Integrative single-cell and multi-omics analysis of ZBTB21-mediated serine metabolism in colorectal cancer: from metabolic reprogramming to immune microenvironment modulation

Abstract

Background

Colorectal cancer (CRC) exhibits pronounced biological diversity, a feature increasingly attributed to alterations in cellular metabolic reprograms. Serine metabolism supports nucleotide synthesis, redox balance, and epigenetic regulation via one-carbon metabolism, yet its role in shaping the tumor cellular interactions and immune landscape at single-cell level remains unclear.

Methods

Single-cell transcriptomic profiles (GSE284449) were jointly analyzed with bulk expression datasets (GSE271719), together with Mendelian randomization (MR)-based inference, to dissect serine metabolism in CRC. High-serine-metabolism (HSM) cell populations were identified using scMetabolism, and robust marker genes were selected through Lasso, random forest, XGBoost, and SVM-RFE machine learning approaches. MR was applied to evaluate causal associations with CRC risk. Functional validation included IHC, qRT-PCR, western blot, and LC–MS metabolomics, while CellChat analysis characterized HSM cell interactions with immune and stromal cells.

Results

ZBTB21 and TRPM2 were identified as core regulators of HSM cells, with ZBTB21 predominantly expressed in monocytes and pro-B cells. MR analysis suggested a potential inverse association between genetically predicted ZBTB21 expression and CRC risk, indicating that ZBTB21 may exert context-dependent effects at the population level. Cell–cell communication analysis suggested that HSM monocytes may interact with fibroblasts through signaling pathways including the MIF–SPP1 axis; however, these interactions are computationally inferred and require further experimental validation. Functional assays showed that ZBTB21 overexpression upregulated PHGDH and SHMT1, increased intracellular NADPH/NADP⁺ ratios, and influenced monocyte-related phenotypes. ZBTB21 levels were markedly increased in CRC samples relative to matched non-tumorous mucosal tissues.

Conclusions

At single-cell resolution, ZBTB21 emerges as a metabolic regulator that strengthens serine biosynthesis and redox homeostasis. While integrative analyses suggest a potential link between ZBTB21-associated metabolic states and immune interactions, further experimental validation is required to establish causal relationships. This integrative framework connects genetic causality, metabolism, and immune interactions, providing mechanistic insights and potential strategies for metabolic-targeted therapies in CRC.

Trial registration: Not applicable.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s00262-026-04427-7.

Introduction

Colorectal cancer (CRC) is ranked among the most frequently diagnosed cancers globally and accounts for a substantial proportion of cancer-associated disease burden and deaths [1, 2]. Metabolic reprogramming has emerged as a central driver of colorectal tumor initiation and disease advancement [3–5]. Serine-centered metabolic programs directly fuel de novo nucleotide production, preserve intracellular redox homeostasis, and influence chromatin regulation through epigenetic modifications [6, 7]. Dysregulation of serine biosynthesis enzymes, such as PHGDH, PSAT1, PSPH, and SHMT1/2, has been observed in CRC tissues, implying that elevated serine flux may facilitate tumor cell proliferation and survival [6, 8, 9]. Despite these observations, most studies are based on bulk transcriptomic analyses, which cannot resolve the cellular heterogeneity within the tumor microenvironment (TME) or distinguish cell-type-specific metabolic dependencies. This gap highlights the necessity for analytical strategies capable of resolving metabolic circuitry at cellular granularity, such as single-cell-level transcriptomic profiling, which permits systematic reconstruction of metabolic interactions across distinct cell compartments [10–12].

Recently, multi-omics integration has been increasingly enabled systematic interrogation of the regulatory interplay connecting metabolic programs with immune activity. Such analyses facilitate the prioritization of candidate genes from integrated transcriptional and metabolic features and enable functional inference of their involvement in serine-related metabolic processes and immune modulation in CRC [6]. This analytical framework extends beyond conventional in vitro and in vivo models while offering translational insights that inform the design of innovative therapeutic interventions. Integrative analytical strategies enable the prioritization of serine metabolism-associated genes and delineate molecular candidates with potential relevance for future metabolic and immunomodulatory therapeutic development [13]. Serine metabolic dysregulation in CRC arises from coordinated crosstalk among multiple interconnected pathways rather than from an isolated regulatory mechanism [6, 14]. By integrating information across diverse omics layers, such approaches can comprehensively evaluate how serine-related metabolic pathways shape immune responses, thus clarifying their collective contribution to CRC initiation and progression [15–17].

This study leveraged publicly accessible colorectal tumor single-cell transcriptomic resources curated in GEO and TISCH, together with large-cohort, variant-level genetic association datasets pertaining to serine metabolic pathways. The analytical framework comprised five sequential modules: (1) two-sample Mendelian randomization to infer causal relationships between core serine metabolic enzymes and CRC susceptibility; (2) single-cell transcriptomic profiling to delineate cell type-resolved expression patterns of PHGDH, PSAT1, PSPH, and SHMT1 across malignant, immune, and stromal compartments; (3) integration of bulk RNA-seq datasets with machine learning algorithms to construct multi-gene risk prediction scores; (4) cell–cell communication and pseudotime trajectory analyses to characterize interactions between serine metabolic cancer cell subsets and stromal or immune populations; and (5) integrative interpretation of metabolic–immune cross talk in CRC pathogenesis. Together, these strategies illuminate the cellular and genetic architecture of serine metabolism in CRC and identify potential therapeutic targets for its prevention and treatment.

Methods

Data acquisition and integrity assessment

Single-cell transcriptomic datasets from CRC tissues and matched non-tumorous tissues were obtained from the Gene Expression Omnibus (GEO, accession GSE284449) [18]). Samples were pre-labeled as tumor (T1-T3) or normal (N1-N3) mucosal tissues based on clinical metadata, providing ground-truth labels for downstream comparative analyses. Raw count matrices were loaded into R (v4.3.1) and processed using Seurat (v4.3.1). Cells with fewer than 200 detected genes, more than 10,000 genes, or a mitochondrial gene fraction exceeding 10% were excluded. Quality metrics—including total RNA counts (nCount_RNA), number of detected genes (nFeature_RNA), and mitochondrial transcripts proportion (percent.mt)—were visualized using scatter and violin plots to assess sample consistency and identify outliers.

