Multi-dimensional profiling of lactylation-related signatures for prognostic and therapeutic insights in bladder cancer

Abstract

Background

Bladder cancer (BCa) is a prevalent malignancy with challenging prognosis and management. Lactylation, a novel post-translational modification, is dysregulated in various cancers. This study investigates lactylation-derived signature for prognostic assessment and treatment decisions in BCa.

Methods

We used immunohistochemistry (IHC) to assess tissue microarrays (TMA) and Western blot to investigate BCa cell lines to evaluate overall lactylation. The Cancer Genome Atlas (TCGA) cohort was categorized into two clusters based on lactylation-related genes (LRGs) using consensus clustering. We analyzed survival rates, clinical features, tumor microenvironment (TME), and responses to immunotherapy and chemotherapy. A 3-gene lactylation-related gene risk score (LRGRS), based on AHNAK, ALDH1A1, and CALR, was developed using least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation, random forest (RF), and support vector machine recursive feature elimination (SVM-RFE), and was validated in Gene Expression Omnibus (GEO) datasets and further characterized in single-cell datasets. Subsequently, LRGRS was analyzed alongside clinical characteristics, mutation profiles, biological functions, immune cell infiltration, immunotherapy responses, and drug sensitivity. Core genes were examined in vitro.

Results

Our findings demonstrated a significant increase of pan-lysine lactylation (Pan-kla) in BCa. Through the integration of TCGA datasets, GEO datasets, single-cell datasets, and in vitro studies, we found and validated an LRGRS model including three genes capable of predicting the prognosis of BCa patients. Significant differences were noted for clinical features, biological functioning, immune cell infiltration, immune checkpoint expression, immunotherapy response, and drug sensitivity among various risk categories.

Conclusions

In conclusion, our research presents LRGRS as an innovative and dependable prognostic biomarker for BCa, capable of precisely and consistently forecasting survival rates, informing personalised therapy for BCa patients, and offering new therapeutic methods for the condition.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12935-026-04364-6.

Introduction

Bladder cancer (BCa) ranks as the second most prevalent malignant tumor of the urinary system and the tenth most prevalent malignancy worldwide [1, 2]. According to the 2020 GLOBOCAN epidemiological report, over 500,000 new cases of BCa are diagnosed annually, with more than 200,000 deaths attributed to the disease. These figures continue to rise due to an ageing population and the accumulation of environmental risk factors [1, 2]. BCa can be categoried into two subtypes based on muscle layer invasion: non-muscle-invasive bladder cancer (NMIBC), comprising approximately 75% of new cases, and muscle-invasive bladder cancer (MIBC) accounting for about 25% [3]. While NMIBC generally has a favorable prognosis, 15%–20% of high-risk cases may progress to MIBC, leading to significantly poorer clinical outcomes [4]. For MIBC, the standard treatment strategy involves neoadjuvant chemotherapy (NAC) with cisplatin followed by radical cystectomy (RC). Some patients experience long-term survival benefits from these treatments [5]. However, due to challenges such as chemotherapy resistance, complications, and the absence of robust clinical predictive markers, a considerable number of patients still face the risk of tumor recurrence or distant metastasis post-surgery [6]. In recent years, novel therapies, particularly immune checkpoint inhibitors (ICI), have expanded treatment options for BCa patients. The US Food and Drug Administration (FDA) has approved several anti-PD-1/PD-L1 agents, including atezolizumab, pembrolizumab, nivolumab, durvalumab, and avelumab, for advanced or metastatic urothelial carcinoma [4]. Large clinical trials, such as IMvigor210, ABACUS, and KEYNOTE-36, have demonstrated that ICI can prolong progression-free survival and improve overall survival in patients unresponsive to traditional chemotherapy or intolerant to platinum-based drugs [7–10]. Despite these advancements, the response rate remains low for the majority of patients. Factors such as the highly heterogeneous tumor microenvironment, immune escape mechanisms, and the lack of reliable stratification markers contributed to significant variations in clinical efficacy. Therefore, identifying molecular characteristics and biomarkers associated with treatment response, alongside large-scale cohort studies, has become crucial for developing precision treatment strategies for BCa.

The Warburg effect is a hallmark metabolic feature of tumor cells, characterized by their reliance on glycolysis for energy production even in the presence of sufficient oxygen, resulting in the accumulation of lactic acid [11]. Recent studies have found that lactic acid serves not only as the end product of glycolysis, but also as a crucial substrate for a novel post-translational protein modification known as lactylation. This process involves the covalent binding of lactic acid molecules to lysine residues on proteins and is regulated by lactyltransferase and de-lactylase enzymes [12, 13]. As research progresses, numerous non-histone lactylation modification sites have been identified in addition to the classic histone lactylation modification. These modifications play important roles in various biological processes, including tumor metabolic reprogramming, immune escape, and drug resistance, collectively facilitating tumor initiation and progression [14–16]. In recent research, the multifaceted roles of lactylation within the tumor microenvironment have been increasingly recognized. For example, CircXRN2 has been shown to inhibit tumor progression driven by H3K18 lactylation through activation of the Hippo signaling pathway in BCa [17]. Single-cell RNA sequencing has further validated the critical role of H3K18 lactylation in gene activation within cisplatin-resistant cell subsets [18]. Moreover, the absence of PFK-1 impairs glycolysis and reduces histone phosphorylation, thereby inhibiting tumor cell invasion and proliferation [19]. Additionally, lactylation can induce tumor-associated macrophages to change to an M2 immunosuppressive phenotype [20]. These findings underscore the complex regulatory functions of lactylation in the initiation, progression, and therapeutic resistance of BCa. However, most current research focuses on the molecular mechanisms underlying tumor development, the prognostic and therapeutic implications of lactylation-related genes in BCa remain largely unexplored.

