Mechanisms of ferroptosis in primary hepatocellular carcinoma and progress of artificial intelligence-based predictive modeling in hepatocellular carcinoma

Abstract

Ferroptosis, an iron-dependent form of programmed cell death, has garnered significant attention in tumor research in recent years. Its core characteristics include aberrant accumulation of lipid peroxides and impairment of antioxidant defense mechanisms, such as dysfunction of glutathione peroxidase 4. These features are closely intertwined with the initiation, progression, and therapeutic resistance of hepatocellular carcinoma (HCC). This review presents a systematic overview of the fundamental molecular mechanisms underlying ferroptosis, encompassing iron metabolism, lipid metabolism, and the antioxidant system. Furthermore, it summarizes the potential applications of targeting ferroptosis in liver cancer treatment, including the mechanisms of action of anticancer agents (e.g., sorafenib) and relevant ferroptosis-related enzymes. Against the backdrop of the growing potential of artificial intelligence (AI) in liver cancer research, various AI-based predictive models for liver cancer are being increasingly developed. On the one hand, this review examines the mechanisms of ferroptosis in HCC to explore novel early detection markers for liver cancer, to provide new insights for the development of AI-based early diagnostic models. On the other hand, it synthesizes the current research progress of existing liver cancer predictive models while summarizing key challenges that AI predictive models may encounter in the diagnosis and treatment of HCC.

Key Words: Ferroptosis; Liver cancer; Sorafenib; Ferroptosis-related enzymes; Artificial intelligence prediction model; Ferroptosis-related noncoding RNAs

Core Tip: The molecular mechanisms of ferroptosis have been systematically investigated, and this review further summarizes the potential applications of targeting ferroptosis in liver cancer treatment. Not only does this review cover the anticancer drugs such as sorafenib and the mechanisms of action of related enzymes, but it also provides a comprehensive overview of the potential therapeutic strategies in this context. In addition, based on the rapid advancement of artificial intelligence-driven prediction models for liver cancer in recent years, this review discusses various types of artificial intelligence-based prediction models, aiming to provide new insights into the development of robust predictive tools for liver cancer.

INTRODUCTION

Liver cancer ranks among the most prevalent malignant tumors in China. Considering the lack of obvious early symptoms, most patients with hepatocellular carcinoma (HCC) are diagnosed in the middle or late stage. This results in a high mortality rate (20/100000-40/100000 people), which is the second highest among malignant tumors. Current treatments for liver cancer include surgery, radiotherapy, molecular targeted drug therapy, chemotherapy, immunotherapy, and traditional Chinese medicine. For patients with early liver cancer, surgery is currently the most effective treatment method. However, most liver cancer patients are in the middle or late stage at the time of diagnosis, meaning that they have already lost the opportunity of radical curative treatment by surgery, so no more than 20% of patients can undergo surgical treatment[1,2]. Presently, liver cancer treatment predominantly relies on multimodal medical therapies, including targeted drug therapy, chemotherapy, immunotherapy, and traditional Chinese medicine[3]. The research and development and application of targeted drugs in liver cancer also face many challenges, such as the lack of individual differences in the application of most vascular targeted drugs, the inability to better apply targeted drugs to achieve individualized treatment[4], resistance to targeted drugs, and issues in research and development of new targeted drugs. There is abundant evidence that only 30% of HCC patients are sensitive to targeted drug therapy, and most of them experience recurrent drug resistance[5]. Thus, finding new biomarkers and therapeutic targets to solve the problem of resistance to molecular-targeted drugs in HCC has recently become a hot spot in liver cancer treatment.

Ferroptosis, a newly defined non-apoptotic form of programmed cell death, was first proposed by Dr. Brent R Stockwell in 2012 to describe an iron-dependent cell death pathway driven by lipid peroxidation and excessive accumulation of reactive oxygen species (ROS). Accumulating studies have confirmed that ferroptosis is involved in various pathophysiological processes in humans, including tumorigenesis, cardiovascular diseases, ischemia-reperfusion injury, and fetal development[6,7]. In recent years, the use of iron death as a tissue-shaping strategy during embryogenesis has been confirmed by demonstrating that iron death accompanies large-scale but spatially restricted cell death events during limb muscle remodeling in the chicken embryo[8]. In the most recent study, it has been shown that iron death stress can transport Fe from cytoplasmic lysosomes vesicle-containing intracellular bacteria via ferroportin (FPN), thereby enhancing macrophage-mediated clearance of intracellular pathogens[9].

In studies exploring the association between ferroptosis and HCC progression, Zhang et al[10] identified that ferroptosis-related genes - including RRM2, AURKA, HELLS, CDC25A, and KIF20A - are closely linked to the initiation and development of HCC. Additionally, resistance to sorafenib, a first-line therapeutic agent for advanced HCC, has been strongly correlated with ferroptosis dysregulation[11]. Considering the close relationship between iron death mechanism and liver cancer development, this article reviews the recent research on iron death and liver cancer. Given the significant role of artificial intelligence (AI) in current medicine, this article also reviews the current status of AI-based prediction models related to liver cancer. Based on the current status of AI-driven prediction models for liver cancer, the aim is to establish a connection between novel liver cancer markers of iron death and AI prediction models, to provide a new construction direction for the AI prediction models of liver cancer and improve the AI prediction models of liver cancer. By integrating novel liver cancer biomarkers with AI predictive frameworks, this project aims to enhance clinical capabilities in diagnosis, prognosis, and treatment decision-making, and facilitate the development of novel targeted therapies for liver cancer.

LIVER CANCER

Infection with hepatitis B virus or hepatitis C virus, aflatoxin exposure, alcohol abuse, and other factors may lead to changes in biological characteristics of the liver during the process of repair after injury, gene mutation, suppression of oncogene expression and oncogenes, excessive cell proliferation, and carcinogenesis. In recent years, tremendous progress has been made in the research of liver cancer at the molecular level. Specifically, it has been shown that the occurrence of liver cancer is related to a variety of related signaling pathways, mainly the Ras/Raf/MEK/extracellular-signal regulated kinase (ERK) signaling pathway, the vascular endothelial growth factor (VEGF) receptor signaling pathway, the epidermal growth factor signaling pathway, and the hepatocyte growth factor/c-Met signaling pathway, which have become important topics in the research of molecular targeted drugs. A variety of molecular-targeted drugs have been developed to provide the best solution for patients with advanced HCC and those who cannot tolerate surgical treatment[12]. Molecular targeted therapy is a promising type of cancer treatment based on targeting specific molecules. Specifically, this approach involves interfering with the function of specific molecules through drugs, to prevent the occurrence and development of HCC. For example, VEGF/VEGF receptor (VEGFR) inhibitors affect liver cancer cell proliferation and apoptosis by inhibiting the phosphatidylinositol 3-kinase/protein kinase B signaling pathway, inhibiting the signal transducer and activator of transcription 3/Janus kinase 2 signaling pathway, affecting the mitogen-activated protein kinases signaling pathway, and activating the caspase signaling pathway. Sorafenib, a representative targeted agent, exerts antitumor effects primarily through two key mechanisms: First, inhibiting tumor cell proliferation by targeting receptor tyrosine kinases including c-Kit, FLT-3, and Raf-1; second, suppressing tumor neovascularization by inhibiting receptor tyrosine kinases such as VEGFR-2, VEGFR-3, and platelet-derived growth factor receptor-α/β[13]. Additionally, a recent study by Sun et al[14] demonstrated that sorafenib induces ferroptosis in HCC cells by inhibiting the p62/Kelch-like ECH-associated protein 1 (Keap1)/nuclear factor erythroid 2-related factor 2 (Nrf2) signaling pathway and activating ferroptosis-related genes. Key signaling pathways and their representative targeted drugs in recent years are summarized below, with detailed information provided in Figure 1 and Table 1.

