Multi-omics approaches for identifying the PANoptosis signature and prognostic model via a multimachine-learning computational framework for intrahepatic cholangiocarcinoma

Abstract

Background and Aims:

The aims of the present study were to characterize the PANoptosis signature in intrahepatic cholangiocarcinoma (ICC) patients, construct a novel model to guide clinical diagnosis and treatment, and further explore the associated molecular mechanisms of drug resistance.

Approach and Results:

In total, 85 PANoptosis-related genes that possess both PANoptosis and multi-omics features were, respectively, screened from transcriptomic data from the OEP001105 public cohort and from transcriptomic and proteomic sequencing data from The First Affiliated Hospital of Chongqing Medical University. A novel framework integrating Cox regression analysis and 5 machine learning algorithms was developed to identify the 5 hub genes (POSTN, SFN, MYOF, HOGA1, and PECR). The subsequently constructed PANoptosis risk score demonstrates outstanding performance in predicting prognosis and clinical translation across multicenter cohorts with multi-omics profiling. Bulk and single-cell transcriptome profiling were used to investigate the tumor microenvironment, emphasizing the crucial role of macrophages in the tumor microenvironment of ICCs. Moreover, a positive spatial correlation of cancer-associated fibroblasts–derived POSTN expression with tumor-associated macrophages infiltration and PD-L1/PD-L2 expression in ICC patients was observed, suggesting that overexpression of POSTN may lead to resistance to immune checkpoint blockade therapy in ICC patients.

Conclusions:

The present study identified a precise prognostic and treatment strategy for ICC patients prone to PANoptosis, investigated the molecular mechanisms of PANoptosis in ICC cells, and highlighted the potential clinical relevance of the PANoptosis risk score in predicting prognosis and therapy response. These findings will help guide clinical treatment strategies for ICC.

INTRODUCTION

Intrahepatic cholangiocarcinoma (ICC) is a relatively uncommon yet highly aggressive malignancy, accounting for 5%–10% of all primary liver cancers.¹ Early diagnosis of ICC remains challenging due to the lack of symptoms in the early stages of this disease. Therefore, systemic drug therapy is one of the preferred treatment options for most patients with ICC. In particular, the application of immune checkpoint inhibitors (ICIs), such as anti-PD-1/L1/L2 monoclonal antibodies, to block the PD-1/L1/L2 signaling pathway is a new strategy for the treatment of malignant tumors, and it has played a significant role in the treatment of various solid tumors.² However, drug resistance due to intertumoral and intratumoral heterogeneity, which is mediated by the complex tumor microenvironment (TME) of ICC, is a crucial factor that hinders the efficiency of ICI therapy in advanced ICC patients. Therefore, developing more effective drug therapy agents and identifying different subgroups of tumor patients may increase the accuracy of drug therapy.

Multi-omics approaches have generated bioinformation from multiple layers at unprecedented scales and resolutions, making them widely used tools in cancer research for understanding molecular subtypes and prognosis-related features.³ Additionally, these applications have significantly improved the clinical efficacy of precision medicine and substantially improved the prognosis of cancer patients. Recently, 3 ICC molecular subtypes, with distinct prognoses, immune inflammation, and immunotherapy sensitivity (chromatin remodeling, metabolism, and chronic inflammation), have been identified via a multidisciplinary approach.⁴ These multi-omics–based molecular subtypes and risk stratifications have the potential to guide clinical treatment decisions, leading to precise personalized treatment. However, the use of multi-omics approaches for precision medicine in ICC requires further enhancement.

PANoptosis is a novel programmed cell death (PCD) pathway associated with cytoplasmic multiprotein complexes called PANoptosomes. Several PANoptosome complexes have been identified, including the ZBP1-PANoptosome, AIM2-PANoptosome, RIPK1-PANoptosome, and NLRP12-PANoptosome, with distinct sensors and regulators.^5–7 For example, Z-DNA binding protein 1 (ZBP1) detects influenza A virus via its Zα domain, facilitating interactions with RIPK3 and other associated proteins, including caspase-8/-6 and RIPK1. This interaction culminates in the formation of the ZBP1-PANoptosome complex, ultimately triggering PANoptosis.⁸ PANoptosis is recognized as a single-cell death-inducing complex that controls all 3 pathways rather than being activated independently of each other.⁹ PANoptosis also promotes crosstalk among multiple forms of PCD (eg, pyroptosis, apoptosis, and necroptosis)^10–12 and interactions with other types of PCD, such as ferroptosis.¹³ The activation of PANoptosis may lead to inflammation and the depletion of immune cells in the TME, weakening the antitumor immune response.⁸ Additionally, PANoptosis-induced inflammation may upregulate the expression of immune suppressive molecules (such as PD-L1), further inhibiting T cell function. Additionally, an increase in immune suppressive cells (such as myeloid-derived suppressor cells and regulatory T cells) triggered by PANoptosis weakens immunotherapy efficacy, leading to drug resistance.^14,15 Hence, PANoptosis-related features may inform prognosis-related stratification. Targeting PANoptosis and the PANoptosome complex may unveil novel treatment targets and clinically significant therapeutic approaches.

The present study utilized multiple bioinformatics approaches to identify PANoptosis-related features of ICCs, providing new insights into the regulatory role of PANoptosis in ICC and enhance the understanding of its heterogeneity, offering guidance for personalized prognosis prediction and treatment.