In this study, we systematically explored the expression patterns of lactylation-related genes (LRGs) in BCa and their roles in disease progression and clinical prognosis. A comprehensive approach was employed, integrating RNA transcriptome analysis, single-cell RNA sequencing, machine learning algorithms, immune infiltration analysis, and functional enrichment analysis. In addition, we developed the Lactylation-Related Gene Risk Score (LRGRS) to quantify the prognostic impact of LRGs in BCa and evaluate its predictive value for clinical outcomes and treatment responses. This study not only advances the understanding of the lactylation regulatory network in BCa, but also may provide a potential prognostic biomarker for the personalized therapeutic strategies for BCa.

Materials and methods

Data acquisition and processing

The present study used the Cancer Genome Atlas (TCGA) database to obtain the transcriptome data, somatic mutations and clinical information for BCa, which serves as the training set. Transcriptional and survival data for 33 distinct tumor types were retrieved from the UCSC Xena website for pan-cancer analysis. The transcriptome data were log2 (TPM + 1)-transformed for subsequent analysis. For validation purpose, RNA-seq data and corresponding clinical data from BCa cohorts (GSE31684, GSE32894) were extracted from the Gene Expression Omnibus (GEO). Samples lacking complete survival data were excluded from the analysis. Batche effects between the TCGA and GEO datasets were corrected using the Combat algorithm from the ‘sva’ R package. Concurrently, single-cell analysis was performed on the GSE135337 dataset, which comprises 7 cases of BCa. Based on previous studies, lactylation-modified writers, erasers and multiple lactylation modification sites were included in our study [12, 13, 21]. Ultimately, 332 LRGs were used for subsequent analysis (Supplementary Material Table S1).

Identifying differentially expressed LRGs

Differentially expressed genes (DEGs) (|logFC| > 1.0 and FDR < 0.05) related to lactylation in tumors and adjacent tissues were obtained using the R package ‘limma’. Additionally, the web-based tool Metascape (http://metascape.org) was utilized to perform gene annotation and functional enrichment analysis of these DEGs.

Consensus clustering analysis and differences between clusters

Consensus clustering analysis and differences between clusters were explored using univariate Cox regression to identify LRGs associated with prognosis. These LRGs were then subjected to consensus clustering analysis using the ‘ConsensusClusterPlus’ R package [22]. Differential gene expression analysis across the clusters was performed with the ‘limma’ R package, applying criteria of an absolute log-fold change (|logFC| > 2.0) and FDR < 0.05. The Kaplan–Meier method was employed to analyze overall survival differences between the two clusters.

Functional enrichment analysis

Functional enrichment analysis of differentially expressed genes was conducted with the R package ‘clusterProfiler’, including Gene Ontology (GO) and Kyoto Encyclopaedia of Genes and Genomes (KEGG) pathways. Additionally, Single-sample Gene Set Enrichment Analysis (ssGSEA) and Gene Set Variation Analysis (GSVA) was conducted using the R package ‘GSVA’ to evaluate pathway activity across various groups. All gene sets were obtained from the Molecular Signature Database (MSigDB; https://www.gsea-msigdb.org/gsea/msigdb).

Immune infiltrate analysis

The ESTIMATE algorithm was employed to evaluate variations in ESTIMATE score, immune score, stromal score, and tumor purity among samples. Additionally, the relative abundance of various cells in BCa was assessed using the CIBERSORT and ssGSEA algorithms, as commonly performed in recent multi-omics bioinformatics studies investigating tumor immune heterogeneity [23, 24]. We also examined differences in the expression of the human leukocyte antigen (HLA) gene family and immune checkpoint genes. Then, based on these findings, we analysed the immune subtype classification of TCGA-BLCA subjects [25, 26].

Predicting response to immunotherapy and chemotherapy

The Tumor Immune Dysfunction and Exclusion (TIDE) algorithm was used to evaluate the response to immune checkpoint inhibitor therapy in the TCGA-BLCA patients [27]. Mutation data were obtained from the TGCA database, and the R package ‘maftools’ was used to calculate tumor mutation burden (TMB), a predictor of immunotherapy response [28]. The Cancer Immunology Atlas (TCIA) webtool provided comprehensive results of immunogenomic analyses, with the immunogenicity of tumors quantitatively scored on a scale ranging from 0 to 10, termed immune phenotyping score (IPS) [29]. The IPS has been demonstrated to have the capacity to predict response to immune checkpoint inhibitors [29]. Furthermore, to investigate the differences in sensitivity predictions for common chemotherapeutic agents between low- and high-risk groups in BCa, the half-inhibitory concentration values (IC50) of the drugs were calculated using the ‘oncoPredict’ R package.

LRGs signature construction and validation

To develop an LRG-based signature, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation, random forest (RF) algorithms and support vector machine recursive feature elimination (SVM-RFE) algorithms were utilised to identify LRGs with notable prognostic significance, as similarly applied in recent machine learning-based feature screening studies [30]. A final group of three genes (AHNAK, ALDH1A1, CALR) was used to construct a risk score. These genes were included into the multivariate Cox regression model to compute the risk score, referred to as the LRGRS. For each subject, LRGRS were calculated as follows: , where “Coeff” represented the coefficients of genes in the Cox regression model and “Exp” meant the mRNA expression levels of the genes. The median score was used as the cut-off value to divide patients into high- and low-risk groups. The same formula was applied to the GEO dataset for validation. The Kaplan–Meier approach was used to examine the overall survival between two risk groups. The precision of the risk score in predicting survival at 1, 3, and 5 years was evaluated using Receiver Operating Characteristic (ROC) curves using the ‘timeROC’ R package. A predictive nomogram was developed using the ‘regplot’ R package, integrating risk score and other clinical characteristics. The efficacy of the nomogram was assessed by calibration curves and Decision Curve Analysis (DCA).