Figure 1 The hepatocarcinogenesis-related signaling pathways and their drugs of action are critical for understanding liver cancer biology and developing targeted therapies. Sorafenib inhibits the expression of c-Myc and ETS transcription factor 1 through multiple rapidly accelerated fibrosarcoma/mitogen-activated protein kinase kinase/extracellular signal-regulated kinase signaling pathways, including the Janus kinase/signal transducer and activator of the transcription pathway, while promoting the expression of pro-death markers like P62. Cabozantinib exerts its anti-liver cancer effects primarily through the phosphatidylinositol 3-kinase/phosphatidylinositol-(3,4)-P2/protein kinase B pathway. Additionally, bevacizumab and apatinib target liver cancer by inhibiting key downstream signaling molecules in the phosphatidylinositol 3-kinase/phosphatidylinositol-(3,4)-P2/protein kinase B pathway, thereby modulating the cellular response to these anti-tumor therapies. These findings underscore the importance of understanding the molecular mechanisms of drug action in the development of more effective liver cancer treatments. EGF: Epidermal growth factor; FGF: Fibroblast growth factor; PTK: Protein tyrosine kinases; RAF: Rapidly accelerated fibrosarcoma; MEK: Mitogen-activated protein kinase kinase; ERK: Extracellular signal-regulated kinase; Elk1: ETS transcription factor 1; Keap1: Kelch-like ECH-associated protein 1; Nrf2: Nuclear factor erythroid 2-related factor 2; VEGF: Vascular endothelial growth factor; PI3K: Phosphatidylinositol 3-kinase; PIP3: Phosphatidylinositol-(3,4,5)-P3; HGF: Hepatocyte growth factor; AKT: Protein kinase B; STAT: Signal transducer and activator of the transcription; DLL4: Delta-like ligand 4.

Table 1 Representative targeted drugs and signaling pathways in liver cancer.

Representative targeted drugs	Related signaling pathways
Sorafenib	Ras/Raf/MEK/ERK signaling pathway
Apatinib, lenvatinib, bevacizumab, ramelimumab	VEGF signaling pathway
Erlotinib, cetuximab	EGF signaling pathway
Cabozantinib	HGF/c-Met signaling pathway

Raf: Rapidly accelerated fibrosarcoma; MEK: Mitogen-activated protein kinase kinase; ERK: Extracellular signal-regulated kinase; EGF: Epidermal growth factor; VEGF: Vascular endothelial growth factor; HGF: Hepatocyte growth factor; c-Met: Mesenchymal-epithelial transition factor.

PROCESS OF IRON DEATH

Ferroptosis is a novel form of regulated cell death. It is dependent on intracellular ferrous ions and centers on lipid peroxidation, which is induced when intracellular oxidative metabolism is abnormally enhanced and/or the operation of antioxidant mechanisms is blocked, so that the intracellular redox homeostasis is imbalanced. Morphologically, ferroptosis is primarily manifested by mitochondrial alterations, including mitochondrial morphological disorganization, changes in membrane potential, iron overload in mitochondrial membranes, and accumulation of lipid ROS[15]. The core mechanisms underlying ferroptosis mainly involve processes such as metabolic dysregulation, lipid peroxidation, and imbalance in the antioxidant system.

Iron metabolism

Iron is an indispensable trace element in the human body. The iron content of the adult body is 4.0-5.0 g. About 70% of the body iron is functional iron, which is present in hemoglobin, myoglobin, and enzyme cofactors, whereas 30% of the body iron is storage iron, which exists mainly in the form of ferritin and ferritin-containing hemoflavin in the liver, spleen, and bone marrow. Normally, people obtain iron through diet and iron is absorbed in the duodenum. After iron ions are bound in plasma transferrin, iron is transported to all parts of the body through blood for cellular utilization. Notably, only about 10% of iron is absorbed through the intestinal epithelium to compensate for daily iron losses, whereas the remaining 90% is recycled through the macrophage-erythroid cell cycle[16]. In the human body, iron is an important component of iron-containing enzymes related to DNA synthesis and repair, cellular respiration, lipid oxidation, and signaling, and is essential for physiological processes such as cell replication, metabolism, and cell growth[17].

Dysregulated iron metabolism is a pivotal driver of ferroptosis, an iron-dependent form of programmed cell death. Iron from dietary sources is absorbed in the small intestine and enters systemic circulation. Meanwhile, senescent red blood cells are phagocytosed and degraded by macrophages in the reticuloendothelial system; during this process, iron is released from the protoporphyrin ring of hemoglobin under the catalysis of heme oxygenase 1 (HO-1) for storage and reuse[18]. In humans, iron in the circulation mainly exists as Fe³⁺ and binds to transferrin (Tf). At a slightly alkaline extracellular pH of around 7.4, a single Tf molecule can bind two Fe³⁺ ions to produce a stable Tf-Fe³⁺ complex[19]. This complex subsequently binds to Tf receptor 1 (TfR1) located on the cell surface and transports Fe³⁺ into the cell[20]. TfR is regulated in multiple ways. While iron regulatory protein (IRP) and hypoxia-inducible factor-1 increase cellular iron uptake by increasing the expression of TfR, heat-shock protein B1 inhibits the expression of this receptor and thereby reduces iron uptake. Notably, the Hippo-YAP/TAZ signaling pathway promotes ferroptosis by increasing labile iron levels - achieved through upregulating TfR1 and downregulating ferritin heavy chain 1[21]. In an acidic environment, Fe³⁺ dissociates from Tf-TfR1 and is reduced to Fe²⁺ by reductase prostate 3 six transmembrane epithelial antigen. It then enters the cytoplasmic destabilized iron pool via divalent metal transporter protein 1, while the remaining Tf-TfR1 complex cycles to the cell surface[22]. In addition, ZRT/IRT-like protein 14 or ZRT/IRT-like protein 8 and lipid carrier protein 2 can mediate the entry of free iron into the cell[22,23]. A portion of intracellular iron is stored in the cytoplasmic iron-unstable pools and is involved in several biochemical processes. Specifically, it serves as the active site of proteins such as ribonucleotide reductase, participates in the catalytic conversion of ribonucleotides to deoxyribonucleotides, and is used in the synthesis of heme and iron-sulfur clusters. When intracellular iron exceeds metabolic demands, excess iron is stored in ferritin - a 24-subunit protein capable of sequestering up to 4500 iron atoms in its hydrotalcite-like core - regulated by IRP1 and IRP2. IRPs increase labile iron levels by repressing ferritin expression and upregulating TfR1[20]. Iron transporter protein (FPN) is the only channel protein capable of transporting Fe²⁺ to the extracellular compartment. FPN is mainly regulated by FPN, which binds FPN in the central cavity between the N and C structural domains, thereby closing the outwardly open FPN and inhibiting the exocytosis of Fe²⁺. Iron itself greatly enhances FPN’s binding to FPN[24]. Fe²⁺ is the main source of ROS production in the cell. Indeed, through Fenton reaction and Haber-Weiss reaction, Fe²⁺ generates a large amount of ROS, leading to the accumulation of peroxides, which ultimately triggers iron death[25]. Ferritin autophagy is also an important way to increase intracellular free Fe²⁺. When the autophagy receptor, nuclear receptor coactivator 4 (NCOA4), is present, ferritin binds to autophagosomes to increase intracellular free Fe²⁺ levels[26], which ultimately leads to the development of iron death. In the human body, a variety of factors such as excess heme and non-heme iron, exogenous supplementation of iron, increased cellular instability of iron, and iron-containing heme enzymes such as lipoxygenase (LOX) and cytochrome P450 may lead to iron accumulation[27]. Accumulated intracellular iron enhances oxidative damage by directly generating excess ROS via the Fenton reaction. Furthermore, iron can upregulate the arachidonic acid lipoxygenases (ALOXs) or EGLN prolyl hydroxylases, which regulate lipid peroxidation and oxygen homeostasis[28]. In summary, intracellular Fe²⁺ levels are a prerequisite for ferroptosis initiation, and factors governing iron homeostasis represent potential targets for regulating ferroptosis. Effective modulation of ferroptosis can be achieved by targeting molecules involved in iron influx and efflux, such as TfR1, FPN, and ferritin.

