Radiogenomics deciphers tumor heterogeneity in hepatocellular carcinoma and combined hepatocellular-cholangiocarcinoma Authors: Yi Wang, Shichen Miao, Xiao Wang, Weiren Liu, Yinghong Shi Published: 2026 ABSTRACT AbstractBackgroundCombined hepatocellular cholangiocarcinoma (cHCC-CCA) is a highly heterogeneous cancer with limited knowledge of molecular features and prognosis. To compare the molecular features of cHCC-CCA with different subtypes of hepatocellular carcinoma (HCC) and to investigate the biological meanings of prognostic radiomic features in cHCC-CCA.MethodsWe performed an integrative analysis to investigate the heterogeneity of cHCC-CCA and HCC by combining transcriptomic and radiomic approaches. RNA sequencing revealed distinct transcriptomic profiles between cHCC-CCA and HCC. For radiomic analysis, an eight-feature radiomic signature was constructed using preoperative contrast-enhanced MRI to predict overall survival in cHCC-CCA, and its prognostic significance was validated in an external dataset. Furthermore, weighted gene co-expression network analysis was employed to correlate radiomic features with underlying biological pathways.ResultsLR-M subtype of cHCC-CCA exhibited more aggressive characteristics and immunosuppressive tumor microenvironment compared with LR-4/5 subtype, HCC with rim arterial-phase hyperenhancement (rim-APHE) and HCC without rim-APHE. The radiomic signature was correlated with overall survival (hazard ratio [HR], 5.47; 95% CI:2.08, 7.23, p = 0.006) in the cHCC-CCA validation cohort. Three types of prognostic radiomic features were correlated with distinct pathways: fatty acid metabolism, signaling pathways regulating pluripotency of stem cells and T cell receptor signaling pathway, respectively. Original_Glszm_GrayLevelNonUniformity might reflect the upregulation of fatty acid metabolism and immune suppression of cHCC-CCA.ConclusionsWe illustrated the heterogeneity of subtypes of HCC and cHCC-CCA, which provides insights into the differentiation of primary liver cancer. Correlation of prognostic radiomic features with transcriptome profiles in cHCC-CCA enables noninvasive evaluation of tumor microenvironment.Supplementary InformationThe online version contains supplementary material available at 10.1007/s12672-026-05101-1.Keywords: Combined hepatocellular cholangiocarcinoma, Tumor microenvironment, Hepatocellular carcinoma, Radiomic, Weighted gene co-expression network analysis Introduction Combined hepatocellular cholangiocarcinoma (cHCC-CCA) is a biophenotypic tumor with occurrence rate of 0.4~14.2% in primary liver cancer [1, 2]. The World Health Organization characterizes classic cHCC-ICC by the clear architectural integration of hepatocytic and cholangiocytic differentiation in the same nodule, independent of the specific ratio between the two tumor cell types [3]. Recently, there has been evidence suggesting that cHCC-CCA may originate from liver progenitor cells which gradually transforms into tumor stem cells [4, 5]. cHCC-CCA expressed multiple tumor stem cell related markers including CK19, CD56 and CD133. The high expression of these molecules correlated with poor prognosis [6]. A multiomic study revealed that the genomic spectrum of two components of cHCC-CCA were different from those of hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma (iCCA) [7]. Previous studies found that the radiologic features of cHCC-CCA were related to molecular characteristics. It can be categorized into LR-4/5 or LR-M according to liver imaging reporting and data system (LI-RADS), which is determined by the proportion of HCC and iCCA components [8, 9]. In addition, radiologic features can also reflect the heterogeneity of HCC. Compared with most HCCs presenting diffuse hyperenhancement in the arterial phase of contrast-enhanced computed tomography (CT) or magnetic resonance imaging (MRI), HCC with rim-like arterial-phase hyperenhancement (rim-APHE) was associated with more aggressive pathologic features and elevated stemness [10]. The molecular characteristics of the subtypes of HCC and cHCC-CCA were variable, conferring different differentiation trajectories of primary liver cancer and their clinical outcomes. With this concern, understanding the heterogeneity of HCC and cHCC-CCA enables clinicians to make tailored treatment plan. Radiomic deciphers the landscape of a whole tumor without selection bias [11]. Multiple quantitative image features derived from tumor images can reflect tumor heterogeneity comprehensively. Radiogenomic analysis reflects the underlying biological function of radiomic features, which bridges the gap between imaging features and genomic information [12]. Hence, radiomic has potential to provide genome-related information noninvasively and can be