Single-cell RNA sequencing analysis

The scRNA-seq data were analysed via the ‘Seurat’ R programme. Quality control included screening cells with a minimum expression of 3 cells, with each cell expressing 500–6000 genes, and limiting mitochondrial gene expression to no more than 10%. Subsequent to filters, 36,407 cells were detected. Data normalisation, scaling, harmonisation, and principal component analysis (PCA) were conducted. The first 13 main components were picked for dimensionality reduction and clustering using the ‘FindClusters’, ‘FindNeighbors’ and ‘RunTSNE’ functions. Cell annotation was performed according to previously published studies [31]. The GSVA and correlation analysis between Hallmark pathway activity and LRGRS intended to clarify possible pathways linked to the identified signature.

Clinical samples and IHC

Shanghai Outdo Biotech Company (Shanghai, China) provided the OD-CT-UrBla03 human BCa Tissue microarray (TMA), which was used for Immunohistochemistry (IHC) for pan-lysine lactylation (Pan-kla). The TMA contained 29 BCa samples and 24 adjacent normal tissues. The IHC staining approach adhered to the prior publication [32]. Pan-kla (PTM BioLab, China) were incubated with TMA for a minimum of 16 h at 4 °C. The cell count and staining intensity of each sample in the TMA were quantitatively assessed using QuPath software (https://qupath.github.io). Ultimately, the H-score for each sample was calculated to assess the intensity of Pan-kla’s staining.

Cell lines and cell culture

Human bladder uroepithelial cell (SV-HUC-1) and human BCa cell (5637) were purchased from the Institute of Cell Biology, Chinese Academy of Sciences (Shanghai, China). SV-HUC-1 cell were cultured in Ham’s F-12 K (Kaighn’s) medium (Gibco, USA). 5637 cell were cultured in RPMI1640 medium (Gibco, USA). Each medium contained 10% foetal bovine serum and 1% penicillin/streptomycin, and all cells were maintained in a water-saturated environment with 5% CO2 at 37 °C.

RNA extraction and qRT‑PCR

Total RNA was extracted from cells utilising RNAiso Plus reagent (TaKaRa, USA) in accordance with the manufacturer’s protocol. cDNA synthesis was performed with the PrimeScript™ RT Reagent Kit (TaKaRa, USA). Quantitative real-time PCR (qRT-PCR) was executed employing the SYBR Green Premix Pro Taq HS qPCR Kit (Accurate Biology, China). Data analysis was conducted using the 2^−△△CT method. The mRNA expression level of a gene was assessed by normalising to GAPDH. The primers utilised in the study are detailed in Supplementary Material Table S2.

Western blot

Proteins were extracted from cell lines using RIPA buffer (Thermo Fisher Scientific, USA) under cold conditions. The protein concentration was measured via a BCA protein assay kit (Thermo Fisher Scientific, USA). Thereafter, the protein samples were applied to SDS-PAGE gels and subsequently transferred to PVDF membranes. Following a blocking procedure with 5% skim milk, the membranes were incubated with primary antibodies at suitable dilutions overnight at 4 °C. Subsequently, secondary antibodies were administered at ambient temperature for roughly 1 h, and the immunoreactivity was detected using the Bio-Lab imaging system. The subsequent main antibodies used were: Anti-L-Lactyl Lysine (PTM-BioLab, China), anti-AHNAK (Proteintech, USA), anti-ALDH1A1(Proteintech, USA), anti-CALR (Proteintech, USA), and anti-β-actin (Cell Signalling Technology, USA). The anti-L-lactyl lysine antibody was used to detect pan-lysine lactylation (Pan-Kla) and recognizes lactylated lysine residues across proteins rather than a specific target protein.

Statistical analysis

All statistical analyses in this research were conducted using R software (version 4.4.1) and GraphPad Prism programme (10.3.1). The threshold for statistical significance was established at P < 0.05.

Results

Global lactylation level was up-regulated in BCa

During the assessment of the TCGA BCa dataset, it was shown that tumor tissue had genetic signature indicative of elevated glycolysis and increased serum lactate levels (Fig. 1A-B). Simultaneously, LDHA and LDHB, a crucial enzyme in lactate metabolism, showed increased expression levels in tumor tissues (Fig. 1C-D). The metabolic features lead to heightened lactate generation, which can be utilized as a substrate to promote lactylation modification [13]. The following IHC examination of TMA revealed that the global lactylation level in BCa tissue was markedly increased relative to normal tissue (Fig. 1E). The total global lactylation level in BCa cell 5637 was much greater than that in normal bladder epithelial cell SV-HUC-1 (Fig. 1F). These results indicated that an elevation in total lactate modification is a prominent characteristic of BCa. In-depth investigation of the molecular processes of lactylation in BCa may provide novel therapeutic strategies for this formidable malignancy.

Fig. 1.

Increased glycolysis, lactate metabolism, and lactylation in BCa. (A-B) ssGSEA score of glycolysis and serum lactate in tumor (n = 399) vs. normal (n = 18) tissues in the TCGA cohort. (C-D) Genes expression of LDHA and LDHB in tumor (n = 399) vs. normal (n = 18) tissues in the TCGA cohort. (E) Immunohistochemical staining visualization of lactylation levels in normal and cancerous tissues. (F) Western blot analysis of lactylation levels in Human bladder uroepithelial cells (SV-HUC-1) and human BCa cells (5637)

Identification of lactylation-related DEGs in BCa

Initially, we used the ‘limma’ R package to discover lactylation-related DEGs in BCa. Through the comparison of cancers and normal tissues, 30 DEGs were found in the TCGA cohort including 399 malignancies and 18 normal tissues, including 21 upregulated and 9 downregulated genes. Differences in gene expression patterns are shown by heatmaps and volcano graphs (Fig. 2A-B). Through functional enrichment analysis using the Metascape database, we found that these DEGs are closely related to DNA Damage/Telomere Stress Induced Senescence, Pid Aurora B pathway, cellular senescence, and regulation of DNA metabolic process (Fig. 2C-E).

Fig. 2.