METHODS

Specimen collection, sequencing data acquisition, and preprocessing

Eighteen pairs of ICC tissues and paracancerous tissues were obtained from the First Affiliated Hospital of Chongqing Medical University. All patients underwent preoperative examination to exclude metastasis, recurrence, secondary tumors, preoperative adjuvant therapy, or other contraindications to surgery. Additionally, 177 pathological paraffin-embedded samples (including 90 ICC samples and 87 adjacent normal liver tissue samples, as well as survival data for 72 patients through postoperative follow-up) from patients who underwent surgery from 2020 to 2023 were used for tissue microarray preparation and subsequent immunohistochemistry (IHC) and in situ hybridization. Informed consent was obtained from all participants, and all samples were confirmed as ICC by postoperative pathology. Ethical approval for this study (NO. 2024-180-01 and NO. 2024-488-01) was provided by the Ethics Committee for Scientific Research at the First Affiliated Hospital of Chongqing Medical University. All research was conducted in accordance with both the Declaration of Helsinki and the Declaration of Istanbul. The RNA-seq data of the above cohorts were converted into log₂ (TPM+1) format for subsequent analysis. The ComBat R package was used to remove the apparent batch effect between the different cohorts.

PAN-RG consensus clustering analysis

Twenty-three PANoptosis-related genes, which play important roles in various solid tumors and impact cancer development and patient prognosis, were identified from published research (Supplemental Table S1, http://links.lww.com/HEP/J768). The ConsensusClusterPlus R package was utilized to perform consensus clustering on the expression profiles of these characteristic genes. Although the clustering algorithm itself is unsupervised, the selection of input genes was based on prior biological knowledge, introducing a degree of supervision at the feature selection stage. The PAM algorithm and Pearson distance metric were employed for 500 bootstraps, with each bootstrap involving clustering analysis of 80% of the training set patient samples. Differential analysis was conducted with the Limma R software package (version 3.40.6) to identify differentially expressed genes between PANoptosis-related gene (PAN-RG) clusters, using a significance threshold of p<0.05 and fold change >1.2.

Univariate Cox regression analysis, machine learning algorithms, and PANRS model construction

A multivariate Cox regression analysis was used to identify survival-related genes for ICC patients, and each ICC patient was assigned a PANoptosis risk score (PANRS) calculated via the following formula:

$${\sum }_{(k=1)}^n{\mathrm{Coef}}_k\times {\exp }_k$$

Here, n is the number of independent prognostic genes; ${\mathrm{Coef}}_{\mathit{k}}$ is the multivariate Cox regression coefficient index of gene k; and ${\exp }_k$ is the mRNA expression level of gene k in the prognostic signature.

Statistical analysis

The statistical analyses were conducted via R software (version 4.2.1) and GraphPad Prism (version 9.4.1). The Wilcoxon signed-rank test was used for continuous variables in 2 groups, and the Kruskal–Wallis test was utilized for 3 or more groups. Pearson chi-square test and Fisher exact test were performed for categorical variables. The Welch corrected t test was used to compare the quantitative reverse transcription polymerase chain reaction result, mean grey value for in situ hybridization, and the average optical density for IHC. p values <0.05 were considered statistically significant. Detailed materials and methods are provided in the Supplemental Tables and Supplemental Methods, http://links.lww.com/HEP/J769.

RESULTS

Identification of PAN-related gene clusters and expression landscape

The overall study flow scheme is depicted in Supplemental Figure S1, http://links.lww.com/HEP/J770. Using the 23 PAN-RGs, which were identified from published research for their established roles in PANoptosis across various solid tumors, 244 ICC patients from the OEP001105 cohort were clustered with the ConsensusClusterPlus R package. The clustering variables were adjusted to achieve optimal results at k=2, ensuring a balance between simplicity and complexity while maintaining biological significance (Figures 1A–C, Supplemental Figures S2 K, L, http://links.lww.com/HEP/J771), and 2 distinct clusters were identified, namely, cluster 1 (127 samples) and cluster 2 (117 samples). Principal component analysis revealed satisfactory PAN-RG cluster distinctions between the 2 groups of ICC patients (Figure 1D), and these findings were supported by gene expression heatmaps displaying differential gene expression between the 2 clusters (Figure 1F).

FIGURE 1.

Identification of the PAN-RGs clusters and PAN-RGs expression landscape. (A) Consensus clustering to identify molecular subtypes was performed using the transcriptomic data (n=244, training cohort). (B) Consensus clustering model with cumulative distribution function (CDF) for k=2–10, k means cluster count. (C) Relative change in the area under the CDF curve for k=2–10. (D) PCA analysis revealed a satisfactory PAN-RG cluster distinction. (E) Relationship between PAN-RG clusters and long-term prognosis in the training cohort (p=0.02). (F) The heatmap shows a significant difference between the 2 PAN-RG clusters. (G) The expression of 23 PAN-RGs between 2 PAN-RG clusters. (H–J) Boxplot showing immune cell abundance in 2 PAN-RG clusters by ESTIMATE, MCPCounter, and TIMER algorithms. Significance was evaluated by Kruskal–Wallis with *p<0.05, **p<0.01, ***p<0.001, and ****p<0.001. (K) Boxplot showing drug sensitivity in 2 PAN-RG clusters by the Prism database. Abbreviations: PAN-RGs, PANoptosis-related genes; PCA, principal component analysis.

