2.1 Clinical summary

We collected iCCA samples and the paired adjacent normal tissues from 116 patients who received liver resection without any anticancer treatments prior to surgery and were pathologically diagnosed with iCCA at Zhongshan Hospital (Shanghai, China), and the follow-up ended in December 2022. We obtained tissue samples and promptly stored them in liquid nitrogen, within 30 min of resection. All collected samples were intratumoral specimens and confirmed by pathologists from Zhongshan Hospital. Additionally, there were two interruptions in perfusion, each lasting for 15 min during the surgical procedure. The sample collection procedure was consistent across all cases. Patients enrolled in this study all signed the informed consent form to allow the use of their data and samples.

The medical ethics committee of Zhongshan Hospital reviewed and approved the use of all the human tissues involved in this research (B2021-611). The baseline clinicopathological characteristics of the 116 patients with iCCA are shown in Supplementary Tables S1-S2.

2.2 Sequencing

2.3 Quality control and quantification of sequencing data

Genomic and transcriptomic data were sequenced for fragment quality using FastQC software (https://www.bioinformatics.babraham.ac.uk/projects/fastqc/) to filter out low-quality reads using the FASTX-Toolkit (v 0.0.13, http://hannonlab.cshl.edu/fastx_toolkit/). WES data were analyzed as follows: the Burrows-Wheeler Alignment Tool [17] (v0.7.17, https://bio-bwa.sourceforge.net/) was used to map all sequenced read segments to the hg38 reference genome, with the main algorithm implemented using either spaced-seed indexing or Burrows-Wheeler transform (BWT) techniques. Data pre-processing, including the removal of repetitive sequences, was performed according to the Best Practices Workflows recommended by the Genome Analysis Toolkit (GATK) [18]. We used GATK (v4.1.5.0, https://gatk.broadinstitute.org/hc/en-us) to identify single nucleotide variants (SNVs) in tumor tissues and adjacent normal tissues, using the GATK-recommended af-only gnomad.hg38.vcf with somatic-hg38_1000g_pon.hg38.vcf as a reference. Variant call format (Vcf) was further performed to screen for normal population germline variants and to annotate them with annotate variation (ANNOVAR) (v2019-10-24, http://www.openbioinformatics.org/annovar). Finally, 18,593 SNVs were obtained by screening variant allele frequency (VAF) > 5% and supporting mutations of reads > 10 in tumor tissues, VAF < 1% and variant reads < 5 in adjacent normal tissues. Tumor cellularity and tumor ploidy were estimated by sequenza-utils and sequenza, as well as copy number detection (allele specificity and total copy number) and quantification using the adjacent normal tissues as references.

We chose Salmon [19] (v1.5.2, https://combine-lab.github.io/salmon/) and the reference genome described above for transcriptomic quantitative analysis. The Salmon tool quantifies reads by matching directly to the transcript (cDNA) sequences through a fast and lightweight quasi-mapping program, with the following two main steps to run: build index and sample quantification. Subsequent gene expression analysis was performed using the R package tximport (v1.22.0, https://bioconductor.org/packages/release/bioc/html/tximport.html) [20].

Raw proteomic data were searched using MaxQuant [21] (v1.6.1.0, https://www.maxquant.org/) against UniProt database reference files (Version: 202002; 20,367 sequences, https://www.uniprot.org/). The relevant parameters were set as follows: carbamidomethyl (C) was set as a fixed modification while oxidations (M) and acetyl (protein N-term) were set as variable modifications. MS/MS tolerance (Ion trap mass spectrometry, ITMS) was set to 0.5 Da. Searches allowed for up to 2 missed lysine/arginine cleavages by trypsin (full specificity). The false discovery rate (FDR) for peptides and proteins was set to less than 0.01. Pollution and reverse library results were excluded.

Further proteomic analysis was performed using intensity-based absolute quantification (iBAQ) values, normalized by quantile and treated with log2. We selected proteins expressed in at least 25% of samples while removing liver-specific genes [22, 23] and imputing missing values using MissForest (v1.5, https://cran.r-project.org/web/packages/missForest/index.html) [24].

2.4 General proteomic and metabolic subtyping

Proteins used for general proteomic molecular subtyping were selected based on their coefficient of variation (CV) among samples. The abundance matrix of the top 1,500 proteins with the highest CV values was used as input features after performing min-max normalization across the samples, and the non-negative matrix factorization (NMF) algorithm (https://cran.r-project.org/web/packages/NMF/index.html) was used for subtyping. When k = 3, cophenetic values reached their maximum and began to decline, while silhouette values ranged between 0.64 and 1.00 (Supplementary Figure S1A); thus, we selected three clusters as candidate proteomic clusters.

At the same time, we also used the protein abundance matrix of the top 1,500 proteins with the highest CV, which was identified in at least 50% of the samples and normalized by rowwise min-max scaling, as the input for NMF. When k = 4, cophenetic values reached their maximum and began to decline, while silhouette values ranged between 0.60 and 1.00 (Supplementary Figure S1B); then, we also chose four clusters as candidate general proteomic clusters. Thus, for the proteomics clusters, with the different thresholds of missing values, we obtained two sets of clusters for iCCA: one had three clusters, and the other had four clusters.

For metabolic subtypes, we mainly tried two strategies. For the protein abundance-based strategy, three metabolic protein abundance matrices were selected as input. We input the metabolic proteins in the Kyoto Encyclopedia of Genes and Genomes (KEGG) database [25], the top 1,500 variant metabolic proteins in the Human-GEMs [26], and the 412 variant proteins overlapping between the top 1,500 variant proteins used in the Human-GEMs and top 1,500 proteins used in the three proteomic clusters. For the gene set variation analysis (GSVA) (v1.38.2, https://bioconductor.org/packages/release/bioc/html/GSVA.html) [27] pathway enrichment score-based strategy, 69 metabolic pathways in the c2.cp.kegg.v7.4.symbols.gmt gene set in MsigDB (https://www.gsea-msigdb.org/gsea/msigdb) were selected as the metabolic gene set. We compared all these input silhouette values with silhouette values ranging from 0.77 to 1.00 (Supplementary Figure S1C) when using the top 1,500 variant metabolic proteins in the human genome-wide model as NMF input. Three metabolic subtypes were defined from this input.