High variable genes were selected with the FindVariableFeatures function, retaining the top 2,000 genes with the highest standardized variance. Principal components, determined by the ElbowPlot method, were used for neighborhood graph (FindNeighbors) and Louvain clustering (FindClusters, resolution = 0.2). For two-dimensional visualization, RunUMAP was applied.

Cell populations were annotated using a combined approach of automated inference and manual validation with canonical markers. Initial cell identities were assigned using SingleR with the Human Primary Cell Atlas as a reference, followed by manual refinement based on established lineage-specific markers. For T cells, we primarily relied on the gold-standard markers CD3D and CD3E. IL7R, which emerged as a top differentially expressed gene for the T cell population in this dataset, was also examined but used with caution given its known expression in pro-B cells. For B cells, while CD19 and CD79A/B are standard markers, MS4A1(CD20) and BANK1 emerged as the most statistically prominent features in this dataset and were used for visualization, complemented by CD79A/B for validation. The pro-B and pre-B cell clusters, representing early developmental stages, were identified through their global correlation with proliferative progenitor profiles in the reference atlas and confirmed by the expression of proliferative markers such as KIF18B. For NK cells and fibroblasts, we used KLRD1 and IGFBP7, respectively. Additional markers, including AP1S1 and RARRES2, were used to characterize stromal and immune subsets. Of note, a cell cluster initially labeled as astrocytes by SingleR was reinterpreted as enteric glial cells, which are physiologically present in the gut wall. While GFAP is a canonical glial marker, its transcript is subject to frequent dropout in scRNA-seq; therefore, we used cluster-specific features including LINC00461 and MEGF10 for transcriptomic characterization. Marker specificity was verified using FeaturePlots and violin plots, with cluster markers required to be expressed in > 30% of cells at an average expression > 0.5.

Differential cell-type analysis and visualization

To examine differences in cellular composition between tumor and normal tissues, metadata extracted from the Seurat object—including sample ID, tissue type, cell type, and cluster information—were grouped and analyzed using the dplyr package. Dot plots were used to display the distribution of lineage-specific genes among cell populations, with dot area reflecting the fraction of cells with detectable expression and color intensity corresponding to mean transcript abundance. Feature plots (FeaturePlot) were used to visualize expression of individual genes across UMAP embeddings, confirming annotation consistency.

Metabolic pathway activity scoring

Metabolic activities of single cells were quantified using gene set scoring approaches. Specifically, pathway activity scores for serine metabolism, glycolysis, TCA cycle, and pentose phosphate pathway were calculated for each cell. Dot plots and boxplots were used to compare pathway activity across annotated cell types. UMAP feature plots were used to map serine metabolic activity across cells, facilitating the detection of metabolically active subpopulations.

Differential expression analysis

To identify genes with significantly altered expression in CRC compared with matched non-tumorous tissues, we performed differential expression analysis (DEA) on bulk transcriptomic data using DESeq2 (v1.40.2). Genes with a false discovery rate (FDR) < 0.05 and an absolute log₂ fold change > 1 were considered statistically significant differentially expressed genes (DEGs). Volcano plots were generated to illustrate the significance and effect size of serine metabolism-associated DEGs, while heatmaps were used to compare expression patterns across individual samples. Candidate genes for subsequent modeling were defined as the intersection of these DEGs and marker genes derived from single-cell serine metabolism-active subpopulations.

Identification of serine metabolism-related core genes

To identify key serine metabolism-related genes with high predictive value for CRC status, we employed four machine learning algorithms: Lasso regression (glmnet v4.1–7), random forest (randomForest v4.7–1.1), extreme gradient boosting (XGBoost v1.7.6), and support vector machine recursive feature elimination (SVM-RFE, e1071 v1.7–13) [19–22]. These models were applied to distinguish CRC samples from matched normal mucosal tissues based on the expression profiles of candidate serine metabolism-related genes, with clinical diagnosis as the response variable. To minimize method-specific bias and enhance biological robustness, we defined core genes as those consistently exhibiting high importance across all four algorithms. This consensus strategy leverages the distinct mathematical principles of each method: Lasso enforces feature sparsity via L1 regularization; random forest and XGBoost utilize ensemble tree-based learning and gradient boosting; and SVM-RFE performs recursive feature elimination using support vector machines. By selecting the intersection of top contributors across these diverse approaches, we prioritized candidates with high cross-model stability, reducing the risk of overfitting or algorithm-specific artifacts. Model performance was evaluated using tenfold cross-validation combined with bootstrap resampling (n = 1,000). Although all four algorithms achieved high predictive accuracy, no single method consistently outperformed the others across all metrics. Therefore, we adopted an ensemble-like consensus approach rather than selecting a single algorithm based on marginal performance differences, thereby leveraging the complementary strengths of multiple frameworks. The identified core genes were further validated in independent external datasets to confirm robustness and generalizability, establishing a foundation for downstream functional and mechanistic analyses.

Model explainability assessment

Shapley Additive Explanations (SHAPs) were applied to elucidate feature contributions in the XGBoost model, with values computed for each input variable. Individual-level contributions to the model were visualized using force plots, showing both the direction and strength of each gene’s effect on predicted outcomes. Pairwise effects among candidate genes were examined using interaction SHAP plots. Partial dependence plots (PDPs) illustrated the population-level association between individual genes and the predicted outcomes, while individual conditional expectation (ICE) plots were generated to show the gene-specific effects on predictions for each sample. The top three core genes were subsequently identified based on their SHAP importance scores and further validated in external datasets to ensure robustness and generalizability.

Intercellular crosstalk analysis

Monocytes in the upper quartile of serine metabolism activity were defined as high-activity cells, and those in the lowest quartile as low-activity cells. Cell–cell communication among annotated clusters was investigated using CellChat (v1.6.1) and CellPhoneDB (v4.1.0). Ligand–receptor interactions were evaluated based on normalized expression matrices, with statistical significance determined by a permutation test (n = 1,000); interactions with p < 0.05 were considered significant. To ensure biological relevance, a ligand or receptor was required to be expressed in at least 10% of cells within its respective cluster for the corresponding interaction to be retained. CellChat analyses were performed using the CellPhoneDB human database, encompassing over 1,900 validated interactions involving secreted signaling, extracellular matrix (ECM)–receptor binding, and cell–cell contacts. CellPhoneDB analyses employed its proprietary database of ligands, receptors, and multimeric complexes. Communication networks and interaction strengths were visualized using the built-in plotting functions of each tool.

