Abstract Background Long noncoding RNAs (lncRNAs) play a critical role in gastric cancer (GC) progression and metastasis. However, research comprehensively exploring tissue-derived lncRNAs for predicting peritoneal recurrence in patients with GC remains limited. This study aims to investigate the transcriptional landscape of lncRNAs in GC with peritoneal metastasis (PM) and to develop an integrated lncRNA-based score to predict peritoneal recurrence in patients with GC after radical gastrectomy. Methods We analyzed the transcriptome profile of lncRNAs in paired peritoneal, primary gastric tumor, and normal tissue specimens from 12 patients with GC in the Sun Yat-sen University Cancer Center (SYSUCC) discovery cohort. Key lncRNAs were identified via interactive analysis with the TCGA database and SYSUCC validation cohort. A score model was constructed using the LASSO regression model and nomogram COX regression and evaluated using receiver operating characteristic curves. The role of lncRNAs in the PM of GC was then examined through wound healing, Transwell, 3D multicellular tumor spheroid invasion, and peritoneal cavity xenograft tumorigenicity assays in mice. Result Five essential lncRNAs were screened and incorporated into the PM risk score to predict peritoneal recurrence-free survival (pRFS). We developed a comprehensive, integrated nomogram score, including the PM risk score, pT, pN, and tumor size, which could effectively predict the 5-year pRFS with an Area under the curve of 0.79 (95% CI: 0.71-0.88). Subsequently, we demonstrated that one of these lncRNAs, CASC15, promoted the invasion and migration of GC cells in vitro and facilitated the PM of GC cells in vivo, initially verifying that lncRNAs contribute to the PM of GC. Mechanistic analysis demonstrated that CASC15 promoted EMT and metastasis by activating the JNK and p38 pathways. Conclusion A lncRNA-based integrated score was constructed in this study to predict peritoneal recurrence in patients clinically. Supplementary Information The online version contains supplementary material available at 10.1186/s12943-024-02196-4. Keywords: Gastric cancer, Transcriptome profiles, Long noncoding RNA, Peritoneal recurrence Introduction Gastric cancer (GC), the fourth most prevalent cancer globally, is often detected at an advanced stage due to its atypical early signs, which poses a significant health risk [[64]1, [65]2]. The peritoneum is a common site for distant metastases and disease recurrence, highlighting its clinical importance [[66]2–[67]4]. Patients with GC who develop peritoneal metastases (PM) have a poor prognosis, with an overall survival rate of < 6 months [[68]5]. Hence, identifying these patients is crucial. Computed tomography (CT) and positron emission tomography (PET) are commonly used to diagnose PM in patients with GC [[69]6, [70]7]. However, their diagnostic sensitivity and accuracy are unsatisfactory [[71]6, [72]8, [73]9]. Clinically, the gold standard for diagnosing PM involves laparoscopic exploration and histopathological examination [[74]7], but laparoscopic exploration is invasive and challenging to implement in all patients with GC. Therefore, identifying accurate biomarkers to predict the risk of PM is critical and clinically promising. Long noncoding RNAs (lncRNAs), which exceed 200 nucleotides, play vital roles in transcriptional regulation, chromatin reorganization, and post-transcriptional regulation [[75]10, [76]11]. Certain lncRNAs are uniquely expressed in specific cell types or tissues, suggesting their potential as biomarkers for specific disease states [[77]12]. For example, Liu et al. identified key malignancy-specific lncRNAs in esophageal squamous cell carcinoma (ESCC) and demonstrated the potential of lncRNAs as non-invasive biomarkers for the early detection of ESCC [[78]13]. Several lncRNAs were also found to be significantly associated with PM in GC. Zhou et al. revealed that the expression of lnc-TRIM28-14 is significantly increased in PM of GC, which improves the sensitivity and specificity of diagnosis in GC with PM [[79]14]. Zhao et al. found that LINC00589 delivered by polyethyleneimine-modified mesoporous silica nanoparticles could suppress the PM of GC in vivo and in vitro [[80]15]. These biomarkers not only facilitate appropriate clinical decisions but also offer potential targets for patient-specific therapies. However, a comprehensive study on the use of tissue-derived lncRNAs to predict peritoneal recurrence in patients with GC remains limited. Therefore, this study aims to explore the transcriptional landscape of lncRNAs in GC with PM and to construct a lncRNA-based integrated score for the prediction of peritoneal recurrence in patients with GC following radical gastrectomy. In this study, we analyzed the transcriptome profiles of lncRNAs in paired peritoneal, primary gastric tumor, and normal tissue specimens from 12 patients with GC. Through interactive analysis with the TCGA database and Sun Yat-sen University Cancer Center (SYSUCC) validation cohort, we identified five key lncRNAs. Utilizing this data, we developed a PM risk score to predict peritoneal recurrence-free survival (pRFS), where higher scores indicate a greater probability of peritoneal recurrence. Furthermore, we constructed a comprehensive, integrated nomogram score, incorporating the PM risk score, pT, pN, and tumor size. This score can effectively predict pRFS, recurrence-free survival (RFS), and overall survival (OS) and identify patients with GC who would benefit from adjuvant chemotherapy. Subsequently, we conducted a comprehensive assessment on a representative lncRNA, CASC15, demonstrating its role in promoting the metastasis of GC cells in vitro and facilitating PM of GC cells in vivo, initially verifying that lncRNAs participate in the PM of GC. In summary, this study introduces an integrated score based on lncRNAs that holds promise for individualized prediction of peritoneal recurrence in patients with GC clinically. Materials and methods Collection and preparation of patient specimens This study was conducted in accordance with the Declaration of Helsinki and approved by the Institutional Review Board (IRB) of Sun Yat-Sen University Cancer Center (SYSUCC, Guangzhou, China) with the IRB