2.5 OS, disease-free survival (DFS), and prognostic gene analysis

To compare survival differences in proteomics clusters and metabolic subtypes, Kaplan-Meier curves and log-rank tests were used. The endpoint of the study was OS. OS was defined as the interval between the date of diagnosis and the date of patient death or the last follow-up. DFS was defined as the interval between the date of surgery and the date of disease recurrence or death for any reason or the last follow-up (whichever occurs first). We calculated the hazard ratio (HR) for each gene and corrected the log-rank P value with the Benjamini-Hochberg (BH) algorithm using geneSA (v0.1.1, https://github.com/huynguyen250896/geneSA) [28].

2.6 Metabolic pathway variation analysis of the three metabolic subtypes

To characterize the unique metabolic function of each metabolic subtype, we performed GSVA and selected c2.cp.kegg.v7.4.Symbols.gmt in MsigDB as the gene set together with the gene set composed of metabolic genes in Human-GEMs. To determine the pathway differences between subtypes, we used the R package Limma (v3.46.0, https://bioconductor.org/packages/release/bioc/html/limma.html) [29] to analyze the upregulated and downregulated pathways of each subtype [P < 0.001, corrected with BH algorithm; log₂fold change (log₂FC) > 0, upregulated; log₂FC < 0, downregulated] compared to the other two subtypes.

We also analyzed the differentially expressed proteins of each subtype, which we identified using DeqMS (v1.8.0, https://www.bioconductor.org/packages/release/bioc/html/DEqMS.html) [30] by comparison with the other two subtypes. Proteins with log₂FC > 0.5 and P < 0.05 were considered upregulated, and those with log₂FC > 1 and P < 0.05 were considered highly upregulated in the subtype.

2.7 Construction of metabolic subtype predictors and metabolic protein signatures of the metabolic subtypes

The proteomic data of this study were used as the training set, and supplemental data from Dong et al. [31] were used as the validation set. We first classified the validation set into three metabolic subtypes using the same metabolic stratification method and aligned the subtype label based on the number of overlapping upregulated proteins with our cohort data. To determine the correspondence between subtypes in the validation set and those in the training set, we compared the number of overlapping differential proteins. In the validation set, we observed that 317 out of the 372 upregulated proteins (log₂FC > 1, P < 0.05) identified in the S1 training set exhibited upregulation in one of the subtypes (log₂FC > 0, P < 0.05). Consequently, we designated this subtype as the “S1 validation set”. The remaining two subtypes were assigned in the same way, with the S2 validation set overlapping 231 of the 295 upregulated proteins in the S2 training set and the S3 validation set overlapping 547 of the 1,148 upregulated proteins in the S3 training set. The identification of differentially expressed proteins in the validation subtypes was performed using limma, with the same comparison method being used in the training set.

We then constructed metabolic subtype predictors and performed molecular screening of the metabolic signature proteins for each metabolic subtype based on the permutation-based feature importance test (PermFIT) method (https://github.com/SkadiEye/deepTL) [32]. Before predictor construction, we performed feature selection for each subtype, screening for highly upregulated metabolic genes in each subtype in our cohort. Here, metabolic genes were genes in c2.cp.kegg.v7.4.symbols.gmt and genes in the Human-GEMs (version 1.8.0). For each subtype, we screened subtype-specific metabolic signatures using PermFIT-support vector machines (SVMs), PermFIT-deep neural networks (DNNs), and PermFIT-random forests (RFs) and validated the predictive ability of signatures in the validation set. To improve the predictive ability, we replaced one signature based on permutation screening with one prognosis-related metabolic protein for each subtype. For the S2 and S3 training sets with multiple highly upregulated proteins with prognostic implications, the metabolic protein with the highest hazard ratio (HR) was selected for S2, and the metabolic protein with the lowest HR was selected for S3. We finally used the set of metabolic protein signatures with the best accuracy, as well as the area under the curve (AUC) evaluation in the validation set for each subtype.

2.8 Rate-limiting enzyme annotation

A union set of 3,745 metabolic genes was extracted from the Human-GEMs and the KEGG database. Then, a total of 2,728 metabolic genes identified in our cohort remained for further analysis. Based on the gene name and the keyword “rate-limiting”, the abstracts and titles of relevant literature were retrieved and collated from PubMed. Here, we only considered metabolic proteins upregulated in each metabolic subtype such that a total of 293 proteins had retrieval results. After the manual check, we eventually screened a total of 112 upregulated metabolic rate-limiting enzymes in the three subtypes (log₂FC > 1, P < 0.05). We classified them according to the KEGG pathway annotations. For glycolysis and the tricarboxylic acid cycle (TCA) pathway, we only focused on the three commonly considered essential rate-limiting enzymes and relaxed log₂FC to > 0.5.

2.9 The prognostic gene-related pathway

Metabolic genes (HR > 1, log-rank P < 0.01) annotated in the Human-GEMs or KEGG database were considered high-risk metabolic genes significantly associated with prognosis; 62 of these genes were identified, and 53 low-risk metabolic genes (HR < 1, log-rank P < 0.001) were significantly associated with prognosis. The genes were submitted as a gene list to Metascape (https://metascape.org/gp/index.html) [33] for pathway enrichment analysis. The ten KEGG pathways with the lowest q value are shown.

2.10 Metabolic network construction and analysis

We used Human-GEMs [26] as a reference metabolic network model consisting of 142 metabolic subsystems, 3,626 genes, 8,379 metabolites of different cellular sublocalizations, and 13,084 metabolic reactions. Proteomic and transcriptomic individual metabolic networks were extracted using the task-driven integrative network inference for tissues algorithm (tINIT) [26, 34]. For transcriptomic data, we followed the tutorial at the website https://sysbiochalmers.github.io/Human-GEM-guide/gem_extraction_old_tINIT/. For proteomic metabolic networks, we used the iBAQ values for proteins identified in at least 25% of samples and set discrete values for the protein abundance of each sample according to the following rule: if there was a missing value, we defined it as “none”; if its rounded percentile was in 0-24%, 25%-74%, or 75%-100%, we defined its abundance grade as “low”, “medium”, or “high”, respectively.