The expression levels of 23 PAN-RGs were compared in the 2 clusters. Cluster 1 presented increased expression of 18 genes, whereas cluster 2 presented increased expression of 2 genes (CASP6 and TAB3) (Figure 1G). The Kaplan–Meier curve indicates a significantly poorer prognosis for C1 patients (p=0.02) (Figure 1E). Subsequent immune infiltration analyses revealed that C1 had significantly higher immune scores, stroma scores, and ESTIMATE scores compared to C2 (Figure 1H). The MCPCounter and TIMER algorithms revealed increased abundances of various immune cell types in the C1 group, along with increased expression levels of immune checkpoint blockade (ICB) genes (Figures 1I, J). Drug sensitivity analysis conducted with the Prism database, C1 exhibited increased sensitivity to gemcitabine, paclitaxel, irinotecan, cobimetinib, dabrafenib, and tivantinib but decreased sensitivity to oxaliplatin, 5-fluorouracil, capecitabine, sorafenib, and afatinib (Figure 1K). We also performed the same clustering analysis on external validation sets E-MTAB-6389 and GSE107943, identifying C1 with poorer prognosis, higher immune infiltration levels, and elevated expression of key ICI molecules, confirming the universality of this clustering (Supplemental Figure S3, http://links.lww.com/HEP/J772).

Identification of PAN-DEGs in PAN-RG clusters via multi-omics approaches

To explore the expression characteristics of genes between different clusters, gene functional analysis of the 2 PANoptosis clusters was performed. GSEA revealed that C1 was enriched mainly in functions, such as apoptosis, inflammatory response, KRAS signaling, and epithelial–mesenchymal transition, which are typically associated with the malignant characteristics of tumor cells, especially in cancer development and progression. However, C2 was enriched primarily in cellular metabolic functional characteristics, such as adipogenesis, bile acid metabolism, fatty acid metabolism, and oxidative phosphorylation, which are often associated with metabolic reprogramming phenotypes of tumor cells (Figure 2A).

FIGURE 2.

Identification of PAN-DEGs in PAN-RG clusters using multi-omics approaches. (A) GSEA analysis in the 2 PAN-RG clusters; *p<0.05, **p<0.01, and ***p<0.001. (B) Volcano plots of DEGs between the 2 PAN-RG clusters. (C) A Venn diagram represents common DEGs from transcriptome and proteomic sequencing of 5 pairs of ICC specimens derived from our center, log₂ FC≥1.2, p<0.05. (E, F, G) BP, CC, and MF categories of the GO annotation diagram of 85 PAN-DEGs. (H) KEGG enrichment analysis of 85 PAN-DEGs. Abbreviations: BP, biological process; CC, cellular component; DEGs, differentially expressed genes; GO, Gene Ontology; GSEA, gene set enrichment analysis; ICC, intrahepatic cholangiocarcinoma; KEGG, Kyoto Encyclopedia of Genes and Genomes; MF, molecular function; PAN-DEGs, PANoptosis-related differentially expressed genes; PAN-RG, PANoptosis-related gene.

Comparing gene expression between the 2 PAN-RG clusters, we identified 3224 upregulated and 1034 downregulated genes, with a fold change of 1.2 and a significance threshold of <0.05 (Figure 2B and Supplemental Table S5, http://links.lww.com/HEP/J827). Furthermore, 425 differentially expressed genes with consistent expression patterns at both the transcriptomic and proteomic levels were identified from 5 pairs of matched ICC tumor and adjacent normal tissues (Figure 2C). Using a Venn diagram, the differentially expressed genes were merged to obtain 85 PANoptosis-related differentially expressed genes (PAN-DEGs), including 24 upregulated genes and 61 downregulated genes (Figure 2D).

GO and KEGG analyses of the 85 PAN-DEGs revealed that their transcription products were located mainly in mitochondria and play a key role in cellular metabolic processes, such as oxidoreductase activity, carboxylic acid binding, tetrapyrrole binding, and lyase activity (Figures 2E–G). These genes were involved in the following biological molecule metabolism processes: carbon metabolism; peroxisome and fatty acid degradation; glyoxylate and dicarboxylate metabolism; and glycine, serine, and threonine metabolism (Figure 2H).

Constructing the PANRS based on multi-omics approaches and multimachine learning methods

To facilitate the clinical application of the above results, a PANRS prognostic model was constructed using various machine learning methods and Cox regression methods. Univariate Cox regression was used to screen 66 genes with prognostic power among 85 PAN-DEGs, and 5 machine learning methods (LASSO, GBM, SVM, decision tree, and random forest) were utilized to screen for hub genes in ICC patients susceptible to PANoptosis. The top 20 key genes with the greatest importance were screened via the decision tree, random forest, and SVM algorithms. By combining the results of the 5 machine learning methods via Venn diagram analysis, 5 PANoptosis-related hub genes (PECR, HOGA1, SFN, POSTN, and MYOF) were identified (Figure 3A; Supplemental Table S6, http://links.lww.com/HEP/J828). Each of these 5 genes had significant prognostic power for the clinical outcome of patients in the training set (Figure 3C) and validation sets (Supplemental Figure S4, http://links.lww.com/HEP/J773).

FIGURE 3.