Identification and functional enrichment analysis of differentially expressed LRGs between normal and tumor sample in the TCGA cohort. (A) Volcano plot of differentially expressed LRGs between normal and tumor sample. (B) Heatmap of differentially expressed LRGs between normal and tumor sample. (C) Enriched terms of differentially expressed LRGs between normal and tumor sample. (D–E) Network of enriched terms colored according to (D) cluster ID (nodes with the same cluster ID were typically close to each other) and (E) p-value (terms with more genes tended to have higher p-values)

Consensus clustering on LRGs in BCa

10 LRGs related with prognosis were discovered by univariate Cox regression analysis and categorised as either ‘risk’ or ‘protective’ genes (Fig. 3A). To underscore the clinical relevance of LRGs, consensus clustering was conducted using the ‘ConsensusClusterPlu’ R package, revealing two distinct patient clusters when k = 2 (Fig. 3B-D). Kaplan-Meier survival analysis demonstrated a substantial survival difference between the two groups (P = 0.0013; Fig. 3E), with Cluster 1 exhibiting a lower survival rate. There were significant distinctions in pathological staging and survival status between the two clusters (Supplementary Material Table S3). Patients in Cluster 1 had higher staging and T grading, along with a higher number of deaths. The two clusters exhibited distinct expression patterns of 10 prognosis-related LRGs (Fig. 3F–G), with Cluster 1 showing elevated expression of ‘risk’ genes and Cluster 2 demonstrating increased expression of ‘protective’ genes.

Fig. 3.

Consensus Clustering based on Prognostic LRGs in the TCGA cohort. (A) Results of the univariate cox analysis. (B-D) Consensus matrix for k = 2 was obtained by applying consensus clustering. (E) Kaplan–Meier survival curves of the two clusters. (F) Genes Expression levels of 10 LRGs in the two clusters. (G) Heatmap showing the associations between clinicopathological features and the expression profiles of 10 LRGs in the two clusters. (H) Volcano plot of differentially expressed genes (DEGs) between two clusters. (I-J) GO and KEGG enrichment analyses for DEGs between two clusters. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Next, in order to acquire an in-depth knowledge of the biological aspects underpinning the two clusters of differences, we identified 151 DEGs between the two clusters (Fig. 3H). Functional analysis by GO and KEGG analysis showed that these DEGs are mainly related to immune functions such as immunoglobulin mediated immune response, humoral immune response, lymphocyte mediated immunity and antimicrobial humoral response (Fig. 3I-J). Furthermore, GSVA research revealed that the two groups had considerable disparities in several biological processes, with Cluster 1 exhibiting a superior GSVA score (Fig. S1). The findings reveal the distinctions in biological roles between the two clusters categorised by LRGs, hence validating the rationale behind this categorisation in BCa.

Immune infiltration and responsiveness to immunotherapy and chemotherapy between two clusters

We used the ESTIMATE, CIBERSORT, and ssGSEA algorithms to illustrate the alterations in the constituents of the tumor immune microenvironment across clusters (Fig. 4A-F). Cluster 1 exhibited higher Stromal, Immune, and ESTIMATE scores, yet had lower Tumor Purity. In addition, most immune cells had higher abundance in Cluster 1. The two clusters also had different immune subtype compositions, with Cluster 1 having a higher proportion of ‘IFN-γdominant’, ‘immune-enriched, fibrotic’ and ‘fibrotic’ (Fig. 4G-H). These findings suggested that there were significant differences in the TIME between different clusters, with Cluster 1 being more complex. TIDE is an algorithm designed to predict the response of immunotherapy, with a higher TIDE score meaning a reduced benefit from immunotherapy and an increased risk of immune escape [27]. In our analysis, Cluster1 exhibited higher “TIDE scores”, “Exclusion scores”, and “Dysfunction scores” (Fig. S2A-C). Meanwhile, Cluster1 had a lower proportion of patients predicted to respond to immunotherapy (Fig. S2D). Meanwhile, multiple HLA family and immune checkpoint genes showed high expression in Cluster 1 (Fig. 4I-J). The observations indicate that Cluster 1 possesses a heightened capacity for immunological evasion and may get less benefit from immunotherapy. In addition, to further evaluate the immune checkpoint treatment response (ICI) of BCa patients, we calculated immunophenoscore (IPS). Our results showed that the IPS score was higher in the high-risk group (ips_ctla4_pos_pd1_pos), indicating that high-risk patients had a higher response rate to ICI (Fig. S2E-H).

Fig. 4.

Immune Infiltration and Immune types across two Clusters in the TCGA cohort. (A-D) ESTIMATE analysis including Stromal Score, immune Score, ESTIMATE Score and Tumor Purity. (E-F) The relative abundance of infiltrating immune cell types in two clusters. (G-H) Differences in immune subtypes between the two clusters. (I-J) Expression differences of HLA family and immune checkpoint genes between the two clusters. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

To further assess the disparities in chemotherapy effectiveness between the two clusters, we examined the variations in IC50 values of six prevalent chemotherapeutic agents in the two clusters. The findings indicated that the IC50 values of cisplatin, rapamycin, and camptothecin in Cluster 1 were much lower than those in Cluster 2, suggesting that patients in Cluster 1 may get more benefit from these three chemotherapeutic agents (Fig. S2I-N).

