Multi-omics and machine learning reveal DPPC as a key contributor to colorectal cancer progression and tumor immune microenvironment remodeling

Abstract

Background

Although previous studies have linked colorectal cancer (CRC) with lipid metabolism and inflammatory signaling, the specific roles of phosphatidylcholine metabolites and their interactions with inflammatory cytokines in the tumor microenvironment remain poorly understood.

Methods

This study utilizes an integrated multi-omics analysis approach. We performed a two-sample Mendelian randomization analysis using data from 1400 metabolites and 91 inflammatory cytokines to investigate their associations with CRC, followed by experimental validation of the findings. Single-cell transcriptomics revealed metabolic state differences, while machine learning constructed a predictive model. SHAP analysis interpreted the model, with spatial transcriptomics validating key findings.

Results

The phosphatidylcholine metabolite DPPC was identified as causally associated with CRC risk. Our results demonstrated DPPC promotes tumor progression by inhibiting TNFSF14 secretion. Our DPPC-based model effectively predicted CRC progression, with SHAP analysis identifying ARL8A, MTUS1, and TMEM184A as key contributors. These findings were validated spatially and translated into a clinical nomogram for prognosis and immunotherapy guidance. In summary, this study highlights the significance of DPPC-mediated regulation within the tumor microenvironment in predicting CRC progression and guiding potential therapeutic strategies.

Supplementary information

The online version contains supplementary material available at 10.1186/s12967-025-07576-y.

Background

colorectal cancer (CRC) accounts for approximately 10% of all cancer cases worldwide and remains one of the leading causes of cancer incidence and mortality in both men and women [1]. CRC develops through a complex interplay of genetic predispositions and environmental influences, including hereditary syndromes such as Lynch syndrome and chronic conditions such as inflammatory bowel disease. The latter encompasses not only the makeup of the gut microbiota but also certain lifestyle factors including smoking, excessive alcohol consumption, and obesity [2].

Metabolic reprogramming is a significant hallmark of cancer [3]. Colorectal cancer originates from intestinal epithelial cells [4]. Beyond absorbing and transporting dietary nutrients, the intestinal epithelium also regulates systemic metabolism by secreting hormones and modulating plasma metabolite levels [5, 6]. Concurrently, intestinal epithelial cells are subjected to endogenous mediators and external stressors, which alter their behavior and functionality [4]. From a metabolic perspective, dysregulation in the metabolic activity of intestinal epithelial cells can facilitate the initiation of epithelial tumors. These cells also contribute to systemic metabolic regulation, thereby modulating plasma metabolite levels. In turn, these circulating metabolites can reciprocally influence the behavior and function of intestinal epithelial cells. Emerging evidence highlights a bidirectional relationship between plasma metabolites and tumorigenesis [7, 8]. For example, some scholars have proved that higher blood sugar levels are significantly associated with an increased risk of colorectal cancer, especially early-onset colorectal cancer (EOCRC) [7].

Phosphatidylcholine, particularly 1,2-dipalmitoyl-gpc (16:0/16:0) (DPPC), is a significant lipid metabolite that exerts a pivotal function within the domain of liposomal drug delivery [9–11]. Moreover, phosphatidylcholine has been shown to regulate immune responses by mitigating endoplasmic reticulum and oxidative stress, thereby enhancing antitumor activity. Additionally, it promotes the growth of immunosuppressive regulatory T cells (Tregs) while inhibiting the generation of pro-inflammatory Th17 cells [12–14]. Furthermore, phosphatidylcholine has been shown in several trials to have therapeutic promise in treating inflammatory bowel disease (IBD), primarily through its ability to strengthen intestinal barrier integrity and attenuate inflammatory signaling, ultimately leading to a reduction in intestinal inflammation [15]. Abnormal metabolism of phosphatidylcholine directly promotes the proliferation and metastasis of CRC cells. Studies indicate that elevated fatty acid synthase (FASN) expression correlates with increased phosphatidylcholine production in CRC tissues, further facilitating tumor cell migration and invasion [16]. Moreover, phosphatidylcholine metabolism enhances tumor cell invasiveness by regulating the expression of epithelial-mesenchymal transition (EMT)-associated proteins, including E-cadherin and occludin [17]. However, the function of phosphatidylcholine in colorectal cancer with regard to the tumour immune microenvironment is not yet understood.

Utilizing genetic variations as instrumental variables (IVs), Mendelian Randomization (MR) is a genetic epidemiology technique that establishes causal relationships between certain exposures and outcomes [18, 19]. When drawing conclusions about the causative relationship between exposure and result, MR uses germline genetic variations as instrumental variables, which makes it less vulnerable to confounding biases and reverse causality than observational research [7]. Recent advancements in artificial intelligence, particularly in machine learning, have led to its broad adoption across various medical disciplines, owing to its remarkable analytical capabilities and predictive accuracy [20, 21]. Boosting, as one of the prominent ensemble techniques in machine learning, enhances predictive performance by sequentially combining multiple weak learners into a robust composite model. Moreover, boosting algorithms can elucidate the decision-making processes of models through specific techniques, such as SHAP (SHapley Additive exPlanations) [22]. This game-theoretic approach interprets the outputs of machine learning models by quantifying the contribution of each feature to individual predictions, thereby enhancing the model’s transparency and credibility [23].

