A novel lactylation-related gene signature deciphers the immunosuppressive microenvironment and stratifies precision therapy in colorectal cancer Authors: Chi Fu, Chengwen Wang Published: 2026 ABSTRACT AbstractBackgroundColorectal cancer (CRC) remains a leading cause of cancer mortality, largely due to the heterogeneity of the tumor microenvironment (TME) and the limited efficacy of immunotherapy in microsatellite stable (MSS) tumors. Histone lactylation, a post-translational modification derived from the Warburg effect, serves as a critical bridge linking metabolic reprogramming to gene regulation and immune evasion; however, its specific prognostic value and clinical implications in CRC remain to be fully elucidated.MethodsIn this study, we systematically analyzed transcriptome profiling data from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO) cohorts, supplemented by single-cell RNA sequencing (scRNA-seq) analysis and Human Protein Atlas (HPA) protein-level validation. By integrating univariate Cox regression, Least Absolute Shrinkage and Selection Operator (LASSO) analysis, and multivariate Cox regression, we constructed a novel lactylation-related gene (LRG) risk signature. We extensively evaluated the association between this risk signature and patient prognosis, immune infiltration patterns, somatic mutations, and therapeutic sensitivity.ResultsA robust 9-gene prognostic signature (DHRS7, SPR, MBD2, RBM17, CSRP2, S100A4, TMSB4X, TKT, COPS4) was identified and corroborated at the protein level. Patients with high risk scores exhibited significantly worse overall survival (OS) across the training and two independent validation cohorts. Immunogenomic and scRNA-seq analyses revealed that high-risk tumors were characterized by an immunosuppressive and stromal-dense microenvironment—with stromal cells exhibiting the highest lactylation risk scores—enriched with regulatory T cells (Tregs), and frequently harbored PIK3CA mutations. Differential expression analysis indicated that this immune exclusion is structurally maintained by enriched extracellular matrix (ECM) organization and TGF-β signaling. Conversely, low-risk tumors displayed an inflamed phenotype with active antitumor immunity. Pharmacogenomic prediction identified distinct therapeutic stratifications: low-risk patients exhibited significant sensitivity to standard chemotherapeutics (fluorouracil, oxaliplatin) and EGFR/HER2 inhibitors (e.g., lapatinib, erlotinib). In contrast, high-risk patients showed specific vulnerabilities to novel targeted agents, including PI3K pathway inhibitors (TG-100-115, XL765), microenvironment-modulating agents (sildenafil, GANT-61), and epigenetic inhibitors (UNC0638).ConclusionWe established a novel lactylation-related risk signature that effectively stratifies CRC patients by prognosis and TME characteristics. By elucidating the crosstalk between metabolic dysregulation, stromal barriers, and immune exclusion, this study provides potential biomarkers and stratified therapeutic strategies—ranging from standard chemotherapy to targeted metabolic and stromal interventions—to optimize precision medicine for CRC patients.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s12672-026-05245-0.Keywords: Colorectal cancer, Lactylation, Tumor microenvironment, Prognostic signature, Precision medicine, Drug sensitivity Introduction Colorectal cancer (CRC) constitutes a formidable global health challenge, consistently ranking as the third most diagnosed malignancy and the second leading cause of cancer-related mortality worldwide [1, 2]. While overall cancer mortality rates have declined due to advances in screening and treatment, a concerning epidemiological shift is emerging with the disease burden increasingly affecting younger populations[3]. Recent statistics reveal that CRC has surpassed lung cancer to become the leading cause of cancer death among men under 50 years of age in the United States, accompanied by an annual incidence increase of 1% to 2% in this demographic [3]. In the realm of clinical management, identifying therapeutic targets within the tumor microenvironment (TME) has become a pivotal strategy for developing effective treatments [4]. Specifically, the advent of immune checkpoint inhibitors (ICIs) has revolutionized the landscape for patients with mismatch repair-deficient (dMMR) or microsatellite instability-high (MSI-H) tumors, offering robust clinical benefits [5]. However, this therapeutic success is sharply limited to a small subgroup, as the vast majority of patients harbor microsatellite stable (MSS) tumors that exhibit primary resistance to current immunotherapies [5, 6]. This significant clinical disparity highlights an urgent need to elucidate novel molecular mechanisms capable of overcoming immune evasion in MSS CRC. Metabolic reprogramming is a hallmark of cancer, wherein tumor cells prioritize glycolysis even under aerobic conditions—a phenomenon known as the Warburg effect—to support their rapid proliferation and survival[7]. This metabolic shift leads to the excessive accumulation of lactate, which was once considered merely a metabolic waste product but is now recognized as a critical signaling molecule and a potent regulator of the tumor microenvironment (TME) [8, 9]. In the context of CRC, elevated lactate levels create an acidic microenvironment that actively reshapes the immune landscape[10]. Recent reviews have highlighted that lactate metabolism in CRC exerts immunosuppressive effects by inhibiting the cytolytic function and proliferation of CD8+ T cells and natural killer (NK) cells, while simultaneously recruiting and activating immunosuppressive cells such as regulatory T cells (Tregs) and M2-polarized tumor-associated macrophages (TAMs) [10, 11]. This lactate-driven metabolic barrier acts in concert with immune checkpoints to facilitate tumor immune evasion and resistance to therapy [11]. A paradigm-shifting discovery in 2019 by Zhang et al. identified histone lactylation (Kla) as a novel