript-selective translation. A third layer deploys regulatory activity by directing protein localization, partner choice and locus access. These operations converge on NOTCH, Wnt-β-catenin, Hedgehog/GLI1, Hippo-YAP and TGF-β/IL-6-STAT3 signaling, but pathway-level convergence does not establish a universal RNA program. We emphasize evidence from serial propagation, limiting dilution, genetic perturbation and rescue, and distinguish functional self-renewal from marker enrichment, sphere formation or tumor shrinkage. Current evidence supports state-specific RNA dependencies in defined HCC stem-cell models, but their universality across HCC subtypes and selectivity over normal liver regeneration remain unresolved.

Key Words: Hepatocellular carcinoma; Cancer stem cells; RNA-centered regulation; Self-renewal; Post-transcriptional control

Core Tip: Here, we examine how RNA molecules and RNA-processing machineries maintain functional self-renewal by controlling self-renewal effectors. Rather than cataloging RNA classes, we organize the evidence by molecular operation. RNA mechanisms first tune effector dosage and availability through transcript repression, stabilization, export, translational access and protein turnover.

INTRODUCTION

Hepatocellular carcinoma (HCC) remains difficult to control because tumor-propagating subpopulations can survive therapy, reconstitute heterogeneous disease and seed relapse. These cancer stem-cell-like populations have been linked to tumor initiation and recurrence[1-3], therapeutic resistance[1,3-5] and aggressive tumor behavior[3,6,7]. However, they do not form a single molecularly defined compartment. Marker-enriched populations such as CD133-positive, EpCAM-positive or CD90-positive cells are useful experimental entry points, but single-cell analyses show substantial phenotypic, transcriptional and functional heterogeneity among cells classified as liver cancer stem cells (CSCs)[8]. A marker or stemness score can therefore enrich for a population of interest, but it cannot by itself explain how self-renewal is maintained.

RNA biology provides a mechanistic route into this problem because self-renewal is not controlled solely by transcriptional programs. A self-renewal effector must first be present at an adequate dose. It must then be translated into the right protein product and deployed in the correct regulatory context. RNA molecules and RNA-processing machineries can intervene at each of these levels. MicroRNAs and RNA-binding proteins alter transcript abundance, stability and accessibility. RNA modifications and editing change transcript fate or coding output, whereas noncoding RNAs can guide, scaffold or redirect protein complexes that control developmental and inflammatory signaling. Selected RNA mechanisms have been linked to tumor-propagating phenotypes in stronger functional systems. These include patient-derived organoids, limiting-dilution xenografts, genetic perturbation, de novo or chemically induced mouse models and mechanistic rescue experiments[1,4,9-12]. These assays move the evidence beyond marker expression alone.

The existing literature, however, should not be read as evidence for one universal RNA program in HCC stem cells. Across the evidence reviewed here, the most recurrent convergence occurs at NOTCH, Wnt-β-catenin, Hedgehog, Hippo and TGF-β-STAT3 signaling[1,3-7,10,11,13-16]. BMP/BMPR1A appears in a more limited, context-specific locus-access mechanism through HAND2-AS1[12] and is therefore treated separately from the recurrent convergence nodes. These RNA mechanisms reach recurrent self-renewal nodes through different molecular routes and in different experimental contexts. The same pathway can be reinforced by microRNA-mediated repression of inhibitors, stabilization of a pathway effector, altered translational output or RNA-guided recruitment of a regulatory complex. Convergence at the pathway level is therefore stronger than evidence for conservation of any individual RNA dependency.

This review organizes RNA-centered regulation around three analytical questions rather than three mutually exclusive biochemical categories: How much functional effector is available, what protein output is produced from the available RNA, and where or how regulatory activity is deployed. MicroRNAs, lncRNAs, circRNAs, snoRNAs, RNA-binding proteins and RNA-processing enzymes are therefore compared according to the proximal RNA-dependent step demonstrated experimentally. A single mechanism may influence more than one downstream layer. We assign each mechanism according to its nearest experimentally resolved RNA-to-effector step and explicitly flag cross-operation cases rather than forcing them into exclusive bins. The framework is therefore intended as a readout-oriented map for comparing heterogeneous evidence, not as a biochemical taxonomy.

CONCEPTUAL AND EVIDENTIARY FRAMEWORK

Defining functional self-renewal and evidentiary strength

The single-cell and marker-level heterogeneity described above sets the evidentiary standard for this review. Marker enrichment is not equivalent to functional self-renewal. Throughout the review, “HCC CSCs” refers to HCC cells with experimentally supported tumor-propagating or serial self-renewal capacity, not to a fixed lineage defined by one surface marker. CD133-, EpCAM- or CD90-enriched fractions may contain more tumor-propagating cells than bulk tumor tissue, but they do not represent a single interchangeable lineage. Single-cell analysis has shown substantial transcriptional and functional variation among cells classified as liver CSCs[8]. Marker enrichment, pluripotency-related factors and computational stemness scores can therefore identify candidate populations. They cannot by themselves demonstrate durable self-renewal.

Self-renewal is used here to mean preservation of tumor-propagating capacity across repeated rounds of growth. Primary sphere or organoid formation provides useful evidence of clonogenic fitness, but it can also reflect survival, aggregation or short-term proliferation. Serial propagation is stronger because it asks whether the phenotype persists after dissociation and replating. Tumor formation at low cell numbers estimates tumor-propagating frequency in vivo, but its interpretation is strongest when paired with genetic perturbation, lineage depletion or repeated propagation. This hierarchy is illustrated by several stronger experimental designs. LGR5-positive cell ablation reduced organoid initiation and tumor growth[2]. METTL16 loss reduced functional CSC frequency and impaired de novo liver tumor initiation[9]. circIPO11 disruption was tested through limiting dilution, patient-derived cells and a genetic knockout model[1]. These assays are stronger than marker enrichment, but they are not absolute proof. Limiting dilution depends on model assumptions and transplantation context. Serial propagation can select for culture-adapted or transplantation-adapted clones. Organoids preserve only part of the in vivo liver niche. They are therefore treated as stronger functional evidence, not as infallible measures of self-renewal.

We evaluate mechanistic evidence separately from phenotypic evidence. An RNA perturbation that reduces sphere formation may be acting through general toxicity, impaired proliferation or altered survival rather than through self-renewal itself. A stronger causal claim requires identification of the relevant RNA-dependent target and, where possible, restoration of that target to rescue the phenotype. This target-and-rescue logic has linked YTHDF1 to NOTCH1 mRNA stability and translation, miR-500a-3p to endogenous inhibitors of STAT3 signaling, and miR-5188 to the FOXO1-β-catenin circuit[4,7,17]. In this review, studies combining a functional CSC assay with limiting dilution, genetic deletion, in vivo depletion, serial propagation or mechanistic rescue are treated as core evidence. A single sphere assay, marker-positive fraction, side population, ALDH activity or bulk xenograft volume is considered supportive rather than definitive. Treatment response is considered only when the same RNA mechanism is also linked to functional CSC behavior through the assays described above[1,4,5].

