On 5 July 2021, mRNA data and clinicopathological profiles of patients with pancreatic cancer were obtained from The Cancer Genome Atlas (TCGA, https://portal.gdc.cancer.gov/). After removal of incomplete information, 182 samples were eventually included in the next data analysis.

The pancreatic cancer tissue samples and clinical information (155 specimens for IHC analysis and 20 specimens for qPCR analysis) originated from the Second Xiangya Hospital of Central South University, and the protocol for the application of the specimens was authorized by the ethics review committee of the institution.

A total of 345 genes in the PI3K/AKT pathway were extracted from the Genome Enrichment Analysis (GSEA, http://www.gsea-msigdb.org/gsea/msigdb/search.jsp). Prognostic and Cox regression analyses using R language yielded 22 PARGs.

ConsensusClusterPlus is a tool for unsupervised class discovery. The pancreatic cancer samples were clustered using the ConsensusClusterPlus package (version 1.56.0). The cumulative distribution function (CDF) index reached an approximate maximum when the number of clusters = 2.

ESTIMATE (Estimation of STromal and Immune cells in MAlignant Tumor tissues using Expression data) is a package for predicting tumor purity and the presence of stromal/immune cells in tumor tissues using gene expression data. The ESTIMATE (version 2.0.0) package was used to assess the immune score, stromal score and tumor purity score for each pancreatic cancer tissue sample.

Single-sample gene set enrichment analysis (ssGSEA), an extension of the GSEA method, calculates enrichment scores for each sample and gene set pair. Quantification of enrichment scores for immune cells and immune function in pancreatic cancer specimens was performed using the ssGSEA method.

The methodology was as described previously (18). Briefly, pancreatic cancer samples from TCGA were randomly divided into training and test groups in a 1:1 ratio using the caret package (version 6.0-91) and further analyzed and validated by Lasso regression analysis and the glmnet R package (version 4.1-3). Lasso is a Cox regression model fit to a single value of lambda by a penalized maximum likelihood method. The model can handle (start, stop) data and stratification as well as sparse design matrices.

The sample was divided into HRS and LRS groups according to the middle of the risk score. Survival and the model were assessed using survival, survminer and timeROC. Kaplan−Meier survival curves and ROC curves for survival were plotted for the training and test groups, respectively. Univariate and multivariate Cox regression analyses confirmed the good independent prognostic value of the model and compared the differences in clinical information across subgroups. Variations in risk scores between subgroups and expression heatmaps for the six PARGs were analyzed with limma (version 3.48.3) and ggpubr (version 0.4.0) software.

pRRophetic uses baseline gene expression levels of cell lines and in vitro drug sensitivity to predict clinical drug response. The sensitivity of the HRS and LRS groups to each chemotherapy drug was assessed using pRRophetic (version 2.0.0) software.

RNA samples were extracted from liquid nitrogen-frozen or fresh human pancreatic cancer using TRIzol reagent (15596026, Invitrogen), followed by reverse transcription of mRNA to cDNA using the HiFiScript cDNA Synthesis Kit (R223-1, Vazyme). Finally, real-time quantitative PCR was performed on a 96-well plate on an ABI Prism 7500 system (Applied BioSystems) using the ChamQ Universal SYBR qPCR Master Mix kit (Q711-02, Vazyme). The 2 ^-ΔΔCt method was used to calculate the relative expression fold change of each gene, and GAPDH was used as an internal control for the comparison of gene expression data. Primer sequences are presented in Supplementary Table 1 .

Two PPP2R3A short hairpin RNA (shRNA) lentiviral vectors and a negative control vector were purchased from GenePharma and transfected into HEK293T cells using lipofectamine 3000 to generate and collect lentiviruses, which were subsequently infected with AsPC-1 and Capan-1 cells and screened using 5 µg/mL puromycin.

Cells were lysed with RIPA buffer, separated by SDS−PAGE, electrotransferred onto PVDF membranes, blocked with 5% skimmed milk, incubated with the indicated primary and secondary antibodies and finally subjected to chemiluminescence detection.

Cells were cultured to logarithmic growth phase, trypsin digested and seeded in 96-well plates at 2000-3000 cells per well for the indicated times. Subsequently, the wells were washed twice with 1× PBS, fixed with 4% PFA for 20 min, dyed with crystal violet 0.1% (C0775, Sigma−Aldrich) for 20 min (20 µl per well), washed with ddH₂O three times and lysed with 10% acetic acid for 10 minutes. Finally, the absorbance was measured at 590 nm.