Constructing the PANRS based on multi-omics approaches and multimachine-learning methods. (A) A Venn diagram showed that 5 machine learning algorithms obtained 5 hub genes (LASSO, GBM, SVM, DecisionTree, and RandomForest). (B) Multivariate Cox analysis showed that POSTN was the most significant gene that could predict the clinical outcome of patients. (C) The Kaplan–Meier plot demonstrates that all 5 hub genes can independently predict the clinical prognosis of ICC patients. (D) Box plots showing the expression levels of 23 PAN-RGs in the high-risk and low-risk PANRS subgroups. (E) The boxplot shows that the PANRS in PAN-RGs cluster 1 are higher, which is consistent with previous research. (F) Distribution of the risk score, gene expression heatmap, and survival status of 244 patients in the OEP001105 and survival analysis of PANRS subgroups based on OS (log-rank test). (G) Distribution of clinical characteristics in high and low-risk groups. (H) Correlation analysis between PANoptosis and 7 other PCDs. Abbreviations: ICC, intrahepatic cholangiocarcinoma; OS, overall survival; PAN-RG, PANoptosis-related gene; PANRS, PANoptosis risk score; PCDs, programmed cell deaths.

The PANRS was determined via multivariate Cox regression (Figure 3B). The training set samples were then divided into high-risk and low-risk groups on the basis of a cutoff value of 2.65 calculated via X-tile software, which revealed significant differences in the expression levels of 15 PAN-RGs between the 2 risk groups (Figure 3D). PAN-RG cluster 1 has greater PANRS than PAN-RG cluster 2 (Figure 3E). Additionally, the Kaplan–Meier curve and time-dependent receiver operating characteristic curve analysis indicated high predictive power for prognosis, with AUC values of 0.76, 0.77, and 0.76 for 1-year, 2-year, and 3-year survival, respectively (Figure 3F), compared to the traditional ICC serological prognostic marker CA19-9, PANRS demonstrated superior prognostic efficacy, achieving consistently higher AUC values across multiple cohorts (Supplemental Figure S5, http://links.lww.com/HEP/J774). Furthermore, we confirm that all 5 genes are essential for distinguishing between good and poor prognosis. The absence of any single gene will diminish the prognostic efficacy of PANRS (Supplemental Table S2, http://links.lww.com/HEP/J775).

We also notice that the high-PANRS subgroup is associated with poor clinicopathological outcomes of patients, such as regional lymph node metastasis, distal metastasis, intrahepatic metastasis, higher TNM stage, and vascular invasion (Figure 3G). Gene set enrichment analysis (GSEA) revealed that genes in the high-PANRS group were associated mainly with malignant phenotypes, including apoptosis, epithelial–mesenchymal transition, glycolysis, hypoxia, IL2/STAT5 signaling, and other pathways related to tumor occurrence and progression. Conversely, benign pathways, such as fatty acid and bile acid metabolism, were enriched in the low-risk group (Supplemental Figure S6, http://links.lww.com/HEP/J776). In addition, the ssGSEA algorithm was used to calculate the current PCD score in the training set samples and compared it with that of the PANRS. The results revealed that PANoptosis was positively correlated with apoptosis, autophagy, necroptosis, pyroptosis, and other phenotypes (Figure 3H and Supplemental Table S3, http://links.lww.com/HEP/J826). These findings confirmed those reported by Samil et al,¹⁶ who demonstrated that the PANoptosome, a key molecule of PANoptosis, can engage 3 key modes of PCD—pyroptosis, apoptosis, and necroptosis—in parallel. Furthermore, previous studies have shown complex connections between PANoptosis and other types of PCD. For example, the liproxstatin-1 (LPT1) ferroptosis inhibitor has been reported to alleviate steatosis and steatohepatitis in metabolic dysfunction–associated fatty liver disease mice by modulating PANoptosis pathways, suggesting a potential link between ferroptosis and PANoptosis.¹³ Additionally, the present study demonstrated significant associations among PANoptosis and malignant phenotypes (such as autophagy, cuproptosis, disulfidptosis, necroptosis, cuproptosis, and pyroptosis) (Supplemental Figure S2B, http://links.lww.com/HEP/J771).

Validation of the PANRS in multiple cohorts via multi-omics approaches

In the transcriptomic external validation cohorts, the PANRS also demonstrated strong predictive accuracy in identifying the clinical prognosis outcomes of ICC patients, and the resulting gene expression heatmaps, Kaplan–Meier curves, and receiver operating characteristic curves were similar to those of the training set (Figures 4A, B). Significant positive correlations were subsequently identified between the Q9BY49, Q86XE5, P31947, Q15063, and Q9NZM1 proteins and their molecular comparators (PECR, HOGA1, SFN, POSTN, and MYOF) (Figure 4C). Multivariate Cox regression was utilized to develop the PANoptosis protein-omics risk score and to validate the predictive ability of the proteomic risk model with AUC values of 0.72, 0.70 and 0.68 for the 1-year, 2-year, and 3-year survival periods, respectively (Figure 4D). Furthermore, to overcome the limited sample sizes that hinder robust model testing on larger cohorts, we employed 5-fold cross-validation to optimize and evaluate the PANRS model, achieving stable predictive performance (Mean AUC: 0.726, 95% CI: 0.661–0.791). This method ensures robustness despite dataset limitations (Supplemental Figure S2H, http://links.lww.com/HEP/J771).

FIGURE 4.

Validation of PANRS across multiple cohorts and multi-omics platforms. (A, B) The PANRS results of the transcriptome validation sets M-ETAB-6389 and GSE107943 showed similar prognostic power as the training set. (C) There was a significant positive correlation between the transcriptome and protein expression levels of the 5 hub genes. (D) The PANoptosis–Protein–RiskScore constructed by the proteomics data of OEP001105 shows good prognostic efficacy. (E) Construction of the nomogram based on the PANRS and TNM stage. (F) Calibration curve of the nomogram for 1-year, 2-year, and 3-year OS. (G) Decision curve analysis (DCA) showing the net benefit by applying the nomogram or PANRS and TNM stage alone. Abbreviations: OS, overall survival; PANRS, PANoptosis risk score.