Scope and analytical framework

We use “RNA-centered” because the mechanisms reviewed here extend beyond the scope of any narrower modifier. “Post-transcriptional” would inadequately capture RNA-guided recruitment of regulatory proteins to chromatin or other mechanisms whose immediate functional consequence is transcriptional or spatial. “Epitranscriptomic” is narrower still because many included mechanisms involve miRNA-mediated repression, RNA-protein scaffolding, RNA export, RNA editing, ribosome biogenesis or tRNA-dependent translation without requiring a covalent RNA modification. “RNA-mediated” is sufficiently broad but does not distinguish a review organized around the proximal RNA-dependent regulatory event from one that simply discusses pathways influenced by RNA. Here, “RNA-centered” therefore denotes mechanisms in which an RNA molecule or RNA-processing machinery constitutes the proximal causal regulator of effector abundance, protein output or regulatory deployment.

Within this definition, we include microRNAs and other noncoding RNAs that regulate transcripts or proteins, RNA-binding proteins that control stability, localization or translation, and enzymes that write, read or edit RNA when functional evidence links them to HCC stem-cell self-renewal. We include RNA maturation or translation machinery only when direct experiments connect it to a CSC-relevant functional endpoint. We exclude differential expression, prognostic association or inclusion in a multigene stemness signature when these observations lack functional validation. We also exclude studies limited to proliferation, migration, bulk xenograft volume or treatment response unless they test a CSC-related functional endpoint. Extracellular-vesicle-mediated RNA transfer and niche-derived RNA signals fall outside the systematic scope of this review.

The article is organized by molecular action rather than RNA class. This avoids treating microRNAs, lncRNAs, circRNAs, snoRNAs, RNA-binding proteins and RNA-modifying enzymes as separate catalogs. The relevant distinction is the operation that each RNA mechanism performs on the self-renewal machinery.

The first operation is control of effector dosage and availability. This includes microRNA-mediated repression, selective transcript stabilization, nuclear export, translational accessibility and RNA-dependent protein turnover. RALYL-dependent stabilization of TGF-β2 mRNA and DIO3OS-dependent restriction of ZEB1 export act at the RNA level. SNORA74A-dependent stabilization of E2F2 and lnc-β-Catm-dependent stabilization of β-catenin act closer to protein persistence. Together, they represent different routes to the same quantitative problem of determining how much functional effector is available to the cell[6,13,15,18].

The second operation is specification of protein output. RNA modifications, editing events and translational machinery determine how efficiently an RNA is converted into protein and, in some cases, what protein product is generated. YTHDF1 and METTL3 regulate productive output from selected m6A-associated transcripts[4,5]. ADAR1 editing changes the coding information and behavior of GLI1[10]. METTL16 links rRNA maturation and ribosome biogenesis to functional CSC frequency[9].

The third operation is deployment of regulatory activity. Some RNAs act less as dosage regulators than as guides for protein localization, partner selection or locus-specific recruitment. SNORA49 alters assembly and promoter recruitment of the HNRNPU-ZC3H18 complex[19]. SNORD88B controls nucleolar localization of WRN[11]. circIPO11 and HAND2-AS1 recruit regulatory proteins to the GLI1 and BMPR1A loci[1,12].

The three operations function as non-exclusive analytical coordinates. We assign each mechanism according to the nearest experimentally resolved RNA-to-effector step: Effector abundance or availability, protein output, or regulatory deployment. When a mechanism spans more than one step, Table 1 records its primary operation and the text describes the secondary consequence. For example, YTHDF1 affects both NOTCH1 mRNA stability and translation but is discussed under protein output because increased productive translation is a major experimentally resolved consequence. SNORA74A organizes an RNA-protein interaction but is placed under dosage control because the demonstrated downstream effect is reduced E2F2 turnover. This assignment rule allows mechanisms with overlapping biochemistry to be compared without assuming that the three operations are physically separable layers[4,10,13]. Alternative splicing, isoform choice and regulated RNA decay remain relevant to HCC biology, but the evidence reviewed here links them less frequently to repeated CSC self-renewal, limiting dilution, serial propagation or mechanistic rescue. We therefore identify them as evidence gaps rather than create parallel categories on the basis of general HCC association. Figures 1, 2 and 3 summarize the three proximal RNA-centered operations reviewed below: Effector dosage and availability, protein output specification and regulatory deployment. Figure 4 integrates these operations into recurrent self-renewal nodes and functional CSC-associated outputs. Table 1 maps the RNA-centered mechanisms discussed in this review onto their primary operation, proximal target, convergent pathway and evidence level.

Figure 1 RNA control of self-renewal effector dosage and availability. Representative routes by which RNA-centered mechanisms alter self-renewal effector dosage or availability in hepatocellular carcinoma cancer stem cell models. A: miR-1246, miR-5188 and miR-500a-3p reduce inhibitory inputs to β-catenin or STAT3 signaling, whereas loss of miR-192-5p supports glycolysis; B: RALYL and ICR preserve selected mRNAs, DIO3OS restricts ZEB1 mRNA export, and CPEB1 regulates SIRT1 mRNA poly(A)-tail length; C: SNORA74A limits E2F2 degradation to promote NOTCH3 transcription, whereas lnc-β-Catm stabilizes β-catenin.

Figure 2 RNA modification, editing and translational machinery specify protein output. A: YTHDF1 increases the stability and translation of m6A-marked NOTCH1 mRNA, whereas METTL3 promotes FZD10 expression; B: ADAR1-mediated A-to-I editing of GLI1 mRNA produces the R701G variant, reducing SUFU restraint and increasing nuclear GLI1 activity; C: METTL16 supports rRNA maturation and ribosome biogenesis, whereas TRMT6-TRMT61A-mediated tRNA m1A favors PPARδ translation and cholesterol-linked Hedgehog signaling.

Figure 3 RNA-guided deployment of self-renewal regulatory activity. Representative mechanisms in which RNA controls regulatory deployment through subnuclear positioning, locus access or RBP redistribution. A: SNORA49 limits HNRNPU-ZC3H18 recruitment to the SOX9 promoter, whereas SNORD88B positions WRN in the nucleolus and attenuates STK4/Hippo signaling; B: circIPO11 recruits TOP1 to the GLI1 promoter, HAND2-AS1 recruits INO80 to the BMPR1A promoter, and lnc-DILC limits IL-6 transcription; C: circZKSCAN1 sequesters FMRP away from CCAR1 mRNA, whereas m6A-modified circRAPGEF1 redirects IGF2BP3 away from ASS1 mRNA.

Figure 4 Integrated model of RNA-centered convergence on functional hepatocellular carcinoma cancer stem cell self-renewal. Integrated model linking the three analytical operations to recurrent self-renewal nodes. The strongest cross-mechanism convergence occurs at NOTCH, Wnt-β-catenin, Hedgehog/GLI1, Hippo-YAP and TGF-β/IL-6-STAT3. BMP/BMPR1A, SOX9 and metabolic circuits are shown as context-specific or less recurrent axes. The model summarizes pathway-level recurrence rather than a universal upstream RNA program.

Table 1 RNA-centered mechanisms linked to hepatocellular carcinoma cancer stem cell self-renewal or stem-like phenotypes.