post-translational modification (PTM) that directly links cellular metabolic status to gene regulation [12]. Derived from lactate, lactyl groups are added to histone lysine residues (eg.H3K18la) via acetyltransferase p300, thereby activating the transcription of specific genes involved in wound healing and tumorigenesis [12]. Since then, the scope of lactylation research has expanded rapidly. Comprehensive reviews and mechanistic studies have elucidated that lactylation is not limited to histones but also affects non-histone proteins, influencing diverse biological processes including glycolysis, immune escape, and therapy resistance across various malignancies [13–17]. Specifically in CRC, aberrant lactylation has been implicated in promoting tumor progression through mechanisms such as enhancing the stability of oncogenic proteins and modulating macrophage polarization [18]. Furthermore, recent investigations into the DNMBP-AS1/miR-30a-5p/PGC1α axis have revealed that metabolic regulators like PGC1α can suppress the Warburg effect and enhance anti-PD-1 efficacy in CRC, further underscoring the critical interplay between metabolism and immunity [19]. Despite these advances in understanding the metabolic and epigenetic landscapes of CRC, significant challenges remain. The molecular heterogeneity of CRC is vast, as evidenced by recent large-scale international studies revealing distinct genomic mutation patterns in early-onset colorectal cancer (EOCRC) compared to late-onset cases, suggesting that age-related metabolic and mutational differences may require tailored therapeutic approaches [20]. While individual targets and mechanisms involving lactylation have been identified, the comprehensive landscape of lactylation-related genes (LRGs) and their collective prognostic value in colon adenocarcinoma (COAD) remain largely unexplored. Currently, a systematic risk model capable of predicting patient outcomes and immunotherapeutic responses based on lactylation status is lacking. Therefore, this study aims to systematically screen prognostic LRGs, construct a robust risk signature, and comprehensively evaluate its association with the immune microenvironment and drug sensitivity. By bridging the gap between metabolic reprogramming and immune evasion, our findings aspire to provide novel biomarkers and therapeutic strategies for optimizing CRC management. Materials and methods Data acquisition and preprocessing Transcriptomic data (HTSeq-FPKM) and corresponding clinical annotations for Colon Adenocarcinoma (COAD) were retrieved from The Cancer Genome Atlas (TCGA) database via the UCSC Xena platform (https://xenabrowser.net/). The TCGA-COAD (449 patients with 39 healthy controls) cohort served as the training set. For external validation, two independent microarray datasets, GSE39582 (566 patients with COAD and 19 healthy controls)and GSE29621(65 patients with COAD), were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). Additionally, to dissect the tumor microenvironment at a single-cell resolution, single-cell RNA sequencing (scRNA-seq) data from the GSE132465 cohort was also acquired from the GEO database.To ensure analytical comparability and mitigate the impact of extreme values, the transcriptomic profiles from each dataset (TCGA, GSE39582, and GSE29621) were preprocessed independently. Specifically, the FPKM values of the TCGA-COAD cohort were converted to transcripts per kilobase million (TPM) and subsequently transformed. For the GEO microarray datasets, the raw matrix data were transformed. During the microarray data processing, probe IDs were mapped to gene symbols based on the corresponding platform annotation files. Probes mapping to multiple genes were discarded, and for genes targeted by multiple probes, the mean expression value was calculated. Identification of lactylation-related prognostic candidates A total of 332 lactylation-related genes (LRGs) were compiled for this study, encompassing the lactylase EP300, delactylases HDAC1-3 and SIRT1-3, and 327 lactylated proteins previously identified in human cell lines[21]. The full gene list is available in Supplementary Table S1. To identify genes with prognostic significance, the expression profiles of these LRGs were extracted from the TCGA training set. Univariate Cox proportional hazards regression analysis was performed using the "survival" R package to screen for genes significantly associated with OS (p<0.05). Simultaneously, Kaplan-Meier survival analysis was conducted to filter genes based on the log-rank test (p<0.05). The intersection of genes identified by both methods was selected as candidate prognostic markers for further modeling. Protein expression validation via HPA To validate the translational expression of the 9 signature genes, immunohistochemistry (IHC) data were retrieved from the Human Protein Atlas (HPA) database (https://www.proteinatlas.org/). Representative IHC images from normal colorectal tissues and colorectal cancer tissues were selected for comparison. Protein expression levels were evaluated based on staining intensity (negative, weak, moderate, or strong) and the quantity of stained cells () to provide protein-level evidence for the transcriptomic findings. Construction and validation of the lactylation-related risk signature To construct a robust and parsimonious prognostic model, the Least Absolute Shrinkage and Selection Operator (LASSO) Cox regression algorithm was applied to the candidate genes using the “glmnet” R package [22]. Ten-fold cross-validation was utilized to determine the optimal penalty parameter (λ ) and minimize the risk of overfitting. Subsequently, a multivariate Cox proportional hazards model was established, and backward stepwise selection based on the Akaike Information Criterion (AIC) was employed to identify the final gene signature. The risk score for each patient was calculated using the following formula: where represents the regression coefficient of the i-th gene derived from the multivariate