HSP90 co-regulates the formation and nuclear distribution of the glycolytic output complex to promote resistance and poor prognosis in gastric cancer patients.

Introduction
Gastric cancer is the fifth most common and lethal malignant tumor globally, and in China, its mortality rate ranks third [1, 2]. Over 80% of patients are diagnosed at an advanced or metastatic stage, missing the best surgical opportunity, with 5-fluorouracil and cisplatin chemotherapy remaining the main clinical treatments. However, resistance and metastasis result in a 5-year survival rate of only 9.4% for advanced gastric cancer. Studies have shown that Cancer Stem Cells (CSCs) and metabolic reprogramming play crucial roles in cancer recurrence and resistance, promoting a series of malignant behaviors, including tumor migration, invasion, and treatment resistance [3]. For example, CSCs have been reported to exhibit enhanced glycolysis as a primary energy source, a process commonly known as the Warburg effect[4]. This metabolic reprogramming in tumors supports the energy expenditure necessary for stemness and malignant behavior by providing ample ATP and serves as a crucial basis for tumor resistance and metastasis [5]. However, the underlying mechanisms by which glycolysis and its products drive resistance remain unclear. Limited studies suggest that the activity, stability, and abundance of glycolytic enzymes are critical [6–8]. For example, the tetramerization of PKM2 promotes the glycolytic reprogramming of glioblastoma stem cells, enhancing resistance to both radiotherapy and chemotherapy [9]. DX39B promotes the metastasis and invasiveness of colorectal cancer by stabilizing PKM2 [10]. Recent studies have found that the glycolytic byproduct lactate promotes tumor progression and resistance through lactylation modifications of specific proteins. For instance, lactylation of the NBS1 protein and the key DNA damage repair protein MRE11 enhances tumor DNA repair capacity, inducing chemotherapy resistance [11]. These studies have deepened the understanding of metabolic regulation in tumors. Therefore, exploring the underlying mechanisms and crosstalk between CSCs and metabolism in tumors, as well as identifying potential targets based on this understanding, is important for addressing resistance and poor prognosis in gastric cancer treatment.
Heat shock protein 90 (HSP90) is a classic molecular chaperone that fascilitates the folding, maturation, and stabilization of client proteins [12]. HSP90 has recently been found to be associated with tumor proliferation, metastasis, and stemness properties [13, 14]. HSP90 is also related to glucose metabolism. For example, inhibiting HSP90 can downregulate PKM2 and PFKP, weakening glycolysis and enhancing the effects of radiotherapy in head and neck cancer [15]. SU086, an HSP90 inhibitor, can disrupt glycolytic metabolism in prostate cancer cells [16]. However, research in this area is still in its early stages, and it remains unclear how HSP90 specifically regulates glycolysis. Previous studies have primarily focused on HSP90 as a molecular chaperone regulating the abundance of glycolytic enzymes, particularly its role in enhancing glycolysis in hepatocellular carcinoma cells via modulation of PKM2 levels [16]. Interestingly, our previous studies have demonstrated that HSP90 can directly interact with glycolytic enzymes, including ENO1 and PKM2, thereby promoting tumor stemness [17]. However, the underlying molecular mechanisms remain to be elucidated. Therefore, uncovering how HSP90 mediates glycolytic metabolic reprogramming through specific regulatory pathways is essential to understanding its functional role in tumor biology.
An increasing number of studies have revealed the connection between the cytoskeleton and cellular metabolism, particularly glycolysis. Many glycolytic enzymes are associated with the cytoskeleton and may be released during increased actin dynamics. For instance, activation of PI3K can modulate the interaction between aldolase and actin, thereby increasing aldolase activity in the cytoplasm [18]. TRIM21 maintains a high glycolytic rate in cancer cells by enhancing the structure of F-actin [19]. Recent studies have indicated that the cytoskeleton can also physically connect to the nuclear lamina, directly influencing cancer cell characteristics such as proliferation, adhesion, cytoskeletal remodeling, and migration [20–23]. Although such studies are limited in number, they highlight the important role of the cytoskeleton in metabolic reprogramming and glycolysis. Notably, existing research is still limited to the regulation of individual glycolytic enzymes by the cytoskeleton, primarily focusing on changes in enzyme abundance and activity, and lacking an understanding of the interactions between glycolytic enzymes. There is also a lack of further research on whether the cytoskeleton can regulate the distribution of glycolytic products.
This study reveals that HSP90 influences the stemness and malignant signatures of gastric cancer cells by regulating the formation and nuclear distribution of glycolytic enzyme complexes. It also validates HSP90’s role in chemotherapy resistance and prognosis in gastric cancer patients, providing new insights for overcoming resistance.

Methods

Cell lines and culture
The Chinese Academy of Medical Sciences (China) provided the human gastric cancer cell lines AGS and HGC27, which were kept at 37 °C in a humidified incubator with 21% O2 and 5% CO2.

Cell transfection for silencing and overexpression of HSP90
Both cell lines were infected with lentivirus containing HSP90-targeting shRNA in order to silence HSP90. The plasmid pcSLenti-CMV-Hsp90AA1-3xFLAG-PGK-Puro-WPRE3 was transfected into the cells to temporarily overexpress HSP90. To assess off-target effects, we included a random shRNA and a negative control (nc) transfection group as negative controls, and used control medium as a blank control. Cells with differential HSP90 expression were enriched via puromycin (Sigma, USA) selection for 2 days to isolate positive clones. Obio Technology (Shanghai) prepared the overexpression plasmid and the shRNA-coding lentivirus. Following preliminary experiments, we selected a viral concentration at a multiplicity of infection (MOI) of 80% as the optimal infection condition.