RNA regulator/machinery	Class	Primary operation	Proximal target or molecular partner	Convergent pathway/functional output	Functional evidence captured in manuscript	Evidence level	Ref.
Effector dosage/availability
miR-338-3p	miRNA	Dosage/threshold control	SOX4 repression	SOX4-linked stemness	Sphere/stemness and functional phenotypes	Supportive	[45]
miR-2117	miRNA	Dosage/threshold control	SOX2 repression	SOX2-linked CSC expansion and chemoresistance	CSC expansion/chemoresistance assays	Supportive	[46]
circ_0000972	circRNA	Transcript-level competition	miR-96-5p/PFN1 axis	Stemness suppression	Stemness phenotypes; not primary deployment evidence	Supportive	[42]
METTL3-SOCS3	m6A writer axis	Dosage/RNA fate	SOCS3 mRNA; JAK2-STAT3 signaling	STAT3 pathway output	Stemness/tumorigenicity assays	Supportive	[35]
SOCS2-AS1	LncRNA	Dosage/ceRNA-like control	miR-454-3p/CPEB1	Stemness suppression axis	CSC phenotype assays	Supportive	[28]
DIO3OS	LncRNA	Transcript availability/export	NONO-mediated ZEB1 mRNA export	ZEB1 protein availability	RNA export and CSC phenotype assays	Supportive	[18]
DUBR	LncRNA	Dosage feedback architecture	miR-520d-5p/CIP2A/E2F1	NOTCH1-associated stemness	Mechanistic feedback and CSC phenotypes	Supportive	[33]
IGF2BP1	m6A reader/RBP	Transcript persistence	MGAT5 mRNA stability	MGAT5-linked CSC phenotype	m6A-dependent binding and stemness assays	Supportive	[27]
RALYL	RNA-binding protein	Transcript persistence	TGF-β2 mRNA stability	PI3K-AKT-STAT3 signaling	Transcript-stability mechanism and CSC phenotypes	Core	[6]
LINC01013	LncRNA	Dosage/ceRNA-like control	miR-6795-5p/FMNL3	CSC features	CSC phenotype assays	Supportive	[29]
MALAT1	LncRNA	Dosage/ceRNA-like control	miR-375/YAP1	YAP1-linked CSC expansion	CSC phenotype assays	Supportive	[30]
miR-192-5p	miRNA	Metabolic threshold control	GLUT1, PFKFB3, c-MYC	Glycolytic permissiveness	Metabolic and stem-like phenotype assays	Supportive	[24]
miR-5188	miRNA	Dosage/pathway threshold control	FOXO1 repression	β-catenin nuclear accumulation; HBV/HBX-linked context	Target-and-rescue logic; pathway mechanism	Core	[7]
miR-613	miRNA	Dosage/threshold control	SOX9 repression	Stemness suppression	CSC expansion assays	Supportive	[47]
miR-26b-5p	miRNA	Marker-associated dosage control	HSPA8 in malignant EpCAM-positive cells	EpCAM-positive malignant compartment	Malignant vs non-malignant EpCAM-positive comparison	Supportive	[21]
miR-365	miRNA	Dosage/threshold control	RAC1 repression	CSC phenotype suppression	CSC phenotype assays	Supportive	[48]
THOR	LncRNA	Dosage/pathway output	β-catenin-associated regulation	Wnt-β-catenin-linked CSC expansion	CSC expansion assays	Supportive	[31]
miR-302a/d	miRNA	Dosage/pathway threshold control	E2F7/AKT-β-catenin signaling	CSC context signaling	Sphere/tumorigenicity-related assays	Supportive	[23]
miR-200b-ZEB1	miRNA circuit	CSC-state composition	ZEB1 circuit	Marker-defined CSC state redistribution	Marker-state and stemness assays	Supportive	[20]
miR-217	miRNA	Dosage/threshold control	DKK1-dependent Wnt regulation	Wnt-linked stem-like traits	Stem-like phenotype assays	Supportive	[43]
miR-500a-3p	miRNA	Dosage/threshold control	SOCS2, SOCS4, PTPN11	STAT3 signaling	Functional CSC assay and target-rescue logic	Core	[17]
MSI2	RNA-binding protein	RBP-linked dosage control; proximal mechanism unresolved	LIN28A downstream component	CSC self-renewal and tumorigenicity	Functional evidence; direct mRNA stabilization unresolved	Supportive	[26]
miR-452	miRNA	Dosage/threshold control	Sox7 repression	Wnt-β-catenin activation	Stem-like phenotype assays	Supportive	[44]
miR-589-5p	miRNA	Marker-associated dosage control	MAP3K8 in CD90-positive cells	CD90-positive compartment dependency	Marker-defined compartment evidence	Supportive	[22]
miR-491	miRNA	Supportive threshold logic	GIT-1/NF-κB/EMT	CSC-like properties	CSC-like phenotype assays	Supportive	[49]
ICR	LncRNA	Transcript persistence/RNA-RNA duplex control	ICAM-1 mRNA stability	ICAM-1 + CSC state; PVTT-associated phenotype	Sphere assays, in vivo ICR inhibition, clinical PVTT correlation	Supportive	[25]
miR-25	miRNA	Supportive threshold logic/apoptotic resistance	PTEN-PI3K-AKT-Bad	Apoptotic resistance in CSC context	Treatment/phenotype-linked evidence	Supportive	[50]
miR-1246	miRNA	Dosage/pathway threshold control	AXIN2 and GSK3β repression	β-catenin destruction machinery	Functional CSC evidence and pathway mechanism	Core	[3]
miR-4461	miRNA	Dosage/threshold control	SIRT1 repression	CSC expansion and chemoresistance	CSC expansion/chemoresistance assays	Supportive	[51]
Protein output/translational competence
YTHDF1	m6A reader	Protein output	m6A-modified NOTCH1 mRNA	NOTCH1 output	Patient-derived organoids, conditional mouse systems, mechanistic rescue	Core	[4]
METTL16	RNA methyltransferase/ribosome biogenesis regulator	Protein output/translational competence	rRNA maturation; ribosome biogenesis; eIF3a	Functional CSC frequency; de novo HCC initiation	Genetic loss and functional CSC assays	Core	[9]
ADAR1	RNA-editing enzyme	Protein output/recoding	GLI1 R701G editing	Hedgehog activity; mitophagy; oxidative phosphorylation	Mechanistic editing evidence plus tumor initiation	Core	[10]
SNHG3	LncRNA	m6A/RBP-linked RNA fate	miR-502-3p; YTHDF3/METTL3; ITGA6	ITGA6-linked CSC self-renewal	Self-renewal phenotypes and molecular axis	Supportive	[37]
ALKBH5-SOX4	RNA demethylase axis	RNA modification/pathway output	SOX4 demethylation; SHH signaling	Hedgehog/SHH activity	CSC phenotype and mechanism assays	Supportive	[36]
METTL3-FZD10	m6A writer axis	Protein output/receptor-level signaling	FZD10 expression	β-catenin and YAP1 signaling	Mechanistic pathway and functional assays	Core	[5]
TRMT6-TRMT61A	tRNA m1A methyltransferase complex	Protein output/tRNA modification	Selected tRNAs; PPARδ translation	Cholesterol synthesis; Hedgehog activation	tRNA modification and CSC-associated output	Core	[38]
CPEB1	RNA-binding protein	Translational accessibility	SIRT1 3′ untranslated region; poly(A)-tail regulation	SIRT1 protein output	Translation-focused mechanism and tumorigenicity phenotypes	Supportive	[32]
EIF5A2	Translation-associated factor	Protein output/translational machinery	c-MYC-miR-29b circuit	CD133-positive HCC stem-like cells	Marker-enriched evidence; weaker than direct translation models	Supportive	[39]
Regulatory deployment/Locus access/RBP redistribution
SNORA49	snoRNA	Regulatory deployment	HNRNPU-ZC3H18 complex; SOX9 promoter access	SOX9/self-renewal transcription	Functional CSC renewal assays; mechanistic complex/Locus evidence	Core	[19]
SNORD88B	snoRNA	Regulatory deployment	WRN nucleolar retention; XRCC5-dependent STK4 repression	Hippo attenuation	Liver cancer-initiating cell self-renewal assays	Core	[11]
circIPO11	circRNA	Regulatory deployment/Locus access	TOP1 recruitment to GLI1 promoter	Hedgehog/GLI1 output	Limiting dilution, patient-derived cells, genetic knockout model	Core	[1]
LINC00324	LncRNA	Regulatory deployment/TF association	PU.1 association; FasL expression	CSC-like properties	Functional evidence less extensive than core locus-access models	Supportive	[40]
circZKSCAN1	circRNA	Regulatory deployment/RBP redistribution	FMRP sequestration; CCAR1 mRNA interaction	β-catenin-dependent transcription	RBP competition and stemness assays	Core	[14]
HAND2-AS1	LncRNA	Regulatory deployment/Locus access	INO80 recruitment to BMPR1A promoter	BMP signaling	Functional and locus-access evidence	Core	[12]
Lnc-DILC	LncRNA	Regulatory deployment/promoter control	Promoter-associated IL-6 transcription	IL-6-STAT3 circuit	Mechanistic promoter and CSC context evidence	Supportive	[16]
LncCAMTA1	LncRNA	Regulatory deployment/Locus-linked repression	CAMTA1 promoter association	CSC-like properties	Locus-linked and CSC-like phenotype evidence	Supportive	[41]
Boundary mechanisms
SNORA74A	snoRNA	Protein persistence/proximity-mediated dosage control	DCAF13-E2F2 interaction; K48-linked E2F2 ubiquitination	NOTCH3 signaling; liver CSC self-renewal	Genetic deletion, self-renewal and hepatocarcinogenesis assays; ASO intervention	Core	[13]
circRAPGEF1	circRNA with m6A-dependent stabilization	Regulatory deployment/RBP redistribution	IGF2BP3 redistribution away from ASS1 mRNA	Aspartate accumulation; S6K-CAD pathway	Mechanistic RNA-RBP competition plus CSC phenotypes	Core	[34]
Lnc-β-Catm	LncRNA	Protein persistence/proximity control	EZH2-β-catenin proximity	β-catenin stabilization	Mechanistic proximity and CSC phenotypes	Core	[15]
DDX3	RNA helicase	RNA-helicase-associated, epigenetically mediated control of tumor-suppressive miRNA networks	Tumor-suppressive miRNA repression	CSC phenotypes	Upstream machinery evidence; boundary to miRNA section	Supportive	[52]

