Development and validation of liquid–liquid phase separation-associated prognostic biomarkers in gastric cancer

Abstract

Objective

Liquid–liquid phase separation (LLPS) is the membrane-less formation of functional assemblies from specific molecules. It plays a key role in tumor growth by governing gene expression. This study aims to identify LLPS-associated molecular subtypes and prognostic biomarkers in gastric cancer (GC).

Methods

LLPS-associated genes, gene expression profiles and clinical data for GC were sourced from public databases. Differential expression analysis, UpSetR package, and Cox regression analyses were employed to pinpoint LLPS-related biomarkers. Consensus clustering was employed to confirm GC molecular subtypes. Mendelian randomization (MR) was applied to uncover causal links between biomarkers and GC risk. The quantitative reverse transcriptase polymerase chain reaction (qRT-PCR) gauged biomarker expression in GC cells (BGC-823, MGC-803, and MKN-45) and normal gastric epithelial cells (GES-1).

Results

Totally 62 LLPS-associated genes were detected. Subsequently, three LLPS-associated subtypes of GC were confirmed, containing subtypes 1, 2 and 3, and patients in subtype 2 had significantly worse overall survival (OS) than subtypes 1 and 3. Furthermore, three LLPS-associated biomarkers (DACT1, PTPN6, and GLE1) were identified for prognostic model construction. Kaplan–Meier (KM) curve suggested that OS was significantly shorter in low-expression of GLE1 and PTPN6 groups than in high-expression groups, respectively, while an opposite trend was observed in DACT1. Notably, MR analysis suggested that DACT1 was associated with the risk of GC. Moreover, qRT-PCR results demonstrated that PTPN6 and GLE1 were significantly higher expressed in GC cell lines compared to GSE-1 cells, whereas the opposite was true for DACT1.

Conclusion

This work detected three LLPS-associated prognostic biomarkers (DACT1, PTPN6, and GLE1) for GC. These findings were expected to lead to the development of more precise and effective therapeutic strategies for GC in the future.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-025-02804-9.

Introduction

Gastric cancer (GC) ranks fifth in terms of global cancer incidence and is the fourth leading cause of cancer-related mortality worldwide [1–3]. With most diagnoses at advanced stages, its 5-year survival rate stands at a low 30% [4, 5]. Effective decision-making for treatment is crucial, necessitating prognostic biomarkers for monitoring and targeted therapy to improve outcomes [6–8]. While debates continue their applicability, certain biomarkers hold significant value in research and clinical practice [4]. Advancements in prognostic biomarkers facilitate personalized medicine and response monitoring in GC [9–11]. Genetic engineering has led to the identification of accessible circulating biomarkers for early diagnosis and prognosis prediction [4, 12]. For example, Shimura et al. [13] found that Helicobacter pylori status combined with urinary TFF1 and ADAM12 protein levels were independent diagnostic biomarkers for GC.

Liquid–liquid phase separation (LLPS) manifests in diverse cellular phases and compartments, serving specialized roles. The components shaping LLPS can vary significantly. LLPS involves biomolecules transitioning from uniform surroundings to create condensed and dilute phases [14]. Proteins engaged in LLPS aggregates often possess intrinsically disordered regions (IDRs), which can facilitate LLPS by enabling weak, nonspecific interactions with multiple targets [15–17]. LLPS is involved in key biological events like autophagy, transcription, chromatin organization, tumorigenesis, and DNA damage response, producing distinct membraneless condensates to execute specialized functions [18–20]. Aberrations in liquid–liquid phase separation have been implicated in several human conditions, encompassing neurodegenerative disorders and cancer [21, 22]. Prior research indicated the potential of LLPS-associated gene signatures as prognostic indicators in hepatocellular carcinoma, gliomas, and ovarian cancer [23–25]. However, LLPS-associated prognostic biomarkers in GC have not been reported.

Mendelian randomization (MR) is a rapidly evolving field rooted in Mendel's inheritance principles. MR employs single nucleotide polymorphisms (SNP) as instrumental variables (IVs) to unveil causal connections between modifiable exposures and critical clinical events [26]. This approach, using genetic variants as instruments, offers a robust method to estimate causal links between exposures and their effects on outcomes, enriching causal insights in epidemiological research. MR proves valuable in investigating causal links between specific risk factors and tumor susceptibility [27–29]. However, the application of MR to explore the causal associations between prognostic biomarkers based on LLPS-associated genes and the risk of GC remains unexplored.