To enhance the clinical applicability of the PANRS, a nomogram based on both the PANRS and clinical characteristics was developed and shows greater net clinical benefit than the TNM stage (Figure 4G).

Correlations between the PANRS and the TME landscape and immune characteristics

To further investigate the immunological characteristics of the different PANRS groups, multiple immune algorithms, including MCPcounter, IPS, ESTIMATE, and ssGSEA, were employed to analyze 28 immune cell types to evaluate the relationship between the PANRS and the tumor immune microenvironment (TIME) (Supplemental Figure S2D, http://links.lww.com/HEP/J771). The high-risk group had greater infiltration of T cells, NK cells, monocytes, myeloid dendritic cells, endothelial cells, neutrophils, and fibroblasts compared to the low-risk group (Figure 5B). The difference in immune infiltration was also confirmed in external validation sets with CIBERSORT, MCPCounter, and ssGSEA algorithm (Supplemental Figure S7, http://links.lww.com/HEP/J780).

FIGURE 5.

The correlation between PANRS and the TIME landscape and immune characteristics. (A) Correlation analysis of PANRS and 5 hub genes with 28 immune cell infiltrations. (B) Twenty-eight immune cell infiltration differences between high and low PANRS risk groups. (C) Infiltration of both macrophages and monocytes was significantly associated with PANRS and 5 hub genes. (D) Analysis of differences in expression levels of major immune checkpoints in high and low PANRS subgroups. (E) Correlation analysis between PANRS and 5 hub genes and expression levels of major immune checkpoints. Abbreviations: PANRS, PANoptosis risk score; TIME, tumor immune microenvironment.

To identify the key immune cells that mediate PANoptosis in the TME, correlation analysis was subsequently conducted, which identified 21 immune cell types that were significantly correlated with the PANRS (Spearman, p<0.05) (Figure 5A). The relationships between the infiltration levels of 28 immune cells in the ICC patient TME and overall survival (OS) were also analyzed, which identified 17 immune cells significantly associated with patient prognosis (Supplemental Figure S8, http://links.lww.com/HEP/J781). By combining the results of the significant differences (Figure 6A), infiltration analysis (Figure 6B), and survival analysis (Supplemental Figure S8, http://links.lww.com/HEP/J781), the Venn diagram identified 13 immune cell types closely related to the PANoptosis characteristics for ICC patients (Supplemental Figure S2E, http://links.lww.com/HEP/J771).

FIGURE 6.

PANoptosis characteristic in the single-cell transcriptome. (A) Violin plot showing the top 2 marker genes in each cell cluster. (B) t-SNE plot showing the cell types identified by marker genes. (C) t-SNE plot showing the gene set score of 5 PANoptosis hub genes in each cell. (D) Box plots indicate AddModuleScore of 5 hub genes for 10 cell types in TME, the line indicates the mean. (E) t-SNE plot showing the expression level of 5 PANoptosis hub genes in each cell. (G) t-SNE plot showing the expression level of the immune checkpoint in each cell. (G) Circos plots showing the MK, SPP1, and ITGB2 signaling pathway networks. Abbreviation: TME, tumor microenvironment.

We noticed that PANRS and nearly all hub genes were highly correlated with the levels of monocyte–macrophage infiltration. Increased monocyte infiltration may lead to lower PANRSs and better prognostic outcomes, whereas enhanced macrophage infiltration may lead to higher PANRSs and poorer survival outcomes (Figure 5C). These findings suggested that monocytes and macrophages are the crucial immune cells that mediate PANoptosis in the TME of ICC.

Characterization of the immunotherapy response and drug resistance related to PANoptosis

The anticancer immune response consists of a series of stepwise events known as the cancer–immunity cycle, in this context, TIP analysis revealed that the high-risk group had stronger activity in releasing cancer cell antigens, presenting cancer antigens, infiltrating immune cells into tumors, and recognizing cancer cells with T cells (steps 1, 2, 5, and 6), which suggested that the high-PANRS group exhibited greater anticancer activity in the functioning of immune cells (Supplemental Figure S2F, http://links.lww.com/HEP/J771).

Among the key immune checkpoints, we observed significant differences in the expression of PD-1, PD-L1, PD-L2, and BILA between the 2 risk groups. Compared to the high-risk group, the low-risk group presented higher levels of PD-1 and BILA but lower levels of PD-L1 and PD-L2 (Figure 5D). The strong correlation between the PANRS and the expression levels of PD-L1 and PD-L2 suggests the differential immunotherapeutic responses among patients in different PANRS subgroups, particularly the anti-PD-L1/L2 immunotherapy (Figure 5E).

Drug sensitivity analysis based on the PRISM database indicated that PANRS did not provide medication guidance for advanced ICC patients in first-line systemic chemotherapy regimens, such as gemcitabine with a platinum-based doublet, gemcitabine plus cisplatin plus S-1, or gemcitabine plus cisplatin plus albumin-bound paclitaxel (nab-paclitaxel) triple therapy, which are widely recommended in international guidelines.^17–19 However, for paclitaxel and various molecular targeted drugs, such as lestaurtinib, sorafenib, lonafarnib, and afatinib, the sensitivity of the high-risk group was significantly greater than that of the low-risk group (Supplemental Figure S9A, http://links.lww.com/HEP/J779).