This article will use Mendelian randomization (MR) analysis to study the relationship among circulating metabolites, inflammatory mediators and CRC. We indicated that the presence of DPPC in the colorectal cancer microenvironment was associated with a suppression of TNFSF14, an inflammatory factor. This observation was subsequently substantiated through experimental verification. In addition, single-cell level analysis of the tumour microenvironment was conducted to elucidate the role of DPPC, leading to the identification of novel biomarkers with potential for predicting colorectal cancer progression [24, 25]. Ultimately, the accuracy of these markers was validated in multiple colorectal cancer cohorts with the help of machine learning algorithms.

Methods

Study design

As illustrated in Fig. 1, the research comprises three primary components. First, we evaluated the causal association between more than 1400 metabolites and CRC. Second, we analyzed the contribution of 91 inflammatory cytokines to CRC. Third, we performed a mediation analysis to evaluate the role of cytokines in the pathway linking metabolites to CRC. In this study, when conducting Mendelian randomization (MR) analysis, we selected single-nucleotide polymorphisms (SNPs) as instrumental variables (IVs). The validity of MR relies on three key assumptions: [1] the IVs are closely linked to the exposure of interest; [2] they are unaffected by any other factors; and [3] their influence on the result only comes from the exposure route [26]. To address potential sources of bias and improve the robustness of our findings, a number of downstream studies were conducted. We used the Steiger directionality test and computed F-statistics to verify the validity and strength of the instrumental variables (IVs). In light of horizontal pleiotropy and possible outliers, we used the MR-PRESSO and MR-Egger regression analyses. Leave-one-out (LOO) sensitivity analysis was also performed to identify influential instrumental variables. In addition, reverse MR was conducted to examine the directionality of causality [27–29]. Supplementary Fig. S1 is the graphical presentation of the research design.

Fig. 1.

Study overview. Step 1 illustrates the causal relationship between metabolites and colorectal cancer (CRC). The results show that some metabolites are causes of CRC occurrence, while CRC has no causal effect on metabolites. Step 2 illustrates the causal relationship between cytokines and CRC. The results show that some cytokines are causes of CRC occurrence, while CRC has no causal effect on cytokines. Step 3 illustrates the mediation analysis of cytokines in the pathway from metabolites to CRC: path c represents the total effect of metabolites on colorectal cancer; path b represents the causal effect of cytokines on colorectal cancer; path a represents the causal effect of metabolites on cytokines

Data sources

We obtained 1400 metabolites from a study conducted by Ferkingstad et [30] and 91 inflammatory cytokines from another study [31]. Spatial transcriptomics [32] and single-cell datasets of CRC patients (GSE225857, GSE16655, GSE205506 and GSE225857) were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/) for analysis. The CRC TCGA data was downloaded as a training group. The bulk RNA-seq datasets GSE17536 (177 tumour samples) and datasets GSE39582 (19 normal and 443 tumour samples) were downloaded for validation. Data for CRC was derived from large-scale GWAS consortia, including N = 185,616 cases.

Instrumental variables selection

Initially, Strong correlations between single-nucleotide polymorphisms (SNPs) and metabolites () were found and chosen for additional examination. The SNPs with a P-value of were chosen as the criteria in order to guarantee that the maximum number of instruments were available for each cytokine. The SNPs that showed linkage disequilibrium (LD) were then removed from the study. To maintain independence among the selected SNPs associated with metabolites, linkage disequilibrium (LD) pruning was conducted by excluding variants with an greater than 0.001 and ensuring a minimum separation of 10,000 kilobases (kb) between SNPs [33].

Mendelian randomization analysis

Using the R package TwoSampleMR, we performed MR analyses of cytokines and metabolites on CRC through MR analysis. Several complementary MR methods were employed, including weighted median, MR-Egger, inverse-variance weighted (IVW), simple mode, and weighted mode approaches. Among these, IVW was considered the primary approach due to its high statistical efficiency and robustness. Causal estimates were expressed as odds ratios (ORs) per standard deviation (SD) increase in metabolite levels. A schematic overview of the MR analysis framework is presented in Fig. 1. To control for multiple testing and reduce the likelihood of false positives, false discovery rate (FDR) adjustment was performed using the Benjamini-Hochberg (B/H) procedure, with statistical significance defined as a B/H-adjusted p-value < 0.05, corresponding to an FDR threshold of 5%.

Mediation analysis

The two-sample MR analyses (Steps 1 and 2 in Fig. 1) were used to select metabolites and cytokines that showed a substantial causal influence on CRC for further mediation assessment. In order to determine if cytokines were the mediators in the pathway from metabolites to CRC, we carried out multiple MR analysis to see whether metabolites had a causal influence on cytokines (step 3, path a, in Fig. 1).

Bidirectional causality analysis

To assess potential bidirectional causal relationships among metabolites, cytokines, and CRC, reverse MR analyses were performed using CRC as the exposure and CRC-associated metabolites or cytokines as outcomes (Steps 1 and 2 in Fig. 1). Instrumental variables (IVs) were defined as SNPs significantly associated with CRC.