implemented multiple times during the follow-up period. Radiomic has been widely applied to distinguish cHCC-CCA with HCC or iCCA [13, 14]. However, the association of prognostic radiomic features with biological pathways in cHCC-CCA remains unclear. In the present study, we performed transcriptomic analysis to illustrate the heterogeneity of cHCC-CCA and HCC based on the imaging category. Prognostic radiomic features of cHCC-CCA were further identified to explore the biological interpretation. Methods Patient cohort preparation We retrospectively identified consecutive cHCC-CCA patients who underwent surgical ablation for radiomic analyses. Two separate datasets in which cHCC-CCA patients: a training set from Zhongshan Hospital of Fudan University (ZS cohort) and an external validation cohort from Affiliated Hospital of Nantong University (NT cohort). Inclusion criteria were as follows: (a) pathologically proven cHCC-CCA according to the 2019 WHO classification criteria and (b) preoperative high-quality contrast-enhanced MR images were obtained within two weeks. Exclusion criteria were as follows: (a) patients who received other treatment before surgical resection and (b) patients with multiple lesions. Besides, clinical data was recorded including age, microvascular invasion (MVI) status, preoperative serum α-fetoprotein (AFP) levels and LI-RADS category. A radiogenomic data set comprised of 25 cHCC-CCA patients (from Zhongshan Hospital of Fudan University, a subset of the radiomic analysis cohort) underwent RNA sequencing. Another RNA sequencing data set of 56 HCC patients was included for exploring the heterogeneity of HCC. 55 HCC patients who received combined tyrosine kinase inhibitor (TKI) and anti-PD-1 antibody treatment and underwent pre-treatment contrast-enhanced MRI were enrolled to investigate the treatment response heterogeneity of HCC with rim-APHE and HCC without rim-APHE. The patient enrollment process was presented in Fig. 1. Flowcharts showing patients recruitment procedure. (A) Description of the cHCC-CCA patients cohorts used in this study. (B) Description of the HCC patients used in this study MRI protocol, imaging analysis and radiomic feature extraction Two radiologists who were blinded to the clinical and pathological data of cHCC-CCAs performed tumor segmentation manually using ITK-SNAP software (version 3.8.0; http://www.itksnap.org). Radiomic features extraction were performed by the PyRadiomics R package. 100 patients were randomly selected from the entire cohort to evaluate the consistency of regions of interest by two radiologists using intra-class correlation coefficient to test reproducibility. Subsequently, a total of 2550 radiomic features for each patient from tumor in arterial phase, portal venous phases and delayed phase were extracted. The radiomic features were comprised of 32 non-texture features (shape and size, and first order statistic) and 75 texture features (Grey Level Co-occurrence Matrix (GLCM), Grey Level Run Length Matrix Features (GLRLM), Gray Level Size Zone Matrix (GLSZM), Neighboring Gray Tone Difference Maxtrix (NGTDM) and Gray Level Dependence Matrix (GLDM)). The values of the radiomic features were standardized by z-score method. Patients in two institutions were both scanned with 1.5T MR scanner (uMR 560, United Imaging Healthcare). Routine liver protocols consisted of transverse T2-weighted imaging (T2WI), T1WI, in-phase and opposed-phase sequences, and diffusion-weighted imaging (DWI, b value = 0, 50, and 500s/mm2). The detailed parameters of each acquisition sequence were summarized in Table S1. Establishment and validation of prognostic radiomic model We adopted a two-step strategy to select prognose-related radiomic features. First, we used univariate Cox regression analysis to identify radiomic features with an adjusted p-value (q-value) < 0.05. Second, the radiomic features after initial selection were included for least absolute shrinkage and selection operator (LASSO) regression with the glmnet R package. “Cox” was set as the family in the algorithm. The coefficients for the final selected radiomic features were applied to calculate the prognostic radiomic score (Radscore) for each patient. To determine the best cutoff value of stratifying patients into low-risk and high-risk groups. We used the surv_cutpoint function of the survminer R package based on the maximal log-rank test. The overall survival (OS) differences were compared by the log-rank test. Kaplan-Meier analysis was used to evaluate the association of survival with prognostic Radscore. Multivariate Cox proportional hazard model was adopted in the training and validation cohorts. Weighted gene co-expression network analysis The RNA sequencing was detailed in supplementary material. We