HCC is characterized by inter-patient heterogeneity arising from diverse etiological factors leaving distinct mutational and epigenetic imprints [8,9]. Host specific variables, including gut microbiome, immune landscape, and germline background, add additional layers of heterogeneity [10]. At the molecular level, oxidative stress, recurrent alteration in oncogenes (CTNNB1, c-Myc, c-Met, TERT, PIK3CA), tumor suppressors (AXIN, RB, PTEN, TP53, KEAP1), and chromatin regulators (ARID1A, ARID2), drive distinct phenotypes, such as the “proliferation” vs. “non-proliferation” subclasses, each carrying unique prognostic implications [11,12].

Equally critical is intra-tumoral heterogeneity (ITH), reflecting the coexistence of genetically and phenotypically distinct subclones within individual tumors. ITH drives clonal evolution, metastasis, recurrence, and therapeutic resistance [13]. Under therapeutic pressure, resistant subclones with survival advantages expand a Darwinian process that underlies the modest benefits of multi-kinase inhibitors and fosters multi-drug resistance. Beyond genetic diversity, ITH drives functional divergence, including differences in proliferative capacity, invasiveness, angiogenic signaling, and immune evasion (Fig. 1). Clinically, this diversity complicates tumor sampling, as a single biopsy rarely captures the complete genomic and immune architecture of a tumor, necessitating multi-regional or liquid biopsy approaches, which remain technically and clinically challenging [14].

Addressing this biological reality requires a shift from static to adaptive therapeutic strategies. Rational combinations—whether simultaneous or sequential—must target multiple vulnerabilities to prevent clonal escape. While recent molecular classifications have categorized HCC by oncogenic pathways (e.g., Wnt/β-catenin) and immune phenotypes (e.g., “inflamed” vs. “excluded”), these frameworks remain largely descriptive. To achieve clinical impact, they must evolve into actionable stratification models that link subtypes to specific biomarkers and therapeutic vulnerabilities. An ideal clinical workflow functions as a dynamic decision tree (Fig. 2), where longitudinal monitoring via liquid biopsies and spatial multi-omics guides real-time adjustments in treatment, shifting from empirical selection to mechanism-based precision medicine [13].

HCC frequently exhibits insensitivity to conventional chemotherapy agents and a propensity to develop multidrug resistance (MDR) rapidly during treatment, leading to reduced survival and poor prognosis. This resistance arises from a complex interplay of molecular and cellular mechanisms (Table 1). Overexpression of ATP-binding cassette (ABC) transporter proteins such as P-glycoprotein (P-gp/ABCB1), Breast Cancer Resistance Protein (BCRP/ABCG2), and multidrug resistance-associated proteins (MRPs/ABCC), actively pump anticancer drugs (adriamycin, paclitaxel, 5-FU, sorafenib) out of tumor cells, thereby diminishing the drug’s therapeutic efficacy and inducing MDR [15].

Dysregulation of programmed cell death (PCD) pathways is a hallmark of MDR. This includes TP53 mutations and the upregulation of anti-apoptotic proteins like Bcl-2 and MCL-1 [16]. PANoptosis is an emerging inflammatory PCD pathway that integrates molecular features of pyroptosis, apoptosis, and necroptosis through the formation of multi-protein PANoptosome complexes. Disruption of PANoptosis related genes can influence tumor progression, immune evasion, and therapy resistance in multiple malignancies, and that PANoptosis linked signatures have potential as biomarkers of prognosis and immunotherapeutic responsiveness in HCC [17]. Furthermore, enhanced DNA repair capabilities (e.g., via XRCC4-like factor in the non-homologous end-joining (NHEJ) pathway allow HCC cells to survive the genomic stress induced by conventional cytotoxic agents [18].

Reversible epigenetic modifications can silence tumor suppressors or activate pro-survival genes. Similarly, while basal autophagy is suppressive, activated autophagy in established tumors helps cells survive metabolic and therapeutic stress.

TME is a complex ecosystem comprising various normal cells (e.g., fibroblasts, endothelial cells, immune cells) and the extracellular matrix (ECM), acts as a protective niche. Specific metabolites (e.g., 27-hydroxycholesterol) or ECM stiffness can impede drug delivery and foster a pro-survival signalling. Furthermore, cancer stem cells (CSCs) (marked by EpCAM or CD133) contribute to recurrence and metastasis due to their inherent resistance and capacity for dormancy (Table 2).