Identification of PANoptosis characteristics in single-cell transcriptomics

To explore PANoptosis-related features at the single-cell level, the scRNA-seq dataset (GEO: GSE138709) was analyzed to identify and characterize different cell subpopulations within ICC tissues (Figures 6A, B and Supplemental Table S7, http://links.lww.com/HEP/J829). Significantly higher scores were observed for the 5 PANoptosis-related hub genes in cholangiocytes (including malignant cells), erythroid cells, and smooth muscle cells (Figures 6C, D), implying a strong association between PANoptosis and these cells, especially cholangiocytes.

Then, we analyze the expression patterns of 5 hub genes across different cell types to explore the role of PANoptosis in the TME of ICC. The results revealed that POSTN was expressed primarily in fibroblasts and smooth muscle cells, whereas SFN, HOGA1, MYOF, and PECR were predominantly expressed in cholangiocytes (Figure 6E). This finding corroborated previous findings indicating that periostin (POTSN), which typically originates from stromal cells, promotes TME remodeling during tumor progression.^20,21

Previous bulk RNA-seq analysis has revealed that the expression levels of the 5 hub genes in ICC patients may indicate potential differential responses to anti-PD-L1/L2 treatment. Therefore, the expression levels of ICB genes were further investigated at the single-cell level. The ICIs significantly associated with the PANRS, such as CD274 (PD-L1) and PDCD1LG2 (PD-L2), were enriched mainly in macrophages (Figure 6F).

Further analyses were focused on cholangiocytes and malignant tumor cells to investigate their interactions with other TME cell types. Cell–cell communication analysis revealed that malignant tumor cells with higher PANRSs communicated with a broader range of TME cells compared to benign cholangiocytes, which had a lower PANRS. Additionally, malignant tumor cells exerted a stronger influence on SPP1, laminin, and MK signaling pathways (Figure 6G).

Validation of PANoptosis hub genes and their roles in cellular functions and ICI resistance

To validate the expression of 5 PANoptosis-related hub genes—POSTN, SFN, MYOF, HOGA1, and PECR—tissue microarrays were constructed from tumor and adjacent normal tissues of 90 ICC patients for in situ hybridization and IHC staining. The results showed significant upregulation of POSTN, SFN, and MYOF, along with downregulation of HOGA1 and PECR in tumor tissues (Figures 7A, B). These findings were consistent with our RNA-seq results and quantitative reverse transcription polymerase chain reaction analysis in 18 pairs of fresh ICC tissues (Supplemental Figure S2G, http://links.lww.com/HEP/J771). Additionally, IHC analysis revealed that POSTN expression significantly correlated with OS and progression-free survival (PFS) (p=0.017, r=−0.280; p=0.022, r=−0.269), while SFN showed correlations with OS and PFS that were not statistically significant, likely due to sample size limitations (Figures 7C, D and Supplemental Table S4, http://links.lww.com/HEP/J777).

FIGURE 7.

Evaluate the differential expression of 5 hub genes in tumor and normal tissues, as well as their correlation with OS and PFS. (A) Assess mRNA expression of 5 hub genes in tissue microarray using ISH, red fluorescence represents staining of the hub gene. (B) Evaluation of the 5 hub genes from 90 paired ICC tumors and adjacent normal tissue by IHC, and calculating dye concentration with AOD. Correlation analysis (Spearman) between the expression levels of hub genes and patient OS (C) and PFS (D); *p<0.05, **p<0.01, and ***p<0.001. Abbreviations: AOD, average optical density; ICC, intrahepatic cholangiocarcinoma; IHC, immunohistochemistry; ISH, in situ hybridization; OS, overall survival; PFS, progression-free survival.

Further, IHC staining for PD-L1 and PD-L2 in 90 ICC tumor samples revealed significant correlations between PD-L1 expression and POSTN (r=0.25, p=0.02) and HOGA1 (r=−0.24, p=0.02), and between PD-L2 and POSTN (r=0.35, p=8.4e−4), SFN (r=0.25, p=0.02), and PECR (r=−0.28, p=8.3e−3), suggesting that these hub genes, especially POSTN, may regulate ICI resistance through immune checkpoint modulation (Figures 8A, B).

FIGURE 8.

The roles of hub genes in cellular functions and ICI resistance. (A, B) Correlation analysis between PD-L1 (A)/PD-L2 (B) and 5 hub genes; (C) CCK-8: absorbance at 450 nm (y-axis) versus incubation time (h) (x-axis). (D, E) Transwell experiment results of 5 hub genes. The y-axis of (E) represents the number of traversing cells. (F, G) The flow cytometry measures cell apoptosis, with the y-axis of (F) representing the overall proportion of apoptotic cells, including early and late apoptosis; *p<0.05, **p<0.01, and ***p<0.001. Abbreviation: ICI, immune checkpoint inhibitor.

Finally, in functional assays, knockdown of SFN and MYOF, or overexpression of HOGA1 and PECR, significantly reduced cell proliferation in ICC cells, with differences becoming significant after 48 hours. However, knockdown of POSTN led to a significant increase in proliferation, suggesting POSTN promotes ICC cell growth (Figure 8C). In the migration assay, knockdown of POSTN (p=0.0339) and SFN (p=0.0012) significantly reduced migration, while knockdown of MYOF (p=0.2492) and overexpression of HOGA1 (p=0.1421) or PECR (p=0.1057) had no significant effect (Figures 8D, E). In the apoptosis assay, POSTN knockdown (p=0.0655), SFN knockdown (p=0.0012), and HOGA1 (p=0.0167) overexpression increased apoptosis, while PECR overexpression (p=0.0047) reduced apoptosis, indicating that POSTN and SFN help ICC cells survive by inhibiting apoptosis, while HOGA1 overexpression promotes cell death (Figures 8F, G and Supplemental Table S9, http://links.lww.com/HEP/J779).