filtered out genes with lower expression levels in cHCC-CCA, and use the standardized gene expression matrix as the input matrix. The input matrix was used to construct a hierarchical clustering tree to discover outliers. The selection of soft threshold was realized by “pickSoftThreshold”, and then the correlation matrix was transformed into adjacency matrix and scale-free, so as to divide genes with similar expression patterns into different modules. Using topological overlap matrix (TOM) to calculate the degree of association between genes, the topological overlap matrix and dissimilarity matrix are obtained. On this basis, we used the dissimilarity matrix to construct the tree graph of hierarchical clustering graph and divide and merge the modules. The division of modules adopts dynamic pruning algorithm. Module epigengene (ME) refers to merging modules with high similarity in the first principal component values. During the merging process, the minimum number of genes in each gene module is 30. Use the ME of each module to perform correlation analysis with z-score normalized radiomic features. Calculate the Spearman correlation coefficient and corresponding p-value between radiomic features and functional modules to elucidate the biological significance of radiomic features. Select modules that were highly related to radiomic features for further analysis, and use the clusterProfile R package for Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis to explore the potential function of genes in key modules. Adjusted p-value < 0.05 was considered statistically significant. Enrichment scores calculation Enrichment scores of the hallmark gene were evaluated using the gene set variation analysis. The repressed T cell score and activated T cell score were calculated based on 35 repressed markers and 28 activated GZMK+CD8+ T cell genes [15], respectively. Marker genes used to compute other signatures included: hepatocyte (ALB, AFP, HAMP, and ARG1), cholangiocyte (CK7, and CK19), iCCA stem cell (ID3, CD33, CD44, and ZEB1), embryonic stem cell (OCT4, SOX2, NANOG, and KLF4), myeloid-derived suppressor cells (MDSCs) (CCR2, CCR5, CD38, CD39, and CD40). Estimation of immune infiltration and tumor purity In order to determine the immune composition, we utilized the R package called ‘CIBERSORT’ and applied it to the normalized expression matrix. The ESTIMATE R package was employed to compute both the immune and stromal scores. Cancer associated fibroblasts (CAFs) were estimated by the EPIC R package. Immunohistochemical staining To assess expression levels of various protein markers, immunohistochemical staining was performed according to standard protocols, as detailed in supplementary material. Statistical analysis Continuous variables with a normal distribution were compared using student’s t-test, while the Mann-Whitney U test was applied for non-normally distributed continuous variables. All statistical tests were two-sided, and a p-value < 0.05 was considered statistically significant. Statistical analyses were performed using R software. Significance levels are denoted as follows: *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001. The complete code used for data analysis and radiomic model construction has been deposited in a publicly accessible repository (https://github.com/wangyi0614/Radiogenomics-analysis-of-cHCC-CCA-and-HCC). Results Transcriptome profiles of HCC without rim-APHE, HCC with rim-APHE and cHCC-CCA Principal component analysis revealed that cHCC-CCA samples were distributed in the middle between HCC and cholangiocarcinoma. Compared to HCC without rim-APHE, the proportion of HCC samples with rim-APHE that were closer to cHCC-CCA was higher (Fig. 2A). GSVA analysis showed that metabolism-related pathways and tumor-related pathways were generally highly expressed in cHCC-CCA, compared with two subtypes of HCC. However, immune-related pathways were enriched in HCC without rim-APHE (Fig. 2B). Considering the heterogeneity of the origins of these three types of liver tumors, we assessed the expression levels of different cell marker genes. We observed that HCC without rim-APHE expressed the highest level of hepatocyte markers, whereas cHCC-CCA expressed the highest level of cholangiocyte and iCCA stem cell markers. Meanwhile, embryonic stem cell markers were found remarkably highly expressed in cHCC-CCA and HCC with rim-APHE (Fig. 2C). The cell subset composition of tumor microenvironment was estimated by utilizing transcriptomic data. The immune cell signatures presented a remarkable difference in tumor microenvironment among three groups. cHCC-CCA was characterized by a higher abundance of M2 macrophages and lower abundance of CD8+ T cells, which indicated the suppressive tumor immune signature (Fig. 2D). cHCC-CCA expressed the lowest activated T cell markers, whereas repressed T cell markers were highly observed in the cHCC-CCA and HCC with rim-APHE (Fig. 2E). Furthermore, we identified 55 consecutive advanced HCC patients who received combined TKI and anti-PD-1 antibody treatment. A total of 55 HCC lesions were analyzed, of which 18 exhibited rim-APHE and 37 did not. Among the 55 patients, 25 achieved an objective response (complete or partial response), while 30 had stable or progressive disease. Notably, rim-APHE was observed in 3 of the 25 responders (12%) and in 15 of the 30 non-responders (50%). This distribution suggests that the absence of rim-APHE may serve as a potential predictor of response to combined targeted therapy and immunotherapy (p = 0.003, Fig. 2F). The transcriptome profile of cHCC-CCA, HCC with rim-APHE and HCC without rim-APHE. (A) Principal component analysis showing distribution of cHCC-CCA, cholangiocarcinoma, HCC with rim-APHE and HCC without rim-APHE using RNA-seq data. (B) GSVA for three groups. (C) Boxplots showing expression of the cell marker genes across different subtypes, including the gene sets of hepatocyte, cholangiocyte, iCCA stem cell and embryonic stem cell. (D) Boxplots showing the 22 types of immune cells infiltration in three groups. (E) Boxplots presenting the activated T cell and repressed T cell scores in three groups. (F) Boxplots showing the responses of HCC with rim-APHE and HCC without rim-APHE to the combined targeted therapy and immunotherapy Distinct molecular characteristics and prognosis of cHCC-CCA subtypes Previous studies suggested that LI-RADS category was an independent risk factor for cHCC-CCA prognosis [16, 17]. cHCC-CCA that was categorized into LR-M was considered to have more proportion of iCCA and worse prognosis compared to LR-4/5 subtype of cHCC-CCA. In the transcriptomic cohort, 12/25, 11/25, and 2/25 were categorized into LR-4/5, LR-M and LR-TIV, respectively. Genes with expression correlated to LR-M subtype were significantly enriched for metabolism-related pathways. Immune-related pathways were enriched in LR-4/5 subtype (Fig. 3A, B). LR-M subtype exhibited higher expression of embryonic stem cell genes. NESTIN as a progenitor cell marker was reported to have diagnostic value and serve as an indicator of poor prognosis [18]. Hence, we evaluated the expression level of NESTIN in the two subtypes of cHCC-CCA. We found that LR-M subtype had higher expression level of NESTIN, further demonstrating the aggressive malignancy of this subtype (Fig. 3C). Besides, the stromal score of LR-M subtype was higher than that of LR-4/5 subtype, while the immune score was lower than LR-4/5 subtype. Therefore, LR-M subtype might have a “cold” tumor immune microenvironment (Fig. 3D). Higher percentage of cells were stained positively for PD-1 and CD8 (PD-1+ CD8+ T cell) in LR-M subtype of cHCC-CCA samples (Fig. 3E). Notably, the prognosis of the LR-M subtype was relatively the worst, while the prognosis of HCC without APHE was relatively the best (Fig. 3F). Heterogeneous analysis of LR-4/5 subtype and LR-M subtype of cHCC-CCA. (A) Volcano plot showing the differentially expressed genes in the LR-M subtype of cHCC-CCA compared with LR-4/5 subtype of cHCC-CCA. (B) Barplots showing the significantly correlated pathways with the LR-4/5 subtype and LR-M subtype of cHCC-CCA. (C) Boxplots showing expression of the cell marker genes across two subtypes of cHCC-CCA, including embryonic stem cell and the transcriptomic expression of NESTIN. (D) Boxplots showing the immune score and stromal score across two subtypes of cHCC-CCA. (E) Representative images of anti-CD8 and anti-PD-1 staining in the LR-4/5 and LR-M subtypes of cHCC-CCA, The percentage of PD-1+ CD8+ T cells was quantified and compared between the two subtypes. (F) Overall survival curves and disease-free survival curves showing the distinct prognosis results of the four groups Biological pathways underlying prognostic radiomic features We used two datasets to evaluate the prognostic value of the radiomic features for cHCC-CCA patients: cHCC-CCA cases from ZS cohort were served as training cohort (n = 120) and NT cohort as external validation cohort (n = 61). Eight radiomic features were finally selected for building prognostic Radscore (Table S2). We split patients from the two cohorts based on their Radscore into high-risk group and low-risk group. Patients stratified by Radscore presented distinct OS differences in the training cohort and external validation cohort (p < 0.0001 and p = 0.016, respectively; Fig. 4A, B). The multivariate Cox proportional hazards model also revealed that high Radscore independently predicted worse overall survival both in training and validation cohorts (Table 