Pseudotime trajectory analysis

Pseudotime analysis was performed to investigate dynamic changes in gene expression among monocytes. High-activity monocytes were reprocessed for dimensionality reduction, followed by pseudotime trajectory inference using Monocle (v2.22.0). Cells with mean expression > 0.1 and empirical dispersion exceeding the fitted dispersion were retained. Dimensionality reduction was conducted using the DDRTree algorithm via reduceDimension, and cells were subsequently ordered along pseudotime trajectories using orderCells. Cell distributions along pseudotime were visualized using the plot_cell_trajectory utility. Expression dynamics of core serine metabolism-related genes were examined across pseudotime to identify stage-specific regulatory patterns.

Exposure data sources

Two categories of exposure variables were considered for causal inference with CRC risk. Whole-blood cis-expression quantitative trait loci (cis-eQTLs) for ZBTB21 were obtained from the eQTLGen Consortium. Instrumental variables (IVs) were defined as single-nucleotide polymorphisms (SNPs) with a false discovery rate (FDR) < 0.05, ensuring strong association with gene expression. Linkage disequilibrium pruning was performed (r² < 0.001, physical distance > 10,000 kb) to reduce correlation among IVs, and variants with F-statistics > 10 were retained to ensure instrument strength. Additionally, SNPs reaching genome-wide significance for circulating serine levels (p < 5 × 10⁻⁸) were extracted from publicly available GWAS datasets, with similar LD pruning and F-statistics criteria applied. All selected IVs were harmonized with outcome data, and strand ambiguity was resolved to maintain consistency across datasets.

Outcome data sources

Outcome data were derived from large-scale CRC GWAS summary statistics comprising tens of thousands of participants of European ancestry. Age, sex, and up to 20 principal components were included to control for population stratification.

Instrumental variable harmonization

Genetic variants showing genome-wide significance for plasma serine concentrations (p < 5 × 10^–8) were initially selected as candidate IVs. To reduce bias from linkage disequilibrium, SNPs were further filtered by LD clumping (r² < 0.001) within a 10,000-kb genomic window. Only IVs with F-statistics > 10 were retained to ensure sufficient instrument strength. Effect alleles were aligned across exposure and outcome datasets, with strand inconsistencies resolved to ensure concordance. Palindromic SNPs with ambiguous strands were either corrected or excluded to prevent misalignment.

Two-sample mendelian randomization analyses

Two independent MR frameworks based on separate exposure-outcome datasets were implemented to assess whether ZBTB21 expression and circulating serine levels exert causal effects on CRC susceptibility. For ZBTB21, cis-eQTLs served as IVs, and the causal effects were primarily estimated using inverse-variance weighted (IVW) regression. A series of complementary diagnostic procedures was applied to assess stability. Directional pleiotropy was examined using MR-Egger, between-instrument heterogeneity was quantified via Cochran’s Q test, and potential outlier variants were screened using MR-PRESSO. A leave-one-out analysis was performed to evaluate the influence of individual SNPs. Consistent causal estimates were further obtained using IVW method, accompanied by the same diagnostic assessments. To explore potential reverse causality, bidirectional MR analyses were conducted to assess whether CRC-associated genetic variants influence ZBTB21 expression or circulating serine levels. This structured approach, with independent evaluation of each exposure and comprehensive sensitivity testing, ensures rigorous and interpretable causal inference regarding the role of ZBTB21 and serine metabolism in CRC risk.

Clinical sample collection and immunohistochemistry (IHC)

Formalin-fixed, paraffin-processed CRC specimens together with paired non-tumorous tissues (n = 8 per group) were cut into 4- to 5-μm sections. Following paraffin removal and stepwise alcohol hydration, heat-induced epitope retrieval was performed in EDTA solution (pH 9.0) using a pressurized high-temperature protocol. Endogenous peroxidase was quenched using 3% hydrogen peroxide in methanol, and non-specific antibody binding was minimized by incubation with 5% bovine serum albumin in PBS. Sections were incubated with anti-ZBTB21 primary antibody (Santa Cruz, sc-518254, 1:100) at 4 ℃ overnight, followed by incubation with a biotin-conjugated secondary antibody and SABC complex at ambient temperature. DAB substrate was used for visualization, and sections were counterstained with hematoxylin, differentiated in acid alcohol, and blued in ammonia water. After staining, sections were dehydrated in graded alcohols, cleared with xylene, and coverslipped with neutral mounting medium. Fluorescence images were acquired using a Nikon Eclipse C1 system, and ZBTB21 levels were quantified based on both signal intensity and the proportion of positively stained cells.

Cell culture and siRNA transfection

Caco2 cells (Procell Life Science & Technology Co., Ltd., Wuhan, China) were maintained in glucose-containing, serine- and glycine-free culture medium. Cells were seeded at a density of 1 × 10⁵ per well in 6-well plates. ZBTB21-specific siRNAs and a non-targeting negative control (GenePharma) were delivered using Lipofectamine 3000 (Thermo Fisher) according to the manufacturer’s instructions. Knockdown efficiency was confirmed by quantitative PCR. ZBTB21 overexpression plasmids were transfected similarly. Four experimental groups were established: Vector control, ZBTB21 overexpression, NC siRNA, and ZBTB21 siRNA.

Cell viability, proliferation, and clonogenic assays

Cell growth was evaluated using the Cell Counting Kit-8 (CCK-8, Biosharp, BS350B). A total of 5 × 10³ cells per well were seeded in 96-well plates and incubated for 12 h before treatments. CCK-8 solution (10% v/v) was added to each well, incubated at 37 ℃ for 2 h, and optical density at 450 nm was measured. Cell proliferation was assessed with the EdU Apollo 567 Kit (RiboBio) following the manufacturer’s protocol. For colony formation, 500 cells per well were plated in 6-well plates, cultured for 10–14 days, then fixed in methanol, and stained with 0.5% crystal violet. Colonies containing more than 50 cells were counted.