Establishment and validation of a LRGs-based signature

In order to establish a novel prognostic model for BCa, based on the abovementioned results of univariate Cox regression analysis, we next performed LASSO, RF and SVF-RFE analysis in the TCGA cohort (Fig. 5A-G). Ultimately, AHNAK, ALDH1A1 and CALR were co-determined and included in the multivariate Cox regression model (Fig. 5H-I). The LRGRS was calculated according to the following equation: LRGRS = (0.27240586 * AHNAKexpression) + (0.05716113 * ALDH1A1 expression) + (0.25369305 * CALR expression). According to the median risk score of the training group, patients were divided into high-risk and low-risk groups. Within the high-risk group, we observed an increased mortality rate and elevated expression of all three ‘risk’ genes (Fig. 5J). Simultaneously, the mulberry diagram indicated that individuals with elevated LRGRS are more prone to being categorised into Cluster 1 and have a worse prognosis (Fig. 5K). The Kaplan-Meier analysis findings indicate that patients in the high-risk group have low survival rates, and those with elevated expression of the ‘risk’ genes demonstrate an unfavourable prognosis (Fig. 5L-O). To evaluate the predictive ability of our prognostic signature, we plotted the ROC curves for 1-, 3- and 5-year overall survival. The area under the curve (AUC) values were 0.62, 0.64 and 0.64, respectively, indicating that our model has good predictive performance (Fig. 5P). In addition, we validated the LRGRS in two independent external validation sets (GSE31684, GSE32894), and the results were similar to those of our training set and showed good predictive ability (Fig. 6A-J). More detailed baseline information for the training and validation sets can be found in Supplementary Material Table S4-6.

Fig. 5.

Establishment of LRGRS in the TCGA cohort. (A-C) LASSO regression analysis identified 8 out of the 10 LRGs as significant prognostic factors. (D-E) RF analysis identified 5 out of the 10 Breg-related genes as significant prognostic factors. (F-G) SVM-RFE analysis identified 7 out of the 10 Breg-related genes as significant prognostic factors. (H) 3 genes were determined as significant genes by both LASSO, RF and SVM-RFE analysis. (I) Multivariate COX regression forest plot for the 3model genes. (J) The distribution of risk score, survival status, and the exoression of the 3 hub genes between the low- and high-risk groups. (K) Sankey diagram correlating clusters, LRGRS groups, and survival status. (L) Kaplan–Meier analysis comparing overall survival between high and low LRGRS groups. (M-O) Kaplan–Meier analysis comparing overall survival between groups with high and low expression of 3 hub genes. (P) ROC curves depicting LRGRS signature’s predictive performance for 1, 3, and 5-year overall survival

Fig. 6.

Validation of LRGRS in the GEO cohort. (A-C) The distribution of risk score, survival status, and the exoression of the 3 hub genes between the low- and high-risk groups in GSE31684 cohort. (D) Kaplan–Meier analysis comparing overall survival between groups with high and low expression of 3 hub genes in GSE31684 cohort. (E) ROC curves depicting LRGRS signature’s predictive performance for 1, 3, and 5-year overall survival in GSE31684 cohort. (F-H) The distribution of risk score, survival status, and the exoression of the 3 hub genes between the low- and high-risk groups in GSE32894 cohort. (I) Kaplan–Meier analysis comparing overall survival between groups with high and low expression of 3 hub genes in GSE32894 cohort. (J) ROC curves depicting LRGRS signature’s predictive performance for 1, 3, and 5-year overall survival in GSE32894 cohort

Subsequently, we performed univariate and multivariate cox regression analyses and subgroup analyses to evaluate the impact of clinical characteristics and LRGRS on BCa survival. The results all showed that LRGRS is an important predictor of BCa (Supplementary Material Table S7-8). We then designed a nomogram using age, stage and LRGRS, which was used to predict the 1-, 3- and 5-year survival probabilities of BCa patients (Fig. 7A). The calibration curve and DCA analysis both proved the accuracy of the nomogram predictions (Fig. 7B-G).

Fig. 7.

BCa Survival Prediction Nomogram Based on LRGRS in the TCGA cohort. (A) Nomogram incorporating age, stage and LRGRS, utilized for 1, 3, and 5-year survival predictions. (B-D) Calibration curves at 1, 3, and 5 years, respectively, demonstrating nomogram’s predictive accuracy. (E-G) DCA evaluating the clinical utility of the nomogram

Functional analyses in the TCGA cohort

Next, we performed enrichment analysis to investigate the potential biological processes of LRGRS. A total of 45 differentially expressed genes were identified between the high-risk group and the low-risk group in the TCGA cohort, including 27 upregulated genes and 18 downregulated genes (Fig. 8A-B). In GO analysis, DEGs were mainly enriched in biological processes such as keratinocyte differentiation, intermediate filament organization and intermediate filament−based process (Fig. 8C). KEGG analysis showed that DEGs were significantly enriched in pathways such as amoebiasis, ECM-receptor interaction and IL-17 signaling pathway (Fig. 8D). Furthermore, GSVA research revealed that the two groups had considerable disparities in several biological processes, with high-risk group exhibiting a superior GSVA score (Fig. 8E).

Fig. 8.

Differences in biological functions between risk subgroups in the TCGA cohort. (A) Volcano plot of DEGs between the low- and high-risk groups. (B) Heatmap of DEGs between the low- and high-risk groups. (C) The GO enrichment of DEGs between the low- and high-risk groups. (D) The KEGG pathway enrichment of DEGs between the low- and high-risk groups. (E) GSVA score of pathway enrichment in the low- and high-risk groups. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Immune infiltration and responsiveness to immunotherapy and chemotherapy in the TCGA cohort

High risk group exhibited higher Stromal, Immune, and ESTIMATE scores, yet had lower Tumor Purity (Fig. 9A-D). In addition, most immune cells had higher abundance in high-risk group (Fig. 9E-F). The two risk groups also had different immune subtype compositions, with high-risk group having a higher proportion of ‘IFN-γdominant’ and ‘fibrotic’ (Fig. 9G-H). These findings suggest that there are significant differences in the TIME between different risk groups. In our analysis, high-risk group exhibited higher “TIDE scores”, “Exclusion scores”, and “Dysfunction scores” (Fig. 10A-C). This means that high-risk groups are unlikely to benefit from immunotherapy. However, there was no statistically significant difference in the response to immunotherapy between the two groups (Fig. 10D). In the meantime, we found that the high-risk group has been shown to have higher expression levels at various immune checkpoints and HLA family genes (Fig. 9I-J). The high-risk group also showed higher IPS (ctla4_pos_pd1_pos) (Fig. 10E-H). This means that the high-risk group is more likely to benefit from immune checkpoint inhibitor therapy. By comparing the differences in IC50 between the two groups for six common drugs, we also found that the high-risk group had lower IC50 values for cisplatin, camptothecin, docetaxel and paclitaxel. Patients in the high-risk group may be more sensitive to treatment with these chemotherapeutic drugs (Fig. 10I-N).