1). Radiogenomic map of correlation between gene modules and radiomic features. (A, B) Kaplan-Meier curves show the overall survival of cHCC-CCA patients in low-risk and high-risk group in the training cohort and validation cohort. (C) WGCNA result demonstrated the correlation between gene modules and eight radiomic features. (D, E) Pathway analysis of hub genes in pink module correlated with F1, brown module correlated with F5, greenyellow module correlated with F8 Table 1Multivariate cox proportional hazards regression analysis for overall survival in the training and validation cohorts of cHCC-CCA patientsVariablesTraining cohortHR (95% CI)p valueValidation cohortHR (95% CI)p value Age (y) ≤ 6011> 601.07 (0.95 ~ 1.20)0.480.89(0.34 ~ 1.87)0.84 AFP ≤ 400 ng/mL11> 400 ng/mL0.63 (0.07 ~ 5.56)0.671.56(1.02 ~ 2.56)0.78 Microvascular invasion Absent11Present2.12 (0.54 ~ 5.67)0.021.87(1.23 ~ 7.85)0.01 Macrovascular invasion Absent11Present14.64 (1.13 ~ 25.65)0.013.29(1.23 ~ 7.47)0.01 LI-RADS LR-4/511LR-M2.38 (1.87 ~ 8.26)0.024.87(2.49 ~ 7.98)0.03LR-TIV5.37 (2.76 ~ 10.23)0.048.34(3.87 ~ 17.29)0.02 Radiomic score Low11High4.31 (2.49 ~ 12.45)0.0025.47 (2.08 ~ 7.23)0.006 Multivariate cox proportional hazards regression analysis for overall survival in the training and validation cohorts of cHCC-CCA patients We obtained 14 gene modules in the WGCNA (Fig. 4C), and radiomic features were treated as phenotype. Pink module was most highly correlated with F1 (r = 0.55, p = 0.01), brown module with F5 (r = 0.43, p = 0.02) and green-yellow module with F8 (r = 0.45, p = 0.02). Subsequently, we obtained hub genes in each gene module. The F1 hub genes were enriched in lipid metabolism (Fig. 4D), F5 hub genes were enriched in pathways regulating pluripotency of stem cells (Fig. 4E) and F8 hub genes were enriched in T cell receptor signaling pathway (Fig. 4F). The results interpreted the diverse biological processes underlying radiomic features. Metabolic reprogramming and clinicopathological characteristic associated with F1 Since pink module was most significantly associated with radiomic features, we analyzed the association of F1(Original_Glszm_GrayLevelNonUniformity) value with cHCC-CCA characteristics in the entire cohort. F1 is a texture feature derived from the Gray Level Size Zone Matrix. It measures the variability of gray-level intensity values across the “size zones” within the tumor. A “size zone” is defined as a group of connected pixels with the same gray level intensity. Structurally, a higher F1 value indicates a significant lack of uniformity in the distribution of intensity values across these zones. Visually, this corresponds to a highly heterogeneous enhancement pattern on MRI, rather than a smooth, uniform signal. Representative cases of HCC patients with lower F1 values or higher in the arterial phase of contrast enhanced MRI were shown in Fig. 5A. Patients were classified into F1_high or F1_low group according to the median value of F1 value. The prognosis result was validated in the training cohort and validation cohort (p = 0.041 and p = 0.015, respectively; Fig. 5B). GSVA and KEGG analysis demonstrated the similar results that F1 was remarkably associated with elevated lipid metabolism and suppressed immune-related pathways (Fig. 5C, D). We found that the LR-4/5 subtype included more F1_low patients, and the LR-M subtype included more F1_high patients (p = 0.012; Fig. 5E). F1_high group took much larger proportion in MVI positive group (p < 0.001; Fig. 5F). Overall, we demonstrated that F1_high group represented more aggressive behavior of cHCC-CCA. F1 identified as a lipid metabolism related radiomic feature. (A) Representative cases of F1_low and F1_high cHCC-CCA in the arterial phase of contrast-enhanced MRI. (B) Survival analysis between two F1 groups in training cohort and validation cohort. (C) Enrichment of pathways in F1_high group compared with F1_low group using GSEA based on KEGG. (D) Heatmap of lipid-associated metabolism pathways calculated by GSVA for F1 groups. (E) Distribution of F1 group in LR-4/5 subtype and LR-M subtype of cHCC-CCA. (F) Distribution of F1 group in MVI− cHCC-CCA and MVI+ cHCC-CCA Distinct tumor microenvironment in F1 groups We further depicted the immune landscape across the 22 immune-related cell types. The results showed that F1_high had a higher abundance of M2 macrophages but lower abundance of CD8+ T cells (Fig. 6A). Immune score was lower but stromal score was higher in F1_high group (Fig. 6B, C). MDSCs as another suppressed immune cell type also had a higher abundance in F1_high group (Fig. 6D). The similar result was also found with CAF abundances (Fig. 6E). cGAS-STING as an innate immunity sensing pathway was weakly activated in F1_high group (Fig. 6F). Furthermore, we