Cell proliferation and cytotoxicity assays
The CCK-8 assay was used to measure cell viability (Dojindo, Japan). TAS-116 (MCE, HY-15785, USA), 2-deoxy-d-glucose (2-DG, MCE, HY-13966, USA), and PBS were used to treat the cells in order to analyze their proliferation. A microplate reader (iMarkTM, Bio-Rad, USA) was used to measure absorbance at 450 nm every 24 h for 96 h in order to assess cell proliferation.
The sensitivity of gastric cancer cells to chemotherapeutic agents was assessed by determining the half-maximal inhibitory concentration (IC50) of cisplatin. The cells were treated with a range of concentrations of the cisplatin. The absorbance at 450 nm was measured following a 72-h treatment period. The inhibition dose–response curves with variable slopes were used to calculate the IC50 values for cisplatin.

Colony formation assay
The number of cells in each well of a 6-well plate was adjusted to 500, and fresh culture medium was replaced every three days until visible colonies appeared. The colonies were first fixed with 4% formaldehyde (Solarbio, China), then stained with 0.1% crystal violet (Solarbio, China), washed, air-dried, and photographed. The number of colonies formed in each group was counted.

Cell invasion and migration sssay
Uncoated Transwell™ microchambers (24-well chambers; pore size 8 μm; Corning, USA) were used to assess migration ability, and Transwell™ microchambers coated with diluted Matrigel (BD Biosciences, USA) were used to assess invasion ability. A 1:10 dilution of Matrigel (BD Biosciences, USA) was used in Transwell™ chambers to assess invasion ability. The upper chambers were seeded with different groups of serum-starved cells (2 × 104) in serum-free medium, while the lower chambers were seeded with complete medium containing fetal bovine serum. After incubation at 37 °C for 24 h, cells that migrated through the membrane were fixed, stained, photographed, and counted. Quantitative analysis was then performed to assess cell migration and invasion.

Sphere formation assay
Prepare 0.8% methylcellulose solution in advance, add 2 mL to each well of a low-adhesion 24-well plate, add 500 cells suspended in methylcellulose solution to each well of each treatment group, culture at 37 °C, 5% CO2 for 7–14 days, take photos, and quantify the number and size of cell spheres using an inverted fluorescence microscope.

Western blot (Wb)
Proteins separated by SDS-PAGE are transferred to membranes, which are then washed four times with TBST for eight minutes each. Then, primary antibody is added, and the membrane is incubated at 4 °C for twenty-four hours. HRP-labeled secondary antibody is added, and the membrane is incubated at room temperature for one hour. Finally, ECL detection reagent is added for development. Primary antibodies for β-actin (4970S), HSP90 (ab13492), and ENO1 (ab227978) are sourced from Abcam in the United States, while Proteintech in China is the source of LDHA (Cat No. 19987-1-AP), PGK1 (Cat No. 17811-1-AP), TPI1 (Cat No. 10713-1-AP), PDHE1α (Cat No. 18068-1-AP), GAPDH (Cat No. 60004-1-Ig) and PKM2 (Cat No. 15822-1-AP). HRP-labeled goat anti-rabbit and anti-mouse IgG are examples of secondary antibodies (Jackson, USA). GLUT1 (Cat No. SC-377228), GPI (Cat No. SC-377228) were from Santa Cruz Biotechnology (USA).

Glucose consumption and lactic acid measurement
Cell culture supernatants were collected and extracellular lactate and glucose levels were measured using lactate and glucose assay kits (Nanjing Bioengineering Institute, China), respectively. Cells were subsequently collected and counted, and the data were normalized to the total number of cells per test.

Co-immunoprecipitation (Co-IP) and mass spectrometry assays
Cell pellets were collected and lysed to extract proteins. To encourage antibody binding to Protein A/G magnetic beads (MCE, Cat. No. HY-K0202), HSP90 antibody (Proteintech, Cat. No. 13171-1-AP) was added. To enable the antibody-magnetic bead complex to capture proteins, protein lysis buffer was added after thorough washing and incubated for the entire night at 4 °C. Proteins were eluted using 1 × SDS-PAGE protein loading buffer (Solarbio, #P1040) and heat denatured following incubation and thorough washing. In order to determine which proteins interacted, the eluted proteins were subsequently analyzed using mass spectrometry. Label-free quantification was used to measure protein abundance. The acquired mass spectrometry data were analyzed by comparison against the SwissProt human database. Identified downstream interacting proteins were subjected to enrichment analysis to determine their biological significance. SDS-PAGE and Western blot analysis were used to confirm the identified interacting proteins. All proteomics experiments were carried out at the Institute of Microbiology, Chinese Academy of Sciences (IMCAS) Mass Spectrometry (MS) Core Facility.