Cox model, and denotes the expression level of the corresponding gene. Patients were stratified into high- and low-risk groups based on the optimal cut-off value determined by the "survminer" R package. Survival differences were evaluated using the Kaplan-Meier method and the log-rank test. Time-dependent Receiver Operating Characteristic (ROC) curves were generated using the "timeROC" R package to assess the predictive accuracy (AUC) of the signature for 1-, 3-, and 5-year OS. The same validation procedures were applied to the external GSE39582 and GSE29621 cohorts. Immune landscape and microenvironment analysis To elucidate the immunological characteristics associated with the risk signature, the CIBERSORT algorithm (https://cibersort.stanford.edu/) was utilized to estimate the relative proportions of 22 tumor-infiltrating immune cell (TIIC) types. Samples with a CIBERSORT p-value < 0.05 were included in the comparison. The stromal, immune, and ESTIMATE scores, which reflect the tumor microenvironment (TME) purity, were calculated using the "ESTIMATE" R package[23]. Furthermore, the activity of the cancer-immunity cycle was evaluated based on the tracking tumor immunophenotype (TIP) web server. Gene Set Variation Analysis (GSVA) was performed using the "GSVA" R package to quantify the enrichment scores of distinct immune-related pathways[24]. The immune subtypes (C1–C6) of COAD samples were retrieved from the study by Thorsson et al. to analyze the distribution of risk scores across different immunophenotypes[25]. Single-cell RNA sequencing analysis To dissect the tumor microenvironment at single-cell resolution, the colorectal cancer scRNA-seq dataset GSE132465 was retrieved from the GEO database and processed using the "Seurat" R package (version 4.4.0)[26]. Following initial data filtering (min.cells = 3, min.features = 200) and quality control (excluding cells with >20% mitochondrial gene expression), a total of 63,689 cells and 33,694 genes were retained. The raw count data were normalized using the ’LogNormalize’ method and subsequently scaled. Principal Component Analysis (PCA) was conducted for linear dimensionality reduction, followed by the application of the Harmony algorithm to correct for cross-sample batch effects[27]. Unsupervised cell clustering was performed based on the top 30 Harmony-adjusted principal components, and the results were visualized utilizing Uniform Manifold Approximation and Projection (UMAP)[28]. Based on canonical marker genes, the resulting clusters were annotated into six major cell lineages (B cells, epithelial cells, mast cells, myeloid cells, stromal cells, and T cells). Finally, to quantify the risk at the single-cell level, a weighted risk score was calculated for each cell by multiplying the normalized gene expression matrix by the corresponding multivariate Cox regression coefficients established in our bulk RNA-seq model. Differential expression and functional enrichment analysis To explore the biological mechanisms underlying the high-risk phenotype, differentially expressed genes (DEGs) between the high- and low-risk groups in the TCGA-COAD cohort were identified using the "limma" R package[29]. The screening thresholds were strictly set as an absolute log2 fold change (|log2FC|)> 0.5 and an adjusted P-value < 0.05. Subsequently, functional enrichment analysis, including Gene Ontology (GO) biological processes and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways, was performed on the identified DEGs using the "clusterProfiler" R package[30] Somatic mutation and tumor mutational burden analysis Somatic mutation data (MuTect2 variant aggregation) for the TCGA-COAD cohort were downloaded from the GDC portal. The "maftools" R package was employed to visualize the mutation landscape and calculate the Tumor Mutational Burden (TMB) for each sample[31]. The waterfall plots were generated to display the top mutated genes in the high- and low-risk groups, and the differential mutation frequencies were compared using Fisher’s exact test. Functional enrichment and therapeutic sensitivity prediction To explore the biological functions and signaling pathways underlying the risk stratification, Gene Set Enrichment Analysis (GSEA) was conducted using the "clusterProfiler" R package[32]. The Kyoto Encyclopedia of Genes and Genomes (KEGG) and Gene Ontology (GO) gene sets were used as references. Pathways with a nominal p-value<0.05 and a false discovery rate (FDR) q< 0.25 were considered significantly enriched. Chemotherapeutic drug sensitivity was predicted using the "oncoPredict" R package[33]. The half-maximal inhibitory concentration (IC50) of drugs from the Genomics of Drug Sensitivity in Cancer (GDSC) database was estimated for each patient using ridge regression[34]. Differences in IC50 values between risk groups were compared to identify potential therapeutic agents. Construction of the predictive nomogram To facilitate clinical application, univariate and multivariate Cox regression analyses were first performed to determine the independence of the risk score from other clinicopathological factors (age, gender, TNM stage). A prognostic nomogram incorporating the risk score and independent clinical predictors was then constructed using the "rms" and "regplot" R packages[35]. The predictive performance of the nomogram was assessed using calibration curves (comparing predicted vs. observed survival rates) and time-dependent ROC curves. Decision Curve Analysis (DCA) was conducted using the "ggDCA" package to evaluate the net clinical benefit of the nomogram compared to traditional staging systems. Statistical analysis All statistical analyses were performed using R software (version 4.2.0). Continuous variables were compared using the Wilcoxon rank-sum test or Kruskal-Wallis test, while categorical variables were analyzed using the Chi-square test or Fisher’s exact test. Correlation analyses were performed using Pearson or Spearman correlation coefficients. A two-sided p-value< 0.05 