Sensitivity analysis

To assess heterogeneity across instrumental SNPs, Cochran’s Q test was applied [34]. Scatter plots were also created to show the relationships between SNPs, exposures, and outcomes. To assess the impact of individual SNPs, leave-one-out (LOO) sensitivity analyses were performed by successively eliminating each SNP and recalculating the IVW estimations. This process was employed to determine whether the total causal estimate was disproportionately affected by any one SNP [29]. Furthermore, potential horizontal pleiotropy was evaluated using both MR-Egger regression and the MR-PRESSO approach. The MR-PRESSO method specifically identified outlier variants and adjusted for horizontal pleiotropy by excluding these outliers from the analysis [27].

Raw data processing and quality control

Each sample was given a unique molecular identification (UMI) count matrix. In the R environment (v4.3.2), the Seurat package (v4.4.0) was used to normalize the data [30]. The normalized values were then transformed using log₂. Quality control (QC) filtering was applied based on four criteria: (i) cells with more than 4000 detected genes were excluded; (ii) cells with fewer than 200 expressed genes were removed; and (iii) cells in which ≥10% of transcripts were mitochondrial in origin were filtered out. The Harmony algorithm was applied to remove the batch effect [35]. Principal component analysis (PCA) was then applied to reduce data dimensionality. Cells exhibiting similar transcriptional profiles were grouped using the FindNeighbors and FindClusters functions. Uniform Manifold Approximation and Projection (UMAP) was then used to project the resultant clusters onto a lower-dimensional environment for display.

Data integration, unsupervised clustering, and cell type annotation

To find genes enriched in each identified cell cluster, differential gene expression analysis was carried out using the Seurat package’s FindAllMarkers function. Marker gene identification followed standard thresholds, including an absolute fold change greater than 0.25, expression in over 25% of cells, and adjusted [36]. Subsequently, the SingleR package (version 2.4.0) was used to annotate clusters based on reference marker gene signatures [37, 38]. These determinations were then subjected to manual verification and correction, with the results verified against the CellMarker database [39].

Cell-cell communication analysis

To explore intercellular interactions within the tumor microenvironment (TME), the CellCall R package (v1.0.7) was used to infer cell-cell communication. Using receptor expression in one cell group and ligand expression in another, the system was able to predict enhanced receptor–ligand interactions across different cell types. This approach enabled the identification of cell type–specific ligand–receptor interactions that were most relevant within the TME.

The prediction of TFs activity by Dorothea

The DoRothEA collection of genes contains transcription factors (TFs) that interact with their targets. DoRothEA regulons integrate multiple sources of evidence for transcription factor (TF)–target interactions, including curated literature, ChIP-seq data, TF binding motifs, and expression-based inference [40]. The DoRothEA method (v1.14.1) was utilized to infer transcription factor (TF) activity at the single-cell level using the log-normalized expression matrix that was acquired from Seurat.

Analysis of spatial transcriptomics data

Spatial transcriptomics (ST) data were processed and shown using the Seurat R package. Data normalization was performed using the SCTransform (SCT) method, followed by integration with functions including SelectIntegrationFeatures, PrepSCTIntegration, FindIntegrationAnchors, and IntegrateData. Unsupervised clustering was then performed to identify spatially similar regions and group cell populations, based on histological annotation (H&E staining) and highly variable genes within each cluster. Cell expression levels in ST data are visualized using SpatialDimPlot and SpatialFeaturePlot functions.

Spatial transcriptomics data used for cell type analysis

Robust Cell Type Decomposition (RCTD) was used to map cell types from reference single-cell RNA sequencing (scRNA-seq) data onto the spatial transcriptomics dataset [41]. Using Seurat’s FindAllMarkers function, cell type-specific marker genes were identified, with an emphasis on markers exhibiting positive fold changes. Subsequently, the reference dataset and the Visium spatial transcriptomics data were both examined in a full-capacity doublet mode setup as part of a comprehensive RCTD analysis process.

Machine learning and predictive interpretation

To construct a robust and generalizable DPPC-based risk score, a range of machine learning algorithms was employed, including Survival Support Vector Machine (survival SVM), CoxBoost, Lasso, Stepwise Cox, Ridge regression, and Gradient Boosting Machine (GBM). To identify the model with the highest average C-index across all validation cohorts, multiple algorithms were implemented using varied configurations. To assess the performance of the risk scoring system, the ‘timeROC’ package was employed to calculate time-dependent Area Under the Curve (AUC) values. Additionally, Cox proportional hazards regression was performed using the survival package in R, confirming the risk score as an independent prognostic factor.

Subsequently, the Shapley Additive exPlanation (SHAP) values were employed to visualise the key features affecting CRC onset and progression. The SHAP algorithm is an explainable artificial intelligence (XAI) method grounded in cooperative game theory, aimed at interpreting the predictions made by machine learning models [42]. In this context, SHAP values conceptualize each feature as a ‘participant’ and quantify its contribution to the model’s predictions [22, 23]. The fundamental principle of SHAP is to assess the significance of each feature by examining all possible combinations of features, thus offering both global and local explanations [43]. This algorithm is predominantly implemented via the SHAP library in Python. This approach facilitated the assessment of how specific features contribute to outcome prediction and revealed the impact of critical variables on overall model effectiveness.