Note: Evidence level follows the hierarchy defined in the CONCEPTUAL AND EVIDENTIARY FRAMEWORK section and uses two categories only. “Core” indicates a cancer stem cell-relevant functional assay combined with at least one higher-stringency feature, such as limiting dilution, serial propagation, genetic depletion or in vivo ablation, patient-derived functional models or mechanistic rescue. “Supportive” indicates evidence based mainly on marker-enriched fractions, sphere formation, bulk tumorigenicity, treatment response or incomplete mechanistic rescue. Cross-operation placement is recorded in the “Primary operation” column and is not an evidence level. CSC: Cancer stem cell; EMT: Epithelial-mesenchymal transition; NF-κB: Nuclear Factor-κB; HCC: Hepatocellular carcinoma; IL-6: Interleukin-6.

RNA control of self-renewal effector dosage and availability

A first layer of RNA-centered regulation concerns whether a self-renewal effector is available at a sufficient functional dose. This section therefore focuses on transcript repression, transcript persistence, subcellular access, translational accessibility and protein half-life. Representative mechanisms in this dosage and availability layer are summarized in Figure 1. The POST-TRANSCRIPTIONAL SPECIFICATION AND PROTEIN OUTPUT section addresses how RNA processing and translational machinery determine the quantity or identity of the protein product generated from available RNA.

MicroRNAs tune pathway thresholds and CSC-state composition

MicroRNAs are most informative in HCC CSC models when they alter the functional threshold of a self-renewal circuit rather than merely change a marker. Here, “threshold” denotes the combined inhibitory pressure, effector availability or metabolic support that determines whether a pathway can sustain self-renewal in a defined cell state. This framing treats microRNAs as modulators of pathway state rather than as a separate catalogue of stemness genes[20].

Three core examples illustrate this logic. miR-1246 represses AXIN2 and GSK3β, weakening the β-catenin destruction machinery[3]. HBX-induced miR-5188 suppresses FOXO1 and promotes β-catenin nuclear accumulation within an HBV-linked regulatory circuit[7]. miR-500a-3p targets SOCS2, SOCS4 and PTPN11, thereby lowering endogenous restraints on STAT3 signaling[17]. In each case, the microRNA changes pathway output by reducing inhibitory pressure rather than by encoding a self-renewal effector.

These studies also show why a one-miRNA/one-target description is mechanistically inadequate. miR-1246 acts on two components of the β-catenin destruction machinery, whereas miR-500a-3p distributes its effect across several STAT3 inhibitors[3,17]. The miR-5188 axis is further conditioned by HBX and a reinforcing c-Jun-dependent circuit[7], making its strongest inference etiologically restricted rather than HCC-wide. The common feature is therefore network-level threshold control, with different microRNAs entering the same self-renewal pathway at different regulatory points.

Marker-defined compartments add a second dimension. miR-589-5p suppresses MAP3K8 in CD90-positive cells, whereas the miR-26b-5p-HSPA8 axis was identified in malignant EpCAM-positive cells and is particularly informative for the selectivity question discussed[21,22]. Separately, miR-302a/d restrains E2F7-dependent AKT-β-catenin signaling in broader CSC settings[23]. These studies support state-specific microRNA dependencies rather than a shared microRNA code across marker-defined HCC CSC populations.

MicroRNAs can also alter metabolic permissiveness. Loss of miR-192-5p increases GLUT1, PFKFB3 and c-MYC and promotes a glycolytic feedback state associated with stem-like behavior[24]. The relevant inference is that one microRNA can coordinate several metabolic determinants that make self-renewal more permissive in a defined model. Additional miRNA-target associations supported mainly by marker enrichment, sphere formation, treatment-response phenotypes or limited tumorigenicity assays are retained in Table 1 but omitted from the main narrative because they neither materially extend this mechanistic model nor meet the core evidence standard defined.