was considered statistically significant.Principal Component Analysis (PCA) was performed using the "prcomp" function in R to visualize the transcriptomic distribution of patients in different risk groups based on the expression profiles of the 9 signature genes. The data were scaled before analysis, and 95% confidence ellipses were generated to assess the spatial separation between high-risk and low-risk categories. Results Identification of prognostic candidates and construction of the risk model To screen for reliable prognostic biomarkers, we started with 332 lactylation-related genes curated from the MSigDB database [21]. After intersecting with genes expressed in the TCGA-COAD cohort (N=488 patients), we performed Univariate Cox regression and Kaplan-Meier analysis (P<0.05 threshold for both), yielding 26 candidate genes (Fig. 1a). These were subjected to LASSO penalized regression with 10-fold cross-validation, where λ was selected at minimum error (), retaining 15 genes with non-zero coefficients (Fig. 1b, c). To optimize model parsimony, we applied backward stepwise selection using the Akaike Information Criterion (AIC) on a multivariate Cox model, which identified a final 9-gene signature: DHRS7, SPR, MBD2, RBM17, CSRP2, S100A4, TMSB4X, TKT and COPS4. The forest plot (Fig. 1D) revealed five risk factors (HR>1) and four protective factors (HR<1), with RBM17 demonstrating the strongest prognostic effect (HR=2.588, P<0.001). The risk score for each patient was calculated as: Of the nine signature genes, RBM17 and TKT were significantly upregulated in tumor versus normal tissues, while DHRS7, MBD2 and COPS4 were significantly downregulated in the TCGA-COAD cohort (Supplementary Figure S1). The remaining four genes (SPR, CSRP2, S100A4, TMSB4X) showed no significant difference between tumor and normal tissues. These expression patterns are consistent with their respective risk or protective coefficients in the multivariate Cox model.Protein-level validation via the HPA database further corroborated these transcriptomic trends (Supplementary Fig S2). IHC staining showed increased intensity for TKT and S100A4 in tumor cells, while DHRS7, MBD2, and COPS4 were downregulated compared to normal glands. Consistent with the bulk RNA-seq data, SPR, CSRP2, TMSB4X, and RBM17 maintained stable protein expression across tissues. Detailed staining parameters and patient IDs are summarized in Supplementary Table S2. Screening of prognostic genes and construction of the risk signature. a Venn diagram showing 26 candidate genes identified by Univariate Cox regression and Kaplan-Meier analysis. b LASSO regression coefficient profiles of the 26 candidate genes; c Selection of the optimal parameter λ in the LASSO model using 10-fold cross-validation; d Forest plot of the final multivariate Cox model showing hazard ratios and 95% confidence intervals Validation of the prognostic performance of the risk model The prognostic performance of the 9-gene lactylation-related signature was evaluated in the TCGA training cohort and two independent validation cohorts. In TCGA, time-dependent ROC analysis yielded AUCs of 0.73, 0.65, and 0.68 for 1-, 3-, and 5-year OS, respectively (Fig. 2a). In GSE39582, AUCs were 0.66, 0.59, and 0.57 across these time points (Fig. 2b), while GSE29621 showed AUCs of 0.70, 0.69, and 0.71 (Fig. 2c). Kaplan-Meier survival analysis consistently showed that patients in the high-risk group had significantly poorer overall survival compared to the low-risk group in all three datasets: TCGA (P < 0.0001, Fig. 2d), GSE39582 (P = 0.023, Fig. 2e), and GSE29621 (P = 0.028,Fig. 2f). Furthermore, the integrated validation panels (g–i) reinforced these survival patterns by depicting risk score distribution, survival status correlation, and hierarchical clustering of nine-gene expression profiles across TCGA (g), GSE39582 (h), and GSE29621 (i) cohorts, revealing consistent differential expression patterns of signature genes between high- and low-risk groups. These cross-cohort findings support the clinical applicability of the signature. Validation of the 9-gene lactylation-related risk signature. A-C Time-dependent ROC curves for OS prediction in TCGA (a), GSE39582 (b), and GSE29621 (c). d–f Kaplan-Meier survival curves stratified by TCGA-derived median risk score. g–i Multi-dimensional validation showing risk score distribution, survival status correlation, and hierarchical clustering of nine-gene expression profiles in TCGA (g), GSE39582 (h), and GSE29621 (i) cohorts. OS, overall survival; ROC, receiver operating characteristic; TCGA, The Cancer Genome Atlas Additionally, PCA showed no distinct spatial separation between risk groups in the TCGA-COAD cohort (Supplementary Fig S3). PC1 and PC2 explained 27.5% and 14.9% of the variance, respectively, with 95% confidence ellipses largely overlapping. The risk signature defines an immunosuppressive tumor microenvironment To delineate the immunogenomic landscape shaped by the 9-gene lactylation-related signature, we integrated analyses of immune infiltration, microenvironment composition, and somatic mutations (Fig. 3). Evaluation of immune cell proportions via CIBERSORT revealed a distinct shift towards an immunosuppressive phenotype in the high-risk group, characterized by significantly elevated regulatory T cells (Tregs). In contrast, the low-risk group was enriched for anti-tumor effector cells, including CD4+ memory resting T cells, CD4+ memory activated T cells, activated dendritic cells, and eosinophils (Fig. 3a). The Risk Signature Defines an Immunosuppressive Tumor Microenvironment. a Box plots of CIBERSORT-estimated proportions of 22 tumor-infiltrating immune cell types across risk groups. b Box plots of ESTIMATE-derived microenvironment scores. c Correlation heatmap between risk score and cancer-immunity cycle steps (1–7). d Correlation heatmap of signature genes and risk score with immune checkpoint molecules. e Box plots of GSVA enrichment scores for immune-related pathways. F