In this study, we conducted bioinformatics analyses using publicly available data to unveil LLPS-related molecular subtypes and prognostic biomarkers in GC. Additionally, qRT-PCR was employed to confirm biomarker expression in GC cells (BGC-823, MGC-803, and MKN-45) and normal gastric epithelial cells (GES-1). This research purposed to deepen the understanding of the molecular complexity of GC and to provide potential genetic biomarkers for predicting poor prognosis in GC.

Materials and methods

Data collection and compilation

We gathered gene expression profiles from two sources: the Gene Expression Omnibus (GEO, http://www.ncbi.nlm.nih.gov/geo/, access on GSE54129 and GSE84437) and The Cancer Genome Atlas (TCGA, stomach adenocarcinomas (STAD) project, https://portal.gdc.cancer.gov). GSE54129 (platform: GPL570) dataset included 111 GC cases and 21 healthy individuals. GSE84437 (platform: GPL6947) dataset consisted of 433 cases and associated survival information. Expression profiles of 380 GC cases and 37 healthy individuals, and associated survival and clinical data were also gathered from TCGA online website. LLPS-associated genes in humans were detected from PhaSepDB [30].

Screening of overlapping genes

Differentially expressed genes (DEGs) of GSE54129 dataset and TCGA-STAD cohort were pinpointed employing package limma [31] (Version 3.62.2) with criteria set at |log₂FoldChange|> 0.5 and P < 0.05. To visualize these findings, the ggplot2 package (Version 3.5.1) was fulfilled to craft volcano plots and clustered heat maps showcasing the DEGs. Subsequently, the UpSetR package (Version 1.4.0) was fulfilled to ascertain the overlap genes between LLPS-associated genes, DEGs from GSE54129 dataset, and DEGs from TCGA-STAD cohort.

Consensus clustering

The “ConsensusClusterPlus” package (Version 1.70.0) was fulfilled to execute consensus clustering on the overlap genes expressed in GC samples from TCGA-STAD cohort. This process led to the identification of distinct molecular subtypes in GC, achieved through the k-means method. Clustering outcomes were ascertained upon the stabilization of Cumulative Distribution Function (CDF) values. Subsequently, Kaplan–Meier (KM) method, in conjunction with the log-rank test, was implemented to dissect and compare patient prognoses across different groups. Principal component analysis (PCA) was fulfilled to evaluate the distribution of samples across distinct subtypes. Furthermore, differences in the expression levels of common immune checkpoints (PDCD1, PDCD1LG2, TIGIT, CTLA4, LAG3, HAVCR2, SIGLEC15, and CD274) among different subtypes were determined employing t-test.

Analysis of immune cell infiltration

CIBERSORT [32] algorithm (Version 0.1.0) along with LM22 matrix was fulfilled to quantify the proportions of 22 distinct immune cell types, leveraging RNA-sequencing (RNA-seq) data sourced from TCGA-STAD cohort. The t-test method was then performed to explore the differences in the proportions of different immune cell types.

Confirmation of LLPS-associated prognostic biomarkers in GC

GC samples from TCGA-STAD cohort were randomly partitioned into training and test groups using a 60% to 40% split ratio. Leveraging the expression profiles of overlapping genes within the training cohort, univariate Cox regression analysis was executed employing survival package (Version 3.8-3). This step aimed to uncover genes linked to GC prognosis, with a significance threshold of P < 0.05. The identified prognosis-associated genes underwent the least absolute shrinkage and selection operator (LASSO) Cox regression within the glmnet package (Version 4.1-8), enabling the extraction of prognostic biomarkers and the formulation of a prognostic model. The risk score computation involved an equation: riskscore = ∑Exp (genei) * Coef (genei), “Exp” represents the expression value of a gene within the context of LASSO Cox regression, while “Coef” signifies the coefficient associated with that specific gene. The training cohort was further classified into high- and low-risk subgroups on the basis of median risk score. Moreover, KM and time-dependent receiver operating characteristic (ROC) methods were adopted to explore the prognostic efficiency of the biomarkers for survival prediction, in addition, the area under the curve (AUC) was evaluated utilizing package survivalROC (Version 1.0.3.1). Also, the correlation between prognostic biomarkers and immune cells was explored.

