LLPS-based classification and a novel prognostic signature reveal NRF1 as a therapeutic target in pancreatic cancer Authors: Weishen Wang, Yi Zhao, Songyao Jiang, Yiwei Zhou, Qinxin Yang, Dan Li, Yu Jiang, Haoda Chen, Xiaomei Tang, Linjie Ren, Jia Liu, Jiabin Jin, Da Fu, Hong Yu Published: 2026 ABSTRACT AbstractBackgroundAberrant liquid–liquid phase separation (LLPS) can alter biomolecular condensate functions and may influence pancreatic tumorigenesis and progression, but the specific role of LLPS regulators in prognosis and the tumor immune microenvironment (TIME) in pancreatic ductal adenocarcinoma (PDAC) remains unclear.MethodsWe integrated transcriptome data of LLPS regulator–related differentially expressed genes (DEGs; n = 298) in a cohort of 176 PDAC patients from TCGA. Three LLPS regulator subtypes (LS1–LS3) were identified through multi-omics analyses, and a prognostic LLPS subtype–related risk model (LRRPC) was developed and validated. Chromatin immunoprecipitation confirmed NRF1 binding to promoters of key risk genes, and in vitro and in vivo experiments assessed the effects of NRF1 targeting on tumor growth.ResultsThe three LLPS regulator subtypes exhibited significant differences in prognosis, clinical features, genomic alterations, TIME patterns and predicted immunotherapy response. The LRRPC signature predicted prognosis and immunotherapy efficacy across cohorts and was associated with tumor biomarkers and immune infiltration. Nuclear Respiratory Factor 1 (NRF1) directly regulated hub genes such as FAM83A, RHOV and ITGB6, promoting PDAC cell proliferation, while its inhibition induced apoptosis and reduced tumor growth.ConclusionsThis study proposes an LLPS-based stratification framework for PDAC, and the LRRPC model provides an LLPS subtype–related risk score that may assist personalized prognostic assessment and immunotherapy stratification. NRF1 emerges as a promising therapeutic candidate whose targeting can inhibit tumor progression in PDAC experimental models and warrants further evaluation.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12935-026-04395-z.Keywords: LLPS, Pancreatic cancer, Immunotherapy, NRF1, Prognostic model Introduction Pancreatic ductal adenocarcinoma (PDAC) is a significant health challenge characterized by one of the lowest 5-year survival rates among all malignancies [1]. Many survival prediction models have been trained by integrating various biomarkers, including protein features and genomic data [2–4]. Given the considerable heterogeneity of PDAC, it is crucial to identify optimal biomarkers for customizing personalized therapies, improving treatment outcomes, and ultimately enhancing the survival rates of patients with PDAC. However, only a few of these models have been translated into routine clinical practice. Liquid-liquid phase separation (LLPS) plays a crucial role in the construction of subcellular compartments. Liquid droplet formation is primarily driven by transient and weak intermolecular forces [5, 6]. These forces are mainly mediated by ‘scaffold’ proteins that drive condensate formation and ‘client’ molecules that are recruited into these condensates [7]. Functional groups in proteins (either as scaffolds or clients) that facilitate such intermolecular forces often include intrinsically disordered regions (IDRs) and/or low complexity domains (LCDs). In the human genome, approximately 44% of encoded proteins contain more than 10% IDR/LCD regions [8, 9]. Hence, “stickers,” the binding points or motifs within IDRs/LCDs, are universally distributed in the human proteome. Potential proteins involved in LLPS can be identified using prediction websites that focus on IDRs/LCDs10. Macromolecular regulators modulate membraneless organelle formation by influencing LLPS [11–14]. They can be categorized into three archetypes based on their interaction strength: volume-exclusion promoters, which facilitate LLPS by displacing proteins; weak-attraction suppressors, which inhibit droplet formation by competing for interactions; and strong-attraction promoters, which stabilize droplets through direct, strong binding [15, 16].This framework offers a unified understanding of how diverse cellular molecules regulate organelle assembly and dynamics. Malignant tumor progression involves various aberrant cellular processes, and recent evidence has highlighted that many of these processes occur within biomolecular condensates [3, 17, 18]. For example, oncogenic receptor tyrosine kinases (RTKs) and the MAPK signaling pathway have been identified as the most frequently altered oncogenic networks [19], and LLPS can increase the interaction time between RTK and its substrate, thereby promoting the proliferation of cancer cells. Epithelial-mesenchymal transition (EMT) can activate the transcriptional coactivators YAP and TAZ to facilitate tumor metastasis. YAP can undergo LLPS in the nucleus, where interaction with partner proteins that provide IDRs promotes condensate assembly and modulates its pro-metastatic transcriptional activity [20]. Furthermore, MYC-mediated LLPS can induce tumor vasculature by upregulating VEGF expression [21]. Our previous study demonstrated that BHLHE40, an upregulated transcription factor in PDAC, can facilitate the linking of the promoter-to-enhancer regions of SREBF1 by forming phase-separated condensates, which increase the growth of PDAC [22]. Based on these observations, we hypothesize that coordinated dysregulation of LLPS regulators in PDAC establishes distinct condensate-mediated transcriptional states that actively drive disease progression. Specifically, altered LLPS regulator patterns may reprogram oncogenic transcriptional networks (such as lipid metabolism and proliferation pathways) and reshape the tumor immune microenvironment, thereby influencing patient prognosis and response to therapy. Accordingly, in this study we systematically investigated patient heterogeneity based on LLPS regulator patterns in PDAC. Using the TCGA-PAAD cohort, which predominantly comprises pancreatic ductal adenocarcinoma cases, we identified three distinct LLPS regulator-based subgroups with divergent survival, clinicopathological features, pathway alterations, genomic attributes, immune signatures and predicted responses to immune checkpoint inhibition. We then developed a deep neural network–based prognostic model (LRRPC) to quantitatively predict survival and immunotherapy benefit and identified the transcription factor NRF1 as a key regulator of high-risk hub genes with translational potential as a therapeutic vulnerability in PDAC. Together, these findings establish an LLPS-centered framework for prognostic stratification and therapeutic targeting in pancreatic cancer. Materials and methods Data acquisition and preprocessing RNA sequencing (RNA-seq) data and corresponding clinical information for 176 patients with pancreatic ductal adenocarcinoma (PDAC) were obtained from The Cancer Genome Atlas (TCGA-PAAD) portal (https://portal.gdc.cancer.gov/). Patients with incomplete survival data were excluded. Additionally, RNA-seq data for 167 normal pancreatic tissues were obtained from the Genotype-Tissue Expression (GTEx) database (https://gtexportal.org/home). All RNA-seq data, initially in fragments per kilobase of exon model per million mapped reads (FPKM) format, were uniformly processed and normalized to log₂(TPM + 1). The IMvigor210 cohort, including expression and clinical data of patients with urothelial cancer treated with an anti-PD-L1 antibody, was obtained from the “IMvigor210CoreBiologies” R package and used to evaluate the association between our signature and immunotherapy response. A comprehensive list of LLPS-related genes was compiled for subsequent analyses (Table S1–S2). Specifically, we first collected LLPS-associated regulators from public LLPS databases (e.g., PhaSepDB [23] and DrLLPS [24]) and key reviews [11, 12, 25] on biomolecular condensates, and