Fig. 9.

Immune Infiltration and Immune types across risk subgroups in the TCGA cohort. (A-D) ESTIMATE analysis including Stromal Score, immune Score, ESTIMATE Score and Tumor Purity. (E-F) The relative abundance of infiltrating immune cell types in the low- and high-risk groups. (G-H) Differences in immune subtypes between the low- and high-risk groups. (I-J) Expression differences of HLA family and immune checkpoint genes between the low- and high-risk groups. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Fig. 10.

Responsiveness to immunotherapy and chemotherapy across risk subgroups in the TCGA cohort. (A-D) TIDE analysis including TIDE score, Dysfunction score, Exclusion scoreand potential immunotherapy responders. (E-H) Evaluation of the immunotherapy response using the IPS. (I-N) The drug sensitivity of Cisplatin, Camptothecin, Docetaxel, Gemcitabine, Rapamycin and Paclitaxel between the low- and high-risk groups. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Tumor mutation burden analysis

We examined the differences in the distributions of somatic mutation characteristics detected between the high-risk group and the low-risk group in the TCGA data set. The mutation rates of the TP53, TTN, ARID1A, KMT2D and MUC16 genes were higher than 20% in both groups of BC patients (Fig. 11A-B). Further subgroup analysis revealed that LRGRS is an important prognostic biomarker in both the mutation and wild-type groups of different genes (Fig. 11C). However, we found that LRGRS and TMB were not significantly correlated, but the combination of the two had a better prognostic prediction (Fig. 11D-G).

Fig. 11.

The mutational landscape associated with LRGRS in the TCGA cohort. (A-B) The mutational frequency and subtypes of the Top 20 genes with the highest mutational frequency in the low- and high-risk groups. (C) Subgroup analysis demonstrates the prognostic value of LRGRS in BCa cases with different gene mutation statuses. (D-E) Correlation analysis between LRGRS and TMB and Differences in TMB between the low- and high-risk groups. (F) Kaplan-Meier survival analysis of TMB. (G) Kaplan-Meier survival analysis of TMB combined with LRGRS

Single-cell transcriptome analysis

Furthermore, we analysed the biological significance of LRGRS at the single-cell level. We analysed seven BLAC single-cell samples from the GSE135337 project, and after quality control, we obtained 36,407 cells for analysis (Fig. 12A-E). First, we performed tSNE dimensionality reduction and clustered all cells into 13 clusters. Then, based on the annotations in the original paper, we identified and annotated five different cell types: Endothelium, Epithelium, Fibroblast, Myeloid and T cell [31] (Fig. 12F-H). Subsequently, we applied the risk model to score each cell type, and we found that three ‘risk’ genes had high expression levels in fibroblasts and that LRGRS was also the highest (Fig. 12I-P). We then performed a GSVA analysis to investigate its biological functions, and the results showed that epithelial mesenchymal transition (EMT) was significantly enriched in fibroblasts and positively correlated with LRGRS (Fig. S3A-B).

Fig. 12.

Examining the distribution of LRGRS in single-cell sequencing data. (A-E) Performing gene filtering, normalization, principal component analysis of single-cell sequencing data. (F) Expression levels of marker genes for each cell cluster. (G-H) 13 cell clusters and 5 core cell subtypes visualized by the tSNE algorithm. (I-L) Distribution of 3 risk genes across several cell types. (M-P) Distribution of LRGRS across several cell types

Pan cancer analysis

Furthermore, we combined the TCGA and GTEX databases to perform a pan-cancer analysis of the three ‘risk’ genes and LRGRS. The results showed that CALR exhibited higher expression levels in most tumor tissues compared to normal tissues. However, the expression of AHNAK and ALDH1A1 varied greatly among different tumor types (Fig. 13A-C). Then, through univariate COX regression analysis, we found that LRGRS can still be used as an important prognostic biomarker in some tumor types, especially in Pancreatic adenocarcinoma (PAAD) (Fig. 13D). In PAAD, all three risk genes showed high expression levels, and cox regression analysis found that AHNAK and LRGRS were both risk factors for it. Therefore, LRGRS may also be a novel prognostic signature in PAAD.

Fig. 13.

Pan-cancer analysis of LRGRS and 3 risk genes. (A-C) Expression of 3 risk genes in tumor tissue and normal tissue across several cancer types. (D) Univariate Cox analysis of LRGRS and 3 risk genes across several cancer types. Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Experimental verification of the LRGRS model gene

In addition, the expression levels of these three ‘risk’ genes were verified in the SV-HUC-1 and 5637 cell lines by qRT-PCR and Western blot (Fig. 14A-I). Compared with SV-HUC-1 normal cells, the mRNA expression level of AHNAK and ALDH1A1 in BCa cells was significantly downregulated, while the expression level of CALR was higher in BCa cells, which was consistent with the results of the TCGA dataset. The protein expression levels of AHNAK was also low in BCa cells. However, no significant differences were observed in the protein expression levels of ALDH1A1 and CALR.

Fig. 14.

Validation of the expression levels of the 3 risk genes in cell lines. (A-C) Expression of the 3 risk genes in normal and tumor samples in the TCGA database. (Normal samples: 18; Tumor samples: 399) (D-I) Expression of the 3 risk genes in 3 human BCa cell lines (5637) and human bladder uroepithelial cells (SV-HUC-1) Statistical significance: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns: not significant

Disscussion

The identification of lactylation has yielded fresh insights into the functional importance of the Warburg effect in tumors, primarily driven by lactate, a by-product of this effect. Recent research has elucidated the molecular processes of lactylation in cancer growth, immunological evasion, and treatment resistance, underscoring its significant role in tumorigenesis and development. As this discipline advances, an increasing amount of research is concentrating on using lactylation properties to create machine learning prognostic models. These algorithms forecast cancer behaviour by examining lactylation traits, offering new possibilities for tailored therapy. At present, prediction models based on lactylation have shown good clinical predictive ability in various tumors, including breast cancer, prostate cancer, gastric cancer, hepatocellular carcinoma, pancreatic cancer and colon cancer, providing an important reference for patient treatment decisions [33–38].