Transcript availability and protein persistence

Transcript abundance is only the first layer of effector availability. A self-renewal transcript must remain stable, escape inappropriate compartmental retention and enter a state that supports protein production. RNA-binding proteins and noncoding RNAs can therefore increase pathway output by preserving selected transcripts rather than globally stabilizing the transcriptome. RALYL prolongs TGF-β2 mRNA persistence and sustains PI3K-AKT-STAT3 signaling[6]. ICR forms an RNA duplex with ICAM-1 mRNA and maintains an ICAM-1-positive stem-like state[25]. MSI2 is also required for liver CSC self-renewal and tumorigenicity, with LIN28A functioning as an essential downstream component[26]. Its proximal operation remains less defined than that of RALYL or ICR. Current evidence does not establish whether MSI2 acts through transcript stabilization, translational control or another RBP-dependent mechanism. It is therefore treated here as supportive evidence for RBP-linked dosage control rather than as a defined mRNA-stabilization mechanism[6,26].

These examples differ in molecular architecture but share selective control over defined RNA targets. RALYL sustains a secreted ligand transcript[6]. ICR protects an adhesion-associated mRNA through RNA-RNA pairing[25]. MSI2 connects an RBP dependency to a downstream stemness regulator[26]. None implies that HCC CSCs simply stabilize all growth-promoting transcripts. Instead, selected RNAs encoding pathway ligands, adhesion molecules or stemness regulators can be kept above a functional threshold, even without a proportional increase in transcription. m6A-dependent transcript stabilization also fits this dosage layer. IGF2BP1 maintains MGAT5 mRNA stability through m6A-dependent binding, providing a reader-mediated example of transcript persistence rather than altered coding output[27].

Supportive lncRNA studies extend the dosage framework through proposed competitive endogenous RNA mechanisms. Here, “ceRNA-like” denotes a model in which a lncRNA binds a shared microRNA and thereby relieves repression of another transcript; we use the qualifier “-like” because effective stoichiometric competition is not always established directly. SOCS2-AS1 and LINC01013 illustrate such routes through miR-454-3p/CPEB1 and miR-6795-5p/FMNL3, respectively[28,29]. MALAT1 and THOR connect lncRNA regulation to YAP1 or β-catenin-associated liver CSC expansion[30,31]. These axes add breadth to the dosage framework without changing the evidence hierarchy.

Availability is also spatial. DIO3OS lowers ZEB1 protein without reducing total ZEB1 mRNA because it interferes with NONO-mediated nuclear export[18]. The transcript remains detectable but is less available to cytoplasmic translation. CPEB1 controls a distinct layer of translational accessibility by binding the SIRT1 3′ untranslated region, altering poly(A)-tail length and restricting translation[32]. These mechanisms explain why whole-cell RNA abundance can mislead functional interpretation. A transcript may appear unchanged by bulk measurement while its access to translation is sharply altered by nuclear retention or 3′ untranslated region-dependent control.

Some RNA-centered circuits extend beyond the transcript and regulate the persistence of the resulting protein. K48-linked polyubiquitin chains commonly mark proteins for proteasomal degradation. SNORA74A binds DCAF13 and limits K48-linked polyubiquitination of E2F2, thereby increasing E2F2 persistence and E2F2-dependent NOTCH3 transcription[13]. Lnc-β-Catm brings EZH2 and β-catenin into proximity, promotes β-catenin modification and reduces its degradation[15]. DUBR links transcript and protein dosage in a feedback architecture. By relieving miR-520d-5p repression of CIP2A, it increases CIP2A, stabilizes E2F1 and reinforces NOTCH1-associated stemness[33]. These mechanisms are not merely downstream consequences of transcription. They set the half-life and effective concentration of proteins that sit close to self-renewal decisions.

The same pathway can therefore be reinforced at multiple dosage checkpoints. Wnt-β-catenin activity can rise through repression of AXIN2 and GSK3β, loss of FOXO1 or direct stabilization of β-catenin[3,7,15]. NOTCH output can be increased through E2F2 stabilization or a DUBR-CIP2A-E2F1 circuit[13,33]. STAT3-linked signaling can be sustained either by removal of endogenous inhibitors or by RALYL-dependent stabilization of TGF-β2 mRNA upstream of PI3K-AKT-STAT3 activation[6,17]. Multiple RNA checkpoints can reinforce the same pathway through context-dependent molecular entry points. Distinct RNA mechanisms may operate in different etiological settings, marker-defined compartments or adaptive states. Control of pathway dose is the recurring theme. Whether the same dosage checkpoints are required across the full spectrum of HCC tumor-propagating cells remains unresolved.

POST-TRANSCRIPTIONAL SPECIFICATION AND PROTEIN OUTPUT

Once a self-renewal transcript is available, the next question is what kind of protein output it produces. RNA abundance alone is a poor proxy for this output. Chemical marks can alter reader binding, transcript persistence and translation. RNA editing can change the coding sequence. rRNA or tRNA modification can reshape the translational apparatus itself. These mechanisms differ in chemistry, but they intervene between RNA information and protein execution. Representative mechanisms that specify protein output are summarized in Figure 2.

RNA modification and recoding

m6A-dependent regulation in HCC stem cells is best understood as substrate-specific control of RNA fate, not as a uniform pro-stemness program. YTHDF1 provides the most direct reader-centered example. YTHDF1 recognizes m6A-modified NOTCH1 mRNA. This increases transcript stability and ribosome-associated translation, raising NOTCH1 output in patient-derived organoid and conditional mouse models[4]. METTL3-dependent FZD10 regulation reaches a related functional endpoint through a writer-centered mechanism. FZD10 expression activates β-catenin and YAP1 signaling, and pathway activity feeds back to reinforce METTL3 expression[5]. Together, these mechanisms underlie substrate-specific m6A regulation, with YTHDF1 interpreting modified NOTCH1 transcripts and METTL3 maintaining productive FZD10 signaling output[4,5]. A related m6A-dependent mechanism, circRAPGEF1 stabilization followed by IGF2BP3 redistribution, is discussed below as an example of RNA-guided RBP deployment rather than direct protein-output specification[34].

Beyond these receptor-level examples, additional m6A-linked mechanisms broaden the substrate range. METTL3 affects SOCS3/JAK2-STAT3 signaling, whereas ALKBH5 links demethylation to SOX4-SHH signaling[35,36]. Other mechanisms involve m6A reader or RBP-dependent RNA fate. SNHG3 connects miR-502-3p with YTHDF3/METTL3-dependent ITGA6 regulation[37]. As discussed above, IGF2BP1-MGAT5 is better treated as m6A-dependent transcript-stability control than as protein-output specification[27]. The demonstrated operation is mRNA persistence rather than altered coding sequence or translational specificity. These studies support a wider m6A/RBP layer. However, their mechanistic depth is not equivalent to the best-supported YTHDF1-NOTCH1, METTL3-FZD10 or METTL16 models.

RNA editing changes output in a more qualitative sense. ADAR1-mediated editing introduces an R701G substitution into GLI1[10]. SUFU normally restrains Hedgehog signaling by binding GLI1, whereas β-TrCP contributes to GLI1 turnover. The R701G variant binds SUFU more weakly and shows reduced β-TrCP-GLI1 interaction, increasing GLI1 stability and favoring its nuclear accumulation. The edited product is not just more abundant. It has altered interaction partners, stability and subcellular behavior. This separates ADAR1-GLI1 from YTHDF1-NOTCH1 and METTL3-FZD10, which primarily change how much functional protein is produced from an existing coding sequence[4,5]. ADAR1-GLI1 changes the molecular identity and activity profile of the effector itself[10]. The edited GLI1 axis links Hedgehog transcriptional activity to PINK1-Parkin-dependent mitophagy, oxidative phosphorylation and tumor initiation[10]. However, the evidence remains narrow. Direct functional support for coding-sequence recoding in HCC stem cells is concentrated in this GLI1 mechanism. Current data do not justify a broader claim that widespread protein recoding is a general property of HCC CSCs.