MR study

The TwoSampleMR package (Version 0.6.12) was employed to retrieve SNP data of biomarkers from the genome‐wide association studies (GWAS, https://gwas.mrcieu.ac.uk/) database, and served as exposure factors. Subsequently, to avoid potential bias due to different populations, genetic background was limited to the European population in MR study. The finn-b-C3-STOMACH dataset was obtained from GWAS, which was collected in 2021 and contained 633 tumor samples, 218,159 control samples, and 16,380,466 SNPs. MR analysis was performed using GC as the outcome variable. The extract instruments function was applied to extract exposure factor SNPs. Then, the extracted SNPs were matched with the outcome data and the data were unified using harmonise data. Finally, MR analysis was performed. The relationships between biomarkers and the risk of GC were evaluated utilizing the inverse variance weighted model (IVW). Furthermore, supplementary sensitivity analyses were carried out utilizing Mendelian randomization-Egger (MR-Egger). P-values less than 0.05 were considered significant.

Cell culture

GC cell lines BGC-823 (#CTCC-007-0383, Mason Cell Technology Co., Zhejiang, China), MKN-45 (#CTCC-ZHYC-0503, Mason Cell Technology Co., Zhejiang, China), MGC-803 (#YS445C, Yaji Biological Co., Shanghai, China), and human normal gastric epithelial cells (GES-1, #CTCC-002-0010, Mason Cell Technology Co., Zhejiang, China) were cultured in Roswell Park Memorial Institute 1640 Medium (#C11875500BT, Gibco, NY, US), supplemented with 10% fetal bovine serum (#10099-141, FBS, Gibco), 100 U/Ml penicillin, and 100 μg/Ml streptomycin, and maintained at 37 °C in a humidified incubator with 5% CO₂. All these cells were identified by Short Tandem Repeat (STR) profiling. The fluid was changed every 2–3 days and routine passaging was performed when the cell density was about 80–90%.

Quantitative reverse transcriptase polymerase chain reaction (qRT-PCR)

Total RNA extraction was performed using TRIzol (#15596018, Invitrogen, CA, US) reagent. SYBR Green PCR Master Mix RNA quantification was conducted using the Nano-Drop 8000 spectrophotometer (Thermo Scientific, MA, US). The SYBR Green PCR Master Mix (Life Interlink Technology Co., Xiamen, China) was applied to determine the nucleic acid amplification reactions. RNA was reverse-transcribed into Cdna using the FastKing Cdna First-Strand Synthesis Kit (#KR118-02, TIANGEN, Germany), following the enclosed guidelines. Subsequently, a Qrt-PCR reaction was fulfilled utilizing SYBR Green PCR Master Mix (#A4004M, Life Interlink Technology Co., Xiamen, China) along with specific primers on a real-time fluorescence quantitative PCR instrument (Bio-Rad, CA, USA). Biomarker expression levels were assessed using the 2^−ΔΔCT method, and GAPDH was served as internal reference. The primer details are listed in Table 1.

Table 1.

The primer sequences

Gene	Primer	Sequence (5′-3′)	Size
DACT1	Forward	TTGAACTGTTTGAGGCGAAGAG	201 bp
	Reverse	ACTGAACACCGAGTTAGAGGAAT
PTPN6	Forward	GGTGTCCACGGTAGCTTCC	127 bp
	Reverse	ACAGGTCATAGAAATCCCCTGAG
GLE1	Forward	TTGAAGGCCCTACGCAGTTC	128 bp
	Reverse	ACCACCCATCCAGAATAAGAAGA
GAPDH	Forward	GAAGGTCGGAGTCAACGGAT	133 bp
	Reverse	CTTCCCGTTCTCAGCCATGT

Statistical analysis

Data analyses were fulfilled by R (Version 4.2.2) and GraphPad 7.0 software. Pearson’s correlation was fulfilled to examine the association between immune cell types and biomarkers. The t-test was conducted to assess variations between distinct subgroups, while log-rank test was employed to compare patient prognoses among distinct subgroups. Correlation between biomarkers and the risk of GC were evaluated utilizing the IVW. Supplementary sensitivity analyses were carried out utilizing MR-Egger. Experimental data were presented as mean ± standard deviation (SD). We adopted One-way ANOVA to evaluate group differences, followed by Tukey’s post-hoc test for comparisons. P < 0.05 served as the cutoff value.