Box plots of risk score distribution across immune subtypes (c1–c4). g–h OncoPrint waterfall plots of top 5 mutated genes in high-risk (g) and low-risk (h) groups. GSVA, gene set variation analysis; *, P < 0.05; **, P < 0.01; ***, P < 0.001; ns, not significant This immunosuppressive cellular landscape in high-risk patients was underpinned by a remodeled microenvironment architecture. ESTIMATE analysis indicated that high-risk tumors possessed significantly lower tumor purity and higher stromal scores, while immune scores were comparable between groups (Fig. 3b). This profile suggests a stroma-dominated milieu that may impede effective anti-tumor immunity. Consistently, the risk score demonstrated significant correlations with key steps of the cancer-immunity cycle (Fig. 3c). It was positively correlated with Step 1 (cancer cell antigen release), Step 4 (CD4+ T cell and DC recruitment), Step 5 (immune cell infiltration), and Step 6 (cancer cell recognition), but negatively correlated with the recruitment of specific effector populations during Step 4, including Th1 cells, neutrophils, Th2 cells, and Tregs, indicating a dysfunctional immune recruitment process. Correlation analysis with immune checkpoints demonstrated that risk score positively correlated with TNFRSF4, CD276, TNFRSF18, and LAIR1, while negatively correlating with HHLA2, CD200R1, and CD44 (Fig. 3d), indicating that patients with high risk scores exhibit elevated expression of multiple inhibitory checkpoint molecules. Pathway enrichment analysis by GSVA corroborated this global immune polarization (Fig. 3e). High-risk tumors were characterized by enhanced activity in pathways related to antigen-presenting cell (APC) co-inhibition and co-stimulation, B cell and dendritic cell activity, CCR signaling, macrophage and mast cell activation, parainflammation, and Type I IFN response. Conversely, low-risk tumors exhibited elevated inflammation-promoting activity, MHC class I expression, and Type II IFN response. Analysis of canonical immune subtypes revealed that C4 (lymphocyte depleted) exhibited significantly lower risk scores compared to c1 (wound healing), c2 (IFN-γ dominant), and c3 (inflammatory) subtypes, while the latter three showed comparable risk profiles (Fig. 3f). Mutational profiling indicated that while APC, TP53, TTN, and KRAS were the most frequently mutated genes in both groups, their frequencies varied. Specifically, the low-risk group exhibited a higher proportion of APC mutations (77%) compared to the high-risk group (67%). Additionally, PIK3CA mutations (31%) were common in the high-risk group, whereas MUC16 mutations were frequent in the low-risk group, suggesting distinct oncogenic alterations despite a shared genomic core. Single-cell transcriptomic validation of the lactylation-related risk signature To investigate the cellular basis of the lactylation-related signature within the tumor microenvironment, we analyzed the scRNA-seq dataset GSE132465. Unsupervised clustering and UMAP visualization categorized the cells into six major lineages (Supplementary Fig S4A). Evaluation of the nine signature genes revealed distinct expression patterns across these cell populations (Supplementary Fig S4B). Specifically, CSRP2 was enriched in the stromal compartment, whereas SPR and TKT were predominantly expressed in epithelial cells. Additional genes, including S100A4 and TMSB4X, showed broader expression across multiple immune and stromal lineages. To determine the contribution of specific cell populations to the high-risk phenotype, the weighted 9-gene risk score was projected at single-cell resolution (Supplementary Fig S4C). Quantitative analysis demonstrated that stromal cells exhibited the highest median risk scores among all identified subsets, whereas epithelial cells displayed the lowest overall scores (Supplementary Fig S4D). Differential expression and pathway enrichment analyses between risk groups To explore the transcriptomic differences associated with the lactylation-related signature, we performed a differential expression analysis between the high- and low-risk patients in the TCGA-COAD cohort. A total of 382 differentially expressed genes (DEGs) were identified, comprising 298 up-regulated and 84 down-regulated genes in the high-risk group (|log2FC| > 0.5, adjusted P < 0.05; Supplementary Fig S5A, Supplementary Table S3).Subsequent Gene Ontology (GO) biological process analysis revealed that these DEGs were primarily enriched in extracellular matrix organization, extracellular structure organization, external encapsulating structure organization, cell-substrate adhesion, and the transforming growth factor beta (TGF-β ) receptor superfamily signaling pathway (Supplementary Fig S5B). Furthermore, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis indicated significant enrichment in pathways including the PI3K-Akt signaling pathway, integrin signaling, focal adhesion, and ECM-receptor interaction (Supplementary Fig S5C). Association between risk score and clinicopathological features Across the TCGA and GSE39582 cohorts, the risk score showed no significant correlation with age or gender (Fig. 4a, b, g, h); similarly, no association with gender was observed in GSE29621 (Fig. 4m). In contrast, the risk score was significantly associated with advanced clinicopathological features. Specifically, it showed significant correlations with T stage, N stage, and pathologic stage across all three cohorts (Fig. 4c, d, f, i, j, l, n–q). Additionally, significant associations with M stage were observed in the TCGA and GSE39582 cohorts (Fig. 4e, k). Association between risk score and clinicopathological features. A–F TCGA cohort: a Age, b Gender, c T Stage, d N Stage, e M Stage, f Pathologic Stage. g–l GSE39582 cohort: g Age, h Gender, i T Stage, j N Stage, k M Stage, l Pathologic Stage. m–q GSE29621 cohort: m Gender, n T Stage, o N Stage, p M Stage, q Pathologic Stage. ns, not significant; *, P<0.05; **, P<0.01; ***, P<0.001 Independent prognostic