then refined this list by prioritizing genes with reported roles in tumor biology and stronger support for LLPS-related functions in the literature. Identification of LLPS subtypes in PDAC First, differentially expressed genes (DEGs) between PDAC and normal samples were identified using a significance threshold of P < 0.05 and |log₂FC| > 1. Subsequently, an intersection analysis was performed among the genes associated with LLPS regulators, tumor-driving genes, and the identified DEGs to define a set of tumor-driven LLPS-related DEGs. The expression profiles of these genes were then utilized for non-negative matrix factorization (NMF) consensus clustering, implemented with the “NMF” R package, to classify patients with PDAC into distinct LLPS subtypes. The optimal number of clusters was determined by assessing cophenetic, dispersion, and silhouette metrics. Clustering stability was additionally visualized using t-distributed stochastic neighbor embedding (t-SNE). Survival differences among LLPS subtypes were analyzed using Kaplan–Meier (K-M) survival curves. The enrichment levels of 50 hallmark gene sets from the Molecular Signatures Database (MSigDB) were quantified across different LLPS subtypes using single-sample Gene Set Enrichment Analysis (ssGSEA). Genomic alteration analysis of LLPS subtypes Somatic mutation profiles in mutation annotation format (MAF) were obtained from the TCGA database. The R package “Maftools” was used to systematically analyze and visualize mutation profiles and frequencies across different LLPS subtypes. Tumor mutation burden (TMB) was calculated as mutations per megabase (mut/Mb). Copy number alteration (CNA) data for patients with PDAC were also retrieved from the TCGA database, and significant amplifications or deletions across the genome were identified using GISTIC2.0 [26]. Characterization of the immune landscape of LLPS subtypes The ESTIMATE algorithm, implemented in the “estimate” R package, was used to calculate immune scores, stromal scores, ESTIMATE scores, and tumor purity. Furthermore, the ssGSEA method was employed to quantify the enrichment scores of 29 immune signatures [27]. The Tumor Immune Dysfunction and Exclusion (TIDE) algorithm (http://tide.dfci.harvard.edu/) was used to assess the potential responses to immune checkpoint inhibitor (ICI) therapy. The cancer immunity cycle was evaluated using Gene Set Variation Analysis (GSVA). Additionally, we analyzed the expression profiles of patients with PDAC using Immunophenoscore (IPS) data obtained from The Cancer Immunome Atlas (TCIA, https://tcia.at/home). Patients with lower TIDE scores or higher IPS values tend to exhibit stronger responses to ICI therapy in multiple cohorts. Unsupervised subclass mapping (https://cloud.genepattern.org/gp) was employed to predict ICI therapy responses by comparing gene expression patterns between PDAC patients and a reference cohort of patients treated with anti-PD-1 and anti-CTLA-4 therapy [28]. To further explore the association between our LLPS-related signature and immunotherapy outcomes, the IMvigor210 cohort was used. We calculated risk scores based on this cohort and evaluated their influence on prognosis and treatment response. Development of an LLPS subtypes-related prognostic signature To identify hub genes associated with different LLPS subtypes, we performed differential expression analysis comparing LS1 vs. LS2, LS1 vs. LS3, and LS2 vs. LS3. DEGs meeting the criteria of P < 0.01 and |log₂FC| > 2 were subjected to a robust feature selection process. First, a bootstrapping univariate Cox regression analysis was performed (1000 iterations on 70% of randomly selected samples) to identify genes with significant prognostic value. Genes that were significant in over 90% of the iterations were selected. These robustly identified hub genes were then further analyzed using LASSO-Cox regression to prevent overfitting and select the most critical predictors. The 176 samples from the TCGA database were divided into a training set (n = 123) and a testing set (n = 53) at a 7:3 ratio. An Artificial Neural Network (ANN) model was constructed in the training set using the “nnet” R package to build the LLPS-related model, with the specified genes as inputs and survival outcomes as outputs. The number of nodes in the hidden layer (size) was set to two, with a maximum of 500 iterations (maxit) and a weight decay parameter to prevent overfitting. The risk score for both the training and testing sets was calculated based on the trained ANN model. The effectiveness of the model, named LRRPC (LLPS subtypes-related Risk model in Pancreatic Cancers), was evaluated using the area under the receiver operating characteristic curve (AUC). Cell culture and reagents KPC cell line was derived from the pancreatic tumor tissue of KPC (LSL-Kras; LSL-Trp53;Pdx-1-Cre) mice. Pancreatic cancer cell lines (AsPC-1, BxPC-3, CFPAC-1, PANC-1, SW1990, Capan-1) and the human pancreatic ductal epithelial cell line (HPNE) were purchased from the American Type Culture Collection (ATCC). The PaTu-8988T cell line was obtained from Procell system. Detailed information including species, sex, tissue of origin, official cell line name, and Research Resource Identifier (RRID) can be found in Supplementary Table S3. Cells were cultured in DMEM (BIOAGRIO, Shanghai, China) with 10% fetal bovine serum and 1% penicillin/streptomycin in a humidified incubator at 37 °C with 5% CO₂. All cell lines have been authenticated by STR profiling, tested negative for mycoplasma contamination, and confirmed to be free from cross-contamination with other cell lines. GEPIA-based expression visualization The Gene Expression Profiling Interactive Analysis (GEPIA) web server (http://gepia.cancer-pku.cn/) was used to analyze RNA sequencing expression data from the TCGA project (179 pancreatic adenocarcinoma samples and 4 normal samples) and the GTEx project (167 normal pancreas samples) using its standardized processing pipeline [29]. Quantitative reverse transcription PCR (qRT-PCR) Total RNA was extracted from harvested cells using TRIzol reagent (Invitrogen, USA). cDNA was synthesized using HiScript® III RT SuperMix for qPCR (+ gDNA wiper) (Vazyme, R323-01). qRT-PCR was performed using AceQ® qPCR SYBR Green Master Mix (Vazyme, Q111-02). All primers were synthesized by Sangon Biotech (Shanghai, China) and are listed in Supplementary Table S4. Chromatin immunoprecipitation (ChIP)-qPCR ChIP-qPCR was performed to detect the binding of transcription factors to specific genomic regions. For each immunoprecipitation, 5 × 10⁶ cells were used. Cells were cross-linked with 1% formaldehyde for 10 min, followed by quenching with glycine for 5 min. Cross-linked chromatin was sonicated to generate fragments of 200–1000 bp, with fragment size confirmed by agarose gel electrophoresis. Anti-NRF1 antibody (CST, #46743) or a control IgG antibody was used for immunoprecipitation. After reverse cross-linking and DNA purification, the purified DNA fragments were analyzed by quantitative real-time PCR using specific ChIP–qPCR primers (listed in Supplementary Table S4). For FAM83A, ITGB6 and RHOV, the ChIP–qPCR amplicon located approximately 1 kb upstream of the annotated TSS was defined as the 0 position, and additional amplicons around this site (shown in Supplementary Fig. S5) were used, including flanking regions that served as local negative controls. Flow cytometric analysis of apoptosis To assess apoptosis, floating and adherent cells were collected, washed, resuspended in binding buffer, and double-stained with an Annexin V-FITC/PI Apoptosis Detection Kit (Vazyme, A211-01). After a 15-min incubation at room temperature in the dark, cells were analyzed by flow cytometry within one hour. Cell counting Kit-8 (CCK-8) assay The CCK-8 assay (Dojindo, Japan) was used to assess cell viability. Cells (1 × 10³ per well) were seeded into 96-well plates and incubated for 24 h before