Lipid metabolism

Ferroptosis is closely linked to imbalances in lipid metabolism, with the formation of lipid peroxides representing the hallmark event that triggers ferroptosis initiation. Lipid metabolism in the context of ferroptosis raises two key questions: Where within the cell lipid peroxidation occurs, and which lipids undergo peroxidation. Currently, the precise subcellular location of lipid peroxidation during ferroptosis remains unclear, though studies suggest involvement of the plasma membrane, endoplasmic reticulum, and nuclear membrane[29]. Both free polyunsaturated fatty acids (PUFAs) and PUFA-containing membrane phospholipids [e.g., phosphatidylethanolamine (PE), phosphatidylcholine, and cardiolipin] can undergo peroxidation[30]. PUFA peroxidation is initiated enzymatically and nonenzymatically in cells. Nonenzymatic peroxidation of lipids is mainly triggered by the generation of hydroxyl and peroxyl radicals from iron via Fenton reaction, whereas peroxidation in an enzymatic manner is mainly driven by LOX[31]. LOX is a heme-iron-free dioxygenase that promotes dioxidation of PUFAs that contain at least two isolated cis-double bonds. It has different isoforms (5-LOX, 12S-LOX, 12R-LOX, 15-LOX-1, 15-LOX-2, and eLOX3)[31]. PUFAs are good substrates for autoxidation due to the presence of readily extractable bisallyl hydrogen atoms, which makes them sensitive to lipid peroxidation[32]. Their oxidation is divided into three main processes, and ALOXs, acyl coenzyme A synthetase long-chain family 4 (ACSL4), and lysophosphatidylcholine acyltransferase 3 are important regulatory enzymes in these processes. Taking arachidonic acid and PE as an example: First, ACSL4 catalyzes the formation of arachidonoyl-CoA; then, lysophosphatidylcholine acyltransferase 3 controls arachidonoyl-CoA esterification to arachidonoyl-phosphatidylethanolamine; and finally, arachidonoyl-phosphatidylethanolamine is oxidized to AA-OOH-PE by ALOXs[30]. This is a series of processes known as the high spontaneous autoxidative nature of unsaturated fatty acids, and the occurrence of this series of enzymatic reactions determines the susceptibility to iron death. Among the different LOX isoforms, 15-LOX-induced oxidation is selective and specific, occurring preferentially in arachidonoyl-phosphatidylethanolamine or adrenergic acid-PE[31]. However, expression of the ALOX gene is barely detectable in some of the cell lines commonly used in iron death studies, suggesting that the role of ALOX family members in iron death is limited, possibly only in tissues expressing ALOX[31]. In addition, Zou et al[32] demonstrated that cytochrome P450 oxidoreductase promotes ferroptosis by enhancing the peroxidation of membrane polyunsaturated phospholipids.

Given that the process of lipid peroxidation plays a key role in iron death, several drugs, such as ferrostatin, liproxstatins, deuterated PUFAs, coenzyme Q10, vitamin E, α-tocopherol, and Trolox[33], have been used to block lipid peroxidation, to reduce the occurrence of related diseases. The end result of lipid peroxidation is an increase in the level of malondialdehyde (MDA), which in vitro affects the activity of the mitochondrial respiratory chain complex and key enzymes in mitochondria, and its production can also exacerbate membrane damage. Therefore, the measured amount of MDA reflects the degree of lipid peroxidation in vivo and thus indirectly reflects the degree of cellular damage. In conclusion, signaling pathways that directly or indirectly regulate lipid metabolism through the regulation of lipid peroxidation-related pathways play an important role in iron death. Recent advances in understanding lipid metabolism and ferroptosis signaling pathways in cancer have identified several key regulatory pathways that influence ferroptosis sensitivity. These include the phosphatidylinositol 3-kinase/protein kinase B/mammalian target of the rapamycin pathway, LKB1/AMP-activated protein kinase pathway, E-cadherin-Hippo-YAP/TAZ pathway, and von Hippel-Lindau/hypoxia-inducible factor pathway[31]. Investigating these pathways holds promise for identifying novel therapeutic targets to advance targeted cancer therapies.

Antioxidant system

In normal cells, oxidative and antioxidant capacities maintain dynamic balance; however, when oxidant production increases and antioxidant capacity is impaired, cells succumb to oxidative damage and die. Three major antioxidant pathways regulate ferroptosis: The cystine/glutamate antiporter [System X(c)-], glutathione peroxidase 4 (GPX4) and glutathione (GSH) axis, the ferroptosis suppressor protein 1 (FSP1)-CoQ10-NAD(P)H axis, and the GTP cyclohydrolase 1 (GCH1)-dihydrofolate reductase (DHFR)-BH4 axis. There are three main antioxidant pathways in iron death, namely the System Xc-/GSH/GPX axis, the FSP1-CoQ10-NAD(P)H axis, and the GCH1-DHFR-BH4 axis, which work by detoxifying lipid peroxides, thereby preventing their accumulation to lethal levels and maintaining cell survival[34]. The System Xc-/GSH/GPX4 axis is the most classical regulatory pathway and constitutes the main defense system against iron death System Xc- (cystine/glutamate reverse transporter protein), a sodium-dependent channel, consists of a disulfide bond between a heavy chain (4F2hc, gene name SLC3A2) and a light chain (xCT, gene name SLC7A11) connected heterodimers. This transporter mediates the uptake of extracellular cystine into cells, which is subsequently converted to cysteine. Cysteine is then utilized by GSH synthetase to synthesize GSH[35]. GSH is an important intracellular antioxidant that plays a central role in the activity of GPX4 and is an important hydroperoxide reduction reaction substrate[36]. GPX4 is one of the 25 selenoproteins containing selenocysteine with potent antioxidant activity[37]. There are three isoforms of GPX4, namely mGPX4, nGPX4, and cGPX4, of which cGPX4 is widely present in cells, where it reduces cell death by resisting complex lipid peroxidation. However, it is not yet clear which isoforms should be targeted to ensure effective cancer treatment[38]. cGPX4 plays a crucial regulatory role in iron death, with a unique function to inhibit lipid peroxidation by catalyzing the conversion of lipid hydroperoxides and 2 GSH to nontoxic lipid alcohols, GSH disulfide, and water. It has recently been found that cGPX4 prevents ferroptosis by scavenging intracellular peroxides and maintaining cellular survival[39]. cGPX4 is widely present in cells, where it can reduce cell death by scavenging intracellular peroxides and maintaining cell survival. SLC7A11, GSH synthetase, and GPX4 are all transcriptional targets of Nrf2. Nfr2 is released from Keap1 and translocates to the nucleus under oxidative stress, where it subsequently interacts with antioxidant response elements to drive the expression of antioxidant genes[40]. Specifically, Nrf2 is regarded as an upstream signal of the pathway and therefore positively correlates with ferroptosis-related gene expression. Nrf2-induced HO-1 also affects GPX4 expression[41]. Nrf2 is also implicated in iron metabolism, and Anandhan et al[42] found that Nrf2 controls the expression of ferritin through the HECT and RCC1-like domain domains, which are contained in the E3 ubiquitin-containing ligase 2, vesicle-associated membrane protein 8, and NCOA4. The HECT and RCC1-like domain domains of E3 ubiquitin-containing ligase 2, vesicle-associated membrane protein 8, and NCOA4 control the synthesis and degradation of ferritin, thereby altering intracellular labile iron pool and determining the susceptibility of cancer cells to iron-induced cell death. Based on the importance of Nrf2 in regulating the process of iron death, a number of studies have shown that a variety of substances regulate iron death through the Nrf2-associated pathway. Tuximab enhances RAS-induced iron death by activating p38 mitogen-activated protein kinases and inhibiting the Nrf2/HO-1 axis[43]; irisin, naringenin, and quercetin inhibit iron death by activating the Nrf2-GPX4 signaling axis[44-46]; and acetylated melatonin A activates the Nrf2/HO-1 signaling pathway to promote iron death in breast cancer[47].