NADPH/NADP⁺ratio measurement

Cellular NADPH/NADP⁺ ratios were quantified using a commercial assay kit (Sigma-Aldrich, MAK187). Cells were lysed and deproteinized using the supplied extraction buffer, and reactions were performed according to the kit protocol. Optical density at 450 nm was recorded, and NADPH/NADP⁺ ratios were calculated relative to total protein determined by BCA assay.

Western blot analysis

Proteins were extracted with RIPA buffer (Beyotime, P0013B) containing protease inhibitors (Biyuntian, S1873) and PMSF (Sigma, ST506). Protein concentrations were measured using the Bradford assay (Beyotime, P0006). For each sample, 30 μg of protein was resolved by SDS–PAGE and transferred to PVDF membranes (Millipore, IPVH00010 or ISEQ15150). After blocking with 5% non-fat milk in TBST, membranes were incubated at 4 ℃ overnight with primary antibodies targeting PHGDH (Proteintech, 14,719–1-AP, 1:1000) and SHMT1 (Proteintech, 30,192–1-AP, 1:1000). HRP-conjugated secondary antibodies (CST, Rabbit #7074 or Mouse #7076, 1:10,000) were applied, and chemiluminescent signals were detected using ECL substrate (Thermo Fisher, K-12045-D50). Signal intensities were analyzed using ImageJ and normalized to GAPDH (Proteintech, 60,004–1-IG, 1:5000).

ELISA analysis

Secreted MIF and OPN levels in culture medium were measured using human ELISA kits (LianKe Bio, MIF: EK1158; OPN: EK1135) according to the manufacturers’ protocols. Standards and appropriately diluted samples (100 µL each) were added to 96-well plates and incubated at 4 ℃ overnight. To reduce non-specific binding, wells were incubated with 1% BSA at room temperature for 1 h. Wells were then incubated with enzyme-linked detection antibodies at 37 ℃ for 1 h, followed by three PBS washes. Substrate was added and allowed to react in the dark for 30 min before the reaction was stopped. Absorbance at 450 nm was recorded using a microplate reader (BioTek, Synergy HTX), and protein concentrations were determined from standard curves. Experiments were performed in triplicate, and statistical significance was evaluated using one-way ANOVA or two-tailed t tests, with p < 0.05 considered significant.

Metabolite analysis by LC–MS

Intracellular metabolites, including serine and other energy-related metabolites, were extracted from Caco2 cells with cold 80% methanol. Samples were centrifuged at 12,000 × g for 10 min at 4 ℃, and the supernatants were evaporated under nitrogen and reconstituted in 50 µL of 50% methanol. Metabolite profiling was conducted using an Agilent 1290 Infinity II UHPLC coupled to a 6545 Q-TOF mass spectrometer. Metabolites were identified and quantified using authentic standards and retention times.

Gene expression and secreted protein analysis

Total RNA was extracted with TRIzol (Invitrogen) and reverse-transcribed to cDNA using the HiScript III cDNA Synthesis Kit (Vazyme). Quantitative PCR was performed on a QuantStudio 6 Flex system with ChamQ SYBR qPCR Master Mix (Vazyme). Relative expression of MIF and SPP1 mRNAs was determined by the 2^−ΔΔCt method and normalized to GAPDH. Corresponding secreted proteins in the culture medium were measured using ELISA kits (R&D Systems) manufacturers’ protocols.

Statistical analyses

All analyses were performed in R (v4.3.1). For continuous data, comparisons were made using either Student’s t test or the Mann–Whitney U test, while categorical data were analyzed with chi-square or Fisher’s exact tests. Two-sample Mendelian randomization was performed via the “TwoSampleMR” R package, applying IVW regression for the main causal estimate and MR-Egger for sensitivity analysis. Cochran’s Q test and the MR-Egger intercept were used to evaluate heterogeneity and pleiotropy, while MR-PRESSO identified outlier variants, and leave-one-out analysis assessed the influence of individual instruments. Multiple comparisons were adjusted using the false discovery rate (FDR), with significance defined as P < 0.05. Single-cell RNA sequencing analyses—including clustering, differential expression, pathway scoring, pseudotime inference, and intercellular communication—were conducted using Seurat, Monocle, CellChat, and CellPhoneDB, with consistent statistical methods applied throughout.

Results

Single-cell data preprocessing and analysis

We applied rigorous filtering to the single-cell RNA-seq dataset of CRC and matched normal samples (GSE284449) to exclude low-quality cells and outliers. Violin plots of QC parameters (Supplementary Fig. 1A) display nFeature_RNA, nCount_RNA, and mitochondrial gene proportion (percent.mt) for normal (N1-N3) and tumor (T1-T3) samples. Scatter plots showed minimal correlation between mitochondrial gene proportion and total RNA content (percent.mt vs. nCount_RNA, Pearson r = -0.09) and confirmed the expected positive correlation between detected gene number and RNA content (nFeature_RNA vs. nCount_RNA, r = 0.85) (Supplementary Fig. 1B).

To identify highly variable genes, we plotted average gene expression against standardized variance and selected the top 2,000 most variable genes for downstream analyses (Supplementary Fig. 1C). Dot plots illustrated sample-specific distributions of major cell types, with circle size representing cell proportion and color intensity reflecting gene expression levels (Supplementary Fig. 1D). UMAP-based dimensionality reduction visualized cell clustering, revealing sample-specific segregation and enabling preliminary identification of transcriptionally distinct subpopulations (Supplementary Fig. 1E).

Principal component analysis of the filtered dataset showed partial separation between tumor and normal cells along PC1 and PC2, indicating underlying transcriptional heterogeneity (Supplementary Fig. 1F). The first 20 principal components captured the majority of biological variation, guiding subsequent neighborhood graph construction and clustering (Fig. 2G).

Fig. 2.