drug treatment for the indicated durations. Following treatment, 10 µL of CCK-8 reagent was added to each well, and after a two-hour incubation, absorbance at 450 nm was measured using a BioTek Gen5 system. Cellular thermal shift assay (CETSA) Approximately 1 × 10⁷ harvested cells were resuspended in 1 ml PBS with protease and phosphatase inhibitors. The cell suspension was incubated with the NRF1-targeting compound WRR139 (at its IC₅₀ concentration) or DMSO as a control for 4 h. The suspension was then divided into 10 equal portions, and each was heated at a specific temperature for 3 min. Proteins were extracted by three freeze-thaw cycles using liquid nitrogen. The resulting protein solution was mixed with 5× protein loading buffer, boiled at 100 °C for 15 min, and analyzed by Western blotting. Western blotting Total protein was extracted by lysing cells on ice with RIPA buffer (Solarbio, Beijing, China) containing protease and phosphatase inhibitors. The supernatant was collected after centrifugation at 12,000 rpm for 15 min. Protein concentration was measured using a bicinchoninic acid (BCA) assay (Thermo Fisher Scientific). Samples were prepared with 5× protein loading buffer, boiled, separated on 12.5% SDS-PAGE gels, and transferred to PVDF membranes. After blocking with 5% non-fat milk, membranes were incubated overnight at 4 °C with primary antibodies: Anti-NRF1 (CST, #46743), FAM83A (1:1000, Proteintech, 20618-1-AP), RHOV (1:1000, Proteintech, 26620-1-AP), ITGB6 (1:1000, Proteintech, 28378-1-AP) and Anti-GAPDH (Proteintech, 60004-1-Ig). Membranes were then incubated with the corresponding secondary antibody for 1 h at room temperature before visualization with a chemiluminescence reagent. Immunohistochemistry (IHC) Formalin-fixed, paraffin-embedded (FFPE) pancreatic cancer tissue sections and tissue microarrays (TMAs) (4-µm) were deparaffinized, rehydrated, and subjected to antigen retrieval. Endogenous peroxidase activity was blocked with 3% H₂O₂. Sections were incubated overnight at 4 °C with primary antibodies: NRF1 (1:200, CST, #46743), FAM83A (1:150, Proteintech, 20618-1-AP), RHOV (1:100, Proteintech, 26620-1-AP), and ITGB6 (1:200, Proteintech, 28378-1-AP). This was followed by incubation with an HRP-conjugated secondary antibody and visualization with DAB (AiFANG Biological, AFZD002) and a hematoxylin counterstain. Protein expression was quantified via the H-SCORE [Σ(Pi × i); where i = intensity (0–3) and Pi = percentage of positive cells], assessed by two pathologists blinded to the sample identity. Colony formation assay Single-cell suspensions (1000 cells/well) were plated in 6-well plates. For drug treatment experiments, cells were exposed to WRR139 at the predetermined IC₅₀ concentration. After 14 days of incubation, cells were fixed with methanol, stained with 0.1% crystal violet (Sigma-Aldrich, C0775), and colonies containing > 50 cells were counted. In vivo experiments KPC cells (5 × 10⁶) were injected subcutaneously into the dorsal flank of C57BL/6 mice. When tumors reached approximately 100 mm³, mice were randomized into treatment groups. Tumor size was measured every 2 days with calipers, and tumor volume was calculated using the formula: volume = 0.5 × length × width². The WRR139 (TargetMol, T80771) treatment group received 10 mg/kg administered via intraperitoneal injection twice weekly (biw). At the study endpoint, mice were euthanized, and tumors were excised and weighed. All animal experiments were approved by the institutional ethics committee. Fluorescence recovery after photobleaching (FRAP) FRAP experiments were performed on a Zeiss LSM 900 confocal microscope with live cells maintained at 37 °C and 5% CO₂. A region of interest (ROI) of approximately 1–2 μm in diameter was bleached using a 488 nm laser at 100% power for 1 s. Fluorescence recovery was recorded at a rate of 2 frames per second for 60 s. Fluorescence intensities were corrected for photobleaching using a nearby unbleached region as a reference, and background was subtracted. Recovery curves were generated and fitted to determine the fluorescence recovery halftime (t₁/₂). Results Identification of LLPS regulator-related DEGs and LLPS subtypes In this study, we investigated the role of LLPS regulators in PDAC using a bioinformatics approach. We first compiled a list of 359 LLPS-related genes from published studies (Additional file 1: Table S1). By intersecting these genes with DEGs identified between PDAC and normal samples (Additional file 2: Table S2), we pinpointed 298 LLPS regulator-related DEGs (P < 0.05, |log₂FC| > 1) (Fig. 1A). To assess the clinical significance of these LLPS regulator genes, we applied NMF to stratify patients based on their transcriptomic profiles. Based on cophenetic, dispersion, and silhouette indicators, 3 was determined to be the optimal number of clusters (Fig. 1B). Subsequently, patients from the TCGA-PAAD cohort were categorized into three distinct LLPS regulator subtypes: LS1 (n = 90), LS2 (n = 21), and LS3 (n = 65) (Fig. 1C). The application of t-SNE further demonstrated a robust and clear separation among the three subtypes (Fig. S1A). Identification of LLPS subtypes and their clinicopathological features. (A) Venn diagram showing the 298 LLPS regulator-related differentially expressed genes (DEGs) identified by comparing PDAC tissues with normal tissues. (B) Non-negative matrix factorization (NMF) rank survey identifying three as the optimal number of clusters. (C) Consensus map of NMF clustering for k = 3. (D) Kaplan-Meier survival analysis of patients stratified by the three LLPS subtypes (LS1, LS2, LS3). (E) Distribution of clinical characteristics (stage, grade, age, diabetes, relapse, radiation) among the LLPS subtypes. (F) Reactome pathway enrichment analysis of significantly upregulated DEGs in each LLPS subtype. (G) Heatmap displaying ssGSEA Z-scores of hallmark pathways across LLPS subtypes. (H) Volcano plots illustrating DEGs (P < 0.05, |log₂FC| > 1) for each LLPS subtype compared to the others Analysis of clinical outcomes revealed significant differences in patient prognosis. The Kaplan-Meier survival curve showed that patients in the LS2 subtype had markedly better survival outcomes compared to those in LS1 and LS3 (Fig. 1D). A heatmap illustrated prominent differences in the expression levels of the 298 LLPS-related DEGs across the subtypes (Fig. S1B). In summary, these findings indicate that LLPS-related transcriptional patterns stratify PDAC into distinct subgroups with different outcomes, prompting further investigation into the underlying mechanisms. Clinicopathological features and pathway analysis across LLPS regulator subtypes To understand the biological differences, we conducted a thorough comparison of the clinicopathological and demographic features of the three LLPS subtypes. There were no significant differences in tumor stage, grade, or age among the three subtypes (Fig. 1E, Fig. S2A). However, patients in the LS2 group exhibited a lower prevalence of diabetes and reduced exposure to radiation. Notably, no patient in this group experienced a relapse during follow-up. Pathway enrichment analysis using the Reactome Knowledgebase revealed distinct functional profiles among the three LLPS regulator subtypes (Fig. 1F). The LS2 group showed stronger enrichment in insulin secretion and other glandular pathways, in line with its lower prevalence of diabetes and better survival. To further elucidate the mechanistic associations, we performed ssGSEA using the Hallmark gene sets and organized the signatures into three major modules—metabolic/proliferative, glandular, and immune‑inflamed/suppressed (Fig. 1G). Compared with LS2 and LS3, LS1 exhibited marked enrichment of glycolysis, MYC targets, E2F and G2M checkpoint, DNA repair and unfolded protein response signatures, indicating a metabolically active and highly proliferative phenotype with enhanced stress‑response signaling. At