Computational analysis of proteomics data
Based on LFQ intensity for quantifying protein expression and after log transformation, differentially expressed proteins (DEP) were selected according to the criterion of |Fold Change|> 1.2, followed by intersection analysis of the DEP identified in the AGS and HGC27 cell lines. Given that protein abundance values often follow a distribution skewed towards high values, we used log2 normalization to standardize the data, which helps reduce the impact of extreme values on the analysis.
Pathway analysis was performed using the clusterProfiler R-package for gene ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG), with pathways identified as significantly enriched based on p < 0.05 and q < 0.25. Gene set enrichment analysis (GSEA) was conducted using KEGG and GOBP gene sets from the Molecular Signatures Database (MSigDB; www.broadinstitute.org/gsea/msigdb).
Protein network analysis was conducted using StringDB (https://www.string-db.org), Cytoscape, and the Cytoscape plugin MCODE [24, 25]. Subclustering analysis in MCODE was performed with the following settings, degree cutoff: 3, max depth: 100, node score cutoff: 0.3, haircut: true, fluff: false, and K-Core: 2. After successful clustering, subclusters with a score > 4 and nodes > 6 were selected as highly correlated protein subclusters. Six highly connected protein subclusters were selected. These subclusters were defined based on the functional roles of the constituent proteins, with the intensity of the dot color representing the log2 fold change of differential proteins, and the size of the nodes representing the number of connections.

Computational analysis of transcriptomic data in GC patients
Tumor sample mRNA expression matrix and additional clinical data were downloaded from the Gene Expression Omnibus database (GEO, https://www.ncbi.nlm.nih.gov/geo/) for datasets GSE1420, GSE29272 and GSE15459, we refer to Combine_GEO cohort. We combined the three datasets, and removed batch effects using the ComBat function (sva R-package). To make the data more suitable for cross-dataset analysis, we used R’s scale function for z-score standardization, with the transformation formula: X_new = (X − mean)/standard deviation.
When the expression values of both HSP90AA1 and HSP90AB1 were below the median, they were defined as the low combined HSP90 expression group. Conversely, when either expression value was higher, they were classified as the high combined HSP90 expression group. The top one-third high expression group and the bottom one-third low expression group were compared using a differential analysis based on the limma R-package. Differentially expressed genes (DEG) were selected according to the criterion of FDR p value < 0.001 and |Fold Change|> 1.15. GSEA of DEG was performed using the KEGG and GOBP gene sets from the MSigDB.

Construction of lactate, glycolysis, and various malignant signatures scoring
The MKI67 mRNA expression values were standardized using z-score normalization and used as the proliferation signature score. The remaining signatures were analyzed using the GSVA R-package for single sample gene set enrichment analysis (ssGSEA), with the resulting ssGSEA scores serving as the corresponding signature scores. The lactate score utilized the HP_INCREASED_SERUM_LACTATE gene set from MSigDB, while glycolysis was assessed using the HALLMARK_GLYCOLYSIS gene set. EMT score was evaluated with a concise standardized EMT marker set proposed by Lee et al., which includes 13 mesenchymal and 3 epithelial marker genes [26]. Stemness score was derived from a collection of 109 tumor stem cell genes compiled by Miranda et al. [27].

Core glycolysis related genes (CoGRG)
We compiled a list of 40 coding genes for glycolytic enzymes and transport proteins involved in 13 potential steps of glycolysis (the ten enzymatic steps of glycolysis, along with the lactate dehydrogenase reaction and two transport events across the plasma membrane) [7]. These 40 genes are defined as CoGRG (Table S3).

Consensus clustering
We performed consensus clustering of the selected 9 CoGRG along with HSP90AA1 and HSP90AB1 using the ConsensusClusterPlus R-package [28]. We determined the optimal number of clusters to be 4 based on the cumulative distribution function (CDF) and the delta area score of the CDF curve.

Construction of HGEO complex score (HGScore) and analysis
We selected HSP90AA1, HSP90AB1, PGK1, PKM2, ENO1, and LDHA as six characteristic genes. By deducting the mean from each patient's expression value and dividing the result by the standard deviation, Z-score normalization was applied to each gene's expression values. Each characteristic gene's normalized expression values were added up to determine the characteristic score. The HGScore of gastric cancer patients was truncated based on quartiles; patients in the lowest 25% were defined as having a low HGScore, while the others were classified into the high HGScore group.
The DEG between the high and low HGScore groups were selected according to the criterion of FDR p value < 0.001 and |Fold Change|> 1.5. We performed GO and KEGG pathway analysis, and pathways were identified as significantly enriched when p < 0.05 and q < 0.25. GSEA of DEG was performed using the KEGG and GOBP gene sets from the MSigDB.

Prognostic analysis of HGScore
In our combined datasets GSE1420, GSE29272, and GSE15459, 137 cases contained complete clinical information (age, sex, stage, histological subtype), we refer to Clinical_GEO cohort. We validated the prognostic value of HGScore by dividing these patients into high HGScore and low HGScore groups according to quartiles.
Kaplan–Meier (KM) curves were employed to compare Overall Survival (OS) times among gastric cancer patients in the aforementioned cohorts. The compareGroups R-package was utilized to analyze the distribution differences of various clinical phenotypes between different groups. The independent prognostic ability of HGScore was validated using multivariate Cox regression analyses. Forest plots for visualization were generated using the forestploter R-package. A nomogram model was constructed with the rms R-package to predict the overall survival rates (OS%) of GC patients based on HGScore and clinical features. Using the R-package regplot, clinical calibration plots were created. Finally, time-dependent receiver operating characteristic (ROC) analysis was used to confirm this nomogram's clinical applicability.