number of G2021-036. We collected paired normal gastric tissue, gastric primary tumor, and PM specimens from 12 GC patients with PM in SYSUCC for transcriptome sequencing. These 12 patients with PM accompanied with tumor hemorrhage or obstruction and thus underwent palliative gastrectomy. We recruited 231 GC patients with pathological stages II and III who underwent radical gastrectomy without prior chemotherapy or radiotherapy in the SYSUCC validation cohort between December 2013 and December 2016. Patients in the SYSUCC validation cohort also met the following criteria: aged 18 years or older, regardless of gender, with a life expectancy of at least 3 months and an adequate follow-up period. Patients with a history of other malignancies within 5 years were excluded. Freshly frozen tumor tissues were obtained from each patient at SYSUCC. The clinical and pathological data used in this study were retrieved through a retrospective review of electronic medical records, and the follow-up department provided long-term follow-up data. Written informed consent was collected from all patients before the study began. High-throughput data processing and lncRNA expression mining Transcriptome sequencing of GC patients with PM in SYSUCC was conducted using a HiSeq4000 platform (Illumina, Inc., San Diego, CA, USA). The processed RNA-seq data of gastric carcinoma tissues of patients with GC in the TCGA cohort were obtained from UCSC Xena ([81]http://xena.ucsc.edu/). The expression levels were quantified using TPM (transcripts per million) for further analyses. LASSO regression analysis and PM risk score establishment LASSO regression was utilized to select potential predictive lncRNAs from a pool of candidates and to establish a PM risk score. The LASSO regression employed an L1 penalty to shrink the coefficients to zero. The penalty parameter λ (or tuning constant) determined the intensity of the penalty. We decreased λ and eased the penalty, allowing more predictors to be incorporated into the model. In this study, we employed 200-fold cross-validation to determine the optimal value of λ. Ultimately, λ was selected using means of the 1-standard error (SE) criterion. The PM risk score for each patient in the SYSUCC and TCGA cohorts was calculated with the following formula: PM risk score = ∑exp(i) * coef(i), where exp(i) denotes the expression level of a lncRNA and coef(i) represents the corresponding coefficient in the LASSO model. Cells culture GES1, AGS, SNU216, HGC27, SNU719, MKN28, and BGC823 cell lines were cultured in RPMI1640 or DMEM medium (Invitrogen, Carlsbad, USA) supplemented with 10% FBS (Gibco, NY, USA) at 37 °C with 5% CO2 in culture dishes (JET BIOFIL, Guangzhou, China). Cell lines were tested negative for Mycoplasma. Antibodies and chemicals The antibodies used in this study include: E-Cadherin (BD Biosciences 610182, used for immunoblotting (IB) at 1:1000), N-Cadherin (BD Biosciences 610921, used for immunoblotting (IB) at 1:1000), β-catenin (Proteintech 51067, used for IB at 1:3000), Vimentin (Proteintech 60330, used for IB at 1:2000), α-smooth muscle actin (Proteintech 14395, used for IB at 1:2000), p38 (Cell Signaling Technology (CST) 9212, used for immunoblotting (IB) at 1:1000), Phospho-p38 (Cell Signaling Technology (CST) 4511, used for immunoblotting (IB) at 1:1000), JNK (Cell Signaling Technology (CST) 9252, used for immunoblotting (IB) at 1:1000), Phospho-JNK (Cell Signaling Technology (CST) 9255, used for immunoblotting (IB) at 1:1000), ERK1/2 (Cell Signaling Technology (CST) 4695, used for immunoblotting (IB) at 1:1000), Phospho-ERK1/2 (Cell Signaling Technology (CST) 4370, used for immunoblotting (IB) at 1:1000), GAPDH (Proteintech 60004, used for IB at 1:1000). The chemicals utilized in this study comprise: Anisomycin (Selleck Chemistry S7409), SB203580 (Selleck Chemistry S1076), SP600125 (Selleck Chemistry S1460). Generation of stable cell lines CASC15 overexpression and shRNAs plasmids (targeting sequences: sh1, TTGGCAGTCACACTTGGACTC; sh2, AAGAAATCAAGCCTGCCCATA) were collected from TSINGKE (Guangzhou, China). We generated lentiviruses by co-transfecting the indicated plasmid with psPAX2 and pMD2.G in HEK293T cells using JetPrime (Polyplus, France). The lentiviral supernatant was collected to infect target cells for stable transduction. Stable cell lines were selected using puromycin for a few days, and their efficiency of knockdown or overexpression was confirmed using RT-PCR. For in vivo bioluminescence imaging, cells were transduced with lentiviral vectors containing Gaussia luciferase (Gluc) and selected using geneticin. Quantitative RT-PCR RNA was extracted from cells or tissues using the TRIzol reagent (Invitrogen, Carlsbad, CA, USA) and reverse-transcribed to cDNA using a Reverse Transcription Kit (TAKARA, Beijing, China). RT-PCR was conducted with SYBR Green SuperMix (Roche, Basel, Switzerland), using GAPDH as an internal control. Supplementary Table 2 lists the primer sequences. Western blotting Cells were lysed with a lysis buffer (50mM pH 7.4 Tris-HCl, 150mM NaCl, 1mM EDTA, 0.5% NP-40, and 10% glycerol) containing protease inhibitor cocktail (CW2200S, CWBIO, China) for 30 min on ice. The protein concentration was detected with BCA Protein Assay Kit (GLPBIO, GK10009) and denatured at 100 degrees for 10 min after adding SDS loading buffer. Then proteins were separated by SDS-PAGE gel and transferred to a polyvinylidene fluoride membrane. Membranes blocked with 5% milk were then detected with specific primary and secondary antibodies and detected by ChemiDoc Touch (Bio-Rad, USA). Transwell assays For Transwell assay, cells (5 × 10^4 for invasion assay or 2 × 10^4 for migration assay) were planted in 100 µL of serum-free 1640 medium into the insert of a 24-well plate Boyden chamber, with or without Matrigel-coated membrane (BD Biosciences, New Jersey, USA). The culture medium containing 20% serum was added to the lower chamber, and cells were incubated for 24 h. The cells were then washed with PBS, fixed with methanol at room temperature, and stained with crystal violet. Cells on the upper surface of the membrane were wiped off, and the fixed cells were photographed under a microscope and counted. Wound-healing assay GC cells were uniformly seeded into 