the same time, LS1 showed prominent activation of stromal and EMT‑related pathways, including TGF‑β signaling, epithelial–mesenchymal transition and KRAS signaling, suggesting a desmoplastic, stroma‑driven tumor program that may contribute to its poor prognosis. In contrast, LS2 was characterized by elevated expression of pancreatic β‑cell and other glandular differentiation signatures (e.g., PANCREAS_BETA_CELLS, APICAL_SURFACE), consistent with a more differentiated, endocrine‑like state that may underlie its favorable clinical outcome. LS3 displayed a distinct pattern dominated by immune‑ and inflammation‑related pathways. It was strongly enriched for interferon-α/γ responses, antigen processing and presentation, complement and multiple cytokine/chemokines signaling signatures, together with higher activity of IL6–JAK–STAT3, TNF–NF-κB and angiogenesis pathways (Fig. 1G). Taken together with the immune-cell infiltration and checkpoint expression profiles shown in Figs. 2 and 3, these features delineate LS3 as an immune-inflamed yet immunosuppressed subtype, whereas LS1 appears to be primarily driven by metabolic/proliferative and stromal/EMT programs with only intermediate activation of adaptive immune responses. Genomic landscape of LLPS subtypes. (A) Oncoplot illustrating the mutation profiles of the top mutated genes across the three LLPS subtypes. (B) GISTIC2.0 plot illustrating copy number gains (red) and losses (blue) across LLPS subtypes. (C) Comparison of copy number alteration (CNA) burdens at the focal and arm levels among LLPS subtypes. (D) Comparison of tumor mutation burden (TMB) across LLPS subtypes Immunological features of LLPS subtypes. (A) Comparison of immune scores, stromal scores, ESTIMATE scores, and tumor purity across LLPS subtypes. (B) Differences in the activity of cancer-immunity cycle steps among LLPS subtypes. (C) Heatmap showing the differential infiltration of 29 immune cell types across LLPS subtypes, quantified by ssGSEA. (D) Expression of key immune checkpoint genes across LLPS subtypes. (E) Immunophenoscore (IPS) predicting response to immunotherapy across LLPS subtypes. (F) Tumor Immune Dysfunction and Exclusion (TIDE) scores across LLPS subtypes. (G) Subclass mapping analysis predicting the response to anti-PD-1 therapy in the three LLPS subtypes Genomic alteration profiles in different LLPS regulator-based subtypes To explore the relationship between LLPS regulators and the genomic landscape of PDAC, we investigated genomic variations across the three subgroups. As expected, KRAS and TP53 were the most frequently mutated genes across all subgroups (Fig. 2A). Notably, mutations in ARID1A, a key component of the mSWI/SNF chromatin remodeling complex [30], were more prevalent in LS2. The copy number profiles also revealed distinct patterns of gains and losses across the subtypes (Fig. 2B). There were clear differences in the CNA load between the subtypes. Compared to LS1 and LS3, LS2 exhibited a lower burden of focal deletions but a higher burden of focal amplifications. At the chromosome arm level, no significant differences were observed (Fig. 2C). Additionally, LS1 had a significantly higher TMB compared to LS2 (Fig. 2D). These findings suggest that LS1 is characterized by greater genomic instability, which may contribute to its unfavorable prognosis. TIME and immunotherapeutic responses in different LLPS subtypes Given that LLPS regulators can modulate the TIME (Tumor Immune Microenvironment)and responses to immunotherapy [31, 32], we investigated these patterns among the three subtypes. We quantified TIME components using the ESTIMATE algorithm. The findings indicated that LS3 displayed the highest immune, stromal, and ESTIMATE scores, and consequently the lowest tumor purity, compared to LS1 and LS2 (Fig. 3A). Analysis of immune cell infiltration and the cancer-immunity cycle revealed notable distinctions. LS3 demonstrated higher activity scores for most steps of the cycle (Fig. 3B). Consistently, ssGSEA showed that LS3 had higher levels of most tumor-infiltrating immune cells (Fig. 3C). This indicates an immune-inflamed tumor phenotype in LS3 which, together with the enrichment of immunosuppressive pathways and cell populations described above, defines an immune-inflamed but immunosuppressed TIME. The expression of immune checkpoint molecules is indicative of potential responses to ICI treatment. LS3 exhibited significantly elevated expression of key immune checkpoints, including CD274 (PD-L1), CTLA4, ICOS and IDO1 (Fig. 3D), further supporting an immune-inflamed but immunosuppressed state. Using the IPS to predict response, LS2 individuals had a higher IPS score, suggesting better baseline immunogenicity (Fig. 3E). Conversely, TIDE analysis revealed a significantly lower TIDE score in LS3 compared with LS1, indicating a reduced likelihood of immune evasion and a potentially better response to ICI (Fig. 3F). Subclass mapping analysis further suggested that the LS3 group was more likely to respond to anti-PD-1 therapy (Bonferroni-corrected P = 0.036) (Fig. 3G). Overall, our results suggest that LLPS regulator patterns can stratify patients into distinct immune phenotypes and may help predict responses to immunotherapy in appropriate clinical contexts. Construction of a deep learning model to stratify patients based on LLPS regulator-related hub genes To develop a quantitative prognostic tool, we constructed a deep learning-based signature. Using LASSO regression analysis, we selected the optimal regularization parameter (λ) and identified eight hub genes (CCDC188, RHOV, KRT7, CA12, SERPINB5, ITGB6, MYEOV, and FAM83A) for model development (Fig. 4A-B). The final model, termed LRRPC (LLPS Regulators subtypes related Risk Model in Pancreatic Cancers), was trained using an ANN. Among the selected genes, CCDC188 was a protective factor, while the other seven were risk factors (Fig. 4C). Development and validation of a deep learning-based prognostic signature (LRRPC). (A) LASSO regression analysis showing the coefficient profiles of prognostic genes. (B) Partial likelihood deviance plot identifying the optimal lambda value for selecting hub genes. (C) Forest plot of the eight hub genes and a schematic of the three-layer artificial neural network model. (D-F) Kaplan-Meier survival curves for high- and low-risk groups in the training cohort (D), testing cohort (E), and the entire TCGA cohort (F). (G) ROC curves demonstrating the predictive efficacy of the LRRPC model. (H) Heatmap showing the distribution of risk scores and clinicopathological features. (I) Alluvial diagram illustrating the relationships among risk groups, LLPS subtypes, and survival status. (J) Correlation heatmap between the LRRPC risk score and cancer hallmark pathways. (K) Comparison of immune scores, stromal scores, ESTIMATE scores, and tumor purity between risk groups. (L) TIDE scores and predicted responder fractions between risk groups. (M) Comparison of LRRPC risk scores between immunotherapy response groups in the IMvigor210 cohort. (N) Kaplan-Meier curves for patients stratified by LRRPC risk score in the IMvigor210 cohort Kaplan–Meier analysis showed that patients in the high-risk group, as defined by the LRRPC score, had significantly poorer overall survival in both the training and testing cohorts (Fig. 4D-F). ROC analysis showed that the LRRPC model had good discrimination in the training cohort (AUC = 0.763) and maintained acceptable predictive performance in the testing cohort (AUC = 0.670) (Fig. 4G). In conclusion, the LRRPC model provides an LLPS subtypes-related risk score that can stratify PDAC patients by prognosis and may complement existing clinicopathological factors in survival assessment. Correlation of the LRRPC model with clinical parameters, pathways, and immunotherapy response We next investigated the clinical and biological relevance of the LRRPC risk score. The alluvial diagram showed that almost