Analysis of tumour-infiltrating immune profile
We used ssGSEA score to represent various immune signature reported in the literature, including tumor-infiltrating lymphocytes (TIL), immune checkpoints (ICP), interferon (IFN), and immune cell cytolytic activity (CYT).
We analyzed the abundance and composition of 22 tumor-infiltrating immune cell types using the online CIBERSORT tool (https://cibersortx.stanford.edu/) [29]. The estimate R-package was utilized for ESTIMATE (Estimation of Stromal and Immune cells in Malignant Tumor tissues using Expression data) analysis to get the stromal scores, immune scores, and estimate scores. 8 immune and 2 non-immune stromal cell populations were quantified in the samples using MCPcounter for deconvolution analysis [30].
The Wilcoxon test was employed to assess differences in various immune signatures between the different HGScore groups. Correlations between HGScore, lactate ssGSEA score, and other immune signatures were assessed using Spearman correlation analysis, with p < 0.05 deemed significant.

Basic information of chemotherapy datasets
GEO Metastatic Cohort (GSE14208, N = 123): This cohort includes 123 patients with metastatic gastric cancer from the National Cancer Center Hospital in Goyang, Korea, all of whom developed resistance after first-line treatment with a combination of cisplatin and 5-FU.
GEO Non-Metastatic Cohort (combining KUGH, YUSH, and KUCM for locally advanced non-metastatic patients, N = 180): We merged expression data from three gastric cancer cohorts in Korea: KUGH cohort (GSE26899), KUCM cohort (GSE26901), and YUHS cohort (GSE13861). These patients are locally advanced, non-metastatic gastric cancer patients who all received adjuvant chemotherapy.
Stage IV Patient Cohort (N = 191): We selected a population of stage IV patients from a total of 441 gastric cancer patients, all of whom received advanced chemotherapy. We performed KM analysis to compare OS, time to progression (TTP), and recurrence-free survival (RFS) differences between high HGScore and low HGScore groups in the aforementioned cohorts of gastric cancer patients.

In vitro drug sensitivity analysis based on cancer genome project (CGP)
We obtained drug sensitivity data and expression matrix of gastric cancer cell line from the Cancer Genome Project (CGP) [31]. Specifically, the cgp2016ExprRma dataset from the pRRophetic R-package was used as the training set to predict drug IC50 values for 441 gastric cancer patients in our study. To explore the relationship between drug sensitivity and the HGScore groups, we performed the following steps: (1) IC50 Prediction: Using the expression matrix of our gastric cancer cohort as input, we applied the pRRophetic::predictPhenotype function to calculate the IC50 values of each drug for each sample. (2) Grouping by HGScore: Patients were divided into two groups (high-HGScore and low-HGScore) based on the median value of HGScore. (3) Statistical Analysis: For each drug, the Wilcoxon rank-sum test was conducted to compare IC50 values between the high- and low-HGScore groups, identifying statistically significant differences in drug sensitivity.

Identification of hub genes associated with HGScore signature
We performed Weighted Gene Co-expression Network Analysis (WGCNA) to identify genes associated with high HGScore and low HGScore signatures [32]. The goodSamplesGenes function was utilized for quality control of genes and samples. After setting the optimal soft thresholding to 6, we constructed a scale-free co-expression network at R2 = 0.9, ultimately identifying 14 co-expression modules. We further analyzed the correlation between these modules and HGScore signature using Spearman correlation tests, considering a p-value of < 0.05 as significant. The labeledHeatmap function was employed to visualize the correlation heatmap between modules and signature.

Connectivity map analysis (CMAP) of drug features
CMAP (https://clue.io) was used to uncover relationships between drugs, compounds, and biological processes, revealing connections among diseases, genes, and small molecules [33, 34]. We imported the 6 feature genes of HGEO Complex into the Connectivity Map database and calculated the compound scores (Connectivity Score) associated with high HGEO Complex gene expression. Compounds with a negative score were considered potential effective drugs with inhibitory effects on the feature. We determined the 50 compounds with the highest negative enrichment scores in the end (Table S7).

Statistical analysis
The mean ± standard deviation (SD) of three separate experiments or counts (percentages, %) are used to present all cell experiment data. Using either the Student's t-test or the Chi-square test, two groups were compared. One-way analysis of variance (ANOVA) was used to compare different groups. We used R version 4.2.0, SPSS 25.0, and GraphPad Prism 8.0 for data analysis and visualization. Statistical significance was defined as a p-value of less than 0.05 (*p < 0.05; **p < 0.01; ***p < 0.001, ****p < 0.0001).