6-well plates and cultured overnight to facilitate adherence. Subsequently, scratches or “wounds” were created on the plates using a 200-µL gun head. The wounds were gently washed with PBS, and serum-free medium was added. Photographs were taken under a microscope as controls. After 24 h of continuous culture, photographs were captured at the same locations for analysis. Cell proliferation detection The GC cells were seeded at a density of 2 × 10^3 cells/well in 96-well plates and cultured in a standard incubator. Cell viability was assessed every 24 h using the CCK-8 reagent (MCE, Shanghai, China), following the instructions of the manufacturer. Colony formation assay Digest cells and seed 500–1000 cells per well into 6-well plates. After 24-hour incubation, cells were cultured in a standard medium or the indicated compounds for 10-14 days continuously until clones were formed. Subsequently, colonies were fixed with methanol for 20 min at room temperature, then stained with 0.5% crystal violet solution for 2 h, washed and dried. The flat plate was photographed with ChemiDoc Touch (Bio-Rad, USA), and the area was calculated by ImageJ software. 3D multicellular tumor spheroids invasion assay We performed 3D co-culture as previously outlined [[82]16]. Briefly, 1 × 10^4 GC cells were embedded and covered with growth medium supplemented with 2% Matrigel. The medium was refreshed every 2 days over the following week. Bright-field images were acquired using the Olympus Cell Sens Standard 1.9, and the invasion area was quantitatively analyzed. In-situ hybridization (ISH) staining The ISH Kit specific for human lncRNA CASC15 was collected from Bioster-Bio (Wuhan, China) and the sequence was list in Supplementary Table [83]S2. The ISH assay was performed according to the manufacturer’s instructions. Briefly, after deparaffinization using xylene, rehydration, and digestion using pepsin, sections were incubated with digoxin-labeled ISH probes overnight. Sections were stained with streptavidin horseradish peroxidase and observed post-chromogenic with 3,3′-diaminobenzidine (DAB). Fluorescence in-situ hydration (FISH) and phalloidin staining CY3-labeled CASC15 FISH probe was synthesized by TSINGKE (Guangzhou, China). Fluorescence in-situ hydration was performed using the RNA FISH Kit (GenePharma, China) according to the manufacturer’s instructions. For phalloidin staining, cells were seeded on confocal dishes, fixed with 4% paraformaldehyde, permeabilized with 0.5% Triton X-100 and stained with phalloidin for 20 min. Nuclei were stained with 4′,6-diamidino-2-phenylindole. Images were acquired using a fluorescence microscope (Olympus FV1000, Tokyo, Japan). Xenograft tumorigenicity assay in mice Five-week-old female NSG or BALB/c nude mice (Vital River Laboratory Animal Technology, Beijing, China) were divided into two or three groups. AGS (1 × 10^7 AGS or 6 × 10^6 SNU719 cells resuspended in 100 mL of sterile PBS with 20% high-concentration Matrigel matrix (CORNING, NY, USA) were injected into the celiac cavity of BALB/c nude mice, while 5 × 10^6 MKN28 cells were injected into NSG mice. At specified time points, 150 µL of 15 mg/mL D-luciferin (GoldBio, cat. no. LUCK-1) was injected into the abdomen of each mouse for tumor observation using a Xenogen IVIS Lumina Series II. The mice euthanized killed 4 weeks after the injection. Statistical analysis Continuous variables, where appropriate, were compared using independent samples, unpaired, two-sided t-tests, or Mann-Whitney U tests. Categorical variables were compared using the χ2 or Fisher’s exact tests. The Kaplan-Meier method and log-rank test were employed to estimate the pRFS, RFS, and OS, and Cox proportional hazard regression was utilized to calculate the HR. All statistical analyses were conducted using R software (version 4.3.2). LASSO regression was performed using the “glmnet” package. Nomogram development was conducted using the “rms” package. Receiver operating characteristic (ROC) curves and AUC values were used to assess the diagnostic performance. The 95% CI for individual AUCs was computed using the “timeROC” package. Survival analyses were conducted on the “survminer” package. All experiments were performed at least three times, and data from a representative experiment were presented. Data are presented as mean ± standard deviation (S.D.). The statistical significance of differences was determined using a two-tailed Student’s t-test or two-way ANOVA, with statistical significance set at P < 0.05 significant. Results Identification of the lncRNA profiling for gastric cancer with peritoneal metastasis We conducted a cross-platform clinical discovery and validation study of lncRNA biomarkers for predicting PM (Fig. [84]1A). Initially, we performed a genome-wide screen to identify potential candidate lncRNA biomarkers using samples from the Sun Yat-sen University Cancer Center (SYSUCC) discovery cohort, comprising paired normal gastric tissue, primary gastric tumors, and PM specimens from 12 GC patients with PM. The heatmap (Fig. [85]1B) illustrates the differentially expressed lncRNAs across these sample types. Subsequently, volcano plots depict differentially expressed lncRNAs between normal gastric tissue and primary tumors (Fig. [86]1C) and normal gastric tissue and PM (Fig. [87]1D). To further identify lncRNAs associated with survival predictive potential, we analyzed the transcriptomic data from the primary tumor and adjacent normal tissue of 414 patients with GC in the TCGA database and eventually identified 10 candidate lncRNAs (Fig. [88]S1A). Fig. [89]S1B illustrates that the prognosis of patients with GC expressing high and low levels of 10 lncRNAs in the TCGA cohort exhibited significant differences. Simultaneously, the expression levels of the 10 lncRNAs in normal gastric tissue, primary tumor, and PM were determined (Fig. [90]1E). Preliminarily, the 10 lncRNAs demonstrated potentials of PM and prognosis prediction. Fig. 1. [91]Fig. 1 [92]Open in a new tab Identification of the lncRNA profiling for gastric cancer with peritoneal metastasis. (A) Flow diagram illustrating the study design. (B) Heatmap depicting differentially expressed lncRNAs among paired normal gastric tissue, primary gastric tumor, and PM identified from the SYSUCC discovery cohort. (C) Volcano plot displaying the log2 (fold change) of significantly differentially expressed lncRNAs between primary gastric tumor and normal gastric tissue. (D) Volcano plot displaying the log2 (fold change) of significantly differentially expressed lncRNAs between PM and normal gastric tissue. (E) Boxplots showing the expression levels of 10 lncRNAs among paired normal gastric tissue, primary gastric tumor, and peritoneal metastasis. Statistical analysis was conducted using the Mann-Whitney U test. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Abbreviations: SYSUCC, Sun Yat-sen University Cancer Center; TCGA, The Cancer Genome Atlas; lncRNA, long non-coding RNA; ANT, adjacent normal tissues; GC, gastric cancer; SRCC, signet-ring-cell carcinoma; PM, peritoneal metastasis Genome-wide discovery of lncRNA associated with PM We collected GC tissues from 231 patients with pathological stages II and III who underwent radical gastrectomy. Total tissue RNA was extracted, and the expression levels of the 10 identified lncRNAs were measured using quantitative RT-PCR assays. Based on the optimal cut-off value, patients were divided into high- and low-expression groups. Table [93]S1 shows the clinical and prognostic data collected from these patients, which constituted the SYSUCC validation cohort to assess the predictive ability for peritoneal recurrence. Among the 10 lncRNAs, LINC00941 was excluded because of its low expression levels and could not be compared. Among the remaining nine lncRNAs, high expression levels of seven lncRNAs indicated shorter pRFS, namely [94]AC104654.2, CASC15, CH17-360D5.2, LINC01094, LINC01614, MIR4435-2HG, and RP11-400N13.3, respectively (Fig. [95]2A). We presented these results using a forest plot and simultaneously explored the ability of the nine lncRNAs to predict RFS and OS (Fig. [96]S2). Fig. 2. [97]Fig. 2 [98]Open in a new tab Genome-wide discovery of peritoneal metastasis-associated lncRNA. (A) Kaplan-Meier curves illustrating pRFS of patients with high and low expression levels of nine lncRNAs expression levels in the SYSUCC validation cohort using quantitative RT-PCR assays. (B) Partial likelihood deviance of seven lncRNAs were identified using the LASSO regression model, with 200-fold cross-validation applied for tuning parameter selection. (C) LASSO coefficients of lncRNAs. Each curve represents an individual lncRNA. (D) Venn diagram displaying the mRNAs co-expressed with the five lncRNAs in the TCGA cohort (Pearson correlation coefficient > 0.1). (E-F) KEGG pathway (E) and GO terms (F) enrichment analysis for mRNAs co-expressed with the five lncRNAs. Abbreviations: pRFS, peritoneal recurrence-free survival; SYSUCC, Sun Yat-sen University Cancer Center; HR, hazard ratio; CI, confidence interval Using LASSO regression, five potential predictors of PM among lncRNAs (CASC15, CH17-360D5.2, LINC01094, MIR4435-2HG, and RP11-400N13.3) were filtered (Fig. [99]2B and C). In the TCGA cohort, we identified 610 mRNAs that were significantly correlated with these five lncRNAs (Fig. [100]2D). Functional enrichment analysis revealed that mRNAs co-expressed with the five lncRNAs were enriched in several known cancer-related pathways, including the PI3K-Akt signaling pathway, ECM-receptor interaction, focal adhesion, and collagen-containing extracellular matrix (Fig. [101]2E and F). These pathways are known to influence cancer metastasis [[102]17] and the matrix composition of metastatic foci [[103]18], underscoring their relevance to the mechanism of PM in GC and highlighting their potential for predicting peritoneal recurrence. Development of a PM risk score to predict peritoneal recurrence in gastric cancer In the discovery phase, we developed a generalized PM risk score by integrating the five lncRNAs to predict peritoneal recurrence. In the validation phase, the PM risk score and clinicopathological characteristics of the 231 patients in the SYSUCC cohort were shown in Fig. [104]3A. The PM risk score was significantly higher in GC patients with PM. Furthermore, we classified patients into high (149) and low (82) PM risk scores. Table [105]S3 displays the baseline covariates of the SYSUCC cohort according to the PM risk score. Kaplan-Meier curves demonstrated that patients with high PM risk score exhibited shorter pRFS (Fig. [106]3B), RFS (Fig. [107]3C), and OS (Fig. [108]3D) with hazard ratios of 2.04 (95% confidence interval, (CI): 1.35–3.07; P < 0.001), 1.93 (95% CI: 1.31–2.85; P < 0.001), and 1.83 (95 CI: 1.21–2.76; P = 0.004), respectively. In addition, the probabilities of peritoneal recurrence and mortality in the high PM risk score group were significantly higher than those in the low PM risk score group (Fig. [109]3E). In the TCGA cohorts, we validated the predictive potential of the PM risk score using OS owing to insufficient clinical information on peritoneal recurrence. The PM risk score and clinicopathological characteristics of the 414 patients in the TCGA cohort were shown in Fig. [110]3F. Kaplan-Meier curves demonstrated that patients with high PM risk score exhibited shorter OS (Fig. [111]3G) with a hazard ratio of 1.82 (high vs. low, 95% CI: 1.23–2.68; P = 0.003). Our results mirrored those in the SYSUCC cohort, with the PM risk score effectively predicting the prognosis of patients with GC. Table [112]S4 presents a summary of the area under the curve (AUC) of the PM risk score for survival in the SYSUCC and TCGA cohorts, which remained unsatisfactory. Fig. 3. [113]Fig. 3 [114]Open in a new tab Development of a PM risk score to predict peritoneal recurrence in gastric cancer. (A) Heatmap depicting the distribution of the PM risk score alongside corresponding clinicopathologic characteristics in the SYSUCC validation cohort. (B-D) Kaplan-Meier curves illustrating the probability of pRFS (B), RFS (C), and OS (D) among patients with high and low PM risk score levels. (E) Boxplot displaying the rates of peritoneal and overall recurrence, as well as death in patients with high and low PM risk score levels. (F) Heatmap illustrating the distribution of the PM risk score with corresponding clinicopathologic characteristics in the TCGA cohort. (G) Kaplan-Meier curves depicting the probability of OS among patients with