Results

Conformation of LLPS-associated genes in GC

To confirm DEGs between GC cases and healthy individuals in GSE54129 dataset and TCGA-STAD cohort, differential expression analyses were performed. In the GSE54129 dataset, a total of 6,207 DEGs were validated, comprising 2,829 up-regulated and 3,378 down-regulated DEGs (Fig. 1A, B, Supplementary Table 1). Similarly, the TCGA-STAD cohort exhibited 5,851 DEGs, encompassing 4,890 up-regulated and 961 down-regulated DEGs (Fig. 1C, D, Supplementary Table 2). We then explored 403 LLPS-associated genes from the PhaSepDB database (Supplementary Table 3). Through preforming UpSetR package, we identified 62 LLPS-associated genes in GC by intersecting 6,207 DEGs in the GSE54129 dataset, 5,851 DEGs in the TCGA-STAD cohort, and 403 LLPS-associated genes (Fig. 1E, F, Supplementary Table 4).

Fig. 1.

Identification of 62 liquid–liquid phase separated (LLPS)-associated genes in gastric cancer (GC). A Volcano plot of differentially expressed genes (DEGs) in GSE54129 dataset. Red dot: upregulation, blue dot: downregulation, grey dot: no significant difference. B Heat map showed the top 30 expressed DEGs in GSE54129 dataset. C Volcano plot of DEGs in The Cancer Genome Atlas of Stomach Adenocarcinoma (TCGA-STAD) cohort. Red dot: upregulation, blue dot: downregulation, grey dot: no significant difference. D Heat map showed the top 30 expressed DEGs in TCGA-STAD cohort. E The number of intersected genes among LLPS-associated genes, GSE54129 dataset, and TCGA-STAD cohort. F The Venn diagram showed the 62 LLPS-associated genes in GC

Confirmation of three distinct LLPS-associated subtypes of GC

Based on the expressions of 62 LLPS-associated genes in 380 GC cases from the TCGA-STAD cohort, we performed consensus clustering to divide GC cases into three different LLPS-associated subtypes (Fig. 2A). Upon varying the clustering parameter (k) across a range of 2 to 10, it becomes evident that the most suitable choice for consensus clustering is when k = 3 (Fig. 2B). PCA revealed a consistent division of samples into the three subtypes with stability (Fig. 2C). Notably, significant variations in survival were observed among these three subtypes (P = 0.00024), with patients in subtypes 1 and 3 having better outcomes than in subtype 2 (Fig. 2D, Supplementary Table 5). Moreover, we explored the expressions of immune checkpoints in three subtypes, the results demonstrated that HAVCR2, LAG3, PDCD1, PDCD1LG2, CTLA4, and TIGIT were significantly different expressed in different subtypes (P < 0.01, Fig. 2E).

Fig. 2.

Identification of liquid–liquid phase separated-associated molecular subtypes and comparison of overall survival (OS) and immune checkpoint expression levels between the three subtypes. A Consensus clustering matrix when k = 3. B Consensus cumulative distribution function (CDF) plot for k = 2 to k = 10. C Principal component analysis of samples in three subtypes. D Kaplan–Meier curve showed significant differences in OS among gastric cancer patients in the three subtypes. E Box plots showed the differences in immune checkpoint expression levels among the three subtypes. ** P < 0.01, *** P < 0.001, **** P < 0.0001

Assessment of immune cell infiltration in three LLPS-associated subtypes

To unveil the immune cell infiltrations of three subtypes, the CIBERSORT algorithm was implemented. The distribution of 22 immune cell types in every sample was exhibited in a histogram (Fig. 3A). The heatmap revealed strong positive correlations between CD8⁺ T cells and M1 macrophages, as well as between activated memory CD4 + T cells and CD8 + T cells (Fig. 3B). M0 macrophages, as well as resting memory CD4⁺ T cells, were remarkably negatively related to CD8⁺ T cells (P < 0.05, Fig. 3B). The proportions of M2 macrophages and resting mast cells were remarkably higher in subtype 2 than in subtypes 1 and 3 (P < 0.05, Fig. 3C).