Establishment and validation of nomogram

A prognostic nomogram was constructed to integrate clinical characteristics and the DPPC score, using the rms package in R, based on independently validated survival outcomes. Each predictor variable in this scoring system is given a precise point value, and the sum of the individual contributions from all the factors determines the overall score for each patient [44]. To assess the performance of the nomogram, receiver operating characteristic (ROC) curves were generated for 1-, 3-, and 5-year survival predictions. Calibration plots were also constructed to compare the predicted probabilities of survival with the actual observed outcomes at the corresponding time points.

Cell lines

All cell lines were maintained at 37 °C in a humidified atmosphere containing 5% CO₂. Short tandem repeat (STR) DNA profiling was used to authenticate each cell line, and PCR testing confirmed the absence of mycoplasma contamination. The American Type Culture Collection (ATCC) provided all of the cell lines utilized in this investigation. Media containing 10% fetal bovine serum (FBS) and 1% penicillin–streptomycin were used to cultivate the cells. Specifically, HCT116, HCT15, RKO, SW480, SW620 and DLD1 were grown in DMEM (Gibco); LoVo were maintained in F12K medium (Gibco); Caco2 cells were cultured in MEM (Gibco); HCT8 and HIEC6 were kept in RPMI 1640 medium (Gibco).

Quantitative real-time PCR

Total RNA was extracted from CRC cells using TRIzol reagent (Invitrogen, USA). After treatment with different concentrations of DPPC (0.1 µg/mL–200 µg/mL; MCE, HY-109506), complementary DNA (cDNA) was synthesized using the TransScript Uni All-in-One First-Strand cDNA Synthesis SuperMix for qPCR (One-Step gDNA Removal) (TransGen Biotech, AU341). Quantitative real-time PCR (qRT-PCR) was performed according to the manufacturer’s instructions using the PerfectStart® Universal Green qPCR SuperMix (TransGen Biotech, AQ631). Primer sequences for TNFSF14 and ACTB (β-actin) are listed in Supplementary Table S6. ACTB served as the internal control, and relative gene expression levels were calculated using the 2^−ΔΔCt method. Differences between two groups were analyzed using the Mann–Whitney U test, and p < 0.05 was considered statistically significant. All graphs were generated using GraphPad Prism 10.

Cell proliferation assay

Human colorectal cancer cell lines were cultured in suspension until the logarithmic growth phase under treatment with DPPC (MCE, HY-109506) at varying concentrations. To reduce edge effects and ensure accurate optical density (OD) measurements, 100 μL of PBS was dispensed into the outer wells of the 96-well plate. Ten microliters of CCK-8 solution were added to each experimental well and incubated in a light-protected environment after the specified treatment duration. The plates were gently shaken and kept at 37 °C for an additional 2 hours. Finally, a TECAN Spark10M microplate reader was used to capture absorbance at 450 nm.

Plate clone formation experiment

To analyze colony formation, 500 exponentially growing cells were placed into each well of a 12-well plate. Following 12–15 days of incubation, colonies containing over 50 cells were enumerated. Colony formation efficiency was determined using the formula: (colony count/initial cell number) × 100.

Results

The causal role of metabolites and cytokines in the development of colorectal cancer

Using a two-sample Mendelian randomization (MR) approach, the possible causative involvement of CRC in influencing metabolic and cytokine profiles was examined. The inverse-variance weighted (IVW) method served as the primary statistical tool. Among 83 candidate metabolites assessed, a subset met nominal significance (). However, after false discovery rate (FDR) correction for multiple testing, only one metabolite—1,2-dipalmitoyl-gpc (16:0/16:0, DPPC)—remained significant at the 0.05 threshold, as shown in Supplementary Table S1 and S2. We found that the increased level of 1,2-dipalmitoyl-gpc (16:0/16:0) associated with CRC onset (β = 0.002, 95% CI = 0.044~0.194, , FDR = 0.001156, Fig. 2A, Supplementary Table S1,S2). As for cytokines, we found that 7 candidate cytokines were identified at a significance of p value＜0.05 (Supplementary Table S1). While none of the cytokines maintained statistical significance after false discovery rate (FDR) adjustment (threshold: ), several exhibited nominal correlations with CRC. In particular, CCL4 levels were reduced in CRC samples (β = −0.001, p = 0.039). Increased expression was noted for CSF1 (β = 0.002, p = 0.023), CXCL10 (β = 0.002, p = 0.019), and GDNF (β = 0.001, p = 0.024). Conversely, IL-2RB (β = −0.002, p = 0.047), IL-4 (β = −0.002, p = 0.041), and LIGHT (also referred to as TNFSF14; β = −0.001, p = 0.019) were observed to be downregulated in CRC patients. Results from leave-one-out (LOO) analyses based on alternative MR methods are shown in Supplementary Fig. S2. Consistent results obtained from four additional MR methods, along with various sensitivity tests, reinforced the reliability of the observed causal associations (Supplementary Table S3). Notably, the MR-Egger intercept suggested an absence of horizontal pleiotropy. Moreover, the findings were further supported by the patterns observed in both scatter plots and funnel diagrams, underscoring the robustness and consistency of the analysis.