high and low PM risk score levels. Abbreviations: SYSUCC, Sun Yat-sen University Cancer Center; TCGA, The Cancer Genome Atlas; PM, peritoneal metastasis; EGJ, esophagogastric junction; pRFS, peritoneal recurrence-free survival; RFS, recurrence-free survival; OS, overall survival; HR, hazard ratio; CI, confidence interval Integrated score improved the performance of the model for peritoneal recurrence prediction We performed univariate and multivariate Cox regression analyses of the pRFS, RFS, and OS in the SYSUCC cohort. Table [115]1 shows that the PM risk score, tumor size, pT stage, pN stage, and adjuvant chemotherapy were identified as independent factors influencing pRFS. To enhance the predictive accuracy of the PM risk score, we comprehensively integrated it with tumor size, pT, and pN stage to develop a competing-risk nomogram model (Fig. [116]4A). This model demonstrated significant ability to predict pRFS, achieving a 5-year AUC of 0.79 (95% CI: 0.71 − 0.88) (Fig. [117]4B). Based on this model, we calculated integrated scores and reclassified patients in the SYSUCC cohort into high (n = 157) and low (n = 74) groups. The difference in survival between high and low integrated score group was more significant than that of the PM risk score (pRFS, hazard ratio (HR): 4.57, 95% CI: 2.72–7.66, P < 0.001; RFS, HR: 3.81, 95% CI: 2.38–6.10, P < 0.001; OS, HR: 4.50, 95% CI: 2.64–7.65; P < 0.001) (Fig. [118]4C). Additionally, the predictive efficacy of the integrated score for pRFS, RFS, and OS significantly outperformed the other four factors at 2, 3, or 5 years (P < 0.05, Fig. [119]4D). Table 1. Uni- and multivariable Cox regression of pRFS, RFS and OS in the SYSUCC cohort Univariable Multivariable HR (95% CI) P value HR (95% CI) P value pRFS  PM risk score (high vs. low) 2.04 (1.35–3.07) < 0.001 1.73 (1.14–2.64) 0.011  Gender (male vs. female) 1.33 (0.90–1.95) 0.153  Age (≥ 65 vs. < 65 years) 1.14 (0.77–1.70) 0.509  Location (non-EGJ vs. EGJ) 0.97 (0.67–1.41) 0.882  Tumor size (≥ 5 vs. < 5 cm) 2.53 (1.64–3.90) < 0.001 1.87 (1.18–2.95) 0.008  Borrmann type (III-IV vs. I-II) 2.65 (1.60–4.38) < 0.001 1.60 (0.93–2.75) 0.086  Differentiated grade (G3 vs. G1/G2) 1.15 (0.80–1.66) 0.453  Lauren (Diffuse vs. Intestinal/mixed) 1.08 (0.75–1.55) 0.672  Nerve invasion (yes vs. no) 1.47 (0.89–2.42) 0.135  Vascular invasion (yes vs. no) 1.52 (1.05–2.19) 0.025 1.18 (0.81–1.72) 0.377  pT stage (T4 vs. T1-3) 2.16 (1.49–3.13) < 0.001 1.49 (1.01–2.21) 0.045  pN stage (N1-3 vs. N0) 2.98 (1.45–6.11) 0.003 2.90 (1.39–6.05) 0.005  Adjuvant chemotherapy (yes vs. no) 0.53 (0.36–0.78) 0.001 0.55 (0.37–0.82) 0.003 RFS  PM risk score (high vs. low) 1.93 (1.31–2.85) < 0.001 1.66 (1.11–2.47) 0.014  Gender (male vs. female) 1.42 (0.97–2.07) 0.072  Age (≥ 65 vs. < 65 years) 1.20 (0.82–1.75) 0.344  Location (non-EGJ vs. EGJ) 1.00 (0.70–1.44) 0.989  Tumor size (≥ 5 vs. < 5 cm) 2.57 (1.68–3.92) < 0.001 1.90 (1.21–2.97) 0.005  Borrmann type (III-IV vs. I-II) 2.40 (1.50–3.84) < 0.001 1.49 (0.90–2.46) 0.123  Differentiated grade (G3 vs. G1/G2) 1.08 (0.76–1.54) 0.679  Lauren (Diffuse vs. Intestinal/mixed) 1.00 (0.70–1.42) 0.985  Nerve invasion (yes vs. no) 1.35 (0.84–2.18) 0.217  Vascular invasion (yes vs. no) 1.52 (1.07–2.17) 0.021 1.23 (0.85–1.77) 0.270  pT stage (T4 vs. T1-3) 2.10 (1.46-3.00) < 0.001 1.49 (1.02–2.18) 0.039  pN stage (N1-3 vs. N0) 2.25 (1.21–4.18) 0.010 2.14 (1.13–4.05) 0.019  Adjuvant chemotherapy (yes vs. no) 0.55 (0.38–0.81) 0.003 0.58 (0.40–0.87) 0.008 OS  PM risk score (high vs. low) 1.83 (1.21–2.76) 0.004 1.57 (1.03–2.41) 0.037  Gender (male vs. female) 1.26 (0.85–1.87) 0.255  Age (≥ 65 vs. < 65 years) 1.23 (0.82–1.84) 0.318  Location (non-EGJ vs. EGJ) 0.91 (0.62–1.33) 0.624  Tumor size (≥ 5 vs. < 5 cm) 2.55 (1.63–3.97) < 0.001 1.89 (1.17–3.05) 0.009  Borrmann type (III-IV vs. I-II) 2.51 (1.49–4.20) < 0.001 1.49 (0.86–2.60) 0.158  Differentiated grade (G3 vs. G1/G2) 1.15 (0.79–1.66) 0.470  Lauren (Diffuse vs. Intestinal/mixed) 1.15 (0.79–1.66) 0.467  Nerve invasion (yes vs. no) 1.54 (0.91–2.62) 0.109  Vascular invasion (yes vs. no) 1.61 (1.10–2.34) 0.013 1.29 (0.88–1.89) 0.188  pT stage (T4 vs. T1-3) 2.01 (1.38–2.93) < 0.001 1.36 (0.91–2.04) 0.131  pN stage (N1-3 vs. N0) 3.08 (1.50–6.33) 0.002 2.80 (1.33–5.87) 0.006  Adjuvant chemotherapy (yes vs. no) 0.46 (0.31–0.69) < 0.001 0.45 (0.30–0.67) < 0.001 [120]Open in a new tab SYSUCC, Sun Yat-Sen University Cancer Center; pRFS, peritoneal recurrence-free survival; RFS, recurrence-free survival; OS, overall survival; HR, hazard ratio; CI, confidence interval; PM, peritoneal metastasis; EGJ, esophagogastric junction Fig. 4. [121]Fig. 4 [122]Open in a new tab Integrated score improved the performance model for peritoneal recurrence prediction (A) An integrated score was constructed using nomogram COX regression based on PM risk score, pathological T and N stage, and tumor size in the SYSUCC validation cohort. (B) Time-dependent receiver operating characteristic (ROC) curves for 2-, 3-, and 5-year pRFS based on the integrated score. (C) Kaplan-Meier curves illustrating the probability of pRFS (left panel), RFS (middle panel), and OS (right panel) among patients with high and low integrated scores. (D) Boxplot comparing the AUC values between the integrated score and other variables. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001. Abbreviations: PM, peritoneal metastasis; pRFS, peritoneal recurrence-free survival; RFS, recurrence-free survival; OS, overall survival; HR, hazard ratio; CI, confidence interval; AUC, area under curve In the subgroup analysis stratified by the integrated score, patients with high integrated scores in the SYSUCC cohort benefited from adjuvant chemotherapy, resulting in longer pRFS (HR: 0.52, 95% CI: 0.34–0.79, P = 0.002), RFS (HR: 0.53, 95% CI: 0.35–0.80, P = 0.003), and OS (HR: 0.40, 95% CI: 0.26–0.62, P < 0.001). Nevertheless, patients with low integrated scores did not benefit from adjuvant chemotherapy, showing no significant differences in pRFS, RFS, or OS (Fig. [123]S4). In the TCGA cohort, we developed another integrated score incorporating the PM risk score, age, and pN stage using the same approach (Table [124]S5, Fig. [125]S3A). The integrated score was equally capable of predicting the survival rate of