These resistance mechanisms rarely operate in isolation; they form a redundant network where targeting a single node often triggers compensatory activation of another. For example, epigenetic changes may upregulate ABC transporters, while an immunosuppressive TME protects resistant CSCs. This inherent redundancy necessitates multi-target combination therapies designed to disrupt several nodes concurrently. By identifying master regulators or common vulnerabilities (Table 2), clinicians can move toward highly personalized, mechanism-based strategies that delay resistance and improve long-term outcomes.

The TME is a highly dynamic ecosystem composed of malignant cells, stroma, extracellular matrix (ECM), and immune populations that actively orchestrates tumor survival, angiogenesis, and immune evasion (Fig. 4). Central to this immunosuppressive landscape is the upregulation of inhibitory checkpoints that induce T cell exhaustion and promote immune tolerance. The PD-1/PD-L1 axis is the most clinically validated pathway; tumor cells exploit PD-L1 expression to deactivate effector T cells directly within the TME (Fig. 4). CTLA-4 (Cytotoxic T-Lymphocyte-Associated protein 4) functions as a critical immune brake primarily during the early priming phase of T cell activation in the lymph nodes, distinct from the peripheral activity of PD-1. LAG-3 impairs CD4⁺ T cell activation via MHC-II trans-endocytosis and enhances Treg-mediated suppression [39]. TIM-3 receptor, expressed on exhausted CD8⁺ T cells, interacts with galectin-9 to inhibit interferon-γ secretion and drive macrophages toward a pro-tumorigenic, immunosuppressive phenotype.

Metabolic plasticity is a defining feature of HCC, reflecting both tumor-intrinsic alterations and adaptation to the diseased liver microenvironment. HCC cells exhibit enhanced glycolysis, altered lipid metabolism, increased glutamine dependency, and dysregulated bile acid signaling [40]. In MASLD/NASH related HCC, lipid accumulation, oxidative stress, and chronic inflammation further complicate tumor metabolism and immune responses. These metabolic states can directly influence sensitivity to systemic therapies, including immunotherapy, underscoring the need for etiology aware treatment strategies [41].

Aberrant DNA methylation silences tumor suppressor genes such as CDKN2A, while deregulated histone modifications, alter chromatin accessibility and oncogenic transcriptional programs. These changes not only promote tumor initiation but also blunt antitumor immunity by repressing immune-stimulatory gene expression. Emerging data highlighted a mechanosensitive checkpoint Osr2 recruits HDAC3 to suppress cytotoxic T cell gene programs, directly connecting biomechanical stress with immune exhaustion [33].

Tyrosine kinase inhibitors (TKIs) were the first systemic agents to demonstrate survival benefit in advanced HCC; however, their efficacy is limited by rapid adaptive resistance. In parallel, ICIs have transformed HCC treatment but exhibit modest response rates when used alone, largely due to an immunosuppressive TME. These complementary limitations have driven the development of TKI-ICI combination regimens, which now form the backbone of first-line systemic therapy (Table 4). The mechanistic rationale underlying this approach lies in the ability of anti-angiogenic TKIs to normalize tumor vasculature, reduce hypoxia, and attenuate immunosuppressive cytokine signaling, thereby facilitating immune cell infiltration and function. When combined with ICI, this vascular and immunological reprogramming enhances T cell-mediated anti-tumor responses and mitigates intrinsic immune resistance (Fig. 5). Clinical validation of this strategy was achieved with the IMbrave150 trial, which established atezolizumab plus bevacizumab as the global first-line standard for unresectable HCC (can’t be removed by surgery). Beyond this regimen, multiple combinations pairing ICI with multi-targeted TKI are under active investigation, reflecting a sustained effort to refine efficacy while balancing toxicity profiles.

The future of HCC treatment lies in integrated, multi-modal strategies that combine molecular targeting, immune modulation, and microenvironmental reprogramming. A conceptual framework involves a sequenced approach:

An RNA agent (e.g., siRNA) silences a primary resistance driver (e.g., CTNNB1).