The FSP1-CoQ10-NAD(P)H axis reduces CoQ10 (also known as ubiquinone) to CoQ10H2 (ubiquinol) via myristoylated, plasma-membrane-localized FSP1 acting as a NADH/NADPH-dependent CoQ10 oxidoreductase that consumes NADH/NADPH[48]. CoQ10 is a lipophilic metabolite that consists of an oxidatively reductively active quinone head group and a long polyisoprene lipid tail, and serves as a key electron transfer carrier in the mitochondrial electron transport chain. It endogenously synthesizes a lipid-soluble antioxidant (CoQ10H2) and acts as a lipophilic free radical-trapping agent in the plasma membrane. Its fully reduced form, namely CoQ10H2, also acts as a free radical-trapping antioxidant for detoxification of lipid peroxyl radicals[49]. FSP1 can act as a vitamin K reductase by consuming NAD(P)H and reducing vitamin K to hydroquinone vitamin K, which acts as an antioxidant to trap free radicals and inhibit lipid peroxidation. Among the various forms of vitamin K, only chlorophyll quinone, menaquinone-4, and menaquinone are effective in inhibiting lipid peroxidation in cells[50]. Pharmacological co-inhibition of FSP1 and GPX4 may represent an effective strategy to sensitize cancer cells - particularly those refractory to GPX4 inhibitors alone - to ferroptosis-inducing chemotherapeutic agents[51].

The GCH1-DHFR-BH4 axis is another GPX4-independent iron-death-regulated signaling axis. It was originally identified by Kraft et al[52], who performed a genome-wide activation screen for a group of genes that inhibit iron death by selectively blocking phospholipid depletion of the two polyunsaturated fatty acyl tails. It primarily consists of GCH1 and its metabolic derivatives, including tetrahydrobiopterin (BH4) and dihydrobiopterin (BH2)[52]. BH4 acts as a cofactor for monooxygenases, i.e., aromatic amino acid hydroxylase and nitric oxide synthase. It is oxidized and forms pterin 4a-methanolamine intermediates during aromatic amino acid hydroxylase-catalyzed reactions and serves as a single-electron donor for the activation and reduction of oxygen during nitric oxide synthase-catalyzed processes[53]. Acting as a redox-active cofactor, BH4 is also involved in the production of nitric oxide, neurotransmitters, and aromatic amino acids[51], and can synergize with vitamin E to reduce the accumulation of ROS and inhibit the oxidative degradation of phospholipids containing PUFA chains[54]. GCH1 is the first enzyme in the BH4 synthetic pathway and is also the most important rate-limiting enzyme that catalyzes the conversion of GTP into 7,8-dihydroneopterin triphosphate (NH2TP) and finally BH4 in response to PTPS and SR. BH2 is the dehydrogenation product of BH4. Soula et al[54] showed experimentally that supplementation of GCH1-knockdown cells with BH2 restored endogenous levels of BH4 and eliminated the sensitivity to RSL3 and ML210, inhibitors of GPX4. DHFR catalyzes the regeneration of BH4, which specifically reduces lipid peroxidation and inhibits iron death or attenuates cellular susceptibility to iron death[55]. BH4 also inhibits the action of DHFR to a certain extent, and the two regulate each other, as well as methotrexate, an inhibitor of DHFR, which has been shown to induce iron death[54]. Within the GCH1-DHFR-BH4 axis, GCH1 and BH4 protect cells from lipid peroxidation during ferroptosis. This pathway operates in parallel with the System Xc-/GSH/GPX axis and the FSP1-CoQ10-NAD(P)H pathway to regulate ferroptosis[56]. The detailed mechanism of ferroptosis is illustrated in Figure 2.

Figure 2 Diagram of the mechanism of development of ferroptosis. The key molecular mechanisms of ferroptosis have been systematically investigated, with a focus on understanding the pathways and mechanisms of iron-related processes. Ferritin, the primary storage protein for iron in the body, plays a central role in iron metabolism. Iron is primarily acquired in the bloodstream as Fe³⁺, which enters cells through the transferrin receptor 1 (TFR1) on the cell surface. The expression of TFR1 is regulated by iron regulatory protein and hypoxia-inducible factor-1, while heat shock protein B1 inhibits TFR1 expression. Fe³⁺ is reduced to Fe²⁺ by the lipid carrier protein 2, which allows free iron to enter the cytoplasmic iron-unstable pools. Additionally, ZRT/IRT-like proteins such as ZRT/IRT-like protein 14 can also mediate iron entry into the cell. Ferritin, the sole channel protein for Fe²⁺ export, is primarily regulated by hemosiderin, and ferritin autophagy is a critical mechanism for maintaining intracellular free Fe²⁺ levels. When autophagy receptor nuclear coactivator 4 is present, ferritin binds to autophagosomes, increasing intracellular free Fe²⁺ levels. Free iron in the cytoplasmic iron-unstable pools generates reactive oxygen species through the Fenton reaction and Haber-Weiss reaction. Enzymatic pathways, such as long-chain family 4 catalyzing arachidonoyl-CoA formation, lysophosphatidylcholine acyltransferase 3 controlling the esterification of arachidonoyl-CoA to arachidonoyl-phosphatidylethanolamine, and arachidonic acid lipoxygenases oxidizing arachidonoyl-phosphatidylethanolamine to AA-OOH-PE, contribute to the generation of reactive oxygen species. The primary antioxidant pathways in ferroptosis include the cystine/glutamate antiporter/glutathione peroxidase 4/glutathione axis, the ferroptosis suppressor protein 1-CoQ10-NAD(P)H axis, and the GTP cyclohydrolase 1-dihydrofolate reductase-BH4 axis. Notably, the ferroptosis suppressor protein 1-CoQ10-NAD(P)H axis and the GTP cyclohydrolase 1-dihydrofolate reductase-BH4 axis function to inhibit fatty acid synthesis, thereby contributing to the oxidative stress response in ferroptosis. Tf: Transferrin; IRP: Iron regulatory protein; HIF1: Hypoxia-inducible factor 1; HSPB1: Heat shock protein B1; TFR1: Transferrin receptor 1; LCN2: Lipid carrier protein 2; GSS: Glutathione synthetase; ZIP14: ZRT/IRT-like protein 14; GSH: Glutathione; GSSG: Glutathione disulfide; FSP1: Ferroptosis suppressor protein 1; GPX4: Glutathione peroxidase 4; FPN: Ferroportin; ROS: Reactive oxygen species; AA: Arachidonic acid; AA-CoA: Arachidonoyl-CoA; AA-PE: Arachidonoyl-phosphatidylethanolamine; ACSL4: Acyl coenzyme A synthetase long-chain family 4; LPCAT3: Lyso-phosphatidylcholine acyltransferase-3; ALOX5: Arachidonic acid lipoxygenases 5; LPO: Lipid peroxidation; DHFR: Dihydrofolate reductase; GCH1: GTP cyclohydrolase 1.

FERROPTOSIS MECHANISIM IN HCC

A growing body of research has highlighted a close association between ferroptosis and tumors. Hao et al[57] conducted deep scRNA-sequencing analysis on 212494 immune cells from peripheral blood, tumor tissue, and adjacent normal tissue in four HCC cases and two non-cancer control cases. They found that inhibiting APOC1 promoted the conversion of M2 macrophages into M1 macrophages via the ferroptosis pathway, thereby reshaping the tumor immune microenvironment and improving the efficacy of anti-programmed cell death protein 1 immunotherapy in HCC. Wang et al[58] used least absolute shrinkage and selection operator and single-factor logistic regression analysis to screen for differentially expressed death-associated genes (CDRGs) and identified seven key genes closely associated with ferroptosis (a form of programmed cell death). They further demonstrated that elevated levels of TRIB3 and NQO1 in blood-derived exosomes could serve as promising diagnostic biomarkers for HCC and predict the response to immunotherapy in HCC patients. Collectively, these findings underscore a strong correlation between ferroptosis and liver cancer.