investigated the immune co-stimulators and immunosuppressed marker genes in the two groups. We found that immune co-stimulators were highly expressed in F1_low group while immunosuppressed marker genes such as BTLA and LAG3 were highly expressed in F1_high group (Fig. 6G). Taken together, F1 was a radiomic feature characterized by synergistic biological effect of immune and metabolism. Landscape of the tumor microenvironment of the F1 groups. (A) Differences in the abundance of 22 immune cell types in F1_high group compared with F1_low group. (B, C) Scores of the immune signature and stromal signature calculated by ESTIMATE between F1 groups. (D, E) Comparison of the abundance of MDSCs and CAFs between F1 groups. (F) Comparison of signature scores of cGAS-STING between F1 groups. (G) mRNA expression levels of immunosuppressed marker genes (left) and immune co-stimulators (right) between F1 groups Discussion We have two objectives in the present study. In the first objective, we reported the heterogeneity of HCC with rim-APHE, HCC without rim-APHE and cHCC-CCA. In the second objective, we selected 8 radiomic features derived from arterial phase and portal venous phase of contrast enhanced MR and integrated them into a prognostic radiomics signature. Radiogenomic analysis was performed to establish correlation between radiomics features and biological pathways. In-depth analysis of HCC subtypes reflected that HCC with rim-APHE was frequent with more stromal component and less immune component. A histopathologic study also consistently showed that HCC with rim-APHE exhibited lower proportion of microvessels and poor differentiation compared with HCC without rim-APHE [19]. These results explained why HCC patients without rim-APHE were more responsive to immune therapy. Jeon et al. also supposed that HCC with rim-APHE was an intermediate subtype between HCC and iCCA [10]. The above results indicate that two subtypes of HCC were characterized with different molecular features. We added cHCC-CCA to the comparison category. Interestingly, our result showed that cHCC-CCA had the highest stromal score. Compared with HCC with rim-APHE, cHCC-CCA has repressed tumor immune microenvironment with more M2 macrophages infiltration and several active metabolism pathways like glycolysis and fatty acid metabolism. These characteristics partly explained the poor survival outcome of cHCC-CCA. More specifically, LR-M subtype of cHCC-CCA had higher stromal score and higher level of stem-like features than LR-4/5 subtype. LR-M subtype may be a more aggressive subtype of cHCC-CCA and may response worse to immune therapy. Our findings suggested the complexity of differentiation of primary liver cancer. Additionally, adopting more granular prognostic or etiological categorizations for HCC could potentially refine the transcriptomic clustering results and provide a more nuanced comparison between HCC and cHCC-CCA [20]. At present, interpretation of biological meaning of radiomic features is mainly composed of genomic data, immunohistochemistry, local pathologic analysis and habitat imaging [21–23]. Correlating radiomic feature with genomic data is highly sensitive to many possible biological functions. There are multiple methods to explore the relationship between radiomic features and biological pathways. One study used bi-clustering method to identify coherently expressed radiomic features and biological pathways [24]. Beig et al. calculated single-sample Gene Set Enrichment Analysis (ssGSEA) scores to correlate radiomic features with biological processes [25]. However, WGCNA serves as an ideal tool allowing for the identification of specific gene modules correlated with tumor features [26, 27]. WGCNA has been demonstrated as a useful tool to investigate the relationship between radiomic features and gene expression. Sun et al. suggested that prognostic radiomics signature of glioblastoma were involved in multiple key pathways like immune regulation, treatment responses, tumor regulation, etc [28]. Wu et al. found that the enhancement patterns of tumor adjacent parenchyma of breast cancer was correlated with biological signaling pathways [29], In our study, we used 25 cHCC-CCA samples to build gene co-expression network, the number of samples is reasonable since WGCNA generally requires more than 15 samples. Three radiomics features were significantly correlated with three distinct gene sets, which were fatty acid metabolism, repressive immune regulation and pathways regulating pluripotency of stem cells. These results indicated that radiomic features had the potential for reflecting intratumor heterogeneity of cHCC-CCA. In addition, activation of these pathways was associated with poor prognosis, which explained why the survival