all patients in the favorable LS2 subtype belonged to the low-risk group, reinforcing the model’s consistency with our initial subtyping (Fig. 4I). The risk score was significantly correlated with specific immune cell populations (Fig. S3A) and key cancer hallmarks such as glycolysis and apoptosis (Fig. 4J). Consistent with an immune-excluded phenotype, the risk score was negatively correlated with immune, stromal, and ESTIMATE scores, and positively correlated with tumor purity (Fig. 4K). In terms of immunotherapy prediction, patients with a high-risk score had higher TIDE scores (indicating a higher likelihood of immune evasion) and a lower predicted probability of responding to ICI therapy (Fig. 4L, Fig. S3B). To further explore these predictions, we applied the LRRPC signature to the IMvigor210 cohort (urothelial cancer patients treated with anti-PD-L1). The non-responder group had a significantly higher risk score (Fig. 4M). Moreover, the low-risk group had a significantly higher survival probability and a greater proportion of responders (CR/PR) (Fig. 4N, Fig. S3C). These findings support the notion that the LRRPC model is associated with ICI response and may have value for predicting immunotherapy benefit, although these results are proof-of-concept in a urothelial carcinoma cohort and will need validation in PDAC-specific immunotherapy datasets. NRF1 modulates hub genes of the LRRPC signature and shows therapeutic potential To explore the regulatory mechanisms of our signature, we focused on three risk-associated hub genes: FAM83A, ITGB6, and RHOV. Their high expression was associated with poor survival in the GEPIA database (Fig. S4A-C). qRT-PCR confirmed their upregulation in PDAC cell lines compared to the normal HPNE cell line (Fig. 5A). Similarly, IHC staining of clinical PDAC samples showed that FAM83A, ITGB6, and RHOV were highly expressed in tumor tissues compared to adjacent normal tissues (Fig. 5B). By intersecting predictions from the GeneCards database with expression data, we identified the transcription factor NRF1 as a potential upstream regulator [33]. NRF1 was highly expressed in PDAC tumor tissue compared to normal tissue (Fig. 5C, Fig. S4D), which we further confirmed by IHC and H-score analysis (Fig. 5D-E). Experimental validation of NRF1 as a transcriptional regulator of FAM83A, ITGB6, and RHOV. (A) qRT-PCR analysis showing mRNA levels of FAM83A, ITGB6, and RHOV in six PDAC cell lines relative to the HPNE cell line (n = 3). (B) Comparison of H-scores from IHC staining of FAM83A, ITGB6, and RHOV in PDAC tumor tissues versus adjacent normal tissues (n = 50). (C) Schematic showing NRF1 as a predicted common transcription factor for FAM83A, RHOV, and ITGB6. (D) Comparison of NRF1 H-scores in tumor versus adjacent normal tissues (n = 50). (E) Representative IHC images of NRF1. (F) qRT-PCR analysis showing altered mRNA levels of the three hub genes after NRF1 overexpression or knockdown (n = 4). (G) Western blot analysis showing corresponding changes in protein expression. (H) ChIP-qPCR analysis confirming NRF1 binding to the promoter regions of FAM83A, RHOV, and ITGB6 (n = 4) To validate this regulatory link, we showed that overexpression or knockdown of NRF1 altered the mRNA and protein levels of the three hub genes (Fig. 5F-G). Critically, ChIP-qPCR experiments confirmed that NRF1 directly binds to the promoter regions of ITGB6, RHOV, and FAM83A (Fig. 5H, Fig. S5A-C). NRF1 promotes the proliferation of pancreatic cancer cells To explore the functional role of NRF1, we performed a series of in vitro and in vivo experiments. Overexpression of NRF1 promoted PDAC cell proliferation, while NRF1 knockdown inhibited it, as measured by CCK-8 and colony formation assays (Fig. 6A-C, Fig. S6A). In a subcutaneous mouse model, tumors formed from KPC cells overexpressing NRF1 grew significantly faster and were larger and heavier at the endpoint compared to the control group (Fig. 6D-G). Real-time cell confluence assays further validated these pro-proliferative effects (Fig. 6H). NRF1 promotes pancreatic cancer proliferation in vitro and in vivo. (A) CCK-8 assay showing the effect of NRF1 overexpression and knockdown on cell viability (n = 4). (B, C) Colony formation assays showing the effect of NRF1 upregulation (B) and downregulation (C) on clonogenic growth (n = 3). (D) Waterfall plot showing tumor volume changes in a KPC cell-derived xenograft (CDX) model with NRF1 overexpression (n = 6). (E, F) Representative images (E) and weights (F) of tumors at the study endpoint (n = 6). (G) Tumor growth curves for the subcutaneous model (n = 6). (H) Real-time cell confluence assay showing the effect of NRF1 modulation on PaTu-8988T cell proliferation (n = 3) Translational potential of targeting NRF1 To explore the therapeutic potential of targeting NRF1, we used WRR139, a known NRF1 inhibitor. We first determined the IC50 of WRR139 in PDAC cell lines (Fig. 7A, Fig. S7A). Treatment with WRR139 induced apoptosis in a dose-dependent manner, as confirmed by flow cytometry (Fig. 7B-C, Fig. S7B), and inhibited cell growth in colony formation assays (Fig. 7D, Fig. S7C). In vivo, treatment with WRR139 significantly impeded the growth of KPC tumors (Fig. 7E-G). IHC analysis of treated tumors showed an upregulation of cleaved PARP, a marker of apoptosis (Fig. 7H-I). WRR139 also inhibited the growth of seven pancreatic cancer patient‑derived organoids (PDOs), and higher NRF1 mRNA expression in individual PDO lines was associated with greater WRR139 sensitivity (Fig. 7J–K). In PDAC cell line models, manipulation of NRF1 levels also altered WRR139 sensitivity. Across multiple pancreatic cancer cell lines, NRF1 knockdown shifted the WRR139 dose–response curves to the right and increased the IC50 values, whereas NRF1 overexpression consistently reduced the IC50 (Fig. 7M). Together with the PDAC-PDO data showing that higher NRF1 mRNA expression correlates with greater WRR139-induced growth inhibition, these results support NRF1 as a key functional determinant of WRR139 response. The NRF1 inhibitor WRR139 suppresses pancreatic cancer growth. (A) Dose-response curves of WRR139 in KPC cells (n = 5). (B, C) Flow cytometry analysis (B) and quantification (C) of apoptosis in KPC cells treated with WRR139 (n = 3). (D) Representative images of colony formation assays in cells treated with WRR139. (E) Tumor growth curves of KPC xenografts treated with vehicle, sh-NRF1, or WRR139 (n = 5). (F) Waterfall plot of tumor volume changes across treatment groups. (G) Tumor weights at the study endpoint (n = 5). (H) Western blot analysis of cleaved PARP in tumor tissues. (I) Representative tumor images from each treatment group. (J, K) Representative images (J) and response curves (K) of PDAC organoids treated with WRR139 (n = 5). (L) Response curves to WRR139 in PaTu-8988t cells with NC or NRF1 overexpression (n = 5). (M) Summary of IC50 values in PDAC cell lines with genetic regulation of NRF1. (N, O) CETSA analysis showing decreased thermal stability of NRF1 after WRR139 treatment (western blot image [N] and quantification curve [O]). (P, Q) ChIP-qPCR analysis confirming that pharmacological and genetic regulation of NRF1 affects its binding to the promoter regions of FAM83A, ITGB6, and RHOV (n = 4) Mechanistically, a CETSA experiment showed that WRR139 treatment decreased the thermal stability of NRF1, while the stability of the loading control GAPDH was unaffected (Fig. 7N-O), supporting direct binding of WRR139 to NRF1. At the same time, WRR139 did not significantly alter total NRF1 protein levels (Fig. S8A-B). Furthermore, WRR139 treatment or knock down of NRF1 reduced the transcription of the target genes FAM83A, RHOV, and ITGB6 (Fig. 7P-Q, Fig. S9A-B), suggesting that it inhibits NRF1’s transcriptional activity. These results highlight targeting NRF1 as a promising therapeutic strategy for PDAC. WRR139 enhances intratumoral CD8⁺ T-cell infiltration