Feature selection and model interpretation across multiple machine learning algorithms. (A) Lasso cross-validation plot showing log(λ) versus binomial deviance; the dashed line marks the selected optimal λ. (B) Lasso coefficient profiles showing gene coefficients versus L1 norm. (C) Random forest error rate across tree numbers. (D) RF variable importance: %IncMSE (left) and IncNodePurity (right). (E) XGBoost cross-validation RMSE versus number of variables. (F) XGBoost feature importance by gene (color-coded by cluster). (G) Venn diagram showing the overlap of candidate genes identified by Lasso, RF, SVM, and XGBoost. (H) SHAP summary plot for XGBoost, showing gene-level contributions to disease prediction (color: expression level; x-axis: SHAP value)

Cell-type-specific expression of key marker genes in CRC

To investigate the cellular context of key immune and stromal markers in CRC, single-cell transcriptomic data were analyzed to classify major populations in tumor and matched normal tissues. Dot plot visualization of canonical markers enabled precise annotation of T cells, B cells, NK cells, monocytes, and fibroblasts (Supplementary Fig. 2A). UMAP projections showed that IL7R, MS4A1, KLRD1, IGFBP7, AP1S1, and RARRES2 exhibited distinct, cell-type-specific expression patterns. IL7R was predominantly observed in T cell clusters, MS4A1 marked B cells, KLRD1 was enriched in NK cells, and IGFBP7 localized mainly to fibroblasts. AP1S1 and RARRES2 showed broader expression, consistent with roles in vesicle transport and chemokine responsiveness, respectively (Supplementary Fig. 2B). REG1B and OTOG displayed restricted expression in specific epithelial-like subsets. Violin plots confirmed the specificity of these markers, with elevated expression in their respective clusters relative to other cell types (Supplementary Fig. 2C).

Single-cell analysis further revealed distinct serine–glycine metabolic programs across immune and stromal lineages. Pro-B and mature B cells showed uniformly high PHGDH/PSAT1 expression; T cells preferentially expressed mitochondrial SHMT2; and NK cells selectively expressed SDSL. Monocytes exhibited stronger PHGDH expression than fibroblasts and smooth muscle cells, whereas SHMT1/2 were co-high in stromal populations. Notably, SDS and SDSL displayed mutually exclusive expression patterns primarily in monocyte and dendritic cell populations (Supplementary Fig. 2D-E), suggesting cell-type-specific transcriptional regulation within the serine catabolic pathway. Together, these data reveal a cell-type-specific organization of the serine–glycine/one-carbon axis, highlighting its potential functional specialization across distinct immune and stromal compartments.

Single-cell metabolic profiling reveals immune cell-specific heterogeneity

Bulk RNA-seq averages signals across heterogeneous immune populations, obscuring rare subsets with distinct metabolic programs. To resolve this, we analyzed single-cell transcriptomes to profile pathway-specific metabolic activity across annotated immune subsets. Dot plot visualization revealed lineage-specific metabolic specialization (Fig. 1A). Metabolism pathway activity scores, including serine metabolism, were calculated using the scMetabolism package based on KEGG gene sets (hsa00260 for serine and threonine metabolism), enabling a holistic assessment of pathway enrichment across immune populations. Serine metabolism was highly enriched in monocytes, whereas glycolysis, TCA cycle, and pentose phosphate pathway showed broader but lower activity in lymphocytes (Fig. 1A). Boxplots confirmed that monocytes exhibited significantly higher serine metabolism scores than T or B cells (Fig. 1B), and UMAP mapping localized high serine metabolic activity to monocyte clusters (Fig. 1C). Comparison of monocytes from tumor (T) versus normal (N) tissues revealed 2,062 DEGs, with elevated levels of S100A8, SPP1, and SLCO4A1, and reduced expression of CFD (Fig. 1D), reflecting inflammation-linked metabolic adaptation. Integration with 23 curated serine metabolism genes revealed 15 overlapping transcripts (Fig. 1E), highlighting a disease-associated perturbation of the serine metabolic axis.

Fig. 1.

Systematic characterization of serine metabolism-associated gene expression and pathway activity in CRC. (A) The dot plot illustrates the mean activity of representative metabolic pathways across major immune cell populations. Dot color represents the average pathway activity, while dot size reflects the fraction of cells exhibiting pathway activity in each population. (B) Boxplots display the distribution of pathway activity scores for the TCA cycle, glycolysis, pentose phosphate pathway, and serine metabolism across immune subsets. (C) UMAP visualization of single-cell serine metabolism activity. High-activity cells are spatially concentrated within monocyte clusters, forming a coherent region of metabolic enrichment. (D) Volcano plot showing differentially expressed genes (DEGs) in monocytes between disease (T) and control (N) groups. (E) Boxplots showing expression differences of 15 candidate genes between normal (N) and tumor (T) tissues across all cells (including immune, stromal, and epithelial compartments).*, **, *** indicate P < 0.05, 0.01, 0.001, respectively, and "ns" indicates no significance. (F) Volcano plot of DEGs between tumor and normal tissues (red: upregulated; blue: downregulated; gray: non-significant). (G) Bubble plot of GO and KEGG enrichment for DEGs. x-axis: GeneRatio; y-axis: GO terms or KEGG pathways.

Differential expression and functional profiling of serine metabolism-related genes in CRC

To characterize the transcriptional landscape of serine metabolism-related genes in CRC, we analyzed 15 candidate genes overlapped between serine metabolic regulators and DEGs identified in CRC. These genes were initially screened using DESeq2 (FDR < 0.05, |log₂FC|> 1), which employs a negative binomial generalized linear model with empirical Bayes shrinkage to account for overdispersion and enhance statistical power in bulk RNA-seq data. Their expression distributions across tumor (T) and normal (N) tissues were further visualized using boxplots, with significance assessed by the Mann–Whitney U test (Fig. 1E, based on all cells). Several genes involved in mitochondrial energy metabolism and redox balance—including NPM1, TIAL1, PDHA1, DLAT, and TRPM2—showed consistent significance across both analytical frameworks (P < 0.05). In contrast, six genes (RBM17, FDX1, SYAP1, LIAS, PDHB, and LIPT1) appeared as “non-significant” (ns) in the boxplots but were nonetheless identified as DEGs by DESeq2. This discrepancy arises from the distinct statistical principles of the two approaches: DESeq2 models read-count distributions with global dispersion sharing and applies FDR correction across thousands of genes, whereas the Mann–Whitney U test is a nonparametric pairwise test that does not incorporate such global information and is more sensitive to within-group variance. Notably, ZBTB21 exhibited robust differential expression across both methods (P < 0.001), reinforcing its reliability as a core candidate in subsequent analyses. Heatmap visualization (Fig. 1F) revealed upregulation expression of NPM1, TRPM2, and ZBTB21, alongside downregulation of DLAT, PDHA1, and FDX1 in tumor tissues. GO and KEGG enrichment analyses (Fig. 1G) indicated that these DEGs were enriched in pathways related to mitochondrial dehydrogenase complex assembly, pyruvate dehydrogenase activity, and oxidative phosphorylation.