Translational machinery and selective protein output

Protein output also depends on the capacity and composition of the translational machinery. METTL16 demonstrates this at the level of ribosome production. Its loss disrupts rRNA maturation, ribosome biogenesis and mRNA translation. It also reduces functional CSC frequency and impairs de novo HCC initiation. eIF3a is one downstream component linking altered translational machinery to self-renewal output[9]. This mechanism differs from transcript-specific m6A regulation. METTL16 does not simply increase one receptor transcript. It affects the cellular apparatus through which many available RNAs can be converted into protein. The evidence supports a requirement for ribosomal competence in defined HCC stem-cell states, but not a blanket claim that all HCC CSCs maintain uniformly elevated global translation[9].

TRMT6-TRMT61A adds another layer by modifying tRNAs rather than ribosomes. Increased m1A modification of selected tRNAs favors PPARδ translation, enhances cholesterol synthesis and activates Hedgehog signaling[38]. This mechanism challenges the assumption that mRNA abundance is sufficient to predict protein output when the translational substrate pool itself is chemically biased. In this case, tRNA modification links metabolic rewiring to developmental signaling through selective enhancement of PPARδ production[38]. Together, METTL16 and TRMT6-TRMT61A show that translational competence is not a single variable. It can depend on ribosome maturation, translation-associated factors or the modification state of tRNAs[9,38].

Even when translational machinery is competent, individual transcripts can still be selectively favored or restrained. YTHDF1 promotes NOTCH1 translation through an m6A-dependent interaction, directly coupling an RNA mark to increased production of a self-renewal receptor[4]. CPEB1, discussed above as a dosage-control mechanism, illustrates the converse operation. Poly(A)-tail regulation restricts SIRT1 translation, reducing protein output without altering mRNA abundance or coding sequence[32]. These examples show that protein output can change substantially without a proportional change in whole-cell mRNA abundance.

EIF5A2 provides supporting, but weaker, evidence for the connection between translation-associated machinery and marker-enriched HCC stem-like states. The EIF5A2 study linked this translation-associated factor to maintenance of CD133-positive HCC cells through a c-MYC-miR-29b circuit[39]. This is relevant to protein-output control, but it is not as direct as YTHDF1-mediated NOTCH1 translation or CPEB1-mediated repression of SIRT1 translation[4,32,39]. The evidence instead supports a two-level model in which HCC stem-cell function depends on translational competence and transcript-specific access to that machinery.

The distinction between global translational capacity and transcript-selective output should remain explicit. METTL16 and tRNA m1A modification alter the machinery through which protein synthesis occurs[9,38]. YTHDF1 and CPEB1 act on identifiable transcripts[4,32]. These layers are likely to interact, but current studies do not establish their relative contribution across different HCC CSC populations. What the evidence does show is that self-renewal programs are not simply transcribed and then executed. RNA processing and translation determine which effector proteins are produced, in what amount and, in the case of editing, with what molecular behavior.

RNA-guided deployment of self-renewal regulatory activity

Not all RNA-centered mechanisms sustain self-renewal by changing how much transcript or protein is present. Some RNAs instead determine where a regulatory protein is deployed, which partner it engages and whether it gains access to a specific genomic locus. This layer is conceptually distinct from dosage control, although the two can intersect. A protein can be abundant and properly translated, yet remain functionally irrelevant. This occurs when it is retained in the wrong compartment, excluded from the required complex or unable to occupy its target site. Representative deployment mechanisms are summarized in Figure 3.

Subnuclear positioning and locus access

Noncanonical snoRNA functions provide strong evidence that RNA can control the subnuclear deployment of self-renewal regulators. SNORA49 acts as a negative regulator of liver CSC renewal by binding HNRNPU and limiting its interaction with ZC3H18. When SNORA49 is low, HNRNPU can engage ZC3H18 and occupy the SOX9 promoter, increasing SOX9 transcription[19]. The key operation is not a change in total HNRNPU abundance. It is control of complex formation and promoter access.

SNORD88B retains WRN in the nucleolus[11]. This positioning is connected to an XRCC5-STK4 transcriptional axis. XRCC5, also known as Ku80, is a WRN-interacting DNA-repair protein that can act as a transcriptional repressor; in this model, XRCC5 occupies the STK4 promoter and suppresses STK4/MST1 expression, thereby weakening Hippo pathway restraint and favoring YAP activity. SNORD88B therefore links nucleolar protein positioning to transcriptional control of a self-renewal pathway[11]. SNORA49 and SNORD88B therefore represent opposite forms of subnuclear deployment. SNORA49 restricts formation and promoter recruitment of a nuclear transcriptional complex. SNORD88B enables nucleolar positioning of a regulatory protein that affects Hippo pathway output[11,19]. SNORA74A, discussed above as a protein-persistence mechanism, illustrates a boundary case. RNA-guided protein proximity can serve dosage regulation rather than produce a primary deployment outcome[13]. These examples support noncanonical snoRNA deployment functions, not a shared snoRNA targeting code[11,13,19].

Long and circular noncoding RNAs can also determine locus access by bringing regulatory proteins to defined genomic sites. circIPO11 recruits TOP1 to the GLI1 promoter, increasing GLI1 transcription and Hedgehog output[1]. HAND2-AS1 recruits the INO80 chromatin-remodeling complex to the BMPR1A promoter and activates BMP signaling[12]. Lnc-DILC acts in the opposite direction by restraining promoter-associated IL-6 transcription and limiting the autocrine IL-6-STAT3 circuit[16]. LINC00324 extends the principle to a sequence-specific transcription factor by associating with PU.1 to promote FasL expression[40]. Its functional evidence is less extensive than that supporting circIPO11 or HAND2-AS1[1,12]. LncCAMTA1 provides another locus-linked example. It physically associates with the CAMTA1 promoter, represses CAMTA1 transcription and promotes CSC-like properties[41].

These studies show that RNA-guided locus access is not intrinsically oncogenic. An RNA can facilitate transcription of GLI1 or BMPR1A, or restrict transcription of IL-6, depending on the recruited protein and target locus[1,12,16]. The evidence also remains mechanism-specific. Current studies do not define a general RNA “address code” explaining how noncoding RNAs recognize chromatin targets across HCC CSC states. What they establish is narrower. Selected RNAs can confer genomic specificity on broadly acting regulatory proteins.

Partner choice and regulatory-complex deployment

RNA can redirect regulatory activity without directly recruiting proteins to chromatin. circZKSCAN1 suppresses HCC stemness through an RBP-sequestration mechanism with a non-obvious regulatory outcome. By sequestering FMRP, it prevents FMRP from associating with CCAR1 mRNA, thereby reducing β-catenin-dependent transcriptional output[14]. The operation is RNA-protein competition, and the functional consequence is altered Wnt-related regulatory activity.

circRAPGEF1 provides a related m6A-dependent example of RNA-guided RBP redistribution. METTL3-mediated m6A modification stabilizes circRAPGEF1 and promotes its interaction with IGF2BP3. The modified circRNA competes with ASS1 mRNA for IGF2BP3 binding[34]. This leads to ASS1 mRNA loss, aspartate accumulation and activation of the S6K-CAD pathway. This operation resembles circZKSCAN1 in that a noncoding RNA redirects an RNA-binding protein among competing RNA substrates. Its downstream consequence, however, is metabolic rather than directly transcriptional[14,34].