Fig. 2.

The result of MR analysis. (A) The relationship between DPPC and CRC. (B) The relationship between LIGHT and DPPC. (C) The relationship between LIGHT and DPPC in different methods

Bidirectional causal effects of colorectal cancer on metabolites and cytokines

As presented in Supplementary Table S4, no evidence of reverse causality was observed between metabolites and cytokines. No valid instrumental variables (IVs) were identified when matching SNPs between CRC and metabolites or cytokines. However, reverse MR analysis revealed a significant causal effect of TNFSF14 (LIGHT) on DPPC levels (OR = 0.9259, 95% CI: 0.8691–0.9864, ; Fig. 2B–C, Supplementary Tables S1 and S4).

Mediation analysis

Causal relationships between both metabolites and cytokines and CRC were identified in this study. Notably, cytokines appeared to mediate the link between metabolites and CRC. A key criterion for establishing mediation is a significant association between the metabolite and the cytokine. We found a causal link between DPPC (associated with CRC) and TNFSF14 (LIGHT, also associated with CRC), as shown in Step 3 of Fig. 1 and Supplementary Table S1. This supports the hypothesis that TNFSF14 (LIGHT) acts as a mediator in the causal pathway from metabolites to colorectal cancer.

DPPC promotes tumor proliferation by inhibiting the secretion of TNFSF14

To confirm the role of DPPC in tumor proliferation, we initially treated CRC cells with varying concentrations of DPPC (0.01 μg/ml-100 μg/ml) and conducted a cell proliferation assay. It has been observed that in colorectal cancer cell lines, including SW620, HCT116, HCT15 and RKO, there is an acceleration in cellular proliferation with the elevation of DPPC concentration (Fig. 3A). Additionally, we performed a plate clone formation experiment, which also demonstrated that the rate of cell growth increased with rising concentrations of DPPC (Fig. 3B). To verify the function of TNFSF14 in tumor cells, we first measured the expression levels of TNFSF14 in normal epithelial cells and various tumor cell lines (Fig. S3). Subsequent qRT-PCR analysis was performed to examine TNFSF14 expression following 48-hour treatment with varying concentrations of DPPC (). The results demonstrated a dose-dependent decrease in TNFSF14 levels (Fig. 3C).

Fig. 3.

DPPC promotes tumor proliferation by inhibiting the secretion of TNFSF14. (A) Cell proliferation assay with varying concentrations of DPPC.(B) Plate clone formation experiment with varying concentrations of DPPC.(C) Quantitative real-time PCR results showing TNFSF14 expression after treatment with different concentrations of DPPC in CRC cell lines

Interpretation of DPPC metabolites in the colorectal cancer patients at the single cell level

To investigate the results of Mendelian randomization at the single-cell level, we utilized single-cell transcriptomic data from 16 samples, comprising 8 colorectal cancer cases and 8 matched controls.Guided by the CellMarker database and the SingleR package, four major immune cell types were identified: T cells, NK cells, B cells, and monocytes (Fig. 4A). The distribution of these immune cell populations in CRC patients versus control samples is shown in Fig. 4B. Notably, CRC patients showed a reduced proportion of T cells and NK cells compared to control samples. By obtaining the set of genes related to DPPC metabolism, transport and regulation from the MSigDB database, we observed that the metabolic level of DPPC was expressed at a lower level in T cells but higher in epithelial cells (Fig. 4C). In the single-cell dataset, TNFSF14 was predominantly expressed in T cells and NK cells (Fig. 4D). Additionally, higher TNFSF14 expression was observed in samples with lower levels of DPPC metabolism (Fig. 4E), consistent with the results of our Mendelian randomization (MR) analyses. Based on these findings, epithelial cells were selected for subsequent investigation.

Fig. 4.

Annotation of scRNA-seq datasets. (A) UMAP of scRNA-seq. (B) Proportion diagram of scRNA-seq of colorectal cancer patients and healthy individuals. (C) The expression of metabolism in different cells. (D) The expression of TNFSF14 in different cells. (E) The expression of TNFSF14 in metabolism_ high and metabolism_ low. (F) GSEA analysis of intersection genes. (G) Analysis of metabolic groups and other cells interactions. (H) Pathway analysis of metabolic groups and immune cell differentiation signals. (I) Identification of transcription factors in metabolism high and metabolism_low. Metabolism_ high: Epithelial_cell with high DPPC metabolism; metabolism low: epithelial cell with low DPPC metabolism; CRC: colorectal cancer; CT: healthy subjects

To further investigate DPPC metabolism in colorectal cancer, we applied the AddModuleScore function to compute DPPC-related metabolic activity at the single-cell level. Based on the resulting scores, epithelial cells were stratified into high and low DPPC metabolism groups, enabling a more detailed exploration of its functional role in CRC. Gene Set Enrichment Analysis (GSEA) revealed that the differentially expressed genes were positively associated with pathways such as primary alcohol metabolic process and phosphatidylethanolamine metabolism, while showing negative associations with pathways including ribonucleoprotein complex biogenesis and peptide metabolic processes (Fig. 4F).