value and nomogram construction To assess the clinical applicability of the risk model, we performed univariate and multivariate Cox regression analyses. The univariate analysis indicated that the risk score was significantly associated with Overall Survival (HR = 2.718, 95% CI: 2.047-−3.609, P < 0.001), as did Age and Stage (Fig. 5a). Multivariate Cox regression further confirmed that the risk score remained an independent prognostic factor (HR = 2.27, 95% CI: 1.67-−3.09, P < 0.001) after adjusting for other clinicopathological features (Fig. 5b). Construction and evaluation of the nomogram. a Univariate Cox regression of risk score and clinical characteristics. b Multivariate Cox regression confirming risk score as an independent prognostic factor. c Nomogram predicting 1-, 3-, and 5-year OS probability. d Calibration curves assessing predicted vs. observed survival. E Time-dependent ROC curves of the nomogram. f–h Decision curve analysis evaluating clinical net benefit for 1-year (f), 3-year (g), and 5-year (h) OS. OS, overall survival; ROC, receiver operating characteristic; DCA, decision curve analysis; HR, hazard ratio; CI, confidence interval; AUC, area under the curve Based on these independent prognostic factors, we constructed a prognostic nomogram (Fig. 5c) to quantitatively predict the 1-, 3-, and 5-year survival probabilities for individual patients. The calibration curves (Fig. 5d) demonstrated excellent agreement between the nomogram-predicted probabilities and the actual observed survival rates. The time-dependent ROC analysis of the nomogram (Fig. 5e) exhibited high AUC values, confirming its predictive accuracy. Finally, Decision Curve Analysis (DCA) was performed to evaluate the clinical utility of the nomogram. As shown in Fig. 5f–h, the DCA curves for 1-, 3-, and 5-year OS demonstrated that the nomogram (red line) provided a superior net benefit compared to the “Treat-All” (grey line), “Treat-None” (black line), and other single factors (e.g. Stage, Age) across a wide range of threshold probabilities, highlighting its potential value in clinical decision-making. Drug sensitivity and functional enrichment analysis across risk groups To assess the clinical utility of the risk signature in guiding treatment, we predicted drug sensitivity using the GDSC database via oncoPredict. The analysis identified distinct therapeutic vulnerabilities associated with each risk subgroup. The low-risk subgroup exhibited significantly lower estimated IC50 values for fluorouracil, oxaliplatin, lapatinib, erlotinib, gefitinib, and neratinib, suggesting enhanced sensitivity to these agents (Fig. 6a–f). The high-risk subgroup showed lower IC50 values for TG-100-115, XL765, UNC0638, sildenafil, GANT-61, and BRD-K33199242, indicating potential vulnerability to these compounds (Fig. 6g–l). GSEA revealed distinct pathway enrichments: the high-risk subgroup was enriched in ECM-receptor interaction, complement and coagulation cascades, calcium signaling pathway, and focal adhesion (Fig. 6m–p). In contrast, the low-risk subgroup showed enrichment in ribosome biogenesis, neutrophil extracellular trap formation, cell cycle, and spliceosome (Fig. 6q–t). Drug sensitivity and functional enrichment analysis across risk groups. a–l Box plots of estimated IC50 values for fluorouracil, oxaliplatin, lapatinib, erlotinib, gefitinib, neratinib, TG-100-115, XL765, UNC0638, sildenafil, GANT-61, and BRD-K33199242. m–p GSEA of pathways enriched in the high-risk subgroup (ECM-receptor interaction, complement and coagulation cascades, calcium signaling, focal adhesion). q–t GSEA of pathways enriched in the low-risk subgroup (ribosome biogenesis, neutrophil extracellular trap formation, cell cycle, spliceosome). IC50, half-maximal inhibitory concentration; GSEA, gene set enrichment analysis; ECM, extracellular matrix Discussion Despite substantial advances in screening and therapeutics, the clinical management of colorectal cancer (CRC) remains challenged by its profound heterogeneity and the limited efficacy of immunotherapy in the majority of patients [1, 3]. While immune checkpoint inhibitors (ICIs) have revolutionized the treatment of mismatch repair-deficient (dMMR) or microsatellite instability-high (MSI-H) tumors, patients with microsatellite stable (MSS) tumors—constituting approximately 85% of CRC cases—typically manifest as “cold tumors” characterized by immune exclusion and primary resistance to ICIs [5, 6]. Deciphering the mechanisms driving this immunosuppressive tumor microenvironment (TME) is critical for overcoming therapeutic bottlenecks. Metabolic reprogramming, particularly the Warburg effect, has emerged as a key driver of TME remodeling [7]. Beyond simple metabolite accumulation, the groundbreaking discovery by Zhang et al. (2019) of histone lactylation (Kla) revealed a direct link between metabolic flux and epigenetic regulation [12]. In this study, we addressed the lack of metabolic-epigenetic biomarkers by constructing a robust 9-gene lactylation-related prognostic signature. This model not only accurately stratifies patient prognosis but also delineates a distinct “high-risk” phenotype characterized by metabolic hyperactivity, immune suppression, and specific therapeutic vulnerabilities, providing a roadmap for precision medicine in CRC. A pivotal finding of our study is the mechanistic link between high lactylation risk scores and the formation of an immunosuppressive TME. Our results show that high-risk tumors are enriched with regulatory T cells (Tregs), while being depleted of cytotoxic lymphocytes. This phenotype strongly validates the “lactate-lactylation-immune” axis. As described by Brand et al. (2016), high levels of tumor-derived lactate and LDHA activity blunt tumor immunosurveillance by inhibiting interferon-γ (IFN-γ ) production in T and NK cells [36]. Furthermore, lactate in the TME acts as a signaling molecule that promotes the stability and suppressive function of Tregs via lactylation-dependent mechanisms, such