patients (5-year OS AUC: 0.79, 95% CI: 0.71–0.88), whereas its predictive effectiveness was not as significant as that of the integrated score in the SYSUCC cohort (Fig. [126]S3B-D). CASC15 promotes metastasis of GC cells in vitro and in vivo To investigate the mechanistic role of lncRNAs as predictive factors for PM and prognosis, we prioritize those lncRNAs with higher hazard ratios. The survival analysis of RP11-426C22.4 and CTD-2227E11.1 did not show statistical differences and were discarded. Among RP11-400N13.3 and CASC15, considering the higher contribution coefficient of CASC15 in the LASSO regression model, we chose CASC15 for further research. Quantitative RT-PCR assays revealed CASC15 expression levels in gastric cancer cell lines and immortalized GES-1 cells (Fig. [127]S5A). Stable knockdown and overexpression models of CASC15 in GC cells were established and confirmed (Fig. [128]S5B). Wound healing, Transwell, and 3D multicellular tumor spheroid invasion assays revealed that CASC15 knockdown significantly inhibited GC cell migration and invasion (Fig. [129]5A-E; all P < 0.05). In contrast, CASC15 overexpression promoted GC cell migration and invasion (Fig. [130]5F-H; all P < 0.05). Additionally, CCK-8 and colony formation assays indicated that CASC15 affected the proliferative capacity (Fig. [131]S5C-E). Notably, we found that CASC15-knockdown cells exhibited significantly increased sensitivity to the first-line chemotherapy drugs for gastric cancer, including 5-FU and oxaliplatin (Fig. [132]S5F-G), consistent with the fact that patients with high integrated scores in the SYSUCC cohort benefited from adjuvant chemotherapy (Fig. [133]S4). Fig. 5. [134]Fig. 5 [135]Open in a new tab CASC15 promotes metastasis of GC cells in vitro. (A-B) Wound-healing assay illustrating the effect of CASC15 knockdown on migration ability of indicated GC cells. Left panels: representative images. Right panel: histograms of percentage migration area. Scale bar, 200 μm (C-D) Transwell assay demonstrating the effect of CASC15 knockdown on the migration (C) and invasion (D) ability of indicated GC cells. Left: representative images. Scale bar, 100 μm. Right: histograms of migratory or invasive cell numbers. (E) 3D multicellular tumor spheroids invasion assay showing the influence of CASC15 knockdown on the invasion ability of indicated GC cells. Left: representative images. Scale bar, 20 μm. Right: Quantification of the invading area normalized to the spheroid body in the control group. (F) Wound-healing assay demonstrating the effect of CASC15 overexpression on the migration ability of indicated GC cells. Left: representative images. Right: histogram of the percentage of migration area. Scale bar, 200 μm. (G) Transwell assay illustrating the effect of CASC15 overexpression on the migration (upper panel) and invasion (lower panel) ability of indicated GC cells. Left: representative images. Scale bar, 100 μm. Right: Quantification of migratory and invasive cell numbers. (H) 3D multicellular tumor spheroids invasion assay showing the effect of CASC15 overexpression on the invasion ability of indicated GC cells. Left: representative images. Scale bar, 20 μm. Right: the histogram shows the ratio of the invading area to the spheroid body normalized by the control group. The red line indicates the spheroid body and invading area. For (A-H), data represent mean ± SEM; the dot plot reflects data points from independent experiments. Statistical analysis was performed using a two-tailed unpaired Student’s t-test. Source data are provided in the Source Data file. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 To assess the ability of CASC15 to promote PM with GC in vivo, we established a xenograft mouse model by inoculating GC cells into the peritoneal cavity of Balb/c nude mice and NOG mice. Real-time monitoring was conducted using an in vivo small-animal fluorescence imaging system (IVIS). IVIS demonstrated a significant reduction in tumorigenicity in the peritoneal cavity of GC cells upon the knockdown of CASC15 compared to that in the control group; the opposite result was observed for overexpression (Fig. [136]6A-E, P < 0.05). Mice were euthanized after 3 or 4 weeks, and the number of abdominal tumors was quantified, ultimately leading to the same result (Fig. [137]S6, P < 0.05). Fig. 6. [138]Fig. 6 [139]Open in a new tab CASC15 promotes peritoneal metastasis of GC cells in vivo. (A-C) IVIS assay depicting the effect of CASC15 knockdown on tumorigenicity in indicated GC cells (n = 5 mice). (A) Representative images of bioluminescence of each experimental group at 3 or 4 weeks. (B) Histograms showing the luminescent intensity of photons. (C) Histograms showing tumor numbers per mouse. The number of tumors was determined by counting the xenografts collected from the abdominal cavity after dissection. (D-E) IVIS assay demonstrating the effect of CASC15 overexpression on the tumorigenicity in indicated GC cells (n = 5 mice). (D) Representative images of bioluminescence of each experimental group at the 4 weeks. (E) Histogram of luminescent intensity of photons (left) and histogram of tumor numbers per mouse (right). The number of tumors was determined by counting the xenografts collected from the abdominal cavity after dissection. (F) The expression of CASC15 was detected by in situ hybridization in primary tumor and metastatic site derived from patients with peritoneal metastases of gastric cancer. Left: representative images. Scale bar, 100 μm. Right: Chart of H-score. (G) Boxplot of relative CASC15 content in plasma of patients with or without peritoneal metastases. For (A-E), data represent mean ± SEM; the dot plot reflects data points from each mouse. Statistical analysis was conducted using a two-tailed unpaired Student’s t-test. For (F), the dot plot reflects data points from each patient and the lines represent paired tissues. Statistical analysis was conducted using a paired t-test. For (G), data represent min to max; the dot plot reflects data points from each patient. Statistical analysis was conducted using a Wilcoxon test Source data are provided in the Source Data file.