We assessed metabolic reactions across all patient subsystems using the Kruskal-Wallis H-test to compare overall differences and the Nemenyi test to compare differences between every two subtypes. BH significance was defined as a corrected P < 0.05. For metabolic task analysis of metabolic networks, we used the CompareMultiModels function in Raven (v2.6.0, https://github.com/SysBioChalmers/RAVEN) [35] and included metabolic tasks to assess whether the task could be passed in an individually specific metabolic network. The enriched tasks in each metabolic subtype were calculated using Fisher's exact test [26, 36].

We converted the proteomic metabolic network from the raven model format to the Cobra (v3.1, http://opencobra.github.io/) model format [37]. We used the sample CbModel function and the artificially centered hit-and-run (ACHR) sampling method to obtain the sampling metabolic fluxes and then took the average value for each reaction. We retained reactions with a non-zero flux value in at least 25% of samples, filled the missing reaction flux value with 0, merged the reactions whose flux similarity was above 0.95 (Spearman correlation coefficients), and merged the reaction sets with the overlapping reactions. The flux values of the merged reaction sets were averaged.

If the d value was less than 0, it indicated the opposite direction to that defined by the model. When screening for upregulated proteins that also had high flux, we replaced condition (iv) with a ratio of the mean metabolic flux of the S2 subtypes to the mean of non-S2 subtypes that was greater than 1, and the corresponding protein was expressed at log₂FC > 0.5 and P < 0.05.

2.11 Cell annotation and metabolic heterogeneity analysis of single-cell RNA-sequencing (scRNA-seq) datasets

We used HRA00863 [10] and GSE151530 [38], two scRNA-seq datasets of iCCA. The annotated cells in the HRA000863 dataset were kindly provided by the authors, and GSE151530 was reannotated using SingleR (v1.4.1, https://bioconductor.org/packages/release/bioc/html/SingleR.html) [39]. To obtain an accurate classification of CD8⁺ T cells in iCCA, we used previously reported markers of CD8⁺ T cell subtypes [40, 41] and the single-cell cluster-based auto annotation toolkit (scCATCH, v3.2.1, https://github.com/ZJUFanLab/scCATCH) [42] for annotation. We classified CD8⁺ T cells into exhausted (including precursor exhausted) CD8⁺ T cells, naive-like CD8⁺ T cells, early activated CD8⁺ T cells, and effector memory CD8⁺ T cells.

Based on a previous study of single-cell pathway activity [43], we performed pathway activity calculations for CD8⁺ T-cell subtypes using the Pagoda2 (v1.0.10, https://github.com/kharchenkolab/pagoda2) algorithm [44], with a gene set that was previously selected from a metabolic subset of 69 metabolic pathways from c2.cp.kegg.v7.4.symbols.gmt in MsigDB. To compare the differences between subtypes, we used the Kruskal-Wallis H-test to compare the overall differences and the Nemenyi test to compare the differences between every two groups with a significance level of P < 0.05.

2.12 Tumor microenvironment analysis

We performed microenvironment composition analysis using transcriptomic transcripts per million (TPM) data from 116 patients with iCCA and data from Dong's cohort [31]. A web version of xCell (https://comphealth.ucsf.edu/app/xcell) was used in this analysis [45].

In parallel, we used the deconvolutional R package BayesPrism (v2.0, https://github.com/Danko-Lab/BayesPrism) [46] and HRA000863 scRNA-seq datasets to deconvolute the transcriptomic count data of our cohort and transcriptomic TPM data of Dong's cohort [31]. Thirteen cell fractions were deconvoluted: tumor cells, B cells, plasma cells, natural killer cells, muco-associated constant T cells, T regulatory cells, CD4⁺ T cells, CD8⁺ T cells, dendritic cells, monocytes, macrophages, endothelial cells, and fibroblasts. Cell fractions for four subtypes of CD8⁺ T cells were also deconvoluted: exhausted CD8⁺ T cells, naive-like CD8⁺ T cells, early activated CD8⁺ T cells, and effector memory CD8⁺ T cells. Tumor transcriptome expression profiles were obtained after deconvolution.

We calculated overall significant differences using the Kruskal-Wallis H-test, and we assessed the differences between every two groups using the Nemenyi test with a significance level of P < 0.05. We normalized the cell fraction with the sum of the four CD8⁺ cell subtypes when analyzing the composition differences of the four CD8⁺ T-cell subtypes in the metabolic subtypes.

The TME-based subtypes [47] were also generated for the comparison with the results of BayesPrism and xCell. Briefly, 500 differential genes of five TME-based subtypes [47] (100 for each) and NTP algorithm [48] were used to decide the TME-based subtypes for our data.

2.13 Correlation between metabolic pathway activity, metabolic genes, and cell composition in the microenvironment

We used 69 metabolic pathways from the c2.cp.kegg.v7.4.symbols.gmt gene set in MSigDB and used GSVA pathway enrichment scores as pathway activity. We calculated three levels of metabolic pathway activity: the proteomic and transcriptomic levels, and the transcriptomic expression level of tumor cells after BayesPrism deconvolution.

We also analyzed correlations at the metabolic gene level, selecting metabolic immune checkpoints, as well as metabolic genes with consistent prognosis in our dataset and in Dong's cohort data, and again, we calculated three levels of correlation.

To obtain robust correlation results, we referred to previous research methods [49] and set the following criteria for the robust correlation: (1) P < 0.05 for 2,000 permutation tests; (2) P < 0.05 for 2,000 permutation tests after removing any one sample.

We assessed the consistency of metabolic pathway correlations with the cell composition between our dataset and Dong's dataset: 170, 377, and 566 correlations in our cohort were robust using the proteomics, transcriptomics and deconvoluted tumor transriptomic data, respectively, with 74 (43.5%), 219 (58.1%), and 348 (61.5%) also being robust in Dong's cohort.

At the levels of metabolic immune checkpoints, for the proteome, the transcriptome, and the deconvoluted tumor transcriptome, 51, 98, and 105 correlations were robust in our cohort, respectively, of which 33 (64.7%), 63 (64.2%), and 73 (69.5%) were also robust in Dong's cohort.