as MOESIN lactylation, as reviewed by Chen et al. [13]. Crucially, integrating our single-cell RNA-seq (scRNA-seq) and differential expression gene (DEG) analyses, we identified that this immunosuppression is structurally maintained by a “stromal barrier”. Our scRNA-seq analysis revealed that stromal cells exhibit the highest lactylation risk scores within the TME. Concurrently, DEG analysis highlighted a significant enrichment in extracellular matrix (ECM) organization and the TGF-β signaling pathway in high-risk tumors. In the acidic, lactate-rich TME, this lactylation-driven stromal remodeling creates a dense physical barrier that spatially excludes cytotoxic T cell infiltration, directly precipitating the "immune-excluded" or "cold tumor" phenotype[37]. Furthermore, recent mechanistic studies have demonstrated how lactylation directly drives cancer-associated fibroblast (CAF) activation and ECM remodeling—the structural basis of our observed stromal barrier [38, 39]. In the lactate-rich TME, tumor- or CAF-derived lactate triggers histone lactylation (e.g., H4K8la) in stromal cells, transcriptionally upregulating fibroblast activation protein (FAP) and promoting ECM deposition, thereby physically excluding CD8+ T cells [38]. This aligns with findings that H4K8 lactylation promotes FAP-mediated ECM remodeling and CD8+ T cell exclusion in solid tumors [38]. Concurrently, lactate-mediated tumor-CAF crosstalk via histone lactylation upregulates genes such as ANTXR1, further driving EMT and stromal barrier formation [39]. Our signature gene S100A4, a well-established master regulator of CAF activation and ECM hyperproduction, likely acts as a downstream effector of EP300-mediated histone lactylation, amplifying TGF-β signaling loops to sustain the immunosuppressive niche.This lactylation-driven stromal barrier is additionally reinforced by non-histone mechanisms: lactate-induced MOESIN lactylation at K72 enhances TGF-β /TGF-β RI/SMAD3 signaling in Tregs [40]. This modification promotes Treg stability and suppressive function, which synergizes with ECM components to create a multi-layered immune-exclusion microenvironment [40]. Upstream, PIK3CA mutations amplify this cascade by fueling glycolytic lactate production, establishing a closed “mutation-metabolism-epigenetics-stroma” loop[41]. Collectively, these insights mechanistically link our 9-gene signature to the immunosuppressive TME, offering not only robust prognostic stratification but also rational targets for stromal-modulating therapies (e.g., GANT-61 for Hedgehog signaling in CAFs or sildenafil for PDE5-mediated TME remodeling). To achieve precise risk stratification, we established a robust lactylation-related prediction model via LASSO and multivariate Cox regression, identifying key genes that collectively orchestrate a malignant phenotype.As detailed in our methodology, the 332 lactylation-related genes evaluated encompass key regulatory enzymes (e.g., the lactylase EP300 and delactylases HDAC1-3) alongside specific lactylated proteins. Consequently, our signature genes are tightly coupled to the epigenetic-metabolic axis. For instance, RBM17, a driver of splicing anomalies [42, 43], may act as a downstream transcriptional target of histone lactylation (e.g., H3K18la) mediated by EP300. Parallel to this, the model incorporates SPR, a critical metabolic enzyme in the tetrahydrobiopterin pathway that provides essential metabolic substrates (polyamines) requisite for sustained tumor expansion [44]. The concordant expression patterns of these critical markers at both the transcriptomic and protein levels, as validated by our HPA database analysis, further substantiate their biological relevance. This epigenetic modification directly links the metabolic hyperactivity of the TME to the stromal encapsulation and hyperproliferation characteristic of high-risk CRC patients. Given the upstream role of PIK3CA mutations in fueling this metabolic-epigenetic cascade, this genomic dependency provides a robust biological rationale for the specific therapeutic vulnerability observed in our drug sensitivity prediction analysis. The high-risk group exhibited significant sensitivity to TG-100-115 (a selective PI3K inhibitor) and XL765 (a dual PI3K/mTOR inhibitor). PIK3CA mutations render the PI3K/AKT/mTOR axis a critical therapeutic target in colorectal malignancies [45]. The predicted efficacy of XL765 aligns with its structural mechanism as a dual inhibitor capable of abrogating this hyperactive signaling [46]. Furthermore, PIK3CA-mutant CRC models have been shown to exhibit superior sensitivity to dual PI3K/mTOR blockade compared to wild-type counterparts, confirming that this genomic alteration serves as a predictive biomarker for therapeutic efficacy [47]. Importantly, dual PI3K/mTOR inhibitors and specific PI3K pathway modulators are currently undergoing active evaluation in clinical trials for solid tumors harboring PIK3CA mutations, thereby answering the call for actionable vulnerabilities in precision medicine. Thus, our findings confirm that targeting this mutation-driven metabolic axis represents a viable precision strategy for high-risk patients. From a clinical perspective, our study delineates actionable strategies for stratified treatment. The low-risk group exhibited significant sensitivity to Fluorouracil, Oxaliplatin, and EGFR/HER2 inhibitors (e.g.,Lapatinib, Erlotinib), validating the model’s capacity to identify patients responsive to standard-of-care regimens. In contrast, the high-risk group displayed a refractory multidrug-resistant phenotype but revealed distinct therapeutic vulnerabilities. Consistent with their specific molecular dysregulations, these patients showed high sensitivity to PI3K pathway inhibitors (TG-100-115, XL765), microenvironment-modulating agents (Sildenafil, GANT-61), and the G9a epigenetic inhibitor UNC0638. Despite the robust predictive capacity of our lactylation-related signature in stratifying CRC prognosis and immune landscapes, several