Ferroptosis in HCC treatment

Numerous studies have demonstrated that various therapeutic agents for liver cancer exert their anti-tumor effects through ferroptosis-related mechanisms, particularly in chemotherapy and anti-tumor immunotherapy. Chemotherapy drugs such as sorafenib and lenvatinib have been reported to have therapeutic effects closely related to iron death mechanisms and are already used as first-line treatment options. In anti-tumor immunotherapy, the combination of FSP1 inhibitors with programmed death ligand-1 further inhibits the progression of liver cancer in mice[59]. Currently, research on ferroptosis in liver cancer remains in its exploratory phase. Gaining a deeper understanding of the ferroptosis mechanisms in liver cancer will significantly advance its treatment.

Ferroptosis in the treatment of HCC with sorafenib: Ferroptosis is closely associated with the proliferation and progression of HCC cells, and it provides a novel inhibitory mechanism for the prevention and treatment of HCC. Current studies in HCC focus on sorafenib-induced iron death and its potential targets[60]. Sorafenib is a novel multitargeted signaling inhibitor that inhibits Raf kinase (Raf-MEK), VEGF, and platelet-derived growth factor receptor (tansforming growth factor-β). It also inhibits tumor cell proliferation by blocking the Raf-MEK-ERK signaling pathway. Notably, sorafenib exerts dual inhibitory effects and multi-targeted blocking actions against primary liver cancer. Recently, Liu et al[61] have experimentally shown that sorafenib can enhance the ubiquitination of FSP1 and induce iron death through the ERK kinase pathway in HCC cells. A large number of studies have shown that iron death is closely related to sorafenib resistance. The signaling pathways of HCC resistance to sorafenib include the Hippo-YAP, LIFR-NAP, and LIFR-NAP signaling axis, the LIFR-nuclear factor kappa B-lipid carrier protein 2 pathway, the Keap1-Nrf2 system, and the ETS1-miR-23a-3p-ACSL4 axis[62], which also suggests that HCC resistance to sorafenib can be reversed by regulating these signaling axes. Relevant studies have shown that the combination of sodium butyrate and sorafenib can alleviate the resistance of HCC to sorafenib by reducing the YEP expression, attenuate the resistance of HCC to sorafenib, reduce the GSH level of HCC cells by inducing YEP phosphorylation, and increase the lipid ROS content, which promotes the iron death of HCC cells. Additionally, metformin and quiescin sulfhydryl oxidase 1 can synergize with sorafenib to enhance the downregulation of Nrf2, counteracting sorafenib resistance in HCC[63,64]. Amiloride reduces cysteine acquisition by inhibiting megacytosis, thereby limiting the development of sorafenib resistance and improving the therapeutic efficacy of sorafenib in HCC[65]. The mechanism of ferroptosis by sorafenib is shown in Figure 3.

Figure 3 Effects of sorafenib, metformin and quiescin sulfhydryl oxidase 1 on ferroptosis. Sorafenib enhances the ubiquitination of ferroptosis suppressor protein 1 through TRIM54, leading to the induction of ferroptosis by decreasing CoQH2. Furthermore, quiescin sulfhydryl oxidase 1 induces ferroptosis by inhibiting epidermal growth factor receptor, which subsequently inhibits nuclear factor erythroid 2-related factor 2. Metformin induces ferroptosis by inhibiting the P62/Kelch-like ECH-associated protein 1/nuclear factor erythroid 2-related factor 2/heme oxygenase 1 pathway. FSP1: Ferroptosis suppressor protein 1; Keap1: Kelch-like ECH-associated protein 1; Nrf2: Nuclear factor erythroid 2-related factor 2; EGFR: Epidermal growth factor receptor; HO-1: Heme oxygenase 1.

Ferroptosis-related enzyme in HCC

By detecting the expression levels of relevant enzymes in liver cancer tissues and paracancerous tissues, it has been shown that a large number of enzymes are abnormally metabolized in HCC. The in-depth exploration of the abnormal metabolism of HCC-related enzymes has also become a research hot spot[66]. To date, experimental studies have identified multiple enzymes with aberrant expression in HCC tissues, which exert regulatory effects on the ferroptosis process of HCC cells. These enzymes include: Phosphoglycerate mutase 1[66], nicotinamide N-methyltransferase, flavin-containing monooxygenase 3, lysine-specific demethylase 4C, 7-dehydrocholesterol reductase[67], ubiquitin-specific peptidase 8[68,69], protein arginine methyltransferase 9[70], HMOX1[71], alpha-enolase[72], fatty acid synthase[73], apurinic/apyrimidinic endonuclease 1[74], GSH S-transferase zeta 1[11], tRNA kinase[75], quiescin sulfhydryl oxidase 1[63], RNA deconjugating enzyme DDX5[76], dual-specificity phosphatase 4[77], MER proto-oncogene tyrosine kinase[78], ubiquitin C terminal hydrolase L3[79], aldo-keto reductase 1C3[80], deubiquitinating enzyme EIF3H[81], ubiquitin-specific protease 24[82], and histidine phosphatase[83]. The role of relevant enzymes in the mechanism of liver cancer on ferroptosis is shown in Figure 4 and Table 2 below.

Figure 4 Effect of related enzymes on iron death in hepatocellular carcinoma cells. Phosphoglycerate mutase 1 inhibits ferroptosis through an energy stress/reactive oxygen species (ROS)-dependent protein kinase B (AKT)/Lipid carrier protein 2 pathway; nicotinamide methyltransferase inhibits ferroptosis via the multifunctional microneedle array/ROS axis; lysine-specific demethylase 4 inhibits ferroptosis via the murine double minute 2/P53/SLC7A11/glutathione peroxidase 4 (GPX4) axis; 7-dehydrocholesterol reductase inhibits ferroptosis through the TMEM147/signal transducer and activator of transcription 2/7-dehydrocholesterol reductase/27HC axis; ubiquitin-specific peptidase 8 (USP8) inhibits ferroptosis through the USP8/O-GlcNAc transferase/SLC7A11 signaling pathway and the Wnt/β-catenin/GPX4 axis; protein arginine methyltransferase 9 inhibits ferroptosis through the hepatitis B virus X protein/protein arginine methyltransferase 9/heat shock protein A8/CD44 axis; α-enolase 1 inhibits ferroptosis through the α-enolase 1-iron regulatory protein 1-Mfrn1 pathway; apurinic/apyrimidinic endonuclease 1 inhibits ferroptosis through the AKT/glycogen synthase kinase-3β/nuclear factor erythroid 2-related factor 2 (Nrf2)/SLC7A11/GPX4 axis; phospho-serine acyl-tRNA kinase inhibits ferroptosis through the thioredoxin reductase/thioredoxin/glutathione/GPX4 pathway; inhibits ferroptosis; dual-specific phosphatase 4 inhibits ferroptosis via the YTHDC1/FTL/ferritin heavy chain 1/Lipid ROS pathway; MER proto-oncogene tyrosine kinase inhibits ferroptosis via the extracellular signal-regulated kinase/SP1/SLC7A11 pathway; ubiquitin C-terminal hydrolase L3 inhibits ferroptosis via the Ub-proteasome/Wnt/β-catenin/GPX4 pathway; aldehyde reductase 1C3 inhibits ferroptosis via the Hippo/YAP/TAZ/SLC7A11 pathway to inhibit ferroptosis; the deubiquitinating enzyme EIF3H inhibits ferroptosis via the O-GlcNAc transferase/Lipid ROS/glutathione pathway; the USP24 promotes ferroptosis via the K48/Beclin1/ferroportin pathway promotes ferroptosis; lysine phosphatase promotes ferroptosis via the lysine phosphatase/phosphatidylinositol-3-kinase/AKT pathway; quiescin sulfhydryl oxidase 1 promotes ferroptosis via the epidermal growth factor receptor/Nrf2/GPX4 pathway; RNA helicase DDX5 promotes ferroptosis via the Wnt/β-catenin/GPX4 signaling pathway to promote ferroptosis; glutathione S-transferase zeta 1 promotes ferroptosis via the Kelch-like ECH-associated protein 1/Nrf2/GPX4 pathway; fatty acid synthase promotes ferroptosis via the fatty acid synthase/hypoxia-inducible factor-1α/SLC7A11 pathway; heme oxygenase 1 promotes ferroptosis through the Nrf2/heme oxygenase 1/GPX4 axis; flavin monooxygenase 3 promotes ferroptosis through the flavin monooxygenase 3/trimethylamine N-oxide/signal transducer and activator of transcription 3/SLC7A11/GPX4 axis. LHPP: Lysine phosphatase; EGFR: Epidermal growth factor receptor; QSOX1: Quiescin sulfhydryl oxidase 1; Nrf2: Nuclear factor erythroid 2-related factor 2; PI3K: Phosphoinositol-3-kinase; AKT: Protein kinase B; STAT2: Signal transducer and activator of transcription 2; GSTZ1: Glutathione S-transferase zeta 1; PSTK: Phospho-serine acyl-tRNA kinase; APE1: Apurinic/apyrimidinic endonuclease 1; MerTK: MER proto-oncogene tyrosine kinase; HMOX1: Heme oxygenase 1; CEs: Cholesterol esters; DHCR7: 7-dehydrocholesterol reductase; GPX4: Glutathione peroxidase 4; LCN2: Lipid carrier protein 2; ERK: Extracellular signal-regulated kinase; ROS: Reactive oxygen species; PRMT9: Protein arginine methyltransferase 9; USP8: Ubiquitin-specific peptidase 8; ENO1: Α-enolase 1; HSPA8: Heat shock protein A8; OGT: O-GlcNAc transferase; FASN: Fatty acid synthase; HIFα: Hypoxia-inducible factor-α; PGAM1: Phosphoglycerate mutase-1; IRP1: Iron regulatory protein 1.