While such regimens offer substantial promise, their complexity introduces challenges related to toxicity, treatment sequencing, and cost. Addressing these barriers will require the development of predictive biomarkers, optimized dosing strategies, and adaptive trial designs. Ultimately, the successful implementation of integrated combination therapies will depend on precision medicine approaches that align therapeutic intensity with tumor biology and patient-specific risk profiles. The vast majority of trials combine ICI, TKI, anti-VEGF agents, and locoregional therapy (Table 6).

A significant number of trials are investigating the combinations in the neoadjuvant (pre-surgery) and adjuvant (post-curative treatment) settings to reduce recurrence rates (Table 6). Many trials are specifically focused on treating patients who have progressed on the current first-line standard of care. Emerging agents target new pathways such as TIGIT (Tiragolumab), LAG-3 (Relatlimab), TIM-3 (Cobolimab), CTLA-4 with enhanced Fc function (BMS-986218), and cellular therapies (CAR-T, CAR-Macrophages). The integration and timing of TACE, TARE (transarterial radioembolization), SBRT (Stereotactic Body Radiation Therapy), and HAIC (Hepatic Artery Infusion Chemotherapy) with systemic therapy is a major area of active investigation (Table 7).

HCC arises from diverse etiologies can reshape tumor biology and therapeutic responsiveness (Table 8). HBV-associated HCC is characterized by viral DNA integration, HBx-mediated malfunction of oncogenic signaling pathways, TERT promoter and TP53 mutations, AKT pathway activation, and epigenetic silencing of tumor suppressor genes, generating a relatively “inflamed” TME with T-cell infiltration and often enhanced responsiveness to ICI [86]. In contrast, HCV-associated HCC arises mainly through chronic inflammation, oxidative stress, and immune disruption rather than direct genomic integration, with frequent CTNNB1 and AXIN1 alterations promoting a WNT/β-catenin driven “immune-excluded” TME and resistance to anti-PD-1 therapy. MASLD/NASH-related HCC exhibits pronounced metabolic reprogramming, lipotoxicity, IL-6/JAK/STAT pathway activation, immune exhaustion, and characteristic genetic variants such as PNPLA3, TM6SF2, and MBOAT7, resulting in a metabolically inflamed, immunosuppressive TME that reduces ICI efficacy but highlights potential vulnerabilities to metabolic and cytokine-targeted therapies [87,88]. Similarly, ALD-associated HCC shows overlapping features with MASLD-HCC, including IL-6/JAK/STAT activation and immunosuppressive TME shaped by chronic alcoholic injury, alongside acetaldehyde-induced DNA damage, oxidative stress, and aberrant methylation patterns that accelerate hepatocarcinogenesis. Finally, aflatoxin-related HCC is defined by a high-frequency TP53 R249S mutation caused by aflatoxin B1 exposure, often synergizing with HBV infection, and generating a genotoxin-driven TME with vulnerabilities in DNA damage repair and cell cycle control [89].

While ICI, particularly combinations such as atezolizumab plus bevacizumab, have redefined first-line therapy for advanced HCC, critical gaps remain that limit their universal efficacy [90]. Notably, chronic viral HCC tends to be “hot” and inflamed, favoring ICI responsiveness, whereas MASH/NASH-HCC often exhibits a “cold” microenvironment, potentially rendering ICI ineffective or even detrimental; however, pivotal trials have largely failed to stratify patients by etiology [89]. Moreover, current studies exclude patients with compromised liver function (Child-Pugh B/C), leaving their safety and efficacy profiles uncertain. High-frequency alterations in the TERT promoter (up to 60%) and Wnt/β-catenin pathway (CTNNB1 mutations, up to 30%) remain largely undruggable, highlighting the need for next-generation targeted therapies, including small-molecule inhibitors or gene-editing approaches, potentially in combination with ICI, to achieve durable responses across HCC subtypes [91]. Collectively, these etiological subtypes illustrate that HCC is not a uniform disease entity; integrating molecular, immunological, and metabolic profiling into precision oncology frameworks enables etiology-stratified biomarker development, individualized immunotherapeutic and metabolic interventions, and rationalized clinical trial design [91].