Other RNAs influence regulatory-complex deployment by controlling protein proximity. Lnc-β-Catm does not primarily act by changing EZH2 abundance. Instead, it creates physical proximity between EZH2 and β-catenin, enabling a modification event that increases β-catenin persistence[15]. SNORA74A represents a related boundary case in which proximity control primarily produces a dosage outcome through altered E2F2 turnover[13]. These mechanisms overlap with dosage control at the level of outcome. Their proximal logic, however, is spatial and relational. RNA determines which proteins meet, which interaction is favored and which regulatory complex becomes productive[13,14].

The distinction between deployment and dosage is therefore analytical rather than absolute. circZKSCAN1 and circRAPGEF1 primarily reallocate RNA-binding proteins among competing RNA substrates[14,34]. Lnc-β-Catm and SNORA74A organize protein interactions that subsequently alter protein persistence[13,15]. SNORA49 and circIPO11 affect transcription by controlling promoter access rather than by directly changing the abundance of the recruited proteins[1,19]. In each case, the RNA mechanism determines regulatory activity by controlling location, partner choice or locus access.

The deployment mechanisms differ across RNA classes in both direction and molecular architecture. CircRNAs can redistribute RNA-binding proteins, lncRNAs can recruit or scaffold regulatory proteins and snoRNAs can alter subnuclear protein positioning. Across these examples, the shared principle is context-specific control of where regulatory proteins act, which partners they engage and which loci or RNA substrates they access[1,11-15,19]. More conventional circRNAs-miRNA-mRNA circuits, such as circ_0000972-miR-96-5p-PFN1[42], fit transcript-level competition and are therefore kept outside the deployment category.

INTEGRATIVE MODEL, SELECTIVITY AND LIMITS OF CURRENT EVIDENCE

The studies reviewed above reveal a sharp asymmetry. The upstream RNA regulators are diverse, but their downstream outputs repeatedly fall on a smaller set of developmental, inflammatory and metabolic pathways. The most parsimonious interpretation is that diverse upstream RNA mechanisms create state-specific vulnerabilities by converging on a smaller set of recurrent self-renewal nodes. Figure 4 summarizes this pathway-level recurrence.

Pathway convergence and biological heterogeneity

The strongest pattern of convergence occurs at NOTCH, Wnt-β-catenin, Hedgehog/GLI1, Hippo-YAP and TGF-β/IL-6-STAT3 signaling. NOTCH activity is increased through E2F2 stabilization, YTHDF1-dependent NOTCH1 mRNA stability and translation, and the DUBR-CIP2A-E2F1 circuit[4,13,33]. Wnt-β-catenin signaling is reinforced through several mechanistically distinct routes, including repression of the β-catenin destruction machinery, altered FOXO1-dependent β-catenin localization, direct stabilization of β-catenin, redistribution of the FMRP-CCAR1 interaction and METTL3-dependent FZD10 expression[3,5,7,14,15,30,31,43,44]. These mechanisms converge on pathway output while acting at different molecular checkpoints.

circIPO11-mediated promoter access, ADAR1-mediated GLI1 recoding, ALKBH5-SOX4-SHH signaling and tRNA m1A-dependent PPARδ translation provide distinct routes to Hedgehog-related output[1,10,36,38]. TGF-β and IL-6-STAT3 signaling is controlled through ligand-mRNA stabilization, miRNA-mediated repression of pathway inhibitors, m6A-linked SOCS3 regulation and promoter-associated IL-6 control[6,16,17,35]. Hippo output is linked to SNORD88B-dependent STK4 repression and FZD10-associated YAP1 activation[5,11]. Metabolic rewiring forms a secondary convergence layer. Glycolysis, cholesterol synthesis and aspartate metabolism create permissive conditions for self-renewal[24,34,38]. They are treated as cross-cutting outputs rather than a fourth operation. Current RNA-centered evidence does not yet define metabolism as a recurrent pathway node across independent RNA classes and HCC contexts.

The recurrent pathways nevertheless represent mechanistically distinct outputs. NOTCH1 and NOTCH3 are not equivalent[4,13]. β-catenin activation through destruction-complex inhibition, nuclear redistribution or protein stabilization may generate different biological dependencies[3,7,15]. Likewise, Hedgehog activation through GLI1 transcription is not the same as Hedgehog activation through edited GLI1 protein or altered cholesterol metabolism[1,10,38]. Individual axes such as SNORA49-SOX9, HAND2-AS1-BMPR1A/BMP and miR-192-5p-linked glycolytic feedback are supported by functional data[12,19,24]. They are not recurrent nodes in the same sense as Wnt, NOTCH or STAT3.

Some of this recurrence may also reflect research bias, because established CSC pathways are more likely to be selected for mechanistic follow-up. Even so, convergence across microRNAs, snoRNAs, lncRNAs, circRNAs, RNA-binding proteins and RNA-modifying enzymes is meaningful. It argues that these pathways occupy high-leverage positions in self-renewal control[1,3-5,7,11,13-16,38]. The evidence therefore supports pathway-level recurrence rather than conservation of any single RNA dependency.

Heterogeneity limits generalization further. Single-cell analysis shows that HCC CSC populations are transcriptionally and functionally heterogeneous[8]. Marker-enriched compartments also point to partly distinct dependencies. CD90-positive cells have been linked to the miR-589-5p-MAP3K8 axis[22]. Malignant EpCAM-positive cells have been linked to reduced miR-26b-5p and HSPA8[21]. CD133-positive cells have been linked to translation-associated circuits such as EIF5A2-c-MYC-miR-29b[39]. Etiology and model system add another layer. The HBX-miR-5188-FOXO1 circuit is most directly relevant to HBV-associated disease[7]. Studies have identified other dependencies in xenograft, chemically induced, oncogene-driven, spontaneous and patient-derived models[1,4,9,11,18]. These systems strengthen causal inference in their own contexts, but they do not establish cross-subtype conservation. Direct evidence for context-dependency, defined as the same RNA mechanism producing divergent functional outcomes across different HCC CSC states, remains limited. The stronger claim supported by current data is model- and etiology-specific dependency. Different upstream RNA regulators are required in different experimental contexts, but current data do not prove that individual RNA regulators switch function across contexts.

Functional selectivity and the therapeutic window

Functional relevance and therapeutic selectivity are distinct evidentiary questions. An RNA dependency may be required for tumor propagation but still be unsuitable for therapy if normal liver cells depend on the same mechanism. Functional evidence alone therefore cannot define a therapeutic window. The strongest studies reviewed here link RNA perturbation to serial propagation, tumor-initiating frequency or de novo tumor formation. Several also use genetic disruption, patient-derived models or mechanistic rescue[1,4,9-12]. These experiments establish functional relevance. Selectivity requires additional evidence.