Results

HSP90-associated proteome enriched in proteins related to glycolysis and cytoskeleton
Using transcriptomic data from GEO, we demonstrated that HSP90AA1 and HSP90AB1 are highly expressed in GC patients (Fig. S2a). The median survival of patients exhibiting high expression is significantly shorter than that of those with low expression (HSP90AA1: 15.4 months vs. 22.27 months; HSP90AB1: 17.4 months vs. 22.27 months) (Fig. S2b). Cell experiments confirmed that HSP90 overexpression promotes cell proliferation, colony formation, and self-renewal (Fig. S2e-g), and enhances migration and invasion abilities (Fig. S2d). In summary, high HSP90 expression enhance the stem cell-like properties of GC cells, significantly reducing patient survival rates.
We identified 937 differentially expressed proteins (DEP) that interact with HSP90 through mass spectrometry analysis of GC cell lines (HGC27 and AGS). These proteins are significantly enriched in molecular chaperone pathways as well as in energy metabolism and cytoskeleton pathways (Fig. 1a, b, Table S1). In the transcriptomic data, 2197 differentially expressed genes (DEG) upregulated in the high HSP90 expression group were also significantly enriched in corresponding pathways (Fig. S3, Table S1). We identified 314 genes that are upregulated at both the mRNA and protein levels, potentially more closely related to tumor stemness and poor prognosis. We analyzed the interaction network of the 314 proteins using STRING and Cytoscape (Fig. 1d), revealing six highly correlated protein subnetworks (Fig. 1d, Table S2). These subnetworks included those related to Glycolytic metabolism (9 proteins), Cytoskeleton and Material Transport (24 proteins), and DNA Replication and Repair (8 proteins). Table 1 presents eight proteins related to the cytoskeleton. These findings provide substantial evidence for the role of HSP90 in regulating glycolysis and tumor resistance. Cell experiments indicated that HSP90 overexpression significantly increased glucose consumption and lactate production, whereas the knockdown of HSP90 had the opposite effect (Fig. 1e). Experiments utilizing HSP90 and glycolysis inhibitors demonstrated that combined drug treatment maximally reduced glycolytic levels, suggesting a synergistic effect between HSP90 and glycolysis (Fig. 1f). In conclusion, the association between HSP90 and glycolysis was validated at both the protein and mRNA levels and was further confirmed through in vitro experiments.

HSP90 regulates the formation and distribution of the glycolytic enzyme output (HGEO) complex
To further investigate the relationship between HSP90 and glycolysis, we analyzed 40 genes involved in Core Glycolysis-Related Genes (CoGRG; Table S3). The results showed that when HSP90 is highly expressed, nine key genes (GLUT1, GPI, TPI1, GAPDH, PGK1, ENO1, PKM2, PDHA1, and LDHA) are significantly upregulated at both the mRNA and protein levels (Fig. 2a), and they are positively correlated with HSP90.
We performed consensus clustering analysis on HSP90 and the nine CoGRG, resulting in four clusters (C1 to C4). The C4 group (characterized by high expression of both HSP90 and CoGRG) exhibited the highest levels of glycolysis and lactate, along with the most pronounced malignant signatures, such as proliferation, epithelial-mesenchymal transition (EMT), and stemness (Fig. 2c, d, Fig. S5d). Further stratified analysis revealed that high HSP90 expression is associated with increased glycolytic activity and malignant signatures, irrespective of CoGRG expression levels. This effect was more pronounced under low glucose conditions (Fig. 2c, d, Fig. S5f-h), suggesting that HSP90 can regulate glycolytic levels independently of glycolytic enzyme expression.
We validated the interaction between HSP90 and glycolytic enzymes in cell lines. Immunoprecipitation (IP) experiments showed that four of the nine enzymes (PGK1, PKM2, ENO1, and LDHA) were immunoprecipitated with HSP90 antibody (Fig. 2e). This indicates that HSP90 can directly interact with these four glycolytic enzymes. To validate this finding, we conducted inhibitory experiments to observe whether this interaction was affected. Interestingly, we found that the use of HSP90 inhibitors significantly decreased the binding capacity between HSP90 and glycolytic enzymes, but did not affect the protein levels of glycolytic enzymes (Fig. 2F), suggesting that HSP90 regulates glycolysis in a distinct way, not by altering the abundance of glycolytic metabolic enzymes in the traditional sense. Further experiments indicated that treatment with HSP90 inhibitors did not affect the protein levels of glycolytic enzymes but did reduce their binding to HSP90 (Fig. 2f). This suggests that HSP90 does not regulate glycolytic levels by influencing the abundance of glycolytic metabolic enzymes but rather functions through alternative mechanisms. To further explore this mechanism, we conducted co-localization analysis, where high-magnification microscopy revealed that HSP90 consistently co-localizes with the four glycolytic enzymes within cells (Fig. 2g). We hypothesize that HSP90 promotes the formation of multi-enzyme complexes involving PGK1, PKM2, ENO1, and LDHA, thereby regulating glycolytic metabolism by influencing protein interactions. This multi-enzyme complex functions at the final stage of glycolysis, where PGK1 and PKM2 catalyzing ATP generation, and ENO1 and LDHA facilitate the production of lactate during the penultimate enzymatic reaction. Therefore, we refer to as the “HSP90-Glycolytic Output Complex” (HGEO Complex).
Notably, high-magnification microscopy revealed that the HGEO complex exhibits a regional distribution within the cell rather than a uniform one, co-localizing not only in the cytoplasm but also in the nucleus (Fig. 2g). We performed fluorescence co-staining with PKM2, ENO1, and LDHA, which are known to localize in the nucleus, and observed for the first time the co-localization of HSP90-PKM2-LDHA and HSP90-ENO1-LDHA in the nucleus (Fig. 2h). Combined with mass spectrometry analysis showing that HSP90-related proteins are enriched in nuclear cytoskeletal proteins (such as LMNA, LMNB1, SYNE1, and TMPO), we further speculate that the HGEO complex may facilitate its regional distribution within the nucleus through interactions with these nuclear cytoskeletal proteins.