*P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001, ns, no significance Next, we collected GC tissues from 19 patients with PM, comprising paired primary gastric tumors and PM specimens, and detected the expression of CASC15 by in-situ hybridization. Consistently, we observed higher expression of CASC15 in metastastic site compared to the primary tumor, further illustrating that CASC15 promotes gastric cancer metastasis (Fig. [140]6F). Moreover, we measured the levels of CASC15 in the plasma of 40 patients and found that patients with PM had higher levels of CASC15 in the plasma, suggesting it can be used as a predictor of PM risk (Fig. [141]6G, P < 0.05). CASC15 promotes the metastasis of GC via the p38 and JNK signaling pathways To search for the underlying mechanism by which CASC15 promotes the metastasis of GC, we first examined the localization of CASC15 in GC cells and found that it was mainly localized in the cytoplasm (Fig. [142]7A). Next, we performed transcriptomic analysis of SNU719 cells with stable CASC15 knockdown compared with control cells. Both KEGG and GO enrichment analyses of downregulated genes in CASC15-knockdown cells revealed enrichment in biological processes related to epithelial-mesenchymal transition (EMT), including extracellular matrix organization, extracellular structure organization, wound healing, cell-substrate junction, focal adhesion (Fig. [143]7B, [144]S7A). Indeed, we observed that N-cadherin, Vimentin and α-smooth muscle actin were mostly downregulated, while E-cadherin and β-catenin were upregulated following CASC15 knockdown. In contrast, opposite results were observed in CASC15-overexpressing cells (Fig. [145]7C-D, [146]S7B). The changes in the transcription levels of these EMT markers were consistent with their protein levels, suggesting that CASC15 influences protein expression by modulating transcription (Fig. [147]S7C). Furthermore, we found that CASC15 knockdown reduced pseudopodia formation, whereas overexpression led to the opposite effect (Fig. [148]S7D). These results suggest that CASC15 promotes the metastasis of GC cells by facilitating EMT. Fig. 7. [149]Fig. 7 [150]Open in a new tab CASC15 promotes the metastasis of GC via the p38 and JNK signaling pathways (A) RNA immuno-FISH revealing the subcellular location of CASC15 in GC cell. (B) KEGG analysis and significantly enriched pathways of changed mRNAs in CASC15-knockdown GC cells compared with control cells. (C-D) Western blotting showing the effect of CASC15 knockdown (C) and overexpression (D) on levels of the key proteins of the epithelial-mesenchymal transition (EMT) process. (E-F) Western blotting showing the effect of CASC15 knockdown (E) and overexpression (F) on levels of the key proteins of the MAPK signaling pathway. (G-H) Transwell assay demonstrating that SP600125 (20 µM 24 h) and SB203580 (15 µM 24 h) reverse the facilitating effect of CASC15 overexpression on GC cells migration and invasion. Left: representative images. Scale bar, 100 μm. Right: Quantification of migratory and invasive cell numbers. (I) Transwell assay demonstrating that anisomycin (100 nM 12 h) reverses the inhibitory effect of CASC15 knockdown on GC cell migration and invasion. Left: representative images. Scale bar, 100 μm. Right: Quantification of migratory and invasive cell numbers. For (G-I), data represent mean ± SEM from three independent experiments; the dot plot reflects data points from independent experiments. Statistical analysis was performed using a two-tailed unpaired Student’s t-test. Source data are provided in the Source Data file. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001 Pathway enrichment analysis of downregulated genes in CASC15-knockdown cells highlighted the MAPK signaling pathway. We further investigated whether the MAPK signaling pathway is responsible for the function of CASC15 and found that p-JNK and p-p38 were remarkably downregulated in CASC15-knockdown cells, and converse results were observed in overexpression cells (Fig. [151]7E-F, [152]S7E). To understand the importance of the p38 and JNK MAPK pathways in CASC15-mediated GC, we conducted rescue experiments using the JNK inhibitor SP600125 and p38 inhibitor SB203580. Obviously, both inhibitors partially counteract the increased migration and invasion capacity observed in CASC15-overexpressing cells (Fig. [153]7G-H). Importantly, treatment with anisomycin, a JNK and p38 activator, restored the capacity of tumor migration and invasion in CASC15-knockdown cells (Fig. [154]7I). These data provide evidence that the role of CASC15 in promoting GC metastasis is dependent on the JNK and p38 signaling pathways. Discussion LncRNAs play a crucial role as biomarkers, including distinct expression patterns across different physiological and pathological conditions, offering vital insights into disease diagnosis and prognosis assessment [[155]19]. They frequently exhibit tissue- and disease-specific traits, enabling a more focused and precise detection [[156]20]. Moreover, their stability and detectability in diverse biological samples enhance their practical applications as biomarkers and candidates for targeted therapies [[157]21]. In this study, we explored the potential of lncRNAs lead in diagnosing PM in GC. We conducted the transcriptome profiling of lncRNAs in paired peritoneal, primary gastric tumor, and normal tissue specimens from 12 GC patients in the SYSUCC discovery cohort. Utilizing interactive analysis with the TCGA database, we identified 10 key lncRNAs. Five candidate lncRNAs were filtered out via cross-validation, least absolute shrinkage, and LASSO regression in the SYSUCC validation cohort. Using this data, we developed a PM risk score to predict pRFS, where higher scores indicate a greater probability of peritoneal recurrence. Subsequently, we constructed a comprehensive, integrated nomogram score, incorporating the PM risk score, pT, pN, and tumor size. This integrated score could effectively predict the pRFS, RFS, and OS of patients with GC after radical gastrectomy, significantly enhancing the predictive performance of adjuvant chemotherapy. Moreover, we conducted an in-depth study on a key lncRNA, CASC15, demonstrating its ability to promote the metastasis of GC cells in vitro and facilitate the PM of GC cells in vivo. We also preliminarily confirmed that CASC15 facilitates the EMT process by