Function analysis of DPPC metabolism in scRNA-seq data

A comprehensive interaction analysis was performed using single-cell RNA sequencing (scRNA-seq) data to examine the crosstalk between high- and low-metabolism cell populations and other cell types. Subsequently, we considered the epithelial cells metabolic group as a receptor or ligand by classifying epithelial cells into metabolic high and low groups and interacting with other cells, and found the corresponding results (Fig. 4G). By comparing the communication signals with the metabolic-high group, we detected Wnt signaling pathway showing dramatic interactions between the metabolic-high group and fibroblasts as the conduction conduit consists of the EFNB1–EPHB3, whereas the ErbB signaling pathway in cancer pathway connecting the fibroblasts to the metabolic-low group consists of the EFNB1–EPHB3 (Fig. 4H). Additionally, the top 20 transcription factors exhibiting the greatest variability between metabolic subgroups were identified. Notably, AR was downregulated in the high-metabolism group, whereas MYC showed elevated expression in the low-metabolism group (Fig. 4I).

Construction of the DPPC risk score by integrated machine learning

Using single-cell RNA-seq data, differentially expressed genes (DEGs) between high and low DPPC metabolic groups were identified via the FindMarkers function (Supplementary Table S5). Following this, univariate Cox analysis identified 80 genes that showed significant correlations with overall survival (see Supplementary Fig. 4). In the training cohort of TCGA-CRC, we constructed DPPC risk score models using 99 combinations of algorithms. StepCox[both]+Ridge was the highest mean C-index of 0.617 in all datasets. A total of 99 prognostic models were evaluated using the TCGA dataset, and their performance was further tested in the GSE17536 and GSE39582 validation cohorts based on concordance index (C-index) calculations (Supplementary Fig. S5). According to Kaplan-Meier curves, patients in the high-risk group had a much lower overall survival rate than those in the low-risk group (Fig. 5A–C). Among all models assessed, the DPPC-based risk score consistently yielded the highest C-index (Fig. 5G). Receiver operating characteristic (ROC) analysis further confirmed its superior prognostic performance compared to conventional clinical variables in the TCGA-CRC cohort, with area under the curve (AUC) values of 0.804, 0.817, and 0.875 for 1-, 3-, and 5-year survival predictions, respectively (Fig. 5D–F).

Fig. 5.

The gene signatures based on machine learning. (A-C) Prognoses of patients in the TCGA (A), GSE17536 (B), and GSE39582 (C) cohorts. (D-F) The ROC of predicting patient survival at 1, 3 and 5 years in the TCGA (D), GSE17536 (E), and GSE39582 (F) cohorts. (G) The C-index values of risk score, age, gender and stage

The shap interpretable model analysis for DPPC risk scores

A summary plot was created to compare the SHAP values for different models and outcomes. Features were displayed on the y-axis in order according to their importance according to the mean SHAP value. The dots represent the individual patients and indicate the absolute value of the feature for that patient (colored blue or red). A SHAP value of 0 means the feature is not influential on outcome. Positive or negative SHAP value indicate positive or negative associations with outcome. In terms of OS results, we show the SHAP for CatBoost and XGBoost. To further explore the most important key genes of DPPC risk model consisting 27 genes, we use the SHAP explanatory model of CatBoost (Fig. 6A–C) and XGboost (Fig. 6D–F). The SHAP heatmap of the two models showed that the top three most important genes among the 27 genes were ARL8A, MTUS1 and TMEM184A (Fig. 6).

Fig. 6.

The shap method explains catboost and XGboost. (A)-(C) Summary plots, beeswarm plots, and heatmap plots of catboost. The gene with the largest contribution in catboost is TMEM 184A. (D)-(F) Summary plots, beeswarm plots, and heatmap plots of XGboost. The gene with the largest contribution in NGboost is ARL8A

The spatial transcriptome analysis reveals key genes for DPPC risk score

We analyzed spatial transcriptomic data from four colorectal cancer patients (ST-P1 to ST-P4) involved in the DPPC study [14]. Based on H&E-stained tissue sections and differentially expressed genes (DEGs) within each cluster, eight major spatial cell types were annotated: B cells, endothelial cells, epithelial cells, smooth muscle cells, macrophages, monocytes, T cells, and NK cells (Fig. 7A). Additionally, guided by the SHAP model, the spatial distributions of the top three key genes—ARL8A, MTUS1, and TMEM184A—from the DPPC risk score model was visualized across the four samples (Supplementary Figure S6).

Fig. 7.