Comprehensive multi-omics profiling is essential for capturing the molecular complexity of HCC. Large-scale public consortia have generated foundational datasets integrating genomics, transcriptomics, epigenomics, proteomics, and clinical annotation. The Cancer Genome Atlas-Liver Hepatocellular Carcinoma (TCGA-LIHC) cohort provides matched genomic, transcriptomic, epigenomic, and proteomic data for approximately 360–400 primary HCC tumors, predominantly derived from patients with viral hepatitis (HBV/HCV) or metabolic liver disease (NAFLD/MASH). Complementing TCGA, the International Cancer Genome Consortium (ICGC) incorporates ethnically and etiologically diverse cohorts from multiple geographic regions, expanding the generalizability of molecular discoveries, as highlighted in Table 9. Transcriptomic datasets from the Gene Expression Omnibus (GEO) and European Nucleotide Archive (ENA) further support validation studies, particularly for TME characterization and drug resistance mechanisms. Proteogenomic data from the Clinical Proteomic Tumor Analysis Consortium (CPTAC) provide direct insights into functional protein states and post-translational modifications linked to prognosis and treatment response.

Single-layer molecular analyses are insufficient to explain HCC pathogenesis or therapeutic resistance. Integrated multi-omics approaches overcome this limitation by combining genomics, transcriptomics, proteomics, epigenomics, and metabolomics to construct a comprehensive molecular portrait of the tumor. Genomic analyses reveal recurrent alterations in oncogenic and tumor suppressor pathways, including Wnt/β-catenin, PI3K/AKT/mTOR, MAPK/ERK, and TP53, which inform patient stratification for targeted therapies and immunotherapy. Transcriptomic profiling captures dynamic gene expression programs associated with proliferation, immune exclusion, metastasis, and drug resistance, while also identifying non-coding RNAs (lncRNAs and miRNAs) as emerging biomarkers and therapeutic targets. Proteomics provides a functional readout of tumor biology by quantifying signaling proteins, immune checkpoint expression, and angiogenic mediators, often outperforming mRNA-based predictors. Epigenomic and metabolomic analyses further uncover regulatory and metabolic rewiring, including aberrant DNA methylation and enhanced glycolysis, fatty acid oxidation, and glutaminolysis.

Crucially, integrated multi-omics signatures outperform single biomarkers in predicting prognosis and therapeutic response. Systems-level analyses have identified robust prognostic programs, including PANoptosis-related gene signatures, as well as actionable vulnerabilities in treatment-refractory HCC, underscoring the translational value of multi-modal integration (Fig. 6) [69].

The high dimensionality and heterogeneity of multi-omics, imaging, and clinical datasets necessitate advanced AI and machine-learning (ML) frameworks capable of extracting clinically actionable insights. Deep learning architectures, particularly convolutional neural networks (CNN) for imaging and deep neural networks (DNNs) for multimodal data fusion now form the computational backbone of precision oncology in HCC (Table 10).

AI models in HCC have achieved clinically meaningful performance across diagnostic, prognostic, and therapeutic domains (Table 11). Radiomics-based classifiers differentiate HCC from benign or cirrhotic nodules with pooled AUCs of approximately 0.86, while survival prediction models achieve C-indices ranging from 0.73 to 0.78 in external validation cohorts [86,102,103]. Integrated multi-omics models further enhance survival prediction, reporting 1-year survival AUCs approaching 0.98 in TCGA-based analyses and improved long-term prognostic accuracy when incorporating key driver genes (FBN1, MAP1B) [104,105].

Despite these advances, generalizability remains a major barrier. Models trained in HBV-predominant populations frequently underperform in MASLD-dominant cohorts, reflecting true biological heterogeneity rather than technical bias [86]. Addressing this challenge requires federated learning, multinational collaboration, and harmonization of imaging protocols, radiomics pipelines, and omics preprocessing.

Static treatment algorithms are insufficient to address tumor evolution and microenvironmental plasticity. Adaptive precision oncology reframes treatment as a dynamic, data-driven process in which therapy evolves in response to longitudinal molecular and immunological changes. This framework integrates baseline molecular stratification, AI-guided therapy selection, serial liquid biopsy surveillance, and dynamic therapeutic adaptation.