activating the JNK and p38 MAPK pathways, thereby promoting gastric cancer metastasis. This study provides initial evidence supporting the involvement of lncRNAs in the PM of GC. To facilitate clinical application, we developed an integrated score based on lncRNAs for personalized prediction of peritoneal recurrence in patients. Meanwhile, CASC15 can be used as a liquid biopsy predictor of PM in GC. PM presents a significant challenge in clinical management and is the primary recurrence pattern in patients with GC after radical resection [158]2, [159]5, [160]22]. Nevertheless, effective clinical approaches to diagnose postoperative peritoneal recurrence remain elusive, particularly due to physiological changes in abdominal tissues, ultimately hindering diagnosis. Traditional methods of imaging detection have limitations in this context. To address this, Lee constructed a novel transcriptomic signature for risk stratification and identification of high-risk patients with peritoneal carcinomatosis [[161]23], Shimura generated an miRNA-based signature with the potential to identify PM in GC patients [[162]24], and Chen depicted a collagen signature-based nomogram that can predict the risk of PM in GC with serosal invasion after radical surgery [[163]25]. However, none of these have been widely clinically applied. We investigated a lncRNA-based model for predicting PM prediction from another perspective in the whole transcriptome and combined it with clinicopathological features to create a more comprehensive integrated score. Furthermore, this score can predict recurrence and overall survival. This may be due to the predominance of PM in the recurrence of GC, which is a significant cause of mortality in patients with GC. The diagnosis of PM, along with the simultaneous diagnosis of GC or during intended radical surgery, is referred to as synchronous PM. Correspondingly, the occurrence of peritoneal recurrence detected after radical resection of gastric cancer is termed metachronous PM. The article initially focuses on lesions with synchronous PM to explore the lncRNA transcriptional landscape and identify key lncRNAs. This approach was chosen partly because patients with synchronous PM can more conveniently access tissue from primary gastric cancer, peritoneal metastases, and normal gastric mucosal simultaneously. Additionally, these patients typically have not undergone any prior treatment, making the sequencing results more reliable and persuasive. Teng et al. investigated the key differentially expressed proteins in both synchronous and metachronous PM, which implies that their underlying pathogenic mechanisms might be analogous [[164]26]. Therefore, we constructed a prediction model from the data of patients with synchronous PM and obtained the results as expected, which could effectively predict peritoneal recurrence of gastric cancer. Notably, our study found that CASC15 was higher in the plasma of patients with PM than those without PM. The above results suggest that CASC15 can predict both synchronous PM and metachronous PM, indicating a broad application prospect. Adjuvant chemotherapy is routinely recommended for patients diagnosed with pathological stages II and III [[165]7, [166]27]. However, not all patients derive equal benefits from adjuvant chemotherapy [[167]28], particularly those with mismatch repair deficiency (dMMR)/microsatellite instability-high (MSI-H) [[168]29], and such patients may require more precise stratification to avoid chemotherapy-related adverse effects and unnecessary financial burdens. Our study revealed that patients with high-integrated scores could significantly benefit from adjuvant chemotherapy, whereas the prognosis of patients with low-integrated scores remained unaffected by chemotherapy administration prognosis. The knockdown of CASC15 increased the sensitivity of gastric cancer cells to first-line chemotherapy drugs, such as 5-FU and oxaliplatin. This suggests that the score can serve as an evaluation indicator after radical resection of GC and CASC15 silencers may possess a chemosensitizing effect. However, potential biases due to the small sample size cannot be ruled out, and further validation through large-scale studies is necessary to confirm these findings. Theoretically, lncRNAs play a pivotal role in predicting PM of GC by mediating its metastatic processes. Previous studies show that lncRNAs can either promote or inhibit PM of GC through pathways such as ferroptosis [[169]30] and metabolic reprogramming of fatty acid [[170]31]. In this study, we selected CASC15 as a representative lncRNA and conducted further exploration. We discovered that CASC15 expression could influence the migration and invasion of GC cells in vitro, as well as the PM of GC cells in vivo. The role of CASC15 in promoting GC metastasis is dependent on the JNK and p38 signaling pathways. This finding provides theoretical support for the role of lncRNAs as predictors of PM of GC. This study had some limitations. For instance, due to insufficient details regarding disease recurrence and the time of patients in the TCGA database, we were unable to validate this score externally. Although our studies found that the lncRNA-based model can serve as a powerful tool to predict the synchronous and metachronous PM of gastric cancer, external validation of other institutions is still needed to validate this model. Additionally, obtaining tumor samples required for this score necessitates invasive procedures. Notably, we identified the potential of CASC15 liquid biopsy to predict peritoneal metastasis of gastric cancer. Moving forward, we plan to conduct a study with a larger sample size to validate this finding further. Conclusions In summary, this study investigated the transcriptional landscape and developed an integrated score based on lncRNAs that can be applied for personalized prediction of peritoneal recurrence and death in clinical settings. This integrated score can significantly predict the benefit of adjuvant chemotherapy, thereby optimizing treatment strategies. Electronic supplementary material Below is the link to the electronic supplementary material. [171]Supplementary Material 1^ (3.6MB, pdf) [172]Supplementary Material 2^ (41.3KB, xlsx) Acknowledgements