DISCUSSION

Generally, due to the complexity of the TME and the heterogeneity of ICC, ICB therapy fails to produce satisfactory outcomes after first-line chemotherapy in advanced patients, resulting in rapid disease progression, metastasis, and increased mortality.^22,23 The methods to enhance the efficacy of drug treatments urgently require exploration. PANoptosis is a recently defined inflammatory PCD process mediated by the PANoptosome and emphasizes the crosstalk and interaction of various PCD pathways, such as pyroptosis, apoptosis, and necroptosis.¹¹ Its biological roles have been preliminarily explored in multiple malignancies; however, it still lacks in-depth research in ICC.

Construction of the PANRS and identification of the PANoptosis signature in ICC

Based on the identified 23 PAN-RGs, our study delineated 2 clusters with distinct PANoptosis characteristics and developed a simplified prognostic assessment (PANRS) based on 5 hub genes. Cluster 1 and the high-risk group are associated with TME changes like inflammation and dysregulated cell processes, leading to a worse prognosis. In contrast, cluster 2 and the low-risk group show metabolic homeostasis, efficient energy metabolism, and better immune function, correlating with lower disease progression (Figure 2A, Supplemental Figure S6, http://links.lww.com/HEP/J828 and Supplemental Table S8, http://links.lww.com/HEP/J781). Also, Cluster and PANRS groups differ in OS prognosis, mutation landscapes, and immune characteristics, linking molecular classification with clinical applications for personalized assessments.

Small molecule drugs (such as tyrosine kinase inhibitors, EGF receptor inhibitors, and mitogen-activated protein kinase inhibitors) may demonstrate superior efficacy in treating ICC with higher PANRSs, which may be attributed to the enrichment of pathways in the high-risk samples that are closely associated with tumor growth, metabolism, and drug resistance. These findings indicated that the PANRS model has the potential to predict the prognosis and sensitivity to targeted therapy in ICC patients.

Then, we validated the crucial roles of 5 hub genes, particularly SFN, in cell proliferation, migration, and apoptosis in ICC cell lines. The regulatory role of this molecule has also been revealed in other solid tumors.^24,25 The knockdown of pro-PANoptosis genes and the overexpression that inhibits PANoptosis paradoxically promote apoptosis in ICC cells further confirms that PANoptosis is an independent form of PCD pathway, distinct from a simple combination of pyroptosis, apoptosis, and necrosis.

Tumor heterogeneity is a crucial factor hindering the effectiveness of drug treatment for tumors and must be supported by the TME. Previous research has shown that perturbations and crosstalk in PANoptosis occur in the TME of human cancers, playing a crucial role in tumor cell immune escape and affecting the effectiveness of immunotherapy.⁸ In our study, the high-risk group showed increased infiltration of T cells, NK cells, monocytes, myeloid dendritic cells, endothelial cells, neutrophils, and fibroblasts, which are crucial for immune surveillance, pathogen response, and potentially tumor suppression.^26–29 These findings indicated a more effective immune-mediated tumor control, as well as a more regulated tissue repair process. Immune surveillance is partially activated; however, the presence of immune evasion mechanisms may hinder the full antitumor effect, increasing the risk of tumor progression in ICC patients prone to PANoptosis. These findings suggest that PANoptosis significantly impacts the TME of ICCs and that the PANRS can serve as a prognostic tool that closely reflects the immune status of ICCs.

Macrophages contribute to panoptosis and impact the survival of ICC patients

Among the immune cells in the TME, a significant correlation of the PANRS was observed with the infiltration abundance of T cells, B cells, NK cells, monocytes, myeloid dendritic cells, and endothelial cells. In particular, higher levels of macrophage infiltration and lower monocyte infiltration in ICC are often associated with a higher PANoptosis signature and worse OS. Single-cell analysis revealed significant expression of the CASP1, NLRP3, and PYCARD PANoptosis-related genes in macrophages (Supplemental Figure S9B, http://links.lww.com/HEP/J779). These molecules are involved in the formation and activation of the inflammasome, which plays a significant role in the inflammatory response within the TME, particularly in macrophages. Specifically, NLRP3 acts as a sensor for the inflammasome, becoming activated in response to stimuli, such as infections, metabolic disturbances, or tissue damage, and it initiates the assembly of the inflammasome complex.³⁰ PYCARD then interacts with NLRP3, forming an inflammasome platform that facilitates the activation of caspase 1 (CASP1). Activated CASP1 serves as the effector molecule of the inflammasome and is responsible for cleaving precursor cytokines (such as pro-IL-1β and pro-IL-18) into their active forms (IL-1β and IL-18), which are subsequently secreted extracellularly. These cytokines enhance the local inflammatory response, promoting tumor cell growth, invasion, and metastasis while inhibiting immune surveillance. This process is associated with tumor progression, immune suppression, and resistance to therapy.^31,32 Additionally, activation of the NLRP3 inflammasome may promote macrophage polarization toward the M2 phenotype, thereby enhancing the malignant behavior of tumors and reducing patient survival.³³ The opposite result has been reported for the cuproptosis pathway in ICC: patients who are prone to cuproptosis exhibit reduced macrophage abundance, as macrophages regulate copper homeostasis in ICC cells through ligand–receptor interactions, such as NRP1–VEGFA and NRP2–VEGFA.³⁴ However, the present study confirmed that patients prone to PANoptosis are less likely to experience cuproptosis (Figure 3J), indicating the harmful impact of PANoptosis-related macrophages on ICC patient survival.