AI-driven platforms integrating multi-omics, radiomics, and clinical data can anticipate resistance mechanisms (often preceding radiographic progression) and recommend timely therapeutic modification. Reinforcement learning (RL) models and graph-based frameworks further enable simulation of tumor evolution under different treatment pressures, informing optimal sequencing strategies [86]. Collectively, these approaches redefine HCC management as an anticipatory and personalized process, with AI serving as the analytical backbone for next-generation precision oncology (Fig. 7).

Algorithms such as YOLOv8 and frameworks (MONAI) extract quantitative features from medical images to enable earlier HCC detection, non-invasive characterization of tumor grade and vascular invasion, and more accurate monitoring of disease progression (Table 12). Furthermore, AI tools (Visiopharm and Aiforia) provide quantitative analysis of the TME from digital pathology images, elucidating immune cell infiltration patterns and spatial relationships that are critical for predicting immunotherapy outcomes. The development of generative AI models promises to create sophisticated in silico simulations of tumor evolution. By integrating the full spectrum of omics data, these models aim to reconstruct the dynamic interplay within the TME, predict responses to novel interventions, and ultimately guide the development of truly personalized, adaptive therapeutic strategies for every HCC patient.

While multi-omics profiling reveals what molecular alterations are present in a tumor, it often homogenizes the sample, obscuring the critical architectural context of where these alterations occur. ST addresses this fundamental limitation by mapping gene expression with cellular-to-subcellular resolution directly within the native tissue architecture [106]. This transformative technology provides an unprecedented view of the TME, enabling the deconstruction of the intricate cellular ecosystems, signaling networks, and metabolic niches that collectively govern HCC progression, immune evasion, and therapeutic response [107]. In HCC, a malignancy defined by profound intra-tumoral heterogeneity, ST is particularly powerful. It moves beyond bulk analysis to delineate distinct tumor subclones coexisting within a single lesion, each with unique molecular features. For instance, it can reveal regions characterized by high expression of angiogenic markers (e.g., VEGF) adjacent to zones expressing immune checkpoint proteins (e.g., PD-L1). This granular mapping is essential for understanding the functional consequences of tumor evolution and for designing therapies that can address the tumor’s multifaceted biology. Furthermore, this technology is revolutionizing our understanding of the immunosuppressive landscape of HCC. By spatially resolving the locations and gene expression profiles of diverse cell populations, researchers can precisely map the “battlegrounds” within the TME. This includes identifying immunosuppressive hubs enriched with regulatory T cells (Tregs), myeloid derived suppressor cells (MDSCs), tumor associated macrophages (TAMs), and visualizing the physical barriers that lead to immune exclusion zones, where cytotoxic T cells are prevented from infiltrating the tumor core. Understanding these spatial dynamics is critical for developing next-generation immunotherapies, such as CAR-T cell therapies or novel checkpoint inhibitors, designed to reprogram the TME and overcome resistance.

The synergy between ST and AI is essential for unlocking the full potential of this data rich technology. The sheer complexity of spatial datasets necessitates sophisticated computational methods for analysis. AI algorithms, particularly deep learning models like convolutional neural networks (CNNs), are being deployed to automate the identification of histological features, quantify the density and spatial distribution of tumor infiltrating lymphocytes (TILs), and define recurrent cellular neighborhoods that correlate with clinical outcomes. Moreover, AI is enabling the development of 3D spatial reconstructions of the TME, providing multi-dimensional insights into the interactions between tumor, stromal, and immune cells. This powerful combination is not only elucidating fundamental mechanisms of immune evasion but is also identifying novel, spatially defined biomarkers and therapeutic targets. Integrating these spatial insights with multi-omics data has moved us towards a new era of precision oncology where therapeutic strategies are tailored not only to the molecular profile of a tumor but also to its unique spatial architecture.