Identification of candidate genes using machine learning approaches

While differential expression analysis identified ZBTB21 as a robustly altered gene in CRC, we further employed a multi-algorithm consensus machine learning strategy to move beyond univariate comparisons and reduce method-specific bias. This approach allowed us to assess the independent predictive contribution of each gene within a competitive modeling framework while prioritizing parsimony and generalizability.

First, Lasso regression with L1 regularization was applied to the 15 serine metabolism-related candidate genes. Lasso forces coefficients of uninformative or redundant predictors to zero, thereby preventing overfitting and multicollinearity and enhancing the model’s ability to generalize to independent datasets—a critical feature for identifying robust biomarkers. To further refine these candidates, we integrated three additional complementary algorithms—random forest (RF), support vector machine (SVM), and XGBoost—each based on distinct mathematical principles. Lasso regression, combined with cross-validation, was used to select the most informative features by identifying the optimal λ, emphasizing key contributing genes while filtering out less relevant ones (Fig. 2A–B). RF analysis showed that model error plateaued around 100 trees, with %IncMSE and IncNodePurity consistently identifying ZBTB21 and TRPM2 as top contributors (Fig. 2C–D). XGBoost analysis confirmed these findings, ranking ZBTB21 and TRPM2 highest in feature importance (Fig. 2E–F). A Venn diagram comparing selected genes across all four methods revealed that only ZBTB21 and TRPM2 were shared (Fig. 2G). To further elucidate the direction and magnitude of feature contributions, SHAP analysis of the XGBoost model was performed, indicating that ZBTB21 and TRPM2 exhibited the strongest positive contributions to disease prediction (Fig. 2H). This consensus across algorithms—each based on distinct mathematical principles—substantially reduces the risk of method-specific bias and reinforces the identification of ZBTB21 and TRPM2 as stable, high-priority candidates for subsequent functional investigation.

Fig. 3.

Mendelian randomization (MR) analysis of candidate genes and CRC risk. (A) Volcano plot of SNP effects on gene expression (x: β value; y: -log10 p value). (B) Colocalization analysis between GWAS and eQTL signals; posterior probability of shared causal variants (H4) shown. (C) MR forest plot of SNP-specific effect estimates; overall effects from IVW and MR-Egger models at the bottom. (D) MR scatter plot displaying SNP associations for exposure (x) versus outcome (y), including IVW, MR-Egger, and weighted median regression lines. (E) Additional MR forest plots for alternative datasets or exposure–outcome pairs. (F) MR sensitivity analyses, including MR-Egger heterogeneity tests to evaluate potential assumption violations

Mendelian randomization analysis of key genes and CRC risk

To explore potential causal links between candidate genes and CRC risk, we conducted MR analyses integrating eQTL and GWAS data. All cis-eQTL summary statistics were derived from colorectal tissues—specifically Colon-Transverse and Colon-Sigmoid—obtained from the Genotype-Tissue Expression (GTEx) database (version 8). This tissue-specific selection is critical because gene regulation is highly context-dependent, and using eQTLs from the disease-relevant tissue (colorectum) enhances the biological validity of causal inference. SNP effect sizes on gene expression were plotted against statistical significance in a volcano plot, highlighting strong associations for ZBTB21, TRPM2, DLAT, PDHA1, and GREM1 (Fig. 3A). Colocalization analysis showed ZBTB21 with the highest posterior probability of shared causal variants (H4), indicating that the genetic regulation of ZBTB21 expression and CRC risk likely arises from the same underlying locus (Fig. 3B). MR analyses suggested a potential inverse association between genetically predicted ZBTB21 expression and CRC risk, with MR-Egger indicating a significant negative effect estimate (Fig. 3C–D). Additional MR forest plots across datasets confirmed consistent effect directions and magnitudes (Fig. 3E), and sensitivity analyses including MR-Egger heterogeneity tests suggested minimal violation of MR assumptions (Fig. 3F), supporting the robustness of the causal inference.

Expression patterns of key genes at the single-cell level

To examine the distribution of TRPM2 and ZBTB21 across different cell populations, we interrogated their transcriptomic profiles at single-cell resolution (Supplementary Fig. 3). UMAP visualization showed TRPM2 was enriched in clusters at the center and upper-right regions, while ZBTB21 exhibited a broader, more diffuse pattern across multiple clusters (Supplementary Fig. 3A). TRPM2 expression was highest in monocytes, whereas ZBTB21 was most abundant in pro-B cells (CD34–) with moderate expression in T and B cells (Supplementary Fig. 3B). Dot plot analysis confirmed TRPM2’s high expression and coverage in monocytes and ZBTB21’s prominent expression in pro-B cells (CD34–) and enteric glial cells (initially labeled as astrocytes by SingleR) (Supplementary Fig. 3C). These results reveal distinct cellular expression landscapes, suggesting roles for TRPM2 and ZBTB21 in immune regulation and differentiation within the CRC microenvironment.

ZBTB21 overexpression promotes CRC cell proliferation

Immunohistochemistry (IHC) analysis was conducted to examine ZBTB21 levels in colorectal tumors and matched adjacent mucosa from eight patients. Figure 4A illustrates a clear upregulation of ZBTB21 in cancerous tissues relative to normal mucosa. IHC images at 200 × and 400 × magnification revealed strong immunoreactivity in the cancer group, with brown staining indicating that ZBTB21 predominantly resides in both the cytoplasm and nuclear compartments of malignant cells (adenocarcinoma cells). These cells were identified by their characteristic morphological features, including cellular atypia, pleomorphism, and disordered glandular or nested architecture, which distinguish them from surrounding interstitial myeloid cells and monocytes. Quantitative analysis of integrated optical density (IOD) further confirmed these observations, showing approximately a 2.0-fold increase in ZBTB21 expression in colorectal tumor samples compared with adjacent non-tumor mucosa (P < 0.001) (Fig. 4A).

Fig. 4.