in KPC tumors To assess whether WRR139 treatment altered the tumor immune microenvironment in vivo, we performed immunohistochemical staining for CD8, FOXP3 and F4/80 on KPC subcutaneous xenografts from vehicle- and WRR139-treated mice. WRR139-treated tumors showed a marked increase in intratumoral CD8⁺ T cells compared with vehicle controls (Fig. 8A–B), whereas the densities of FOXP3⁺ regulatory T cells and F4/80⁺ macrophages remained high in both groups and were not significantly changed by treatment (Fig. 8B, Fig. S10A-B). These data suggest that pharmacologic inhibition of the NRF1 axis can enhance CD8⁺ T-cell infiltration into KPC tumors, while the myeloid-rich, immunosuppressive microenvironment characteristic of this model is largely preserved. WRR139 treatment induces CD8 + T cell inflammation, and NRF1 undergoes LLPS. (A) Representative HE and CD8 + IHC images of vehicle- and WRR139-treated KPC CDX tumors. (B) Quantification of CD8+, FOXP3+, and F4/80 + cells in vehicle- and WRR139-treated KPC CDX tumors. (C) FRAP experiment showing fluorescence recovery of full-length NRF1. (D) Quantification of FRAP data from (C). (E) Prediction of intrinsically disordered regions (IDRs) in NRF1 using PONDR. (F) Representative images showing condensates in cell lines overexpressing full-length or truncated NRF1. (G, H) Percentage of cells with puncta (G) and puncta numbers (H) in cell lines overexpressing full-length or truncated NRF1. (I) qRT-PCR analysis of ITGB6, RHOV, and FAM83A mRNA levels in cell lines with NRF1 knockdown and/or overexpression of full-length or truncated NRF1. (J) CCK-8 assay showing effects of NRF1 knockdown and/or overexpression of full-length or truncated NRF1 on cell proliferation. (K) Dose-response curves of WRR139 in cell lines with NRF1 knockdown and/or overexpression of full-length or truncated NRF1 NRF1 itself undergoes LLPS and its C-terminal IDR contributes to function Intriguingly, NRF1, the regulator of our hub genes, was predicted to contain intrinsically disordered regions (IDRs). To test whether NRF1 can form phase-separated condensates, we performed FRAP experiments in HEK-293T cells, where full-length NRF1 readily formed nuclear droplets that showed rapid fluorescence recovery after photobleaching (Fig. 8C-D). PONDR is a widely used tool for predicting IDRs [10], and VL-XT analysis indicated that NRF1 contains two IDRs at residues 31–119 and 357–492, corresponding to the N-terminal and C-terminal IDRs, respectively (Fig. 8E). To directly test whether the predicted IDRs are required for NRF1 condensate formation, we generated a series of truncation mutants deleting either the N-terminal or C-terminal IDR. NRF1 with the N-terminal IDR deleted (Δ31–119) behaved similarly to the wild-type protein, whereas the construct lacking the C-terminal IDR (Δ357–492, hereafter referred to as NRF1-ΔIDR) failed to form typical NRF1 nuclear droplets/condensates under the same conditions (Fig. 8F-H). These results indicate that the 357–492 aa segment is a key IDR that mediates the condensate-forming ability of NRF1. To assess the functional relevance of this C-terminal IDR, we performed rescue experiments in NRF1 knockdown Patu8988T cells using NRF1 WT and NRF1 ΔIDR. Re-expression of NRF1 WT markedly restored the mRNA levels of FAM83A, RHOV and ITGB6, whereas NRF1 ΔIDR provided only minimal rescue compared with the shNRF1 + empty vector group (Fig. 8I). Consistently, NRF1 WT substantially rescued the proliferation defect and enhanced sensitivity to the NRF1 targeting compound WRR139, while NRF1 ΔIDR only weakly affected cell growth and WRR139 response (Fig. 8J–K). These data suggest that the C-terminal IDR, which is required for NRF1 condensate formation, facilitates NRF1 dependent target gene activation, cell proliferation and drug response, although additional LLPS independent mechanisms cannot be excluded. Discussion Liquid-liquid phase separation (LLPS) is a fundamental biophysical process that spatiotemporally organizes biomolecules, leading to the formation of membraneless condensates that are dynamically regulated by cellular cues [9].LLPS regulators are crucial for probing the biology of these organelles and hold significant therapeutic potential for aggregation-related diseases [34]. As both LLPS regulators and biomolecular condensates influence multiple aspects of cell fate, there is growing interest in the clinical application of LLPS in cancer [35–37]. A comprehensive study of LLPS regulators in pancreatic ductal adenocarcinoma (PDAC) holds promise for stratifying patients for prognosis and immunotherapy. Our study demonstrates that LLPS regulator-based biomarkers can effectively stratify patients into subgroups with distinct survival rates and predicted responses to immunotherapy. We further characterized these subgroups from clinical, genomic, and pathway perspectives. We developed LRRPC, an ANN model based on LLPS regulator subtype-related genes, to quantitatively stratify patients. Finally, in vitro and in vivo experiments validated that targeting NRF1, a key transcriptional regulator of our signature, is a promising strategy to impede PDAC progression. While patients in the LS1 and LS3 groups exhibited worse survival rates than those in the LS2 group, their prognoses were not statistically different from each other. Nevertheless, these two subtypes showed clearly distinct biological and immune features. LS1 was characterized by strong activation of metabolic and proliferative programs (e.g., glycolysis, MYC and E2F targets, DNA repair) together with enrichment of stromal and EMT-related pathways such as extracellular matrix organization and TGF-β signaling, suggesting a stroma- and proliferation-driven phenotype. By contrast, LS3 displayed pronounced enrichment of immune- and inflammation-related pathways, including interferon responses, antigen processing and presentation, and multiple cytokine/chemokine axes, as well as higher activity of T-cell and cytotoxic signatures but also increased signatures of Tregs, MDSCs and M2-like macrophages. LLPS can influence both innate and adaptive immunity by promoting the formation of signaling condensates and thereby regulating immune pathways [38]. Together with the elevated expression of multiple immune checkpoints (PD-1, PD-L1, CTLA4, LAG3, TIGIT, etc.), these findings delineate LS3 as an “immune-inflamed but immunosuppressed” subtype, whereas LS1 tumors appear to be primarily driven by metabolic/proliferative and stromal/EMT programs with only intermediate adaptive immune engagement. Improving the efficacy of immunotherapy in PDAC is a major clinical challenge [39–42], and this LLPS-based stratification provides a framework to link distinct LLPS regulator patterns to specific tumor microenvironment states. Compared to LS1, LS3 displayed higher expression of the immune checkpoint CD274 (PD-L1), in line with bioinformatic predictions of a relatively better response to immune checkpoint inhibitor (ICI) therapy. Other robust biomarkers, such as IPS and TIDE scores, also supported the conclusion that LS3 represents an “immune-inflamed but immunosuppressed” subtype with reduced likelihood of primary immune escape under checkpoint blockade. However, given that PDAC as a whole remains largely refractory to current ICIs, these predictions should be regarded as hypothesis-generating and will require validation in transcriptome-profiled PDAC immunotherapy cohorts. The LS2 group was distinguished by its notably improved survival outcomes. Considering the low incidence of diabetes and high enrichment of insulin secretion pathways and pancreas β‑cell markers, LS2 can be characterized as a “favorable prognosis, glandular/enriched and low‑diabetes comorbidity” subtype. LLPS in biomacromolecules can drive metabolic disorders, and its role in the diabetes–PDAC link may involve islet hormones promoting carcinogenesis [43, 44] or metabolic drugs such as metformin