Table 2 Enzymes involved in the process of hepatocellular carcinoma and their role in ferroptosis.

Related enzyme	Relevant impacts	Ferroptosis
PGAM1	Energy stress/ROS-dependent AKT/LCN2 pathway	Inhibitory
NNMT	MNA/ROS axis	Inhibitory
FMO3	FMO3/TMAO/STAT3/SLC7A11/GPX4 axis	Promotion
KDM4C	MDM2/P53/SLC7A11/GPX4 axis	Inhibitory
DHCR7	TMEM147/STAT2/DHCR7/27HC axis	Inhibitory
USP8	USP8/OGT/SLC7A11 signaling pathway, Wnt/β-catenin/GPX4 axis	Inhibitory
PRMT9	HBx/PRMT9/HSPA8/CD44 axis	Inhibitory
HMOX1	Nrf2/HO-1/GPX4 axis	Promotion
ENO1	ENO1-IRP1-Mfrn1 pass-through	Inhibitory
FASN	FASN/HIF1α/SLC7A11 pass-through	Promotion
APE1	AKT/GSK3β/Nrf2/SLC7A11/GPX4 axis	Inhibitory
GSTZ1	Keap1/Nrf2/GPX4 pathway	Promotion
PSTK	TrxR/Trx/GSH/GPX4 pathway	Inhibitory
QSOX1	EGFR/Nrf2/GPX4 pathway	Promotion
RNA helicase DDX5	Wnt/β-catenin/GPX4 signaling pathway	Promotion
DUSP4	YTHDC1/FTL/FTH1/Lipid ROS pathway	Inhibitory
MerTK	ERK/SP1/SLC7A11 pathway	Inhibitory
UCHL3	Ub-proteasome/Wnt/β-catenin/GPX4 pathway	Inhibitory
AKR1C3	Hippo/YAP/TAZ/SLC7A11 pathway	Inhibitory
Deubiquitinating enzyme (EIF3H)	OGT/Lipid ROS/GSH pathway	Inhibitory
USP24	K48/Beclin1/FPN pathway	Promotion
LHPP	LHPP/P13K/AKT pathway	Promotion

PGAM1: Phosphoglycerate translocase 1; ROS: Reactive oxygen species; AKT: Protein kinase B; LCN2: Lipid carrier protein 2; NNMT: Nicotinamide transmethylase; MNA: Multifunctional microneedle array; FMO3: Flavin monooxygenase 3; TMAO: Trimethylamine N-oxide; STAT3: Signal transducer and activator of transcription 3; GPX4: Glutathione peroxidase 4; KDM4C: Lysine specific demethylase 4; MDM2: Murine double minute 2; DHCR7: 7-dehydrocholesterol reductase; USP8: Ubiquitin specific peptidase 8; OGT: O-GlcNAc transferase; PRMT9: Arginine methyltransferase 9; HSPA8: Heat shock protein A8; HMOX1: Hemoglobin oxygenase 1; Nrf2: Nuclear factor erythroid 2-related factor 2; HO-1: Heme oxygenase 1; ENO1: Α-enolase 1; IRP1: Iron regulatory protein 1; FASN: Fatty acid synthase; HIF1α: Hypoxia-inducible factor 1-alpha; APE1: Apurinic/apyrimidinic endonuclease 1; GSK3β: Glycogen synthase kinase-3β; GSTZ1: Glutathione S-transferase zeta 1; Keap1: Kelch-like ECH-associated protein 1; PSTK: Phosphoseryl-tRNA kinase; TrxR: Thioredoxin reductase; Trx: Thioredoxin; GSH: Glutathione; QSOX1: Quiescin sulfhydryl oxidase 1; EGFR: Epidermal growth factor receptor; FTH1: Ferritin heavy chain 1; DUSP4: Dual-specific phosphatase 4; MerTK: MER proto-oncogene tyrosine kinase; ERK: Extracellular signal-regulated kinase; UCHL3: Ubiquitin C terminal hydrolase L3; AKR1C3: Aldehyde-keto reductase 1C3; USP24: Ubiquitin-specific proteases 24; LHPP: Lysine phosphatase; P13K: Phosphatidylinositol-3-kinase.

AI-BASED PREDICTION MODELS FOR LIVER CANCER

In recent years, with the rapid advancement of AI technology, the construction of liver cancer prediction models using big data analysis and pattern recognition has emerged as a research hotspot. AI techniques primarily encompass machine learning (ML) and deep learning (DL). ML consists of computers that run models with repeated iterations to gradually improve the performance of a specific task, such as categorizing the results. DL modeling is a subtype of ML based on a neural network structure inspired by the neuroanatomy of the human brain[84]. Both AI techniques provide strong support for early diagnosis, treatment decisions, and prognosis of liver cancer. While classical ML techniques have been applied in the medical field for decades and rely mainly on researcher-defined features, DL has become dominant in liver cancer research. Compared to ML, DL incorporates thousands of additional free parameters, enabling it to handle complex data more effectively[85]. The construction of AI-based prediction models for liver cancer mainly involves seven steps: Data collection, data preprocessing, feature extraction and selection, model selection and training, model validation and evaluation, model optimization and adjustment, external validation, and multicenter study. A variety of AI-based prediction models for liver cancer have been constructed, such as a diagnosis prediction model based on tumor markers, early diagnosis models for HCC based on computed tomography (CT) or magnetic resonance imaging images, prediction models for the risk of HCC occurrence based on immune indicators, and prediction models for various posttreatment complications in HCC patients. There have been various reviews of relevant liver cancer prediction models. Bo et al[86] reviewed AI-based radiology in liver cancer diagnosis, individualized treatment, and survival prognosis, and emphasized the potential biological significance of radiomics features. Calderaro et al[84] reviewed relevant AI-based prediction models for liver cancer in terms of radiology, histopathology, and biomarkers. Wang et al[87] analyzed and reviewed a large number of AI prediction models for liver cancer published in or before 2023, highlighting the great potential of AI in liver cancer diagnosis, treatment, and prognosis. However, most previous reviews on AI prediction models for liver cancer suffer from limited timeliness and relatively narrow focuses. Given the rapid development of AI technology in recent years and its significant role in liver cancer prediction, this study summarizes the latest research in this field, describes the potential challenges of applying AI in liver cancer management, and proposes future development directions. Specifically, we reviewed articles published in the past five years on the construction of AI prediction models for liver cancer and summarized their findings. A schematic diagram of the model construction process is presented in Figure 5, and representative AI prediction models developed in recent years are listed in Table 3[88-117].