HGEO complex reflects the glycolytic state and malignant signatures in GC patients
Using transcriptomic data from 441 GC patients, we constructed a composite score, referred to as HGScore, based on the expression of HSP90AA1/HSP90AB1 and four core glycolytic enzymes. Our analysis revealed a significant positive correlation between HGScore and levels of glycolysis and lactate, with R values of 0.66 and 0.58, respectively (p < 0.001; Fig. 3b, c). This finding suggests that HGScore effectively reflects the glycolytic metabolic state of tumors.
Further investigation revealed that HGScore-related metabolic characteristics are strongly associated with aggressive tumor phenotypes. Significant positive correlations were found between HGScore and markers of proliferation (MKI67), EMT, and stemness (Fig. S6d). Furthermore, patients in the high HGScore group exhibited significantly elevated levels of these malignant signatures (p < 0.001; Fig. S6e). Furthermore, we found that elevated glycolysis and lactate levels are associated with significantly higher malignant signatures compared to reduced glycolysis and lactate levels. This confirms that the influence of HGScore on malignant signatures is mediated through glycolysis and lactate (Fig. S5h-i).
To validate whether inhibiting the formation of HGEO complex could reduce stemness in GC cell lines, we conducted experiments showing that inhibition of HSP90, suppression of glycolysis, and dual inhibition significantly reduce the proliferation, migration, invasion, and self-renewal capabilities of the cell line. (Fig. 3g–j). These results underscore the close relationship between the HGEO complex and the metabolic characteristics of tumors, suggesting that it plays a critical role in driving the development of stemness characteristics in GC. Targeting this complex may offer novel therapeutic strategies for treating gastric cancer.

HGEO complex as an independent prognostic factor for GC patients
We further investigated whether the malignant signatures reflected by HGScore affects the prognosis of GC patients. The high HGScore group exhibited a significantly worse prognosis in both the entire cohort (N = 441) and the clinically complete cohort (N = 137) (p-values of 0.05 and 0.003, respectively, Fig. 4b, c). Additionally, there were significant differences in the distribution of HGScore among different histological subtypes (p < 0.001) and tumor stages (p < 0.05), with intestinal-type gastric cancer and stage III-IV patients showing higher HGScore values (Fig. 4a). Stratified analysis indicated that combining HGScore with clinical factors improved survival predictions for GC patients, particularly in stage III-IV and diffuse gastric cancer (Fig. 4d, e).
To further validate the ability of HGScore as an independent prognostic indicator, we conducted univariate (uniCox) and multivariate Cox regression analyses. The results indicated that HGScore is the only independent prognostic factor for overall survival (OS) aside from clinical staging (Fig. 4f). The nomogram constructed by combining all clinical factors demonstrated optimal predictive performance for HGScore and clinical staging (Fig. 5g), with areas under the curve (AUC) for 1-year, 3-year, and 5-year survival predictions being 0.72, 0.72, and 0.70, respectively. These findings indicate that HGScore possesses significant value in prognostic assessment (Fig. 4h).

HGEO complex contributes to resistance against targeted and chemotherapeutic drugs
To explore the relationship between HGScore and the response to anticancer drugs, we analyzed drug sensitivity data from gastric cancer cell lines in the CGP database. The volcano plot illustrates that cell lines with high HGScore exhibit resistance to 52 out of 199 drugs and increased sensitivity to 11 drugs (Fig. 5a, Table S5). Common chemotherapeutic agents, such as Cisplatin and Cytarabine, demonstrated widespread resistance (Fig. 5b). However, cell lines with high HGScore display sensitivity to the HSP90 inhibitor 17AAG and the PARP inhibitor Rucaparib (Fig. 5b), indicating their potential therapeutic value in patients with high HGScore gastric cancer.
To validate the clinical significance of HGScore, we analyzed two gastric cancer patient cohorts. In the metastatic cohort (N = 123), high HGScore was correlated with a shorter time to progression (TTP) of 3.9 months and an overall survival (OS) of 8.35 months (Fig. 5c). In the non-metastatic cohort (N = 180), patients with high HGScore exhibited no significant improvement in relapse-free survival (RFS) following adjuvant chemotherapy (Fig. 5d–f), indicating that chemotherapy resistance associated with HGScore may impede the benefits of adjuvant chemotherapy. In vitro experiments demonstrated that HSP90-overexpressing cells exhibit greater resistance to cisplatin, while HSP90 knockdown significantly reduced the IC50 of cisplatin (Fig. 5g). Inhibiting either HSP90 or glycolysis, as well as employing dual inhibition, improved chemotherapy sensitivity (Fig. 5h), indicating that inhibiting the formation of the HGEO complex can significantly enhance chemotherapy sensitivity in gastric cancer (Fig. 5h).
To identify potential therapeutic compounds, we performed Connectivity Map (CMAP) analysis to obtain compounds associated with elevated HGEO complex gene expression. The results revealed that negatively enriched compounds were mainly PI3K/AKT/mTOR and PARP inhibitors. This suggests that these drugs may inhibit high-HGScore patients, offering strategic support for treating this patient subgroup (Figure S7a-b, Table S7). Similarly, GSEA analysis indicated significant activation of the PI3K/ AKT/mTOR pathway and DNA repair mechanisms in tumors with high HGScore (GSEA/MSigDB; GOBP; FDR < 0.05), thereby reinforcing the clinical relevance of this finding.