At the level of metabolic prognosis-related genes, for the proteome, the transcriptome, and the deconvoluted tumor transcriptome, 117, 121 and 165 correlations were robust in our cohort, respectively, of which 58 (49.6%), 66 (54.5%), and 108 (65.5%) were also robust in Dong's cohort.

2.14 Analysis of driver genes and potential functional mechanisms

We used sysSVM2 (https://github.com/ciccalab/sysSVM2) [50] to analyze driver mutations at the individual level in iCCA, selecting 10 candidate driver genes per patient for driver mutation enrichment analysis of metabolic subtypes. In S3, the numbers of patients with BAP1 mutations and IDH1 mutations were 16 and 12, respectively, and a total of 24 patients out of 49 patients. Differentially expressed proteins were identified in patients of S3 with BAP1 mutations (S3_BAP1mut), IDH1 mutations (S3_IDH1mut), and no BAP1 or IDH1 mutations (S3_other) by comparing the protein abundance profiles of these patients to those of patients in non-S3 subtypes (log₂FC > 1, P < 0.05 after BH correction). Metascape was used to assess the functions of S3_BAP1mut, S3_IDH1mut, and S3_other.

2.15 Analysis of cis-regulatory effects

We identified significantly amplified or deleted genes using GISTIC2 (2.0.23) with parameters conf set at 0.99, ta and td set at 0.4, and brlen set at 0.7 [51]. We used iProfun (v0.2.0, https://github.com/songxiaoyu/iProFun) [52] to screen potential cis-regulations, with age, sex and hepatitis B surface antigen as covariates. The copy number variation (CNV), proteomic and transcriptomic expression matrix after z score normalization were used as inputs, the permutation number was set to 1,000, and the FDR was set to 0.1. We then selected proteins in significant deletion and amplification regions with cis-regulation of CNV-mRNA-protein. We performed functional enrichment analysis using Metascape for proteins in the two regions.

2.16 Cell lines and cell culture

Human iCCA cell lines RBE and HuCCT1 were maintained in RPMI-1640 (BasalMedia, Shanghai, China) supplemented with 10% fetal bovine serum (FBS, 10099141C, Gibco, Waltham, MA, USA), 100 U/mL penicillin and streptomycin (100×, 15140122, Gibco) at 37°C with 5% CO₂. All cells were used within ten passages after thawing. The cell lines were authenticated by short tandem repeat profiling before the experiments started and showed no mycoplasma contamination.

2.17 Lentiviral transfection

Specific diacylglycerol kinase α (DGKA) lentiviral short hairpin RNAs (shRNAs) were synthesized by OBiO Technology (Shanghai, China). The shRNAs were cloned and inserted into the GL401 (pCLenti-U6 -CMV-Puro-WPRE) vector, and an empty vector was used as the negative control. RBE and HuCCT1 cells were transfected following the manufacturer's instructions to knock down the expression of DGKA.

2.18 5-Ethynyl-2’-deoxyuridine (EdU) assay

After transfection, iCCA cells were seeded on 12-well plates at a density of 3 × 10⁵ cells/well. The EdU-594 Kit (C0078S, Beyotime, Shanghai, China) was used to detect the proliferation of the indicated iCCA cells. Cells were treated with EdU and Hoechst dyes. A fluorescence microscope (Olympus, Tokyo, Japan) was utilized to observe and record the experimental results.

2.19 Cell proliferation analysis

For cell viability assays, iCCA cells were digested and seeded in 96-well plates (1,000 cells/well) in 5 replicates. The next day, 30 μmol/L ritanserin (HY-10791, MedChemExpress, Monmouth Junction, NJ, USA), 100 μmol/L phosphatidic acid (PA, 840857, Merck, Darmstadt, Germany), or both were added to the culture medium of each treatment group. Then, the cells were cultured in the respective medium for another 96 h. Cell survival was calculated using cell counting kit-8 (CCK8, CK04, Donjindo, Kumamoto, Japan) by measuring the absorbance at 450 nm using FlexStation 3 (Molecular Devices, San Jose, CA, USA).

2.20 Colony formation assay

For colony formation assays, cells were seeded in 6-well plates (1,000 cells/well) in triplicate and incubated for approximately 2 weeks. In the ritanserin group, cells were treated with 30 μmol/L ritanserin, and an equal volume of dimethyl sulfoxide (DMSO, 41640, Sigma-Aldrich, St Louis, MO, USA) was added to the control group. Then, the cells were fixed with 4% paraformaldehyde and dyed with Giemsa stain. The number of colonies in each well was counted by ImageJ software (National Institutes of Health, Bethesda, MD, USA).

2.21 Tissue microarray (TMA) experiment

A TMA containing 90 iCCA samples was established. After ruling out 2 samples with missing data, 88 samples were included in the present study (Supplementary Table S3). Immunohistochemical (IHC) staining was conducted with IHC reagents from Fuzhou Maixin Biotech Co., Ltd. (Fuzhou, Fujian, China). Briefly, TMA was deparaffinized and rehydrated before antigen retrieval, followed by blocking with 3% hydrogen peroxide and 1× Animal-Free Blocking Solution. Then, the sections were incubated with primary antibodies (DGKA, 11547-1-AP, Proteintech Group, Rosemont, IL, USA) overnight at 4°C. The next day, the sections were incubated with the secondary antibody at 37°C for 1 h, and 3,3 N-diaminobenzidine tetrahydrochloride (DAB) staining was conducted. Finally, they were counterstained with hematoxylin for 2 min. Using the PANNORAMIC 1000 slice scanner (3DHISTECH Ltd., Budapest, Hungary), the tissue sections were scanned and imaged. Quant Center 2.1 analysis software (3DHISTECH Ltd.) was used to calculate the H-score of the target area. H-score = ∑(PI × I) = (percentage of cells of weak intensity × 1) + (percentage of cells of moderate intensity × 2) + (percentage of cells of strong intensity × 3), where PI represents the positive area ratio and I represents the staining intensity. The median H-score was defined as the threshold value to distinguish the low and high DGKA expression groups.