We assess selectivity at four levels. First, the RNA dependency should be enriched in tumor-propagating cells relative to bulk malignant cells and normal regenerative cells. Second, the same perturbation should have a greater functional effect in the malignant compartment. Third, liver homeostasis and injury-induced regeneration should remain preserved. Fourth, an effective exposure window should be achievable through dose or delivery. Most studies reviewed here provide evidence for the first level. Far fewer directly address the other three.

The miR-26b-5p/HSPA8 study provides an informative comparison at the first level. The investigators compared EpCAM-positive and EpCAM-negative cells from HCC with EpCAM-positive cells from non-cancerous, non-cirrhotic liver[21]. HSPA8 was enriched and miR-26b-5p was reduced in the malignant EpCAM-positive population relative to both comparator populations. Functional perturbation in malignant cells also linked reduced miR-26b-5p to increased spheroid formation, migration, invasion, tumorigenicity and the EpCAM-positive fraction. These findings identify a molecular difference between malignant EpCAM-positive cells and a relevant normal EpCAM-positive population. They also support a functional role for the miR-26b-5p/HSPA8 axis in malignant cells. However, the same perturbation was not performed in normal EpCAM-positive cells. The normal samples also came from non-cancerous, non-cirrhotic liver rather than regenerating liver. This study therefore provides stronger evidence for molecular selectivity than for functional or regenerative selectivity.

METTL16 addresses a different part of the framework and provides stronger in vivo evidence. Xue et al[9] generated liver-specific Mettl16 conditional knockout mice by crossing Mettl16-floxed mice with Albumin-Cre mice. Neither heterozygous nor homozygous liver-specific deletion significantly changed survival, body weight, liver weight or gross liver size. Liver histology also showed little change, and cleaved caspase-3 was not markedly increased. Homozygous deletion did alter the hepatic immune-cell composition. CD4-positive T cells and NK cells were moderately reduced, whereas macrophages increased[9]. In the same study, Mettl16 loss markedly suppressed tumor development in a hydrodynamic tail-vein injection model of de novo hepatocarcinogenesis. METTL16 depletion also reduced CSC frequency in limiting-dilution assays[9]. These results indicate that HCC initiation and CSC maintenance are more sensitive to Mettl16 loss than baseline postnatal liver growth and homeostasis in this model.

This separation remains incomplete. The conditional knockout experiments examined baseline liver development and homeostasis. They did not test regenerative demand after partial hepatectomy or chronic inflammatory, fibrotic, cholestatic or metabolic injury. They also did not directly measure serial self-renewal of normal hepatic progenitors. Preserved liver architecture after Mettl16 deletion therefore supports a partial therapeutic window. It does not establish that METTL16 is dispensable during liver regeneration. The two studies consequently address different components of selectivity. The miR-26b-5p/HSPA8 study provides a malignant vs normal EpCAM-positive molecular comparison. The METTL16 study provides in vivo evidence that genetic loss affects hepatocarcinogenesis more strongly than baseline liver homeostasis.

Pathway convergence makes this distinction especially important. NOTCH, Wnt/β-catenin, Hedgehog, Hippo, STAT3 and BMP also participate in developmental or regenerative biology. Convergence on these pathways therefore does not itself create a therapeutic window. Selectivity may instead arise at the level of the upstream RNA regulator if malignant cells depend more strongly on that regulator than normal regenerative cells. It may also arise through delivery if tumor tissue receives greater effective exposure than normal or regenerating liver. These possibilities require matched perturbation in malignant and normal compartments. They also require injury-regeneration models and tumor vs normal pharmacodynamic measurements. Lineage-resolved functional assays would further show whether normal progenitor activity is preserved.

Current intervention studies establish therapeutic accessibility more often than durable selectivity. RNA replacement, antisense oligonucleotides, antagomirs, siRNAs and lipid or nanoparticle delivery can suppress several RNA-centered dependencies reviewed here[1,4,7,11-13,19,34]. Several strategies have also been combined with pathway inhibition or systemic therapy[1,4,7,11,12,19]. These findings support experimental feasibility. They do not show that tumor-propagating cells have been durably removed. Reduced sphere formation or tumor volume can reflect transient growth suppression without elimination of residual CSCs. Stronger evidence would combine post-treatment CSC-frequency measurements or serial transplantation with liver regeneration challenges and tumor vs normal biodistribution. Among the examples reviewed here, METTL16 provides one of the strongest partial in vivo separations between HCC dependence and baseline liver tolerance. The miR-26b-5p/HSPA8 study provides one of the clearest molecular comparisons between malignant and normal EpCAM-positive populations. Neither study yet demonstrates a complete therapeutic window.

What remains unresolved

The central unresolved question is stability of dependency. HCC CSCs may not occupy a fixed hierarchy. If tumor cells can move between marker-defined or signaling-defined states, inhibition of one RNA axis may remove the dominant state while selecting for another tumor-propagating population. Current studies rarely track post-treatment state transitions, test recurrent tumors for alternative RNA dependencies or combine perturbation with lineage tracing across serial propagation.

A second unresolved issue is conservation across disease contexts. HBV-associated, metabolic, alcohol-associated and non-cirrhotic HCC may converge on overlapping signaling pathways while using different upstream RNA mechanisms. Pooled HCC cohorts can obscure this distinction. A clinically useful RNA dependency will need to be retained in a definable patient subset, not merely detectable in a mixed tumor collection.

A separate evidence gap concerns RNA-processing mechanisms that are plausible but under-supported in functional CSC assays. Alternative splicing, isoform choice and regulated RNA decay remain relevant. However, the available corpus contains less direct evidence linking these processes to repeated HCC stem-cell self-renewal and tumor propagation. They should remain identified gaps rather than be promoted into parallel mechanistic categories on the basis of general HCC association.

The standard for the next phase should be functional selectivity. A credible RNA target should reduce tumor-propagating frequency, remain required during serial propagation, withstand compensation by alternative stem-like states and spare normal liver regeneration. Until those conditions are met, RNA-centered mechanisms should be treated as model-, etiology- and state-specific vulnerabilities of defined HCC CSC states. Mechanisms that have not met these criteria are best interpreted as model-, etiology- or state-specific vulnerabilities rather than broadly actionable HCC CSC targets.

CONCLUSION

The current literature supports several experimentally compelling RNA dependencies in defined HCC CSC states, but it does not yet support an HCC-wide RNA stemness program. The more consequential conclusion is methodological: Candidate RNA regulators should be judged by functional self-renewal, mechanistic causality, context conservation and normal-liver selectivity rather than by differential expression, marker enrichment or primary sphere formation alone. Under this standard, the field contains a growing set of mechanistically resolved dependencies but very few fully demonstrated therapeutic windows.

Progress now depends on experiments that connect these dimensions rather than expanding the list of RNA regulators. Matched malignant and normal regenerative models, injury-regeneration challenges, serial post-treatment propagation, residual tumor-propagating frequency and longitudinal analysis of state replacement will be especially informative. The central translational question is therefore whether an RNA dependency remains necessary as HCC cells change state and whether its inhibition durably reduces tumor-propagating capacity while preserving liver regeneration. Mechanisms that satisfy those criteria would move RNA-centered HCC CSC biology from a map of context-specific dependencies toward a basis for patient-stratified therapeutic intervention.