Fig. 3.

Immune cell infiltration in the three liquid–liquid phase separated-associated subtypes. A Histogram showed the composition of 22 immune cell types in each sample. B The association of 22 immune cell types was evaluated. Blue: positive correlation; orange: negative correlation. * P < 0.05. C Box plots of the immune cell abundance in the three subtypes

Development of a prognostic model based on three LLPS-associated biomarkers

For prognostic gene exploration in GC, 62 LLPS-associated genes underwent univariate Cox regression, resulting in the identification of 10 LLPS-associated prognostic genes with P < 0.05 (Fig. 4A). The 10 LLPS-associated genes were subsequently subjected to LASSO regression, revealing the presence of three prognostic biomarkers (DACT1, PTPN6, and GLE1) (Fig. 4B). In the training group, the risk score for each sample was computed utilizing the equation: risk score = 0.0473 * DACT1 + (−0.0869) * PTPN6 + (−0.0785) * GLE1. KM curve suggested considerably poorer overall survival (OS) among GC cases in the high-risk subgroup (P = 0.018, Fig. 4C, Supplementary Table 5). The AUC values at 1-year, 3-year, and 5-year were 0.645, 0.623, and 0.562, respectively (Fig. 4D). Furthermore, the test group was used for internal validation, while the GSE84437 dataset served as an external validation set to assess both the prediction efficiency and the robustness of the prognostic model. In the test group, OS was significantly shorter in high-risk subgroup than in low-risk subgroup (P = 0.022, Fig. 4E, Supplementary Table 5). The prognostic model demonstrated well predictive efficiency, as reflected by the AUC values for 1-year, 3-year, and 5-year OS were 0.622, 0.621, and 0.686, respectively (Fig. 4F). In the GSE84437 dataset, OS was remarkably shorter in high-risk subgroup than in low-risk subgroup (P = 0.014, Fig. 4G). The AUCs for 1-year, 3-year, and 5-year OS were 0.567, 0.617, and 0.554, respectively (Fig. 4H).

Fig. 4.

Development and validation of a prognostic model. A The forest plot of univariate Cox regression analysis. B LASSO regression analysis of 62 liquid–liquid phase separated-associated genes. Kaplan–Meier (KM) analyses of the overall survival (OS) between the high- and low-risk subgroups in the training group (C), test group (E), and GSE84437 dataset (G). Receiver operating characteristic curves to predict the OS of patients with gastric cancer at 1-year, 3-year, and 5-year in the training group (D), test group (F), and GSE84437 dataset (H)

Risk stratification analysis

Moreover, to unravel the prognostic performance of the prognostic model for clinical characteristics (gender and age), risk stratification in GC patients was conducted. GC cases in training set were stratified into elder and younger subgroups by the median age (median age = 65). GC cases with high-risk scores exhibited poorer OS than those with low-risk scores in elder (P = 0.037), younger (P = 0.0075), female (P = 0.047), and male (P = 0.023) subgroups (Fig. 5A, Supplementary Table 5). The biomarkers demonstrated excellent predictive performance on female and male subgroups for AUC at 5-year OS were 0.736 and 0.71, respectively (Fig. 5B). Also, the correlation between GC subtypes and prognostic model was analyzed (Fig. 5C), and the OS was significantly shorter in low-expression of GLE1 (P = 0.00076) and PTPN6 (P = 0.006) groups than in high-expression groups (Fig. 5D, Supplementary Table 5), respectively, while an opposite trend was observed in DACT1 (P = 0.019). Also, correlation analysis revealed that DACT1, GLE1, and PTPN6 were both significantly correlated with resting mast cells (P < 0.05, Fig. 5E).

Fig. 5.