2.22 Western blotting

Using radioimmunoprecipitation assay buffer (RIPA) lysis buffer solution (89900, Thermo Fisher Scientific) containing phosphatase and protease inhibitor cocktail, cell lines were dissolved to extract protein. Protein samples were separated by polyacrylamide gel electrophoresis and transferred to polyvinylidene fluoride (PVDF) membranes (IPVH00010, Sigma-Aldrich). After blocking with 5% bovine serum albumin (BSA) solution, the PVDF membrane was incubated with primary antibodies, including DGKA, extracellular regulated protein kinases 1/2 (ERK1/2, T40071, Abmart, Shanghai, China), phospho-ERK1 (T202/Y204) + ERK2 (T185/Y187) (T40072, Abmart), mitogen-activated protein kinase 1/2 (MEK1/2, 11049-1-AP, Proteintech Group), phospho-MEK1/2 (Ser217/221, 9154, Cell Signaling Technology, Danvers, MA, USA), and glyceraldehyde-3-phosphate dehydrogenase (GAPDH, 51332, Cell Signaling Technology), followed by incubation with secondary antibodies (7074, 7076, Cell Signaling Technology) conjugated with horseradish peroxidase (HRP). Enhanced chemiluminescence (ECL) Western blotting substrate (WBKLS0500, Sigma-Aldrich) was used to visualize the signal.

2.23 Bulk RNA-seq analysis for cell lines

TRIzol reagent (15596026, Thermo Fisher Scientific) was used to extract total RNA from DGKA-knockdown iCCA cells and control cells. After ensuring the purity and integrity of all RNA samples, the RNA library was then constructed. Sequencing was performed on the Illumina PE150. We defined significance for downstream analysis as fold change > 1 and P < 0.05. Functional enrichment analysis of KEGG was conducted using ClusterProfiler (v3.4.4, https://github.com/YuLab-SMU/clusterProfiler). Gene set enrichment analysis (GSEA) was carried out with GSEA software (v4.0.3, Broad Institute, Cambridge, MA, USA).

2.24 Statistical methods

The chi-square test and Fisher's exact test were used to analyze categorical variables. Kruskal-Wallis's H-test and Student's t-test were utilized to calculate statistical differences for continuous variables, and Spearman and Pearson's correlations were used for continuous variables versus continuous variables. Survival analysis was performed using the Kaplan-Meier method, the log-rank test, and Cox proportional hazards regression analysis. All statistical tests were two-sided, and a P value < 0.05 was considered statistically significant.

3.1 Metabolic subtypes and protein signatures of iCCA

In our study, we enrolled 116 patients with iCCA and performed WES and quantification of the transcriptome and proteome (Figure 1A). We attempted several approaches to identify the metabolic subtypes in our cohort of iCCA patients. Based on the NMF method and the protein abundance of metabolic genes in Human-GEMs, three metabolic subtypes (S1, S2, and S3) were identified (Figure 1B, Supplementary Table S1). We further explored the relationship between these three subtypes and potential clinical influencing factors. Notably, HBV infection and diabetes were not significantly associated with any subtype, while obese patients tended to be subtype S1 (Supplementary Table S2).

Out of the three metabolic subtypes, individuals belonging to subtype S1 exhibited the highest survival rate (median OS > 60 months), whereas those in subtype S2 suffered from the shortest duration of survival (median OS = 20.5 months) (Figure 1C). General proteomic clusters were also identified for comparison with the metabolic subtypes (Supplementary Table S1). Interestingly, there was a strong concordance between them, indicating a good correspondence between individual changes in the metabolic protein expression profile and the overall protein expression profile (Supplementary Figure S2A). We compared the differences in proteins used in general proteomic subtyping and metabolic subtyping, both of which resulted in three subtypes. A total of 412 proteins were shared. We used these 412 proteins alone for metabolic subtyping again. Its consistency with the Human-GEMs gene-based method was weaker than that with the KEGG metabolic gene-based method but stronger than that with the GSVA-based method (Supplementary Figure S2B). Other metabolic subtyping results also showed significant differences in OS (Supplementary Figure S2C).

To further explore the molecule-level functional characteristics of each subtype, we submitted the highly upregulated genes in each subtype to Metascape [33] for pathway and biological process analysis (Supplementary Table S4). The top 5 most enriched pathways for each subtype are shown in Figure 1D. The results revealed that metabolism-related pathways, such as small molecule biosynthesis and valine, leucine and isoleucine degradation, were enriched in S1, whereas S2 showed pathways predominantly related to metastasis, and S3-enriched pathways were predominantly related to proliferation. This finding was consistent with the different prognostic characteristics of each subtype.

We also used GSVA to perform a preliminary characterization of the metabolic functional differences among the three metabolic subtypes S1, S2 and S3. Consistent with the above results, the subtype S1 exhibited a wider range of metabolic activities compared to the other two subtypes. The proteins in many metabolic pathways and subsystems of amino acid metabolism, carbohydrate metabolism, energy metabolism, lipid metabolism, and other classes of amino acid metabolism were enriched in S1 (Figure 1E). The pathways of glycosaminoglycan biosynthesis-keratan sulfate, O-glycan biosynthesis, protein modification subsystem, and inositol phosphate metabolism were enriched in S2. Glycosaminoglycan biosynthesis-keratan sulfate [53], O-glycan biosynthesis [54] and inositol phosphate metabolism [55] were reported to be associated with tumor metastasis, which led to the worst prognostic phenotype of S2. The biosynthesis pathways of valine, leucine, isoleucine, and aminoacyl-tRNA and the pyrimidine metabolism pathway were enriched in S3 (Figure 1E), consistent with the vigorous proliferation phenotype mentioned above (Figure 1D).

For potential applications in the future, we constructed a metabolic subtype prediction model using the Permutation-FIT method [32] to screen for the most representative metabolic signature proteins for each metabolic subtype. We used our cohort data as the training set and the proteomic data from the Dong cohort [31] as the validation set (Supplementary Figure S2D-E). The AUC of all three subtypes in our validation set reached above 0.85, indicating the reliability of the model and the representativeness of the selected signatures for metabolic subtypes. The accuracy of the predictor reached 0.88 for S1 and 0.77 for S3 (Figure 1F). The metabolic protein signatures in S1 and S2 were involved in a variety of metabolic functions, and NUDT12, ACSS1, and NMNAT3 signatures in S3 were all associated with the process of nucleotide biosynthesis (Figure 1G).