Immune characteristics analysis of DPPC risk scores. (A) Correlation between DPPC risk scores and immune cells. (B) Differential expression of immune-related genes between high and low DPPC risk scores. (C) The TME scores of DPPC risk groups. (D) The tide between high and low DPPC risk scores. *p < 0.05. **p < 0.01. ***p < 0.001

Immunotherapy prediction and immune characteristics related to DPPC risk scores

The DPPC risk score was calculated using seven independent computational tools. The resulting analyses revealed strong correlations between the risk score and immune cell types, including CD8⁺ T cells, activated NK cells, and B cells (Fig. 7A). High risk group expressed more CD274, CTLA4, IDO1 than low risk group (Fig. 7B), while high risk group displayed higher ImmuneScore, StromalScore and ESTIMATEScore than low risk group (Fig. 7C).This study further assessed the potential clinical benefit of immunotherapy across stratified risk groups. Elevated TIDE (Tumor Immune Dysfunction and Exclusion) scores were indicative of increased immune evasion, implying diminished responsiveness to immunotherapy. Higher TIDE (Tumor Immune Dysfunction and Exclusion) scores reflected enhanced immune escape and reduced sensitivity to immunotherapy. Importantly, individuals classified in the low-risk category exhibited markedly lower TIDE scores compared to those in the high-risk group, indicating a greater probability of responding to immune checkpoint inhibitor (ICI) therapy (Fig. 7D). Conversely, patients with elevated TIDE scores were anticipated to derive limited benefit from ICI therapy.

Establishment of a nomogram based on DPPC risk scores

As illustrated in Fig. 8A, the tumor stage and the DPPC risk score showed substantial relationships with CRC outcomes, according to the univariate Cox analysis. Subsequent multivariate analysis further confirmed that the DPPC risk score functioned as an independent prognostic indicator, remaining unaffected by other clinicopathological parameters (Fig. 8B). Considering its prognostic significance, a nomogram was developed by combining the DPPC risk score with clinical parameters to forecast 1-, 3-, and 5-year overall survival (OS) in CRC patients (Fig. 8C). Within the TCGA cohort, the model demonstrated strong predictive performance, as reflected by high AUC values across all evaluated time points (Fig. 8D). Calibration analysis demonstrated close concordance between predicted and actual outcomes, indicating that the nomogram closely approximates an optimal predictive tool (Fig. 8E). Collectively, these results underscore the potential clinical relevance of the DPPC-based nomogram in CRC prognosis.

Fig. 8.

Construction and validation of a nomogram. (A-B) Univariate and multivariate cox analysis of DPPC risk scores. (C) Nomogram for predicting the 1-, 3-, and 5-year os of CRC patients in TCGA cohort. (D) ROC curves for predicting the 1-, 3-, and 5-year results in TCGA. (E) Calibration curves of the nomogram for predicting of 1-, 3-, and 5-year os in the TCGA. **p < 0.01. ***p < 0.001

Discussion

The adverse prognosis of CRC was confirmed in this study by demonstrating the downregulation of TNFSF14 through DPPC using a Mendelian randomization approach with dual samples. The expression and transportation of DPPC in different immune cells within the tumor microenvironment of CRC patients were determined through analysis of single-cell signaling pathway data. Further validation was conducted through metabolic pathway analysis, vector pseudo-time series analysis, and bulk immune infiltration analysis. These results align with previous studies highlighting the role of TNFSF14 in modulating immune responses and promoting tumor progression [45].

In cell signaling pathways, 1,2-dipalmitoyl-gpc (DPPC) functions as a constituent of the cellular membrane [46] and potentially regulates downstream gene expression, such as TNFSF14, by modulating the activity of membrane receptors and signaling molecules. Previous studies have demonstrated the robust binding affinity between DPPC and lung surfactant [47], indicating specific ligand-binding characteristics that support our hypothesis. Furthermore, DPPC participates in membrane lipid metabolism through its involvement in catalytic reactions mediated by phospholipase C, leading to the generation of dipalmitoylglycerophosphate and dipalmitoylglycerophosphodiester [48]. This process subsequently impacts other lipid-mediated signaling pathways like phospholipase A2 (PLA2) [49], which activates arachidonic acid (AA) production further metabolized into proinflammatory mediators including prostaglandins and lipoxins. Previous studies have demonstrated that PD-1 signaling suppresses phospholipase 1 expression and facilitates ferroptosis in intratumoral T cells [50]. These results imply that the modulation of anti-tumor immune responses might be significantly influenced by phospholipid metabolism.

Considering the potential of DPPC as a drug delivery vehicle in colorectal cancer therapy, its identification as a prognostic biomarker may carry important clinical implications [51]. Therapeutic strategies aimed at DPPC or its downstream effectors may represent a promising avenue for improving prognosis in individuals with colorectal cancer. Moreover, these results highlight the importance of integrating genetic components in prognostic biomarker assessment, as genetic variation provides key evidence for elucidating causal links between molecular signatures and clinical endpoints.

TNFSF14 (also known as LIGHT) has been implicated in CRC progression, mediating the link between inflammatory signaling pathways and tumor development. TNFSF14 may function as a potential biomarker for CRC progression, given its positive correlation with pro-inflammatory cytokine expression in tumor tissues. Moreover, alterations in the TNFSF14 gene have been associated with heightened risk for developing colorectal cancer. TNFSF14 contributes to immune regulation by facilitating the recruitment and activation of immune cell populations—including macrophages and T lymphocytes—via its receptor HVEM (TNFRSF14), and is also implicated in the impairment of intestinal barrier function [52]. Our single-cell analysis revealed elevated TNFSF14 expression in tumor-associated NK and T cells. Consistently, previous studies have reported increased TNFSF14 expression in NK and T cells from the peripheral blood of CRC patients. The high expression of TNFSF14 in NK/T cell may be driven by increased NF-κB signaling.TNFSF14 may serve as a mediator linking inflammatory signaling pathways to CRC development, potentially through activation of the NF-κB pathway, which has been associated with suppression of tumor growth and invasion [52]. Together with our findings, these results support the role of TNFSF14 as a key cytokine in the pathogenesis of colorectal cancer.