LMNA regulates nuclear HGEO complex, enhancing drug resistance and poor prognosis
To explore the potential mechanisms underlying the differences in HGEO complex signature, we identified differential hub genes between the high HGScore and low HGScore groups through Weighted Gene Co-expression Network Analysis (WGCNA). A total of 14 co-expression modules were identified, with the MEblue module exhibiting the highest correlation with subtype differentiation (R = 0.79) (Fig. S8a-b). In this module, 252 genes were significantly enriched in pathways related to proliferation, drug resistance, and DNA repair (Fig. S8c), which are closely associated with tumor resistance. Furthermore, pathways associated with the cytoskeleton, nuclear chromosomes, and nuclear localization were enriched, suggesting active nuclear metabolism. GSEA further confirmed the positive enrichment of these hub genes in similar pathways (Fig. S8d, Table S6).
Based on the results of HSP90 proteomics analysis (enriched in nuclear cytoskeletal proteins such as LMNA) and immunofluorescence indicating the co-localization of the HGEO complex in the nucleus (Fig. 2j–h), we hypothesize that HSP90 mediates the nuclear distribution of the glycolytic enzyme complex through nuclear cytoskeletal proteins (such as LMNA). This mechanism regulates glycolysis and its metabolic products, enhancing nuclear metabolism, including chromatin remodeling and DNA repair, ultimately affecting chemoresistance and prognosis. To validate this hypothesis, we analyzed the nuclear lamina protein LMNA. In a cohort of 191 stage IV GC patients who had all undergone chemotherapy, cross-analysis of LMNA and HGScore indicated that patients in the Hsocre_HLM group (high level of both HGScore and LMNA) experienced the worst prognosis, with a median survival time significantly lower than that of patients in the high HGScore only or high LMNA only groups (7.9 months vs. 7.09 months vs. 6.6 months, Fig. 6a, b). This finding indicates that LMNA and HGScore have a synergistic effect on prognostic prediction. Further stratified analysis revealed that LMNA significantly distinguished prognosis only in the high HGScore group (p = 0.01, Fig. 6b), with no significant effect observed in the low HGScore group (p = 0.11, Fig. 6c). This suggests that LMNA acts as an auxiliary factor of the HGEO complex.
The previous results indicate that the effect of HGScore on malignant signatures (Fig. S5h-i) and prognosis (Fig. S4) is mediated by glycolysis and its product, lactate. Subsequently, we validated the influence of HGScore and LMNA on the prognostic prediction of lactate. Cross-analysis indicated that the combination of LAC with HGScore significantly improved prognostic stratification (p = 0.009). Patients in the Hsocre_HLA group (high level of both HGScore and LAC) experienced the worst prognosis, with a median survival of only 7.03 months. Additionally, we found that, regardless of LAC status, high HGScore could further distinguish GC patients with poorer prognosis (Fig. 6g, h). A similar result was also observed in the cross-analysis of LAC and LMNA, where LMNA further enhanced the prognostic effect of LAC. Patients in the HLA_HLM group (high level of LAC and LMNA) experienced the worst prognosis, with a median survival of only 8.8 months and HR = 4.29 (Fig. 6k). Interestingly, we found that although LMNA could further refine the prognosis of LAC, regardless of its status (Fig. 6i–m), this effect was particularly pronounced in the high LAC state, with HR of 2.27 and 1.91, respectively (Fig. 6j).
These results suggest that LMNA, as an auxiliary factor, can cooperate with HGScore to enhance the impact of glycolysis on the prognosis of GC patients. Considering the function and localization of LMNA, this process may be related to the distribution of the HGEO complex and lactate products within the nucleus, ultimately affecting the regulatory role of lactate on prognosis. Additionally, as all stage IV patients received advanced chemotherapy, these findings also offer insights into the efficacy of chemotherapy in GC patients.