3.2 Significant and wide metabolic heterogeneity among different metabolic subtypes

To explore the heterogeneity of metabolic characteristics in different metabolic subtypes, three aspects, including “rate-limiting enzymes”, “prognosis-related metabolic genes” and “metabolic network model”, were involved in further analysis. Although the correlations between proteomic and transcriptomic expression of metabolic genes were higher than those of non-metabolic genes, the median Spearman coefficient was only 0.49, indicating that it is of unique value to use proteomic data to characterize metabolic subtypes and their heterogeneity since there was a closer relationship between proteome and phenotype (Supplementary Figure S3A).

First, we annotated metabolic rate-limiting enzymes using differential protein expression in metabolic subtypes by collecting and reviewing the related literature (Supplementary Table S5). We found that the rate-limiting enzymes involved in the fatty acid degradation pathway in S1 were upregulated, while the rate-limiting enzymes upregulated in S2 and S3 were proteins involved in the glycolysis/gluconeogenesis pathway. Hexokinase 1(HK1), HK2, and HK3, the enzymes for converting glucose to 6-phosphate fructose in the first rate-limiting glycolysis step, were upregulated in S2, while muscle phosphofructokinase (PFKM) and PFKP, the enzymes for fructose 1,6-diphosphate production in the rate-limiting step of glycolysis, were upregulated in S3. Subtype-specific variation in rate-limiting enzymes suggests the key heterogeneity in energy utilization for iCCA.

All rate-limiting enzymes in the pyrimidine metabolism pathway were upregulated in S3, except for ecto-5’-nucleotidase (NT5E) and ectonucleoside triphosphate diphosphohydrolase 1(ENTPD1), which were upregulated in S2 (Figure 2A). CD39 (ENTPD1) binds extracellular adenosine triphosphate (ATP) and hydrolyses it to adenosine monophosphate (AMP) [56]; another extracellular nuclease, CD73 (NT5E), hydrolyses AMP to adenosine, and the abnormally high concentration of extracellular adenosine promotes tumor proliferation through various immunosuppressive mechanisms; furthermore, inhibition of CD73 and CD39 can enhance antitumor immunity [56]. Since the S2 subtype had the worst prognosis, we further focused on other specific rate-limiting enzymes of this subtype. The rate-limiting enzymes highly upregulated in S2, such as proline dehydrogenase (PRODH), glutaminase 2 (GLS2), and hephaestin (HEPH), were involved in oxidative stress, while arachidonate 5-Lipoxygenase (ALOX5), phospholipase A2, group IVA (PLA2G4A), and cytosolic non-specific dipeptidase 1 (CNDP1) were involved in inflammatory responses. Additionally, CNDP1 and glutamine-fructose-6-phosphate aminotransferase 2 (GFPT2) were involved in the regulation of protein glycosylation (Supplementary Figure S3B).

Second, we analyzed the metabolic genes associated with different prognostic phenotypes (Figure 2B and Supplementary Figure S3C). Pathway enrichment results showed that the metabolic pathways associated with poor prognosis in iCCA were central carbon metabolism in cancer, choline metabolism in cancer, various types of N-glycan biosynthesis, etc. These metabolic genes were also widely involved in important signaling pathways, including the hypoxia-inducible factor-1 (HIF-1) signaling pathway and the Fc epsilon receptor I (FcεRI) signaling pathway. Notably, among the 20 metabolic genes associated with unfavorable prognosis, 12 genes were upregulated in S2 (Figure 2B).

Third, to further analyze the metabolic heterogeneity of iCCA from a global view, we constructed an individual-level tissue-specific metabolic network based on the tINIT algorithm. We compared the differences in the number of reactions in the metabolic subsystems within S1, S2, and S3. There were 136 subsystems in the individual-specific proteomics metabolic network of iCCA, and 48 metabolic subsystems had subtype differences (Supplementary Table S6). Subsystems generally had a higher number of metabolic reactions in S1 and a lower number of metabolic reactions in S2. The heterogeneous subsystems among subtypes revealed by the reaction numbers, such as alanine, aspartate, glutamate, and pyrimidine metabolism, were consistent with previous GSVA results, but there were also inconsistencies in that the number of N-glycoside synthesis metabolic reactions in S2 was not significantly higher than that in other subtypes (Figure 2C).

The metabolic task is a set of reactions needed to convert a specific metabolite into a specific product. The differences in metabolic tasks between patients at the proteomic level were greater than those based on the transcriptome. A total of 256 metabolic tasks were involved, of which 67 were different across patients at the proteome level, 27 were different at the transcriptome level, and 25 were identified as different metabolic tasks in both the transcriptome and proteome (Supplementary Table S7). A total of 21 metabolic tasks were found to have metabolic intersubtype differences at two omics levels, and these tasks were widely related to the metabolism processes of amino acids, sugars, energy lipids, coenzymes and vitamins, and nucleic acids (Figure 2D).

Multiple metabolic tasks in S1 showed significant enrichment, while most of the tasks were extensively depleted in S2. The proteomics and transcriptomics together showed that glycocholate and taurocholic acid de novo synthesis and secretion were significantly enriched in S1; conversely, proline de novo synthesis, NAD(+) de novo synthesis, and NADP(+) de novo synthesis were significantly deficient in S1. Arginine and tyrosine de novo metabolic tasks were absent in S2 (Figure 2D). Proteomic expression evidence verified that arginine and tyrosine de novo synthesis were significantly depleted in S2, as well as proline de novo synthesis in S1 (Figure 2E); in addition, some genes shared by NAD(+) de novo synthesis reactions had lower expression in S1 than in other subtypes (Supplementary Figure S3D). Additionally, the proteomic data showed that the NAD(+) and NADP(+) de novo synthesis tasks were enriched in S2, while the alanine degradation task was enriched in S3 and absent in S1 (Figure 2D). Metabolic tasks enriched or depleted in different subtypes indicated metabolic reprogramming in the formation of subtypes of iCCA. Depleted metabolic tasks in S1, as well as enriched metabolic tasks in S2 or S3, may be associated with poor prognosis and may be potential metabolic targets.