In this study, we first performed ssGSEA analysis based on the TCGA database. Compared with normal tissue, BCa tissue exhibited a high glycolysis and high lactate gene signature. This is consistent with previous research [39]. At the same time, the key enzymes of lactate metabolism, LDHA and LDHB, also showed varying degrees of upregulation. These results collectively demonstrated the characteristics of high lactate metabolism in BCa, and we therefore speculated that BCa may also be highly lactylation. Through further experimental verification, we found significantly increased levels of Pan-kla in BCa tissues and cell lines. Based on these results, we further constructed a prediction model based on LRGs to investigate its potential role in treatment strategies and immunotherapy responses.

Initially, we discovered 10 prognosis-associated LRGs that classified BCa patients in the TCGA dataset into two different groups. The two clusters exhibited notable disparities in prognosis, clinical characteristics, risk gene expression, biological activities, and immunological features. Cluster 1 exhibited substantial enrichment in epithelial-mesenchymal transition (EMT) and PI3K/AKT/mTOR signalling pathways, which enhanced the proliferation, motility, and invasion capabilities of tumor cells [40, 41]. Upon further investigation of the tumor microenvironment (TME), we discovered that Cluster 1 was abundant in numerous immunosuppressive cells (including myeloid-derived suppressor cells and regulatory T cells) and immune checkpoints, suggesting that this environment is conducive to immune evasion and may diminish the efficacy of immunotherapy [42]. This view is further supported by the higher TIDE, Dysfunction and Exclusion scores [27]. These occurrences may contribute to the unfavourable prognosis of Cluster 1. Cisplatin serves as a primary drug for adjuvant therapy in BCa [5]. We also found that Cluster 1 had a higher sensitivity to cisplatin, which suggests that BCa patients in Cluster 1 are more likely to benefit from chemotherapy than immunotherapy.

Additionally, we evaluated three genes (AHANK, ALDH1A1, and CALR) using three distinct machine learning methodologies to develop the LRGRS. The resulting LRGRS signature showed good predictive performance in the TCGA cohort and in two independent external GEO validation sets. Patients with high LRGRS were similar to those in Cluster 1 and both showed a poor prognosis. In addition, we developed a nomogram chart using LRGRS and clinical features and showed good clinical predictive value for BCa. These findings suggest that LRGRS may have potential clinical utility and warrant consideration in the context of previously reported prognostic models for bladder cancer. Several prognostic models for bladder cancer have been developed on the basis of mutation profiles, immune-related signatures, or integrated molecular features [43–46]. In this context, the LRGRS developed in our study appears to show comparable prognostic value to existing molecular models. Importantly, its potential advantage may lie not only in predictive capacity, but also in its biological relevance, as it is based on lactylation-related genes and may therefore reflect a distinct dimension of lactate-associated metabolic in bladder cancer. Nevertheless, because previously published models were established in different patient populations and analytical settings, direct superiority of LRGRS cannot be inferred from cross-study comparison alone, and further head-to-head validation is warranted.

To further interpret the biological relevance of LRGRS, it is important to consider how its three component genes may be linked to lactylation through distinct mechanisms in BCa. AHNAK is a large scaffold protein involved in calcium channel regulation, membrane repair, lipid metabolism, embryonic development, and inflammatory responses [47]. Previous studies have shown that AHNAK expression is reduced in BCa tissues, consistent with our findings, and may serve as a potential biomarker for distinguishing bladder urothelial carcinoma from benign urothelial diseases [48]. Notably, recent lactylation proteomic studies have identified multiple lysine lactylation sites in AHNAK, suggesting that it may be more directly involved in lactylation-associated signaling or structural regulation [49]. In contrast, the association of ALDH1A1 with lactylation appears to be more metabolic than substrate-based. As a member of the aldehyde dehydrogenase family, ALDH1A1 is widely regarded as an oncogenic factor in multiple malignancies [50]. In BCa, it is highly expressed and strongly associated with higher tumor grade, invasiveness, recurrence, progression, and poor prognosis, while ALDH1A1-positive cells exhibit enhanced clonogenicity, self-renewal, and tumor-initiating capacity [51]. Emerging evidence further suggests that ALDH1A1 can promote glycolysis and lactate accumulation through the ZBTB7B–LDHA axis, thereby potentially creating a metabolic milieu favorable for lysine lactylation and lactylation-associated transcriptional reprogramming [52]. Compared with AHNAK and ALDH1A1, the mechanistic link between CALR and lactylation appears to be more indirect and may involve cellular stress adaptation. CALR is an endoplasmic reticulum resident protein involved in protein folding, immunity, calcium signaling and cell transformation [53–55]. It is significantly expressed in BCa, has been proposed as a urinary biomarker for disease detection, and its altered expression has been shown to influence BCa progression both in vitro and in vivo [56, 57]. Given its established roles in endoplasmic reticulum stress, hypoxia adaptation, and tumor immune regulation, CALR may function as a stress-response node associated with lactate/lactylation-related remodeling, although direct evidence for functional CALR lactylation remains limited.