Figure 5 Liver cancer prediction models and associated databases are primarily constructed based on three directions: Images, tissue pathology grading, and protein markers. Commonly used algorithms include deep neural network, gradient boosting machine, deep convolutional neural network, you only look once, logistic regression, multilayer perceptron, and compressed linear approximation model. During model training, data are mainly drawn from eight sources: Genetic data, pathology slides, radiology images, laboratory tests, temporal data, medical history, clinical data, and ferroptosis-related proteins. LR: Logistic regression; YOLO: You only look once; DCNN: Deep convolutional neural network; GBM: Gradient boosting machine; DNN: Deep neural network; CLAM: Compressed linear approximation model; MLP: Multilayer perceptron; AI: Artificial intelligence.

Table 3 A review related to artificial intelligence prediction models for liver cancer.

Ref.	Model name	Model type	Aim	Dataset	Model checking	Major limitation
[88]	Hybrid model	Logistic regression	MVI	Total: 773. Training set: 334. Internal test set: 142. External test set: 141	Internal test set: AUC = 0.86. External test set: AUC = 0.84	Selection bias; most enrolled patients had a virus-related HCC
[89]	DECT	Logistic regression	MTM	Total: 262. Training set: 146. Internal test set: 35. External test set: 81	Internal test set: AUC = 0.87. External test set: AUC = 0.89	Different centers; overfitting
[90]	ABRS-P	CLAM	Biomarker of sensitivity to atezolizumab-bevacizumab	Total: 122. ABRS-P-high: 74. ABRS-P-low: 48		Retrospective
[91]	AI-based pathology models	CLAM	Predict the activation of 6 immune gene	Total: 336	AUC: 0.78-0.91	Needs further validation with clinical data
[92]	PLAN-B-DF	GBM	HCC prediction	Training set: 4188. Internal test set: 1397. External test set: 2883	CI: 0.91	Impose radiation exposure; generalizability limited
[93]	AI prediction model for liver cancer recurrence	MLP	Evaluate the survival of patients with HCC	Total: 912	AUC: 0.862	Pack of longitudinal data
[94]	UBE2S related model	MLP	Evaluate the survival of patients with HCC	Total: 370. Training set: 224. Test set: 146	The 1-, 2-, 3-year survival AUC values were 076, 0.72, 0.68
[95]	PAGE-B	Cox regression	Risk prediction tool	Total: 2963	CI: 0.77	The lack of systematic screening information on HDV coinfection
[96]	SCHMOWDER	SCHMOWDER	Predicting patient survival after HCC recurrence and surgery	The transplant cohort: 300. The resection cohort: 169	The transplant cohort: CI = 0.83 (RFS); CI = 0.87 (DSS). The resection cohort: CI = 0.64 (RFS); CI = 0.77 (DSS)
[97]	MoRAL-AI	DNN	Recurrence	Total: 563. Training set: 349. Test set: 214	CI: 0.75	Single-region HBV cohort
[98]	R3-AFP	Logistic	Validation	Total: 508	CI: 0.75	Restricted to SiLVER trial
[99]	Reticulin-CNN	CNN	Prognosis	Total: 105	CI > 0.7	Small sample; missing HBV/HCV data
[100]	IR-lncRNA	Cox/LASSO	Recurrence	Total: 319. Training set: 161. Test set: 158	CI: 0.732	Needs external validation
[101]	CT-DCNN	DCNN	Diagnosis	Training set: 7512. Internal test set: 385. External test set: 556	The internal test set: AUC = 0.887. The external test set: AUC = 0.883	Central China predominance
[102]	TIL score	Cox	Quantify	Training set: 124. Test set 1: 82. Test set 2: 54	Training set: CI = 0.770. Test set 1: CI = 0.769. Test set 2: CI = 0.712	Slide alignment and selection bias
[103]	DSFR	Logistic regression	Predict early recurrence	Total: 208. Training set: 180. Test set: 28	Training set: AUC = 0.782. Test set: AUC = 0.744	Small sample
[104]	Aiforia	AI-based histological model	Histological outcome	Total: 101	rs = 0.72
[105]	LDA	Discriminant analysis	MVI	Total: 140. Training set: 98. Test set: 42	Training set: AUC = 0.995. Test set: AUC = 0.913	Small size; 2-center variability
[106]	Random forest model	Random forest model	Waitlist Dropout	Total: 15444	C-statistic: 0.74	Retrospective; not externally validated
[107]	Multiple ML models	ML (DT, SVM, NN, etc.)	Survival	Total: 393	Early-stage: Recall = 91% (6 months). Advanced: Accuracy = 92% (3 years)	Small, single center, retrospective
[108]	CNN	CNN	Predicting the outcome of ICIs treatment	Training set: PD = 197; PR = 271; SD = 342	F1 score 698%
	YOLO	YOLO
[109]	Cox PH model	Cox regression	Risk	Total: 790	AUC = 0.86	Retrospective; small sample
[110]	CART	Decision tree	HCC	Training set: 55	AUC: 0.950	Small sample; no validation
[111]	nVR	Radiomics	Recurrence	Training set: 130. Test set: 57	Training set: AUC = 0.759. Test set: AUC = 0.765	Retrospective; TACE method bias
[112]	3D-ResNet50	Deep learning	Grade	Total: 858. Training set: 524. Validation set: 131. External test set: 65. Internal temporal test set: 138	Training set: AUC = 0.82. Validation set: AUC = 0.825. External test set: AUC = 0.78. Internal temporal test set: AUC = 0.81	Retrospective; single-phase imaging
[113]	DCNN	Deep learning	Prognosis	Total: 236	Training set: CI = 0.735 (RFS); CI = 0.712 (OS). Test set: CI = 0.718 (RFS); CI = 0.740 (OS)	Single-center; small sample
[114]	MAPL-5	ML	HCC	Training set: 1182. External test set: 562	Training set: AUC = 0.784; balanced accuracy = 0.712. External test set: AUC = 0.862; balanced accuracy = 0.771	Needs external validation
[115]	LightGBM	Radiomics	Biomarker	Training set: 424. Test set: 102	Training set: AUC = 0.866. Test set: AUC = 0.824	Small sample; no multicenter
[116]	Random forest	Radiomics	Grade	Training set: 137. External test set: 28	Training set: AUC = 0.80. External test set: AUC = 0.70	Single-vendor MRI; limited external data
[117]	CRNN	Deep learning	Survival	Total: 207	Training set: 0.777. Test set: 0.704	Small sample; excluded non-lung metastases

MVI: Microvascular invasion; AUC: Area under the curve; HCC: Hepatocellular carcinoma; DECT: Dual-energy computed tomography; MTM: Macrotrabecular-massive; ABRS: Atezolizumab-bevacizumab response signature; CLAM: Clustering-constrained attention multiple-instance learning; AI: Artificial intelligence; GBM: Gradient boosting machine; CI: Confidence interval; MLP: Multilayer perceptron; RFS: Relapse-free survival; DSS: Disease-specific survival; DNN: Deep neural network; HBV: Hepatitis B virus; AFP: Alpha-fetoprotein; CNN: Convolutional neural network; HCV: Hepatitis C virus; LASSO: Least absolute shrinkage and selection operator; DCNN: Deep convolutional neural network; TIL: Tumor-infiltrating lymphocyte; DSFR: Deep semantic segmentation feature-based radiomics; rs: R²d; LDA: Linear Discriminant Analysis; ML: Machine learning; DT: Decision tree; SVM: Support vector machine; NN: Neural network; YOLO: You only look once; ICIs: Immune checkpoint inhibitors; PD: Progressive disease; PR: Partial response; SD: Stable disease; CART: Classification and regression trees; TACE: Transarterial chemoembolization; OS: Overall survival; MRI: Magnetic resonance imaging.