Metabolic fluxes are the immediate rates at which chemical reactions occur, and metabolic fluxes reflect the characteristics of individual reactions from a global metabolic perspective. We simulated the fluxes of metabolic reactions in the metabolic network based on the ACHR method. The simulated metabolic flux data were dimensionally reduced and were found to differ between S1 and S2 (Figure 2F). S2 generally had smaller metabolic fluxes, such as fatty acid oxidation, than other subtypes (Figure 2G). We found that some reactions in prostaglandin metabolism and synthesis and interleukin metabolism in S2 had higher fluxes than other subtypes (Supplementary Figure S3E). Proteins involved in S2 with significantly high fluxes were also upregulated in S2, such as lecithin-cholesterol acyltransferase (LCAT) and solute carrier family 3 member 2 (SLC3A2) (Supplementary Figure S3E). HK1, HK2, HK3, glycerol-3-phosphate dehydrogenase 2 (GPD2), aldehyde dehydrogenase 1 family member L2 (ALDH1L2), and ALDH3B1 were also found to be upregulated and were found to be involved in the high flux reaction (Figure 2H), but most of the upregulation of the metabolic proteins in S2 was not reflected in the simulated fluxes (125 upregulated metabolic proteins in S2). Thus, the sampling fluxes based on the individual metabolic model indicated that from a global metabolic perspective, except for a small portion of upregulated proteins involved in highly active metabolic reactions, the activity of the entire S2 metabolic reaction was significantly lower than the other two subtypes.

3.3 Analysis of microenvironmental composition and metabolic associations in iCCA

The analysis of the microenvironmental composition of different subtypes showed obvious heterogeneity in cell type fractions. Based on the Kruskal-Wallis H-test, we found that the immune and stromal scores (xCell) of the three metabolic subtypes in S3 were significantly lower than those of the other two subtypes (Supplementary Table S8). BayesPrism results supported the findings based on the xCell method as follows: BayesPrism analysis showed that the tumor cell fraction of S3 was significantly higher than that of the other two subtypes (Nemenyi test, P < 0.05); the results of both methods consistently showed that monocytes and macrophages made up significantly smaller proportions in S3 than in the other subtypes (Nemenyi test, P < 0.05); BayesPrism results showed that the proportions of CD8⁺ T cells, dendritic cell (DC), and regulatory T (Treg) cells were significantly higher in S1 than in the other subtypes (Nemenyi test P < 0.05). The characteristics of TME-based subtypes [47] were consistent with our findings: the “immune classical” subtype was enriched in S1 (P = 0.003) and the non-inflamed TME-based “Hepatic stem-like” subtype was enriched in S3 (P < 0.001) (Supplementary Table S9).

Due to the importance of CD8⁺ T cells for iCCA [57, 58], we analyzed the heterogeneity of CD8⁺ T cell subtypes among different patients using the public iCCA single-cell dataset, as well as the BayesPrism algorithm. We first classified CD8⁺ T cells from the public single-cell dataset into the following four subtypes: exhausted CD8⁺ T cells, naive-like CD8⁺ T cells, early activated CD8⁺ T cells, and effector memory CD8⁺ T cells (Supplementary Figure S4A). Both scRNA-seq datasets consistently showed metabolic heterogeneity in the pathways of glycolysis/gluconeogenesis, pyrimidine metabolism, one carbon pool by folate, cysteine and methionine metabolism, and oxidative phosphorylation in the CD8⁺ T-cell subtypes. We found significant differences in all of these pathways between effector memory CD8⁺ T cells and exhausted CD8⁺ T cells. Except for the higher level of oxidative phosphorylation in the effector memory CD8⁺ T cells, the activity of all other metabolic pathways in effector memory CD8⁺ T cells was lower than that in exhausted CD8⁺ T cells (Figure 3A and Supplementary Figure S4B).

Then, we used BayesPrism to deconvolute the cell fractions of CD8⁺ T cell subtypes in each patient in the cohort, and the Kruskal-Wallis H-test revealed that the proportion of CD8⁺ T cells of different subtypes did not differ significantly in patients of different subtypes, but after normalization within CD8⁺ T cells, effector memory CD8⁺ T cells made up a significantly higher proportion in S1 than in S2 (Figure 3B) and naive-like CD8⁺ T cells made up a significantly higher proportion in S2 than in S1 (Figure 3C), indicating the potential trend of internal changes in CD8⁺ T cell status in the microenvironment of iCCA cells with different metabolic states.

To investigate the association between the microenvironment and metabolism in iCCA, we analyzed the correlation of metabolic pathway activity, represented by metabolic gene expression level, with the non-tumor cell fraction. A total of 49 robust correlations were identified between metabolic pathway activities and non-tumor cell fractions, with endothelial cells associated with the activities of the most metabolic pathways at the proteomic level, followed by CD4⁺ T cells, CD8⁺ T cells, and mucosa-associated constant T cells. In contrast, the arachidonic acid metabolic pathway was associated with the greatest proportion of non-tumor cells, followed by the glycosylphosphatidylinositol (GPI)-anchored biosynthetic pathway (Supplementary Figure S4C).

From the single-gene viewpoint, we first analyzed the correlation of 12 metabolic immune checkpoints with the proportion of microenvironmental cells, 6 of which had robust correlations (Figure 3D and Supplementary Table S10). Then, more generally, we analyzed correlations of 24 prognostic metabolic proteins with the fraction of cells in the tumor microenvironment (TME). We identified 42 robust correlations, among which HK2 was associated with the largest proportion of non-tumor cells in 8 cell types, followed by procollagen-lysine, 2-oxoglutarate 5-dioxygenase 1 (PLOD1). Notably, these two proteins were highly upregulated in S2. Effector memory CD8⁺ T cells were negatively associated with aspartate beta-hydroxylase (ASPH) and HK2 and positively associated with ALDH9A1, and early activated CD8⁺ T cells were negatively associated with PLOD1. In addition, DGKA was positively correlated with fibroblasts and monocytes (Figure 3E and Supplementary Table S11). We compared the overlap of the transcriptome and the tumor transcriptome after deconvolution with the proteome, and we found that the correlation results of metabolic immune checkpoint and prognostic genes in different omics were more consistent with metabolic pathways (Supplementary Figure S4D-F).