It is noteworthy that the mRNA and protein expression patterns of ALDH1A1 and CALR were not fully concordant in our study. Such discordance is not uncommon, because mRNA abundance does not always directly predict protein abundance, and may be influenced by post-transcriptional regulation, translational efficiency, protein stability, and degradation pathways [58, 59]. For ALDH1A1, this inconsistency may in part reflect post-translational regulation of protein turnover, as protein modifications have been reported to affect ALDH1A1 stability and degradation in bladder cancer [60]. For CALR, an endoplasmic reticulum-resident Ca2+-binding chaperone involved in calcium homeostasis, protein folding, and quality control, protein abundance may also be influenced by proteostasis-related regulation and cellular stress adaptation [54, 61]. In addition, technical factors, including antibody specificity/sensitivity and the semi-quantitative nature of western blotting, may also have contributed to the absence of an apparent protein-level difference [62, 63]. Therefore, the observed discordance may result from both biological regulation and methodological limitations, and further validation using additional approaches will be needed.

To investigate the potential biological roles of LRGRS within the TME, we conducted a functional enrichment analysis of DEGs across high- and low-risk groups. The findings revealed substantial enrichment in ECM-receptor interactions. The remodelling of the extracellular matrix (ECM) can facilitate tumor progression by activating signalling pathways, such as PI3K/Akt, through interactions with cell surface receptors, thereby enhancing the invasion and metastasis of tumor cells and promoting angiogenesis [64]. Further analysis of immune infiltration indicated that the high-risk group exhibited significant infiltration by myeloid-derived suppressor cells and regulatory T cells, alongside elevated expression levels of immune checkpoints and higher TIDE scores, suggesting a marked immunosuppressive phenotype. This finding may also have implications for immunotherapy response. TIDE reflects T-cell dysfunction and T-cell exclusion, and a higher TIDE score generally indicates a lower likelihood of benefiting from immune checkpoint blockade [27]. The lactate–lactylation axis may contribute to this phenotype by promoting an immunosuppressive tumor microenvironment, including macrophage polarization toward an immune-suppressive state, enhanced Treg-associated immune suppression, impaired cytotoxic T-cell function, and fibroblast/ECM remodeling that restricts effective T-cell infiltration [13, 65–68]. Therefore, the increased immunosuppressive cell infiltration and higher TIDE scores observed in the high-risk group suggest that lactylation-related remodeling may be associated with a more immune-excluded and immunotherapy-resistant state.

Additionally, the high-risk group demonstrated enrichment in EMT and the PI3K/AKT/mTOR pathway, indicating enhanced invasive and metastatic capabilities as well as treatment resistance [40, 41]. In single-cell research, we identified that LRGRS is markedly concentrated in fibroblasts and is strongly correlated with EMT. Fibroblasts are an essential component of the stroma and can release growth factors, inflammatory ligands and ECM proteins, hence increasing cancer development, treatment resistance and immunological rejection [69]. In summary, the LRGRS high-risk group demonstrated a tumor microenvironment marked by immune evasion and increased invasiveness, potentially serving as a critical mechanism contributing to unfavourable prognosis. This offers significant insights for further investigating the prospective utility of this scoring system in personalised treatment and immunotherapy response forecasting. Additionally, we examined the relationship between the IC50 values of various frequently utilised chemotherapeutic agents and the LRGRS. These findings may yield valuable perspectives for combination therapy or for identifying BCa patients who exhibit heightened sensitivity to specific chemotherapeutic drugs. From a translational perspective, LRGRS may have clinical value as a molecular risk-stratification tool. It may also be further explored as a noninvasive biomarker for liquid biopsy, particularly in bladder cancer, where urine-based assays are especially attractive [70]. Moreover, its association with an immunosuppressive microenvironment and predicted treatment response suggests possible value in treatment guidance. However, these applications remain speculative and require validation in prospective clinical cohorts and liquid-biopsy samples before clinical implementation.

Nonetheless, our approach has inherent limitations. Firstly, our investigation is a retrospective analysis utilising the publicly accessible TCGA database. While the prognostic model has been corroborated in two independent cohorts, additional prospective clinical studies are requisite. Secondly, the intricate mechanism underlying the interaction between the lactylation prognostic model and immune cell infiltration remains inadequately elucidated. Further experimentation is essential to investigate and validate this relationship and its molecular mechanisms. Lastly, additional studies are necessary to examine the potential roles of these three lactylation-associated genes in the development and progression of BCa.

Conclusion

In conclusion, this study reveals the critical role of LRGs in prognostic judgement and treatment decision-making in BCa. We constructed and validated a three-gene LRGRS model that can effectively risk stratify patients and predict survival outcomes, and revealed significant differences in molecular features and immune microenvironment among different subtypes. The above findings provide new insights into the mechanisms of the TME in BCs and offer promising biomarkers and potential targets for achieving individualised therapy.

Author contributions

Hongjun Zhao: Conceptualization, Data curation, Formal Analysis, Investigation, Methodology, Resources, Software, Validation, Visualization, Writing – original draft, Writing – review & editing. Kangjing Li: Methodology, Writing – review & editing. Hexin Xu: Methodology, Writing – review & editing. Junhao Liao: Methodology, Writing – review & editing. Chubiao Zhuo: Methodology, Writing – review & editing. Qianqing Wang: Methodology, Writing – review & editing. Xiaoping Liu: Data curation, Methodology, Writing – review & editing. Zhanfang Kang: Methodology, Writing – review & editing. Baoyi Zhu: Methodology, Writing – review & editing. Xiangmao Lai: Methodology, Writing – review & editing. Sihua Zhu: Methodology, Writing – review & editing. Yuying Zhang: Project administration, Supervision, Writing – review & editing. Jianwen Zeng: Funding acquisition, Project administration, Supervision, Writing – review & editing.

Funding

This work was supported in part by grants from the National Natural Science Foundation of China (No.82270815,82104170), the Natural Science Foundation of Guangdong Province (No.2024A1515013233), the special funds of the clinical research from the Affiliated Qingyuan Hospital (Qingyuan People’s Hospital), Guangzhou Medical University (No.QYRYCRC2023016).

Data availability

No datasets were generated or analysed during the current study.

Declarations

Competing interests

The authors declare no competing interests.