Construction of traditional AI models

The roles of AI-driven models mainly include diagnosing uncertain liver nodules, determining eligibility for transarterial chemoembolization, assessing microvascular invasion (MVI) in HCC patients, predicting the risk of HCC recurrence, evaluating the effectiveness of HCC medication, identifying postoperative complications, and the pathological classification and diagnosis of HCC[86]. In the development of AI models for liver cancer, an increasing number of evaluation metrics are being integrated into training processes, thereby enhancing the sensitivity, specificity, and area under the curve (AUC) of predictions. For example, Xia et al[88] studied 918 HCC patients who underwent liver resection at four medical centers and the Taiwan Cancer Imaging Archive, integrating pathological diagnoses and existing clinical data. Patients underwent non-enhanced and contrast-enhanced abdominal CT scans within 8 weeks prior to liver resection. The model was constructed using preoperative multiphase CT images to predict the status of hepatic MVI. The AUC values for both the internal and external test sets were greater than 0.8, demonstrating good predictive performance. In addition, CT radiomic features such as pseudocapsule, two-trait predictor of venous invasion, and peritumoral enhancement were identified as important predictors of MVI. Piñero et al[98] analyzed 508 patients who were eligible for treatment by incorporating the tumor marker alpha-fetoprotein into the gold standard explant-based models for liver transplant recurrence. The R3-AFP model effectively stratified the risk of progression from chronic hepatitis B to HCC. They constructed the PLAN-B-DF model using imaging markers such as the abdominal visceral fat-to-total fat volume ratio, total fat-to-trunk volume ratio, spleen volume, liver volume, liver-to-spleen Hounsfield unit ratio, and muscle Hounsfield units. The c-index of the PLAN-B-DF model reached 0.91, indicating the potential for improving HCC risk stratification. Han et al[116] constructed four AI models based on hepatobiliary phase, portal venous phase, T1-weighted imaging, and T2-weighted imaging using 833 HCC radiomics features and gadolinium-enhanced magnetic resonance imaging to predict the pathological grade of HCC. Finally, through internal validation, they found that only the AUC of the hepatobiliary phase had a high value (above 0.8). However, the performance in external validation requires further improvement. Cheng et al[118] analyzed samples composed of surgical and biopsy specimens using four deep neural networks, namely ResNet50, InceptionV3, Xception, and Ensemble. The resulting stem cell nodule AI model (HnAIM) achieved an overall AUC of 0.935 in the external validation cohort and demonstrated high consistency with the opinions of pathologists and relevant subspecialty physicians.

Limitations of AI models

Current AI models still have certain limitations. For instance, most models are retrospective and lack the capability for dynamic risk assessment. They often fail to account for competing risk events (such as decompensated liver function and mortality), which may lead to inaccurate evaluations of the actual risk of HCC. Additionally, many models are developed based on small sample sizes, with subject selection frequently lacking representativeness. Moreover, sample processing protocols are insufficiently standardized across different research centers. Most AI models are still based on the classical ML and are single models. These limitations are also the main problems of the AI models in practice today. While some bias is inevitable, the following methods can help address these issues in the future. For example, a standardized and unified approach should be adopted for image processing; extensive external validations of AI models should be carried out in a timely manner; and appropriate AI models should be selected for training. The convolutional neural network model ResNet-18 has proven effective in predicting MVI through the preoperative CT images of HCC, while the YOLO model is more conducive to the identification and localization of the object in real time[108]. Some AI models were constructed by combining multiple data. For example, Zhou et al[105] compared the imaging histology model with a hybrid model of imaging histology and clinical features, which showed high accuracy and sensitivity in the training and test sets. Furthermore, incorporating novel tumor markers with high specificity and sensitivity into model training can further enhance performance.

AI model based on ferroptosis

At present, there are numerous methods for detecting iron death-related biomarkers, which greatly facilitate the development of AI models in this area. For example, liquid biopsy can detect lipid peroxidation products (such as MDA and 4-hydroxynonenal), iron metabolism markers (such as serum iron and ferritin), and inflammation-related factors, thereby assisting in assessing the occurrence of ferroptosis and monitoring treatment responses. In addition, by integrating multi-omics data, liquid biopsy can use ML and other analytical methods to stratify patients, identifying individuals who may be sensitive to ferroptosis therapy, thereby supporting precision cancer treatment[119]. Monitoring metabolites produced during ferroptosis via positron emission tomography imaging also offers new insights for noninvasive detection[120]. Multiplexed imaging techniques are modern counterparts to histological analyses and aim to detect a given set of cell types and their state based on target markers[121]. In addition, given the establishment of single-cell RNA sequencing methods[122], the development of AI models for HCC based on transcriptomics has become a promising field[123]. For example, noncoding RNAs regulating ferroptosis, including microRNAs, long noncoding RNAs, and circular RNAs, can serve as one of the analytical standards for AI model construction. Among these, microRNAs primarily regulate the ferroptosis process in HCC by targeting GPX4, ACLS4, and ferritin heavy chain 1. Long noncoding RNAs and circular RNAs primarily target SLC7A11 and GPX4 to regulate the ferroptosis process in HCC[124]. Chen et al[125] constructed an HCC prediction model using two ferroptosis-related mRNAs (SLC1A5 and SLC7A11) and eight ferroptosis-related long noncoding RNAs (AC245297.3, MYLK-AS1, NRAV, SREBF2-AS1, AL031985.3, ZFPM2-AS1, AC015908.3, and MSC-AS1) to predict HCC, which demonstrated a good AUC value. These RNAs were shown to outperform pathological features. In addition to noncoding RNAs, a variety of HCC cell drug resistance- and cell metabolism-related enzymes are embodied in iron death-related molecular pathways, and it may be worthwhile to try to adopt iron death-related markers to provide new training features for the AI model of HCC. Moreover, it may be useful to adopt a multi-data combination to improve the accuracy of the prediction model. Additionally, dynamic changes in ferroptosis markers can be used to effectively assess therapeutic efficacy and prognostic changes[126].

SUMMARY AND PROSPECTS

Ferroptosis, a novel nonapoptotic form of programmed cell death, has recently emerged as a research hot spot. Its core processes primarily involve abnormal iron metabolism, dysregulated lipid metabolism, and impaired antioxidant systems, with multiple signaling pathways governing its regulation. HCC is a highly aggressive malignancy, and most patients are diagnosed at an advanced stage, rendering them ineligible for radical surgical intervention. Currently, targeted drugs play a pivotal role in the treatment of advanced HCC; however, challenges such as drug resistance and the need for individualized therapy remain unresolved. Notably, since most therapeutic agents for liver cancer exert their antitumor effects by regulating ferroptosis, exploring the crosstalk between ferroptosis and liver cancer holds significant implications for HCC treatment.

Nevertheless, due to the complexity of ferroptosis regulatory networks and the intricate pathogenesis of HCC, the mechanisms underlying ferroptosis in HCC remain incompletely elucidated, leaving several critical questions unanswered. For example, it has not been fully explored whether ferroptosis exerts stage-specific roles during HCC progression. In addition, therapeutic agents targeting ferroptosis-related molecules are still in the early developmental stages, and relevant clinical trials have yet to be initiated in a timely manner. To address these gaps, a comprehensive investigation into the roles and mechanisms of ferroptosis in HCC pathogenesis is warranted, integrating perspectives from metabolism, immunity, and genetics. Given the significance of ferroptosis markers in HCC diagnosis, prognosis, and treatment response assessment, AI-driven prediction models constructed using these markers are expected to provide robust support for early diagnosis, precise therapy, and prognostic evaluation of HCC. In the future, further optimization of HCC-related AI models using ferroptosis markers, coupled with validation of their clinical efficacy, will facilitate their translation into clinical management of HCC.

CONCLUSION

In this review, we summarize current understanding of ferroptosis mechanisms, their association with HCC, and the promising potential of AI-based prediction models for HCC, while also addressing the challenges that need to be overcome. Our aim is to offer novel insights for developing HCC AI-driven prediction models based on ferroptosis mechanisms and markers and to inspire future research into targeted therapeutic agents for HCC and the construction of more accurate AI models, ultimately benefiting HCC patients.