3.4 The potential genetic drivers of metabolic and immune features in iCCA

After profiling the metabolic and immune features of different iCCA subtypes, we sought to explore the potential genetic drivers of the metabolic and immune phenotypic features. To this end, we analyzed the role of driver genes and CNV cis-regulation for each metabolic subtype.

In our cohort, a total of 18,593 somatic mutations were identified, among which ARID1A, titin (TTN), BAP1, TP53, and IDH1 had the highest mutation frequency. The region with significant copy number amplification was 19p13.2, and the region with significant deletion was 1p36.11 (Figure 4A and Supplementary Table S12). The median mutation burden was 0.61/MB, and some mutations were mutually exclusive and cooccurring, e.g., ARID1A mutations co-occurring with mucin 16 (MUC16) and ephrin type-A receptor 2 (EPHA2) mutations (Supplementary Figure S5A-B).

Genes with driver mutations in at least 10 individuals were screened for subtype enrichment analysis. The results showed that BAP1 and IDH1 mutations were significantly enriched in S3 (P < 0.001, P = 0.017) and absent in S2 (P = 0.001, P = 0.007), while the remaining driver mutations showed no trend of subtype enrichment (Figure 4B). We divided S3 into S3_BAP1mut, S3_IDH1mut, and S3_Other and compared them with each other to identify differential proteins. In terms of the overlapping upregulated proteins, S3_BAP1mut was more unique than S3_IDH1mut (Figure 4C).

Shared upregulated proteins in S3 exhibited the basic features of S3 subtypes, including histone modification and DNA damage repair. In contrast, the shared proteins in S3_BAP1mut and S3_IDH1mut were involved in biological functions, including bile acid synthesis and stromal adhesion-dependent cell spreading (Supplementary Figure S5C-D).

Next, we analyzed the role of CNVs for each metabolic subtype. A total of 2,503 genes were found to be cis-regulated by iProFun, and 348 genes were mutated at the chromosome arm level; of these genes, 116 were amplified genes, 232 were deleted genes, and 107 were metabolism-related genes, accounting for 30% of the total. Among metabolism-related genes, 34 were amplified, and 73 were deleted (Figure 4D).

The amplification events mainly affected protein phosphorylation, intracellular lipid catabolism, and adenosine diphosphate ribosylation; the deletion events mainly affected carbohydrate derivative biosynthesis processes, protein phosphorylation, cellular modification of amino acid biosynthetic processes, and nucleobase-containing small molecule metabolic processes (Figure 4E).

We found a significant enrichment of chromatin deletion and amplification events in S3 and fewer chromatin deletion and amplification events in S2 (Figure 4F). We screened for genes that were significantly amplified or deleted in S3 with corresponding differentially up- and downregulated proteins. Our analysis revealed that most of these proteins were involved in inflammatory response-related processes; these proteins typically exhibited opposite expression trends in S2 (Figure 4F).

3.5 DGKA inhibition prevented iCCA cell proliferation in vitro

Multiomics analysis revealed that patients with the S2 metabolic subtype had the shortest survival, and DGKA, the protein encoded by DGKA, was the signature protein in the S2 subtype. Through IHC analysis, we found that patients with high DGKA expression had shorter OS (median OS = 33.7 months) and DFS (median DFS = 19.6 months) (Figure 5A-C). Multivariate survival analysis indicated that DGKA was an independent prognostic factor for the OS and DFS of patients with iCCA (Supplementary Figure S6A). To determine the role of DGKA in iCCA progression, we knocked down the expression of DGKA in the iCCA cell lines RBE and HuCCT1 by two different shRNAs (shDGKA #1 and shDGKA #2) (Figure 5D and Supplementary Figure S6B). CCK8 assays showed reduced proliferation of iCCA cells with DGKA knockdown, and colony formation assays and EdU staining revealed similar findings (Figure 5E-G). Meanwhile, we utilized ritanserin, a selective DGKA inhibitor, to inhibit the expression of DGKA in iCCA cells. Likewise, colony formation assays and EdU staining indicated that inhibition of DGKA reduced the proliferation of iCCA cells (Supplementary Figure S6C-D). Therefore, these results showed that DGKA inhibition prevented iCCA cell proliferation in vitro.

3.6 DGKA regulated iCCA cell proliferation via the PA-MEK/ERK pathway

Previous studies reported that DGKA could phosphorylate diacylglycerol (DG) to generate PA [59, 60]. PA, a lipid secondary messenger, plays a role in several critical biological events by regulating downstream signaling pathways [61, 62]. Therefore, we speculated that DGKA regulates iCCA proliferation via the metabolism of PA and the regulation of signaling pathways. Due to the lack of significant differences in the effects of different shRNAs on iCCA cell proliferation, we proceeded with further research using shDGKA #1. Next, we added exogenous PA to the medium of DGKA-knockdown cells, and an increase in the proliferation of iCCA cells was observed by CCK8 assays (Figure 6A). Furthermore, cell viability assays indicated that exogenous PA rescued the viability of ritanserin-treated iCCA cells (Figure 6B). To further explore the mechanism of DGKA in iCCA, RNA-seq was performed in DGKA-knockdown cells. A total of 4,870 differentially expressed genes (DEGs) were found (Supplementary Figure S7A).

Consistent with the results of functional assays, GSEA showed that negative regulation of cell proliferation was enriched in the DGKA-knockdown group (Supplementary Figure S7B-D). Furthermore, KEGG enrichment analysis revealed that the MAPK signaling pathway was significantly enriched based on all DEGs (Figure 6C). Thus, we assessed the expression of p-MEK1/2, MEK1/2, p-ERK1/2, and ERK1/2 in DGKA-knockdown iCCA cells. The results showed that DGKA knockdown reduced the expression of MEK1/2 and ERK1/2 phosphorylation in iCCA cells (Figure 6D). In contrast, exogenous PA rescued phosphorylation of MEK1/2 and ERK1/2 in DGKA-knockdown iCCA cells (Figure 6E). These results suggested that DGKA regulated iCCA cell proliferation via the PA-MEK/ERK pathway.