Stratification analysis of the prognostic risk score. A Kaplan–Meier plots of overall survival (OS) by risk subgroups for patients with gastric cancer in the subgroups according to age (elder and younger) and gender (female and male). B Receiver operating characteristic curves showed the prognostic values of risk score in the elder, younger, female, and male subgroups. C Correlation between GC subtypes and prognostic model. D Kaplan–Meier (KM) analyses of the OS between the high- and low-expression prognostic biomarkers groups. E The correlation between prognostic biomarkers and immune cells

Exploration of the potential association between prognostic biomarkers and GC

To reveal the biomarkers and GC risk, MR analysis was performed. The SNP characteristics of DACT1 and GC were presented in Supplementary Table 6. The causal impact of DACT1 level and OC risk was assessed and presented in Fig. 6A, B, the results suggested that DACT1 was associated with the risk of GC (IVW, OR = 1.149, 95% CI   1.041–1.269, P = 0.006). The MR–Egger method presented no statistical significance (OR = 1.114, 95% CI  0.969–1.280, p = 0.138). The funnel plot exhibited a roughly symmetrical causal effect (Fig. 6C), suggesting a significant causal relationship for all SNP characteristics calculations. There were no causal relationships between PTPN6 and GC risk (Supplementary Table 7, Fig. 6D–F), as well as GLE1 and GC risk (Supplementary Table 8, Fig. 6G–I).

Fig. 6.

Mendelian randomization (MR) study. A Scatter plot showed the causal effect of DACT1 on the risk of gastric cancer (GC). B Forest plot showed the causal effect of each single nucleotide polymorphism of DACT1 on the risk of GC. C Funnel plot to visualize overall heterogeneity of MR estimates for the effect of DACT1 on GC. D Scatter plot showed the causal effect of GLE1 on the risk of GC. E Forest plot showed the causal effect of each single nucleotide polymorphism of GLE1 on the risk of GC. F Funnel plot to visualize overall heterogeneity of MR estimates for the effect of GLE1 on GC. G Scatter plot showed the causal effect of PTPN6 on the risk of GC. H Forest plot showed the causal effect of each single nucleotide polymorphism of PTPN6 on the risk of GC. I Funnel plot to visualize overall heterogeneity of MR estimates for the effect of PTPN6 on GC

Validation of the expressions of three biomarkers by qRT-PCR

This study further assessed the expression levels of three prognostic biomarkers in GC cell lines (BGC-823, MKN-45, and MGC-803) and GES-1 cells by qRT-PCR. Significantly, the expression of PTPN6 and GLE1 were notably higher in BGC-823, MKN-45, and MGC-803 cell lines than in GSE-1 cells (P < 0.01, Fig. 7). DACT1 expression was notably lower in the BGC-823, MKN-45, and MGC-803 cell lines compared to GSE-1 cells (P < 0.01, Fig. 7).

Fig. 7.

Validation of the expression levels of DACT1, PTPN6, and GLE1 in gastric cancer cell lines (BGC-823, MKN-45, and MGC-803) and normal gastric epithelial cells (GSE-1) using qRT-PCR. ** P < 0.01

Discussion

LLPS condensates play roles in diverse biological processes, such as regulating gene expression, organizing chromatin, forming cytoskeletal networks, directing cell fate determinants, facilitating autophagic removal of misfolded proteins, assembling signaling complexes, and creating synaptic signaling assemblies [33]. The potential prognostic significance of LLPS-associated molecules has attracted growing attention. Qiu et al. introduced a LLPS-related gene biomarker for predicting outcomes in epithelial OC [25]. Tang et al. reported a LLPS-associated diagnostic and prognostic signature for glioma [23]. Our study identified novel LLPS-related molecular subtypes and prognostic biomarkers for GC.

Consensus clustering provides an efficient way to analyze and detect distinct patient clusters within extensive datasets [34]. Therefore, this study employed this unsupervised algorithm to confirm three distinct LLPS-related molecular subtypes of GC based on LLPS-associated genes. We further found patients in subtype 2 had significantly worse OS compared to subtypes 1 and 3. In addition, subtype 2 showed higher proportions of M2 macrophages and mast resting cells than subtypes 1 and 3. M2 macrophages have been reported to contribute to tumor invasion, angiogenesis, and immunosuppression, leading to unfavorable outcomes [35–37]. Previous studies have suggested that M2 macrophage and mast cells were correlated with the unfavorable prognosis of GC [38, 39]. We explored immune cell infiltration in discrete LLPS-related molecular subtypes, also unveiling increased proportions of M2 macrophages and mast resting cells in subtypes associated with poorer OS. Moreover, the expressions of immune checkpoints were explored in three subtypes, the results demonstrated that HAVCR2, LAG3, PDCD1, PDCD1LG2, CTLA4, and TIGIT were significantly different expressed in different subtypes, which provides new insights into the development of novel immune checkpoint inhibitors for the treatment of GC.