risk coefficients of three radiomics were positive. Since F1 had the strongest correlation with gene set of pink module (r = 0.55) compared with F5 with brown module (r = 0.43) and F7 with yellow module (r = 0.45), We selected radiomic feature F1 for in-depth analysis based on its association with intra-tumoral structural heterogeneity. Given that fatty acid metabolism is pivotal in HCC evolution and that lipid metabolism intermediates are known to modulate immune responses within the tumor microenvironment, we further investigated the biological correlates of F1. The results suggest that F1 may represent a non-invasive imaging biomarker associated with the metabolic and immune heterogeneity of the tumor microenvironment. Generally, we found high agreement between radiomics and genomics with the validation of immunohistochemistry. Other studies also explored the value of radiomic to predict tumor microenvironment. Jiang et al. reported that peritumoral heterogeneity of triple-negative breast cancer was associated with immune suppression [30]. Li et al. constructed prognostic MRI radiomic model in predicting the proportion of macrophage infiltration in gliomas [31]. Moreover, multiple studies demonstrated that radiomic signature had the potential to serve as biomarkers for predicting immunotherapy response in melanoma, non-small cell lung cancer, gastric cancer, hepatocellular carcinoma, etc [32–35]. The major limitation of our study is the relatively small number of cases that were collected for RNA-seq due to rarity of the cHCC-CCA. Therefore, our results need to be validated in the collection of a large number of this disease to avoid the risk of overfitting. Another limitation is that the tumor microenvironment of the subtypes of HCC and cHCC-CCA should be deciphered more accurately with single-cell sequencing which can better illustrate intratumor heterogeneity. Lastly, clinical validation of our results in cHCC-CCA patients from immunotherapy cohort should be completed. In conclusion, we illustrated the heterogeneity of subtypes of HCC and cHCC-CCA, which provide insights into the differentiation of primary liver cancer. In addition, for the first time, we established the correlation of radiomic features with transcriptome levels in cHCC-CCA. We identified a radiomic feature which can differentiate heterogeneity in both metabolic and immunologic characteristics within tumor. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1. Abbreviations Hepatocellular carcinoma Combined hepatocellular cholangiocarcinoma Weighted gene co-expression network analysis Liver imaging reporting and data system Overall survival Gene set variation analysis Arterial-phase hyperenhancement Myeloid-derived suppressor cell Cancer associated fibroblasts Intrahepatic cholangiocarcinoma Least absolute shrinkage and selection operator Tyrosine kinase inhibitor Computed tomography Magnetic resonance imaging α-Fetoprotein Microvascular invasion Author contributions Experimental design: WY, MSC and WX. Collection and curation of radiological and clinical data: WY. Data analysis and interpretation: MSC and WX. Project supervision: LWR and SYH. Manuscript writing: WY and MSC. All authors contributed to the article and approved the final manuscript. Funding This work was supported by grants from the National Natural Science Foundation of China (grant number 82171897, 82272078), Shanghai Municipal Health Commission (grant number 202240152), Youth Foundation of Shanghai Municipal Health Commission (grant number 20204Y0346) and Science and Technology Commission of Shanghai Municipality (grant number 23Y11907400). Data availability The data that support the findings of this study are available from the corresponding author upon reasonable request. Declarations Ethics approval and consent to participate This study was conducted in accordance with the ethical principles of the Declaration of Helsinki. The study was approved by the Research Ethics Committee of Zhongshan Hospital and Affiliated Hospital of Nantong University. The study was conducted as a retrospective analysis of anonymized routinely collected clinical and laboratory data. Due to the retrospective design of the study and the use of anonymized patient data, the requirement for informed consent was waived by the Research Ethics Committee of Zhongshan Hospital and Affiliated Hospital of Nantong University. Consent for publication All authors have reviewed and approved the final manuscript and consent to its publication. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Yi Wang, Shichen Miao and Xiao Wang contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement The data that support the findings of this study are available from the corresponding author upon reasonable request.