Practical Constraints on Spatial Transcriptomics (ST)

While techniques such as ST and multi-modal AI modeling represent the scientific apex of precision oncology, their clinical implementation faces significant practical barriers. Despite its revolutionary power to map cellular interactions within the TME, the clinical utility of ST is currently hampered by high per-sample cost, low throughput compared to bulk sequencing, and the lack of standardized computational pipelines. These factors restrict its use primarily to specialized research institutions, delaying its deployment as a routine clinical tool. Although many AI models show high performance in silico, integrating them into diverse clinical settings is challenging. Deployment requires significant IT infrastructure investment, data harmonization efforts across multiple imaging systems, and ongoing regulatory approval and maintenance. This high initial and operational cost often creates a disparity, limiting the availability of these advanced predictive tools to large academic medical centers, thereby potentially increasing existing health disparities. The foundational data for multi-omics and ST requires high-quality tissue. HCC often presents in patients with cirrhosis, making repeated, invasive tumor biopsies risky and limiting the amount of available tissue. This constraint forces reliance on less precise liquid biopsy methods or radiomics, highlighting a fundamental physical limitation of data acquisition in this disease setting.

Despite an expanding catalog of oncogenic drivers, immune regulators, epigenetic modifiers, and metabolic dependencies in HCC, only a limited number of discoveries have translated into approved therapies. This translational gap largely reflects the limitations of preclinical models (like standard cell lines and basic CDX models) that inadequately recapitulate the complexity of human TME, including tumor heterogeneity, immune contexture, vascular remodeling, and underlying liver dysfunction [108].

Successful examples, such as the clinical validation of immune checkpoint inhibition combined with anti-angiogenic therapy, illustrate the importance of aligning mechanistic rationale with clinically representative models and biomarker-driven trial design [109]. Future translational research must prioritize advanced preclinical systems such as patient-derived organoids, spatially resolved co-culture models, and humanized mouse platforms that faithfully mimic the TME and chronic liver disease context [110]. These models are essential for improving therapeutic predictability and reducing late-stage clinical attrition.

The intrinsic molecular diversity of HCC and the inevitability of adaptive resistance necessitate a transition from uniform treatment algorithms to individualized therapeutic strategies. Limited tissue availability and spatial heterogeneity remain major barriers to precision medicine; however, the expanding use of liquid biopsy offers a scalable solution for non-invasive, longitudinal tumor profiling. Circulating tumor DNA and RNA analyses enable real-time monitoring of clonal evolution, emergent resistance mechanisms, and immune dynamics, supporting adaptive treatment modification [111]. The clinical utility of liquid biopsy will depend on the development of robust biomarkers capable of guiding therapeutic selection and response assessment. AI-driven analytical frameworks are poised to play a pivotal role by integrating multi-omics, imaging, and clinical data to enhance signal detection, reduce noise, and identify composite molecular signatures that define clinically relevant HCC subtypes. Such approaches move beyond single biomarkers toward multidimensional predictors with greater clinical robustness.

Overcoming therapeutic resistance remains a critical objective. As a single pathway rarely drives resistance in HCC; rather, compensatory signaling networks enable tumor persistence under selective pressure. This complexity underscores the need for rationally designed combination strategies that target shared regulatory nodes or convergent vulnerabilities. Emerging RNA-targeted therapeutics provide a promising orthogonal modality by enabling direct modulation of oncogenic transcripts and resistance-associated regulators [112,113].

PANoptosis is recognized as a bridge between inflammatory signaling and cell death. Its imbalance can suppress antigen presentation, alter cytokine production, and create an immunosuppressive TME. Multi-omics analyses have identified PANoptosis associated regulators and gene signatures linked to aggressive, treatment-refractory HCC phenotypes. Therapeutically activating PANoptosis has the potential to convert “cold” tumors into “hot” immunogenic lesions, thereby enhancing tumor immunogenicity, sensitize tumors to immune checkpoint blockade, and synergize with kinase inhibitors [114]. Integrating PANoptosis associated biomarkers into immunotherapy stratification models may refine patient selection and help mitigate the heterogeneity in treatment outcomes. This strategy exemplifies how systems-level biological insights can inform mechanism based therapeutic innovation.

Precision oncology in HCC therefore requires a shift from histology- and stage-based treatment toward etiology- and molecularly stratified decision-making. AI and ML are essential enablers of this transition. By integrating etiological background, multi-omics profiles, imaging features, and longitudinal clinical data, AI models can stratify patients into biologically coherent subgroups and predict optimal therapeutic sequencing [88]. Future clinical trials must adopt stratified designs that evaluate therapeutic efficacy within defined etiological-molecular subtypes, rather than treating HCC as a homogeneous entity [95]. Such trial paradigms will be critical for maximizing the benefit of emerging systemic and combination therapies.