Moreover, we established a prognostic model on the basis of three LLPS-associated genes, including DACT1, PTPN6, and GLE1. KM curve suggested considerably poorer OS among GC cases in the high-risk subgroup. The AUC values at 1-year, 3-year, and 5-year in train, test, and GSE84437 datasets indicated that the constructed prognostic model had good prediction efficiency and the robustness. The qRT-PCR analysis suggested that PTPN6 and GLE1 were higher expressed in GC cells (MKN-45, MGC-803, and BGC-823) than in GES-1 cells, nevertheless, the opposite was true for DACT1. Also, the OS was significantly shorter in low-expression of GLE1 and PTPN6 groups than in high-expression groups, respectively, while an opposite trend was observed in DACT1. PTPN6 was reported as a major regulator of inflammation [40]. Liu et al. demonstrated that increased PTPN6 expression enhanced the migration, invasion, and proliferation of colon cancer cells [41]. Published studies also demonstrated that the PTPN6 expression was associated with worse OS of myelodysplastic syndrome and breast cancer [42, 43]. GLE1 plays a pivotal role in gene expression regulation by overseeing mRNA export from the nucleus and is also believed to influence the initiation of translation [44]. GLE1 is engaged in supporting correct export of nuclear mRNA, assisting translation, and responding to cellular stress [45–47]. Moreover, MR study showed that DACT1 was associated with the risk of GC. DACT1, positioned within the chromosomal region 14q23.1, was first identified as an inhibitor of the Wnt/β-catenin and JNK signaling pathways, closely linked to Dishevelled (Dvl) [48, 49]. Guo et al. found methylation of DACT1 was related to poor prognosis and progression of esophageal squamous cell carcinoma [50]. A study highlighted DACT1 could advance colorectal cancer tumorigenicity and invasion by upholding β-catenin stability [51]. In addition, studies have shown that DACT1 expression is reduced in non-small cell lung cancer and gastrointestinal mesenchymal tumor [52, 53]. We also found DACT1 was lower expressed in GC cells. Altogether, PTPN6, GLE1, and DACT1 had the potential to be prognostic biomarkers for GC, and the expression levels of these genes might associate with the prognosis and progression of patients with GC.

There are limitations to our study. First, protein levels of PTPN6, GLE1, and DACT1 should be explored in GC. Second, the mechanisms of PTPN6, GLE1, and DACT1 in GC required further experimental studies. Third, the findings were required to validate in vivo to better guide clinical practice. In addition, the AUC values of the 5-year in the training and validation sets are less than 0.6, largely due to the relatively small sample size in the 5-year. Thus, more datasets should be used to validate the prognostic model constructed in this study. Finally, additional robustness tests such as MR-PRESSO, F-statistics evaluation, and heterogeneity assessment should be performed in MR analysis.

Conclusion

Taken together, this work proposed three LLPS-associated subtypes and three prognostic biomarkers (PTPN6, GLE1, and DACT1) for GC. These findings hold promise for advancing the prognosis and treatment of GC through potentially enhancing the development of more precise and effective therapeutic strategies in the future.

Author contributions

The authors confirm contribution to the paper as follows: study conception and design: DL; data collection: ML; analysis and interpretation of results: DL; statistical analysis: DL; draft manuscript preparation: JD and ML; Revision of manuscript for important intellectual content: JD. All authors reviewed the results and approved the final version of the manuscript.

Funding

Not applicable.

Data availability

The dataset analyzed in the current study were obtained from the TCGA-LIHC project and GEO database (accession number: GSE84437 and GSE54129).

Declarations

Ethics approval

Not applicable.

Competing interest

The authors declare no competing interests.