ZBTB21 is upregulated in CRC and promotes cancer cell proliferation. (A) IHC images and quantitative depicting ZBTB21 in CRC tumors and matched normal tissues. Left: 200 × ; right: 400 × magnification. Scale bars: 100 µm and 50 µm. Brown signals indicate ZBTB21 mainly in the cytoplasm and nucleus. Data are presented as mean ± SEM (IOD of mean), n = 6. (B) ZBTB21 transcript levels in CRC cells were measured by qRT-PCR after overexpression or siRNA silencing, with results shown as mean ± SEM. (C) Cell viability was assessed by CCK-8 at 0, 24, 48, 72, and 96 h in CRC cells treated with vector control, ZBTB21 plasmid, NC siRNA, or ZBTB21 siRNA, with data shown as mean ± SEM. (D) EdU incorporation assay showing the proliferation of CRC cells transfected with different plasmids or siRNAs (red: EdU-positive proliferating cells; blue: DAPI-stained nuclei), accompanied by quantitative analysis. Scale bar = 50 µm. Data are presented as mean ± SEM. (E) Colony formation assay showing clonogenic capacity of CRC cells under different treatments, with quantification of colony numbers. Data are presented as mean ± SEM (n= 3). *P < 0.05, **P < 0.01, ***P < 0.001 vs. control group. n = 3

qRT-PCR analysis showed that ZBTB21 overexpression markedly increased ZBTB21 mRNA levels in Caco2 CRC cells, whereas ZBTB21 siRNA efficiently suppressed its expression (Fig. 4B). Consistent results from CCK-8 experiments indicated a progressive increase in cell proliferation upon ZBTB21 overexpression, whereas silencing ZBTB21 led to reduced proliferation over time (Fig. 4C). EdU labeling further demonstrated a marked increase in proliferating cells in ZBTB21-overexpressing samples relative to the control group (Fig. 4D). Colony formation experiments showed that elevating ZBTB21 enhanced the clonogenic potential of cells, while its knockdown led to fewer colonies (Fig. 4E). Together, these results suggest that ZBTB21 facilitates the growth of CRC cells. Importantly, all in vitro experiments were conducted in medium lacking exogenous serine and glycine, a condition deliberately designed to limit external amino acid supply and sensitize cells to endogenous serine biosynthesis. Under these conditions, the regulatory effects of ZBTB21 on serine metabolic flux and cellular proliferation were accentuated, underscoring its role in modulating the intrinsic serine biosynthesis pathway.

ZBTB21 orchestrates metabolic reprogramming to support redox homeostasis and anabolic capacity

To assess the impact of ZBTB21 on cellular metabolism, we performed untargeted LC–MS–based metabolomic profiling. ZBTB21 overexpression substantially elevated the levels of five major metabolite classes-including organic acid, amino acids, carbohydrates, carboxylic acids, and organic oxygen compounds, whereas ZBTB21 knockdown maintained low metabolite abundances comparable to control cells, indicating a positive correlation between ZBTB21 expression and global metabolite accumulation (Fig. 5A). Metabolite quantification revealed that ZBTB21 overexpression substantially increased intracellular serine levels, whereas knockdown reduced serine abundance relative to controls, suggesting that ZBTB21 positively modulates serine metabolism (Fig. 5B). KEGG (hsa)-based pathway enrichment of differential metabolites highlighted significant enrichment in central carbon and amino acid metabolism pathways, including pyruvate metabolism, tricarboxylic acid (TCA) cycle, glycolysis/gluconeogenesis, the pentose phosphate pathway, and amino acid biosynthesis (Fig. 5C). Consistent with a role in metabolic regulation, ZBTB21 overexpression markedly increased the NADPH/NADP⁺ ratio, whereas knockdown significantly reduced it, indicating enhanced reducing capacity (Fig. 5D). Western blot analysis revealed that ZBTB21 positively regulated PHGDH and SHMT1, two critical enzymes linking serine biosynthesis to one-carbon flux (Fig. 5E). qPCR further confirmed transcriptional upregulation of PHGDH and SHMT1 upon ZBTB21 overexpression, with no change in MIF expression, suggesting selective regulation of metabolic rather than inflammatory genes (Fig. 5F). Functionally, ZBTB21 overexpression reduced intracellular reactive oxygen species (ROS) levels, consistent with increased NADPH-mediated antioxidant activity, while knockdown elevated ROS accumulation. Additionally, ZBTB21 overexpression suppressed MIF secretion and SPP1 expression, indicating that metabolic rewiring by ZBTB21 also dampens inflammatory signaling (Fig. 5G). Collectively, these results demonstrate that ZBTB21 reprograms central carbon and amino acid metabolism—including serine biosynthesis, glycolysis, and the TCA cycle—to support NADPH production, maintain redox homeostasis, and modulate immune-related signaling (Fig. 5H).

Fig. 5.

ZBTB21 orchestrates metabolic reprogramming to support redox homeostasis and immune modulation. (A) Relative abundance distribution of major metabolite classes in cells subjected to different treatments. (B) ZBTB21 regulates intracellular serine levels. Statistical significance was assessed by one-way ANOVA. (C) KEGG pathway enrichment of differential metabolites, highlighting central carbon and amino acid metabolism. (D) Quantification of the NADPH/NADP ⁺ ratio in cells under different treatments, including vector control (vector), ZBTB21 overexpression (ZBTB21 OE), and ZBTB21 knockdown (ZBTB21 siRNA). (E) Western blot analysis of PHGDH and SHMT1 protein levels across treatment groups. Representative blots and densitometric quantification are shown, with β-actin used as a loading control. (F) Relative mRNA levels of SPHK1, PHGDH, SHMT1, and MIF were measured by qPCR and normalized to GAPDH. (G) ELISA assays quantifying the secretion of MIF and SPP1. Experiments were conducted in triplicate, and statistical significance was assessed using one-way ANOVA with Tukey’s post hoc test (p < 0.001, n = 3). (H) Schematic model illustrating how ZBTB21-mediated reprogramming of serine metabolism shapes the tumor immune microenvironment. ZBTB21 as a central regulator of redox homeostasis and serine metabolism, which orchestrates the expression of PHGDH and SHMT1 to sustain NADPH production and antioxidant defense while restraining proinflammatory cytokine release

Discussion

CRC continues to rank among the most prevalent malignancies globally, highlighting the importance of uncovering the molecular drivers underlying tumor progression. This study revealed ZBTB21 and TRPM2 as principal regulators of the HSM cell subset, with ZBTB21 showing preferential expression in monocytes and pro-B cells. By combining scRNA-seq, bulk transcriptomic data, and multiple computational approaches, we conducted an in-depth investigation of serine metabolism across diverse CRC cell populations. This integrative strategy allowed identification of key cellular and molecular factors underlying metabolic reprogramming, an area that has received limited prior attention. Functional validation, including IHC, qRT-PCR, western blot, and LC–MS metabolomics, confirmed ZBTB21 upregulation in CRC tissues and its regulatory role in serine metabolism, suggesting that it could serve as a candidate for interventions aimed at modulating cellular metabolism and remodeling the TME.