Discussion
Previous studies have demonstrated that glycolytic enzymes can assemble into non-covalent complexes, such as the PGK1/PGAM1/ENO1 complex formed by lncRNA NEAT1 and the PGK1/PGAM1/ENO1/PKM2/LDHA complex assembled by gLINC [35, 36]. Compared to free enzymes, the formation of these complexes inhibits the diffusion of intermediate products, allowing the products from preceding enzymes to be directly transferred to the active site of subsequent enzymes, thereby facilitating efficient glycolysis. Recent studies have indicated that in a carcinogenic environment, HSP90 can acquire new functions as a scaffolding platform by forming higher-order oligomeric structures, known as chaperones, which participate in the binding and assembly of proteins and protein complexes, thereby enhancing tumor survival and providing metabolic advantages [12, 37, 38].
In this study, cellular experiments revealed that HSP90 co-localizes with four enzymes at the final stage of glycolysis (PGK1, PKM2, ENO1, LDHA) within the cell, exhibiting a localized distribution; however, it does not influence the abundance of glycolytic enzymes. Clinical cohorts have further demonstrated that even in cases of low expression of glycolytic enzymes, GC patients with high HSP90 expression exhibit elevated levels of glycolysis and lactate. In summary, our study reveals a novel mechanism through which HSP90 regulates glycolytic metabolism: HSP90 functions as a chaperone that regulates the formation of the PGK1-PKM2-ENO1-LDHA glycolytic enzyme complex, thereby promoting the generation of glycolytic products.
Some studies suggest that the limited diffusion rate of ATP within cells results in glycolytic enzymes accumulating in regions of high energy demand during energy stress [39]. In our study, patients with high HGScore demonstrated the activation of various nuclear pathways associated with drug resistance (Fig. S7c-d), which require substantial and rapid energy supply. Considering that some independent studies have reported the potential role of nuclear HSP90 and nuclear glycolytic enzymes in drug resistance, evidence indicates that HSP90 modulates lamin-A nuclear-cytoplasmic distribution via interaction, thereby enhancing DNA damage repair and chemotherapy resistance in ovarian cancer cell [40]. Additionally, MYG1 facilitates PKM2 phosphorylation by HSP90, promoting its nuclear accumulation and driving glycolysis in colorectal cancer cells [41]. We hypothesize that energy demand drives the formation of the HGEO complex in the nucleus and promotes in situ ATP generation. To validate this hypothesis, we were the first to observe the co-localization of HSP90-PKM2-LDHA and HSP90-ENO1-LDHA within the nucleus using fluorescence microscopy (Fig. 2j–h), confirming the presence of the HGEO complex in this organelle.
Previous research has indicated that recruiting glycolytic enzymes via the cytoskeleton can rapidly enhance cellular metabolic flux, bypassing the time-consuming processes involved in transcriptional activation and new enzyme synthesis [42]. To explore the mechanism of HGEO complex formation in the nucleus, we focused on nuclear cytoskeletal proteins that interact with HSP90 (Table 1). Among the four nuclear cytoskeletal proteins, LMNA has been extensively reported to maintain the structural integrity of the nuclear cytoskeletal, prevent DNA damage, and contribute to drug resistance in tumor cells [43–45]. Therefore, we chose LMNA as a representative for our study (Fig. 6), and the results demonstrated that the HGEO complex and lactate can synergistically promote drug resistance in GC patients through LMNA.
The clinical value of the HGEO complex in gastric cancer patients was systematically assessed, revealing that those with high HGScore exhibited elevated malignant signatures associated with proliferation, EMT, and stemness. Furthermore, HGScore was identified as an independent prognostic predictor. Notably, the HGEO complex facilitates resistance to chemotherapy. Gastric cancer patients with high HGScore find it more challenging to benefit from both first-line and neoadjuvant chemotherapy. Moreover, the HGEO complex significantly contributes to resistance and poor prognosis by regulating its nuclear distribution through interactions with LMNA. The HGEO Complex defines a subgroup of gastric cancer patients characterized by a highly malignant phenotype and strong resistance to therapy. These patients exhibit more active glycolytic metabolism, derive limited benefit from existing treatments, and have significantly shorter survival. Future studies should further investigate the potential application of HGScore in therapeutic stratification, particularly in developing individualized treatment strategies for high-risk patients. Specifically, clinicians could use HGScore to identify patients with poor chemotherapy responses, thereby optimizing treatment regimens. Targeting the formation of the HGEO Complex may represent an effective therapeutic strategy for this patient subgroup.
In summary, we provide the first evidence that HSP90 can serve as a key protein, regulates the formation of a unique production complex of glycolytic enzymes and mediates their nuclear localization through LMNA. In future research, the molecular mechanisms of this process will continue to be investigated, including the co-localization of the HGEO complex with LMNA, and the distribution, abundance, regulation, and epigenetic functions of lactate and ATP in the nucleus.
Additionally, although the HGEO complex is associated with resistance, the combined inhibition of HSP90 and glycolysis was found to exhibit substantial efficacy in vitro in gastric cancer cells. Furthermore, PARP inhibitors and PI3K/ AKT /mTOR pathway inhibitors may play a therapeutic role in GC patients exhibiting high levels of the HGEO complex. Future research should investigate the efficacy of the aforementioned inhibitors and combination therapy strategies, thus providing new treatment options for GC patients.
In conclusion, this study preliminarily reveals the key role of nuclear spatial distribution in regulating tumor resistance. In recent years, spatial multi-omics technologies, such as Spatial CITE-seq and multi-modal tri-omics, have demonstrated significant potential in elucidating complex intracellular dynamics [46–48]. These techniques offer single-cell resolution integration of proteomics, transcriptomics, and metabolomics data, providing a novel perspective for understanding the nuclear-cytoplasmic localization of protein complexes and their spatiotemporal relationships with cytoskeletal dynamics. Moving forward, these advanced methods are expected to provide substantial support for a deeper understanding of the dynamic regulatory mechanisms of protein complexes.

Conclusion
Our research reveals that HSP90 and LMNA mediate the formation and nuclear distribution of the glycolytic multi-enzyme complex, representing a novel mechanism that exacerbates poor chemotherapy efficacy and reduces survival in gastric cancer patients. These findings suggest the potential for improving chemotherapy outcomes through interventions in the HGEO complex and its associated signaling pathways, such as PI3K/ AKT/mTOR.

Supplementary Information