suppressing oncogenes like HNF4G to improve PDAC outcomes. The conflicting clinical evidence underscores the need for rigorous trials. Our findings for the LS2 group suggest a potential interplay among LLPS, diabetes, and pancreatic cancer outcomes that merits further mechanistic investigation. Multiple transcriptomic classification schemes have been proposed for PDAC, including “classical/glandular” versus “basal-like/squamous” or “quasi-mesenchymal” subtypes, as well as “immunogenic” or “immune-rich” categories based on immune infiltration patterns [45–48]. Within this context, LS1, with its strong activation of MYC, E2F and EMT/stromal programs, resembles previously described basal-like or quasi-mesenchymal states; LS2, characterized by glandular differentiation, insulin-secretion pathways and favorable outcomes, overlaps with classical/glandular phenotypes; and LS3, defined by an immune-inflamed yet immunosuppressed microenvironment and higher expression of multiple immune checkpoints, shares features with immunogenic or immune-rich subtypes. Rather than replacing these established frameworks, our LLPS-based stratification provides a complementary perspective centered on condensate-related regulators, linking LLPS-associated transcriptional programs to tumor microenvironment states and potential therapeutic vulnerabilities. Beyond subtype definition, numerous prognostic or predictive models for PDAC have been proposed based on histology, transcriptomics or specific signaling pathways [49], including hypoxia-related signatures [50], exosome-based liquid biopsy markers [51] and DNA-methylation-driven classifiers [52]. Our LRRPC model adds an LLPS-centered dimension to this landscape by integrating clinicopathological features, cancer-related biomarkers, genomic profiles and immunotherapy-response surrogates into a single risk score. The LRRPC model demonstrated strong correlations with immune signatures and cancer hallmarks, showing high predictive efficacy for both prognosis and immunotherapy response. The eight hub genes included CCDC188, RHOV, KRT7, MYEOV, SERPINB5, ITGB6, CA12, and FAM83A. Three of these genes (FAM83A, RHOV, and ITGB6) were further validated, confirming their elevated expression in tumor tissues and their association with unfavorable outcomes. RHOV belongs to the RAS-related small GTPase superfamily and is involved in apoptosis, cell differentiation, and migration [53, 54]. Although RHOV is an unfavorable prognostic marker in PDAC, its precise role in tumorigenesis was previously unclear. Integrin Subunit Beta 6 (ITGB6) has been linked to apoptosis, and its upregulation correlates with advanced histologic grade and poorer prognosis in PDAC [55]. Additionally, Family with sequence similarity 83 member A (FAM83A) is involved in cell proliferation and the EGFR signaling pathway. Inhibition of FAM83A has been shown to suppress pancreatic tumorigenesis, suggesting its potential as a therapeutic target [56]. FAM83A has also been reported to promote tumors through EMT and immune infiltration [57, 58]. To validate that inhibiting these risk factors could lead to favorable outcomes, we identified NRF1 as a key transcriptional regulator of FAM83A, RHOV and ITGB6. NRF1 (Nuclear respiratory factor 1) has been widely recognized as a central regulator of mitochondrial biogenesis and respiratory chain gene expression. In normal tissues, NRF1 controls the transcription of nuclear encoded mitochondrial genes and oxidative phosphorylation programs, thereby supporting cellular energy homeostasis [59–61]. In the cancer context, NRF1 overexpression has been shown to increase mitochondrial biogenesis and restrain TGF β driven epithelial–mesenchymal transition (EMT) [62], thereby suppressing proliferation and migration, suggesting that NRF1 can modulate tumor behavior via mitochondrial and signaling pathways in a context dependent manner. Consistent with this notion, our ChIP-qPCR, qRT-PCR and Western blot experiments showed that NRF1 directly regulates the expression of FAM83A, RHOV and ITGB6 by binding to their promoter regions, and these genes belong to the high-risk arm of our LRRPC signature. Given that these are risk factors, we hypothesized that NRF1 itself would promote PDAC proliferation, which was confirmed through in vivo and in vitro experiments showing that NRF1 overexpression enhanced, whereas NRF1 knockdown or pharmacological targeting reduced, pancreatic cancer cell growth. For translational validation, we investigated whether targeting NRF1 could attenuate PDAC growth. WRR139, a small-molecule NRF1-targeting compound, decreased the thermal stability of NRF1 in a CETSA assay and exhibited a cytotoxic effect on PDAC cells. Both pharmacological inhibition with WRR139 and genetic knockdown of NRF1 significantly impeded the growth of pancreatic cancer in vivo and in vitro. Interestingly, NRF1 contains IDRs, suggesting its potential to undergo LLPS, and our FRAP experiments confirmed that NRF1 can form dynamic nuclear condensates, whereas a C-terminal IDR-deleted mutant (NRF1-ΔIDR, Δ357–492) failed to form such droplets. Functionally, NRF1-ΔIDR retained the CNC–bZIP DNA-binding domain but showed a markedly reduced ability to activate LRRPC hub genes, to promote cell proliferation and to modulate WRR139 sensitivity compared with full-length NRF1. These findings support a model in which the C-terminal LLPS-prone IDR facilitates NRF1-dependent transcription and pro-survival functions in PDAC. At the same time, deletion of an entire IDR is a relatively coarse perturbation that may also influence other properties of NRF1 (such as conformation or cofactor interactions), so we refrain from concluding that LLPS per se is the only mechanism required. This limitation and the need for more refined LLPS-defective mutants are important directions for future work. This study has several limitations. First, the LLPS regulator–based subtypes and the LRRPC signature were developed through retrospective analyses of publicly available cohorts with a limited sample size, and prospective, multicenter validation will be required before clinical implementation. Second, due to data availability constraints, the predictive capability of LRRPC for immunotherapy was evaluated in the IMvigor210 urothelial carcinoma cohort; this should be regarded strictly as a proof-of-concept application to an independent ICI-treated dataset rather than as prognostic validation in PDAC, and its performance needs to be confirmed in PDAC-specific immunotherapy cohorts once such data become available. Finally, although we experimentally validated a key regulatory axis involving NRF1 and three LRRPC hub genes, additional work using refined LLPS-defective mutants, combination treatments and diverse in vivo models will be required to fully elucidate the underlying mechanisms and to translate these findings into clinical strategies. Conclusion In this study, we categorized patients with pancreatic ductal adenocarcinoma into three distinct LLPS regulator‑based subtypes, each with unique clinical, genomic and immunological characteristics. LS1 was defined by metabolic/proliferative and stromal/EMT programs, LS2 by glandular differentiation and favorable outcomes, and LS3 by an immune‑inflamed yet immunosuppressed tumor microenvironment. We developed a prognostic signature, LRRPC, to quantitatively assess patient risk and to provide a basis for personalized treatment decisions. Our experimental validation confirmed that the transcription factor NRF1 directly regulates three LRRPC‑related risk factors—FAM83A, RHOV and ITGB6—and that targeting NRF1 with the NRF1‑targeting compound WRR139 shows anti‑tumor activity in vitro and in vivo. We further discovered that NRF1 itself undergoes LLPS and that its C‑terminal intrinsically disordered region is required for condensate formation and full transcriptional activity, providing a potential mechanistic link between LLPS condensates and transcriptional control of oncogenic programs. Taken together, these findings offer an LLPS‑centered framework for individualized prognostic prediction and generate testable hypotheses for improving immunotherapy strategies in patients with PDAC. Translational impact statement This study identifies liquid-liquid phase separation (LLPS) regulators as key factors in pancreatic cancer prognosis and immune response, offering novel classification and biomarkers to guide personalized treatment and immunotherapy strategies, with NRF1 emerging as a potential therapeutic target. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1: Fig. S1. Additional characterization of LLPS subtypes. (A) t-SNE plot visualizing the distinct clustering of LLPS subtypes based on the expression profiles of the 298 LLPS-related DEGs. (B) Heatmap showing the expression levels of the 298 prognostic LLPS-related DEGs across the three LLPS subtypes. Fig. S2. Additional clinical characteristics of LLPS subtypes. (A) Comparison of gender, alcohol consumption, tumor site, and TNM staging across the three LLPS subtypes. : Fig. S3. Additional immune and treatment response analyses of the LRRPC model. (A) Correlation between the LRRPC risk score and 29 immune signatures. (B) Proportion of predicted immunotherapy responders and non-responders in the high- and low-risk groups. (C) Distribution of observed immunotherapy responses (CR/PR, SD, PD) in the IMvigor210 cohort stratified by risk group. Fig. S4. Bioinformatic analysis of LRRPC hub genes and NRF1. Fig. S5. Location of primers used for ChIP-PCR. (A–C) Predicted NRF1 binding regions and negative control regions at the FAM83A (A), RHOV (B), and ITGB6 (C) promoters. Fig. S6. Quantitative analysis of colony formation assay. (A) Quantification of colony numbers showing the effect of NRF1 modulation on PDAC cell clonogenicity (n=3). Fig. S7. Dose-dependent effects of WRR139 on pancreatic cancer cells. (A) Dose-dependent decrease in viability of PDAC cells after WRR139 treatment (n=5). (B) Quantification of flow cytometry data showing increased apoptosis with WRR139 treatment (n=3). (C) Quantification of colony formation assays showing the inhibitory effect of WRR139 (n=3). Fig. S8. Molecular effects of WRR139 treatment. (A) Western blot showing NRF1 protein levels are unaffected by WRR139. (B) qRT-PCR analysis showing WRR139 reduces the mRNA levels of RHOV, ITGB6, and FAM83A (n=4). Fig. S9. Genetic and pharmacological regulation of NRF1 affects its binding to target regions. (A, B) ChIP-PCR data including IgG controls related to Fig. 7 (P, Q). Fig. S10. WRR139 does not affect F4/80+ or FOXP3+ cell numbers. (A) Representative images of F4/80 and FOXP3 IHC in KPC CDX tumors. Table S1. List of liquid-liquid phase separation related genes. Table S2. List of tumor development genes. Table S3. Detailed information of cell lines. Table S4. The primers sequence used for Chip-PCR and qRT-PCR. Table S5. Predictive information for upstream transcription factors. Acknowledgements We are grateful to Dr. Baiyong Shen for his valuable support and insightful suggestions on project design. We also thank Dr. Shuyu Zhai, Dr. Yizhi Cao, and Dr. Dongjie Chen for their support in specific areas of this study. Abbreviations Alzheimer’s Disease Artificial Neural Network Area Under the Curve Carbonic Anhydrase 12 Cellular Thermal Shift Assay Chromatin Immunoprecipitation Copy Number Alteration Differentially Expressed Genes Epithelial-Mesenchymal Transition Family With Sequence Similarity 83 Member A Fragments Per Kilobase of exon model per Million mapped reads Gene Expression Profiling Interactive Analysis Genotype-Tissue Expression Immune Checkpoint Inhibitor Immunophenoscore Integrin Subunit Beta 6 Kaplan–Meier Liquid–Liquid Phase Separation LLPS Regulators-related Risk Model in Pancreatic Cancers Mutation Annotation Format Molecular Signatures Database Non-negative Matrix Factorization Nuclear Respiratory Factor 1 Pancreatic Cancer RNA sequencing Receiver Operating Characteristic single-sample Gene Set Enrichment Analysis Type 2 Diabetes Mellitus The Cancer Genome Atlas-Pancreatic Adenocarcinoma The Cancer Immunome Atlas Tumor Immune Dysfunction and Exclusion Tumor Immune Microenvironment Tumor Mutation Burden t-distributed Stochastic Neighbor Embedding Author contributions HY, DF and JBJ conceived and directed the study. All authors performed literature collection, analyzed and interpreted data. All authors contributed to the final version of the manuscript and have agreed on the journal to which the article will be submitted, gave final approval of the version to be published, and agree to be accountable for all aspects of the work. Weishen Wang, Yi Zhao, Songyao Jiang and Yiwei Zhou contributed equally to the work. Funding This study was supported partly by grants from National Natural Science Foundation (81972214, 82373348), Shanghai Natural Science Foundation (21140903500), Project Foundation of Taizhou School of Clinical Medicine, Nanjing Medical University (TZKY20220204 and TZKY20220205). Data availability The raw data are available from the authors upon request. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Weishen Wang, Yi Zhao, Songyao Jiang and Yiwei Zhou contributed equally to this work. Contributor Information Jiabin Jin, Email: jjb11501@rjh.com.cn. Da Fu, Email: fuda@shsmu.edu.cn. Hong Yu, Email: yuhong@njmu.edu.cn. References Associated Data Supplementary Materials Supplementary Material 1: Fig. S1. Additional characterization of LLPS subtypes. (A) t-SNE plot visualizing the distinct clustering of LLPS subtypes based on the expression profiles of the 298 LLPS-related DEGs. (B) Heatmap showing the expression levels of the 298 prognostic LLPS-related DEGs across the three LLPS subtypes. Fig. S2. Additional clinical characteristics of LLPS subtypes. (A) Comparison of gender, alcohol consumption, tumor site, and TNM staging across the three LLPS subtypes. : Fig. S3. Additional immune and treatment response analyses of the LRRPC model. (A) Correlation between the LRRPC risk score and 29 immune signatures. (B) Proportion of predicted immunotherapy responders and non-responders in the high- and low-risk groups. (C) Distribution of observed immunotherapy responses (CR/PR, SD, PD) in the IMvigor210 cohort stratified by risk group. Fig. S4. Bioinformatic analysis of LRRPC hub genes and NRF1. Fig. S5. Location of primers used for ChIP-PCR. (A–C) Predicted NRF1 binding regions and negative control regions at the FAM83A (A), RHOV (B), and ITGB6 (C) promoters. Fig. S6. Quantitative analysis of colony formation assay. (A) Quantification of colony numbers showing the effect of NRF1 modulation on PDAC cell clonogenicity (n=3). Fig. S7. Dose-dependent effects of WRR139 on pancreatic cancer cells. (A) Dose-dependent decrease in viability of PDAC cells after WRR139 treatment (n=5). (B) Quantification of flow cytometry data showing increased apoptosis with WRR139 treatment (n=3). (C) Quantification of colony formation assays showing the inhibitory effect of WRR139 (n=3). Fig. S8. Molecular effects of WRR139 treatment. (A) Western blot showing NRF1 protein levels are unaffected by WRR139. (B) qRT-PCR analysis showing WRR139 reduces the mRNA levels of RHOV, ITGB6, and FAM83A (n=4). Fig. S9. Genetic and pharmacological regulation of NRF1 affects its binding to target regions. (A, B) ChIP-PCR data including IgG controls related to Fig. 7 (P, Q). Fig. S10. WRR139 does not affect F4/80+ or FOXP3+ cell numbers. (A) Representative images of F4/80 and FOXP3 IHC in KPC CDX tumors. Table S1. List of liquid-liquid phase separation related genes. Table S2. List of tumor development genes. Table S3. Detailed information of cell lines. Table S4. The primers sequence used for Chip-PCR and qRT-PCR. Table S5. Predictive information for upstream transcription factors. Data Availability Statement The raw data are available from the authors upon request.