intrinsic limitations warrant careful consideration. Regarding the model’s performance, we observed a slight attenuation in the ROC predictive accuracy within the GSE39582 external validation cohort compared to the TCGA training set. However, it is crucial to note that the signature maintained strict statistical significance in Kaplan-Meier survival stratification across all distinct validation cohorts (e.g., P = 0.023 for GSE39582). This confirms that while unavoidable batch effects and inherent platform disparities (microarray vs. RNA-sequencing) may influence absolute predictive metrics, the biological robustness and generalizability of the signature across diverse patient populations remain highly reliable. From a therapeutic perspective, our pharmacogenomic predictions utilizing the GDSC database are fundamentally based on 2D cell line screening data. We explicitly acknowledge that standard in vitro 2D models cannot fully recapitulate the complex, multi-cellular architecture and physical barriers of the human tumor microenvironment (TME), including stromal support, extracellular matrix deposition, and immune cell interactions [48, 49]. This inherent limitation may lead to discrepancies between predicted and actual clinical responses, as 2D assays show only modest predictive power for patient outcomes. To translate these specific predictions (such as the efficacy of microenvironment-modulating agents or PI3K inhibitors) into actionable clinical practice, future validations must be conducted in more physiologically relevant models, including Patient-Derived Organoids (PDOs) or 3D heterotypic co-culture systems—models that have demonstrated superior concordance with clinical efficacy in colorectal cancer[48].Furthermore, because our findings are primarily grounded in retrospective bioinformatic profiling, the current study lacks prospective clinical validation, even though it was cross-validated across multiple independent transcriptomic and single-cell cohorts. Moving forward, large-scale, prospective, real-world multi-center clinical cohorts are imperative. Alongside clinical validation, rigorous in vivo and in vitro functional assays—such as targeted knockdown or overexpression of key signature genes (e.g., S100A4)—are required to conclusively substantiate the proposed “lactylation-TGF-β -stroma” mechanistic axis and its direct causal role in CRC immune evasion. Conclusion Our study indicates that the novel lactylation-related gene signature has great potential as both a prognostic biomarker and therapeutic target in colorectal cancer. By utilizing data from multiple public databases, we can gain insights into the complex interplay between metabolic dysregulation and the immune microenvironment, and accurately predict the prognosis of CRC patients. This enables the identification of specific subgroups that are likely to benefit from personalized treatment approaches, such as standard chemotherapy and EGFR blockade for low-risk patients, versus targeted metabolic and stromal interventions for high-risk patients. Overall, our research highlights the significance of the lactylation-related signature in improving patient stratification and guiding precision treatment strategies for refractory CRC patients. Electronic Supplementary Material Below is the link to the electronic supplementary material. Additional file 1. Supplementary Figure S1. Additional file 2. Supplementary Figure S2. Additional file 3. Supplementary Figure S3. Additional file 4. Supplementary Figure S4. Additional file 5. Supplementary Figure S5. Additional file 6. Supplementary Table S1. Additional file 7. Supplementary Table S2. Additional file 7. Supplementary Table S3. Funding The author declare that they have received no funding. Data availability The datasets presented in this study can be found in online repositories. The names of the repositories and accession numbers are detailed in the article. The RNA-sequencing data and clinical phenotypes for the training cohort were obtained from The Cancer Genome Atlas (TCGA) database via the UCSC Xena platform under the dataset identifier TCGA-COAD (https://xenabrowser.net/). The validation datasets generated and analyzed during the current study are available in the Gene Expression Omnibus (GEO) repository with accession numbers GSE39582 and GSE29621, and GSE132465 (https://www.ncbi.nlm.nih.gov/geo/). Declarations Ethics approval and consent to participate Not applicable. This study utilized only publicly available, de-identified transcriptomic and single-cell RNA sequencing datasets from the TCGA and GEO databases and did not involve any human participants, animal experiments, or identifiable personal information. Consent for publication Not applicable. Competing interests The authors have no relevant financial or non-financial interests to disclose. Footnotes Publisher's Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Chi Fu and Chengwen Wang contributed equally to this work. References Associated Data Supplementary Materials Additional file 1. Supplementary Figure S1. Additional file 2. Supplementary Figure S2. Additional file 3. Supplementary Figure S3. Additional file 4. Supplementary Figure S4. Additional file 5. Supplementary Figure S5. Additional file 6. Supplementary Table S1. Additional file 7. Supplementary Table S2. Additional file 7. Supplementary Table S3. Data Availability Statement The datasets presented in this study can be found in online repositories. The names of the repositories and accession numbers are detailed in the article. The RNA-sequencing data and clinical phenotypes for the training cohort were obtained from The Cancer Genome Atlas (TCGA) database via the UCSC Xena platform under the dataset identifier TCGA-COAD (https://xenabrowser.net/). The validation datasets generated and analyzed during the current study are available in the Gene Expression Omnibus (GEO) repository with accession numbers GSE39582 and GSE29621, and GSE132465 (https://www.ncbi.nlm.nih.gov/geo/).