Using SHAP analysis to interpret the DPPC risk score model, we identified ARL8A, MTUS1, and TMEM184A as key contributors among the hub genes. Notably, ARL8A emerged as a potential clinical biomarker associated with colorectal cancer metastasis. This study demonstrates the influence of transcription factors, including ARL8A, and their regulatory mechanisms affecting gene expression through copy number variation and DNA methylation [53]. A key tumor suppressor gene in colorectal cancer, MTUS1 is known to have decreased expression in tumor samples when compared to matched neighboring normal tissues [54]. Reduced MTUS1 expression has been associated with poor overall survival in patients with colorectal cancer, demonstrating its dual use as a prognostic and diagnostic marker [55]. Moreover, MTUS1‘s involvement in cancer-related signaling pathways underscores its significance in colorectal cancer pathophysiology and possible therapeutic targeting [56]. Although there are no studies directly reporting the function of TMEM184A in colorectal cancer, its important homolog TMEM184B has been extensively studied. TMEM184B is implicated in colorectal cancer, with genetic ancestry influencing its mutation frequency. Studies indicate a higher likelihood of mutations in genes like TMEM184B associated with African ancestry. Tumors in Latino populations exhibit fewer TMEM184B mutations compared to non-Latino individuals, highlighting the importance of considering ethnic backgrounds in understanding colorectal cancer’s genomic landscape and developing tailored precision medicine strategies [57]. In summary, ARL8A, MTUS1 and TMEM184A are all may relate to the pathogenesis of colorectal cancer, indicating their potential as new targets for colorectal cancer diagnosis and treatment in the future.

A significant advantage of this research is the application of Mendelian randomization (MR) to determine the causal links between inflammatory variables, metabolites, and CRC. The MR design minimizes bias from confounding variables, thereby strengthening causal inference. And this result was verified through qRT-PCR analysis, cell proliferation assay, and plate clone formation experiment. Furthermore, we employed Single-cell analysis, ML, spatial transcriptomice analysis and SHAP methods to identify key genes in CRC, enhancing the reliability of our findings. Nomograms were developed based on the DPPC risk score model to provide guidance for clinical practice. However, several limitations should be acknowledged. Although Mendelian randomization provides strong evidence for causality, its results may still be influenced by potential biases such as horizontal pleiotropy and linkage disequilibrium.

Conclusions

Mendelian randomization studies investigate intricate causal relationships among metabolites, inflammatory factors, and CRC. The findings underscore TNFSF14‘s protective role and DPPCs, contribution as a risk factor in CRC development. Moreover, ARL8A, MTUS1, and TMEM184A have been pinpointed as pivotal genes in the metabolic risk model associated with DPPC. Importantly, DPPC and its associated key genes in risk score models are anticipated to emerge as novel targets for diagnosing CRC and implementing therapeutic strategies.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

XL, ZHJ, and HFD contributed to the conceptualization of the study. XL, CW, SYC, and JYC were responsible for data curation, while XL, ZHJ, HFD, and HY performed the formal analysis. Data analysis and methodology were conducted by XL and HFD. XL, CW, and HFD carried out the investigation. Project administration was managed by CHZ and HXC. Resources were provided by XL, ZHJ, HFD, and HZ. XL, ZHJ, and HFD drafted the original manuscript, and XL, ZHJ, HFD, and HY helped with the editing and revision of the paper.The completed work was reviewed and approved by all authors.

Funding

We gratefully acknowledge the GEO and GWAS databases for providing the data resources used in this study. This work was supported by the National Natural Science Foundation of China (Grant No. 82573049); the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2025B1515020095); the Shenzhen Fundamental Research Program (Grant No. JCYJ20240813150212017); the Shenzhen Municipal Bureau of Human Resources and Social Security (Grant No. 381058); the Research Supporting Start-up Fund for Associate Researcher of The Seventh Affiliated Hospital, Sun Yat-sen University (Grant No. ZSQYRSSFAR0008); Supported by Shenzhen Science and Technology Innovation Committee (Grant No. KJZD20240903100904007); the Science and Technology Planning Project of Guangdong Province supporting the Guangdong Provincial Key Laboratory for Prevention and Treatment of Digestive System Malignant Tumors (2021) (Grant No. 2021B1212040006); the Sanming Project of Medicine in Shenzhen (Grant Nos. SZSM202411013 and SZSM202411023); and the Shenzhen Clinical Research Center for Gastroenterology (Gastrointestinal Surgery) (Grant No. LCYSSQ20220823091203008).

Data availability

All datasets generated or analyzed in this study are deposited in publicly available repositories. Specific repository names and accession numbers can be found in the main text of the article.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Competing interests

The authors declare no conflict of interest.