Role of the POSTN PANoptosis hub gene in ICB resistance

Immune checkpoints are crucial regulatory pathways that prevent the immune system from indiscriminately attacking cells.³⁵ However, due to complex intertumoral and intratumoral heterogeneity, ICC exploits immune checkpoints, which stimulate antitumor immunity evasion, ultimately resulting in a low response rate in cancer treatment.³⁶ The present results suggested that patients in high-PANoptosis risk subgroups have higher reactivities to immunotherapy, especially anti-PD-L1/L2 immune therapy.

Our in vitro studies confirm that knocking down the oncogene POSTN leads to increased proliferation of ICC cell lines. This paradox suggests that POSTN does not simply promote tumorigenesis in ICC cells; rather, interactions among various cells in the TME may be the culprit. Meanwhile, some scholars have reported that cancer-associated fibroblasts (CAFs) and tumor-associated macrophages (TAMs) are associated with poor prognosis in various cancers and may induce resistance to various ICB therapies;^37–39 however, the underlying mechanism has not been fully elucidated. By utilizing bulk and single-cell sequencing, our study revealed an abundance correlation between CAF-derived POSTN and TAMs, as well as a positive correlation between POSTN expression and PD-L1/PD-L2 expression in the TME of ICC patients.

Previous studies have proposed that the overexpression of periostin (POSTN), a multifunctional extracellular protein, plays a crucial role in inflammatory disorders, tumorigenesis, microvascular invasion, tumor differentiation, and lymph node metastasis in various solid cancers, indicating poor survival.⁴⁰ Furthermore, Utispan et al²¹ reported that high periostin levels distinguish CCA from other related liver diseases and induce CCA cell proliferation and invasion. In the present study, POSTN was identified as the most significant hub gene, and it was spatially correlated with the infiltration level of stromal cells, such as smooth muscle cells and fibroblasts. We also confirmed that periostin was significantly negatively correlated with ICC patient OS and PFS, as well as the expression level of PD-L1 and PD-L2 (Figures 7C, D and 8A, B).

Overall, our findings indicate a novel CAFs–POSTN–TAMs axis: the overexpression of CAF-derived periostin (POSTN) recruits TAMs, which secrete anti-inflammatory cytokines and exosomes, as well as increase the PD-L1/L2 immune checkpoint ligands, ultimately leading to the resistance of tumor cells to ICB therapy. Wei et al²⁰ reported that compared with blockade of only periostin or the PD-1 immune checkpoint, combined blockade enhances the efficacy of immunotherapy by weakening the crosstalk among tumor cells, CAFs, and immune cells, thereby subsequently inhibiting the growth of colorectal cancer. Similarly, targeting POSTN may represent a potential approach to achieve favorable clinical outcomes for patients with ICC.

CONCLUSIONS

By utilizing multimachine learning and multi-omics sequencing, our study comprehensively explored PANoptosis features in ICC for the first time. First, a PANRS model was developed and validated to predict PANoptosis in ICC patients, with excellent prediction efficiency and clinical significance. Second, the PANoptosis-related molecular mechanisms and TME landscape of ICC were investigated, emphasizing the role of macrophages in the TME of ICC. Finally, we propose the novel CAFs–POSTN–TAMs axis: the overexpression of periostin (POSTN) from CAFs was demonstrated to recruit TAMs, which then secrete anti-inflammatory cytokines and exosomes to increase the PD-L1/L2 surface immune checkpoint ligands, ultimately leading to tumor cell resistance to ICB therapy and poor clinical outcomes. These findings shed light on the role of PANoptosis in ICC and its potential implications for predicting the prognosis of patients and guiding drug treatment strategies.

Limitations

ICC is a relatively rare subtype of liver cancer, and as such, publicly available datasets, particularly those focused on ICB therapy, are limited. Moreover, the heterogeneity in sequencing platforms, clinical data, and patient characteristics across the available datasets may introduce variability, potentially affecting the accuracy and generalizability of our analysis. Future studies involving larger, ICB-specific cohorts are needed to strengthen our findings. Furthermore, although the consensus clustering in our research is an unsupervised method, the selection of PANoptosis-related genes was guided by prior knowledge, introducing a degree of supervision. This may influence the objectivity of subgroup identification, and future studies using genome-wide, data-driven clustering approaches are warranted to validate our classification framework. Although the PANRS shows clinical relevance in prognosis prediction, this study was not designed or powered as a biomarker validation study. Future prospective studies are needed to assess its translational potential.

Supplementary Material

hep-83-466-s001.xlsx ^{(151.8KB, xlsx)}

hep-83-466-s002.xlsx ^{(5.4MB, xlsx)}

hep-83-466-s003.xlsx ^{(16.2KB, xlsx)}

hep-83-466-s004.xlsx ^{(267.6KB, xlsx)}

hep-83-466-s005.xlsx ^{(11.7KB, xlsx)}

hep-83-466-s006.docx ^{(35.5KB, docx)}

hep-83-466-s007.pdf ^{(323.4KB, pdf)}

hep-83-466-s008.xlsx ^{(18.5KB, xlsx)}

hep-83-466-s009.xlsx ^{(53.8KB, xlsx)}

hep-83-466-s010.pdf ^{(468KB, pdf)}

hep-83-466-s011.pdf ^{(354.2KB, pdf)}