Our results provide novel insights into serine metabolism within CRC, revealing cell-type-specific patterns that were not captured by earlier studies focusing on generalized mechanisms. While metabolic reprogramming has been implicated in tumor reprogramming and immune escape, we reveal ZBTB21 as a critical regulator in monocytes and pro-B cells, highlighting a potential avenue for precision-targeted interventions. While immune and stromal components are known contributors to TME remodeling, how discrete metabolic regulators like ZBTB21 coordinate metabolic–immune interactions between cell populations is still largely unexplored. Prior work has linked the MIF–SPP1 axis in immune–stromal communication and tumor progression. Here, we uncover a previously unrecognized mechanism whereby ZBTB21-regulated metabolic programs in HSM monocytes drives a “metabolically activated monocyte–fibroblast” network via MIF–SPP1 signaling, highlighting its critical role in shaping the tumor immune microenvironment [10, 23–26]. Compared with earlier work that relied on bulk-level analyses and limited mechanistic resolution, this study clarifies the contribution of metabolically altered immune cell states to colorectal cancer progression. Through the combined use of single-cell and bulk transcriptomic datasets, together with data-driven modeling approaches, we delineated functionally relevant cell subsets and gene signatures, supporting their translational value in tumor microenvironment-oriented diagnosis and intervention.

MR analysis suggested a potential causal involvement of ZBTB21 in CRC risk, which warrants further investigation, as functional experiments showed ZBTB21 promotes HSM cell proliferation and metabolic rewiring. Mechanistically, ZBTB21 modulates the serine biosynthesis pathway through upregulation of PHGDH and SHMT1, enhancing intracellular NADPH/NADP⁺ ratios and maintaining redox balance in HSM monocytes. This metabolic reprogramming promotes monocyte recruitment and polarization via the MIF–SPP1 axis, thereby driving immune–stromal crosstalk and shaping the tumor microenvironment [27, 28]. HSM monocytes emerged as a central cellular component in CRC, with ZBTB21 prominently enriched within this population. Our findings indicate that ZBTB21 participates in the regulation serine metabolic pathways while simultaneously integrating metabolic programs with immune-related signaling. Elevated ZBTB21 expression was closely linked to increased serine pathway activity in both monocytes and pro-B cells. Distinct from earlier reports, this work positions ZBTB21 as a molecular bridge connecting metabolic remodeling and immune communication, characterized by pronounced cell-type-restricted expression and a unique role in reshaping the tumor microenvironment. Experimental validation through immunohistochemistry, quantitative PCR, immunoblotting, and LC–MS-based metabolomic profiling consistently demonstrated substantial induction of ZBTB21 together with its downstream enzymes PHGDH and SHMT1, supporting its functional involvement in regulating monocyte infiltration and phenotypic polarization through the MIF-SPP1 signaling cascade.

SHMT1 facilitates serine-derived one-carbon flux, generating glycine and reducing folate intermediates that support nucleotide biosynthesis and couple serine utilization to NAD(P)H-mediated redox homeostasis. Changes in SHMT1 activity can disrupt the balance between biosynthesis and antioxidant defense, affecting cellular resilience under metabolic or oxidative stress[29]. PHGDH functions as a key control point in endogenous serine biosynthesis, providing the serine supply required to fuel SHMT-mediated one-carbon metabolic flux. Its dysregulation shows stage- and context-specific effects across cancers. PHGDH upregulation enhances anabolic precursor availability and strengthens NADPH-mediated redox buffering, directly linking serine metabolic rewiring to cellular stress resilience [30, 31]. Collectively, these mechanistic insights highlight ZBTB21 as a potential therapeutic target in CRC, given its coordinated regulation of serine metabolism and immune remodeling.

Notwithstanding these advances, several aspects merit additional exploration. First, while ZBTB21 was shown to influence serine metabolic programs and immune interactions, the underlying molecular circuitry driving these effects has yet to be fully delineated. Second, the present work largely focused on transcriptional activity and metabolic readouts; potential regulation of ZBTB21 at post-transcriptional levels (such as phosphorylation or ubiquitin-mediated turnover) and post-translational layers (including acetylation and glycosylation) was not addressed. Third, while in vitro and tissue validation confirmed the relevance of ZBTB21, in vivo models and clinical studies are needed to evaluate its diagnostic and therapeutic potential. Expanding future investigations to include longitudinal cohorts, spatial transcriptomics, and single-cell multi-omics will be essential for fully elucidating ZBTB21-mediated metabolic–immune networks.

Conclusion

In summary, we defined a single-cell-level signature of high serine metabolism in CRC and identified ZBTB21 as a key regulatory factor through integrative transcriptomic and learning-based analyses. ZBTB21 links serine metabolic activity to immune regulation, particularly in monocytes and pro-B cells, thereby contributing to tumor microenvironment remodeling and CRC progression.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yimei Jiang , Xiaopin Ji and Ren Zhao contributed to the study conception and design. All authors collected the data and performed the data analysis. All authors contributed to the interpretation of the data and the completion of figures and tables. All authors contributed to the drafting of the article and final approval of the submitted version.

Funding

None.

Data availability

The datasets generated and analyzed during the present study are available from the corresponding author on reasonable request.

Declarations

Conflict of interest

The authors declare no competing interests.

Ethics approval and consent to participate

This study was approved by Ruijin Hospital Ethics Committee, Shanghai JiaoTong University School of Medicine (reference number: 2022–336). All procedures performed in studies involving human participants were in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

Informed consent

All data published here are under the consent for publication. Written informed consent was obtained from all individual participants included in the study.