The procedure was performed essentially as previously described (19). Briefly, cells were treated with trypsin and seeded in 12-well plates (1500 per well with AsPC-1; 500 per well with Capan-1), and the medium was changed every 3 days up to visible colony formation. Subsequently, the cells were washed 2 times with 1 x PBS, fixed in 4% PFA, dyed with 0.1% crystal violet, washed 3 times with 1 x PBS and finally photographed, and the clone number was counted.

After trypsin digestion, cells were seeded in 6-well plates (2×10⁵ per well with AsPC-1; 1×10⁵ per well with Capan-1), cultured until the cells formed a confluent monolayer, scribed with a 200 μl pipette tip, rinsed twice with 1× PBS, incubated in medium containing 1% FBS for the indicated time, and photographed under a microscope. The wound width was measured, and the percentage of slit closure was calculated using ImageJ software.

The procedure was performed essentially as previously described (20). Matrigel was thawed and diluted with serum-free medium (ratio 1:8), followed by the addition of 40 µl of the dilution mixture to the upper chamber (performed on ice). Subsequently, 0.1 ml of 1-1.5 × 10⁵ cell suspension was injected into the upper compartment (1% FBS), and 0.8 ml of culture medium was injected into the lower compartment (10% FBS). After incubation for the indicated time, the chambers were removed, fixed in methanol, stained with 1% crystal violet and photographed with a microscope after the cells were removed from the upper compartment with a cotton swab.

All animal experiments were performed in accordance with China Food and Drug Administration guidelines. Protocols were reviewed and approved by the Department of Laboratory Animals, Central South University. Four- to six-week-old BALB/c nu/nu female mice were purchased from SJA Laboratory Animal Co., Ltd. A total of 1×10⁶ AsPC-1 cells transfected with PPP2R3A shRNA lentivirus (KD) or empty vector (NC) were injected subcutaneously into the right flank of BALB/c nu/nu female mice. Tumor sizes were measured every three days. Twenty-four days later, the mice were sacrificed, and the tumors were dissected and weighed. Tumor volume was calculated using the equation V (mm³) = (a×b²)/2, where a is the largest diameter and b is the smallest diameter.

The procedure was performed essentially as previously described (21). Briefly, primary antibodies against PPP2R3A (ab218165, Abcam) and Ki-67 (#9449, CST) were diluted 1:250 and subsequently added to tissue sections incubated overnight at 4°C. After three washes in PBS, tissue sections were stained using a nonbiotin horseradish peroxidase detection system (ZSGB-BIO), and hematoxylin-stained nuclei and IHC staining were performed by two independent pathologists for scoring. To evaluate PPP2R3A and Ki-67, a semiquantitative scoring criterion was used in which both the staining intensity and positive areas were recorded. The staining index (values 0-12) was calculated from the product of the positive staining intensity (weak, 1; moderate, 2; strong, 3) and the proportion of positive cells of interest (0%, 0; <25%, 1; 26-50%, 2; 51-75%, 3; >76%, 4).

Statistical analyses were performed by R software (version 4.1.3) or GraphPad Prism 8. The chi-squared test was used to analyze the clinicopathological characteristics of patients in the training and test groups, and unpaired two-tailed Student’s t tests were used to compare differences between groups. (* p < 0.05; ** p < 0.01; *** p < 0.001).

To investigate the role of PARGs in pancreatic cancer clustering, we performed univariate Cox regression analysis from 346 PARGs and selected 22 PARGs with significant hazard ratios for clustering ( Figure 1A ). The results demonstrated that pancreatic cancer patients were grouped into two clusters, C1 and C2 ( Figures 1B–D ). Subsequently, survival analysis demonstrated a shorter OS for C2 compared with C1 ( Figure 1E ), suggesting that PARGs are closely associated with pancreatic cancer prognosis.

We further investigated the relationship between clustering and PD-L1 or CTLA-4. The two immune checkpoint proteins were differentially expressed in C1 and C2 ( Figures 2A, B ), with PD-L1 highly expressed in C2 and CTLA-4 highly expressed in C1.

Immune and ESTIMATE scores were higher in C1 than in C2 ( Figures 2C–E ). Furthermore, antitumor immune cell populations such as naive B cells, plasma cells and CD8 T cells were more abundant in C1. In contrast, activated NK cells, regulatory T cells (Tregs), and M0 and M2 macrophages were higher in C2 ( Figure 2F ).

To construct a PARG-based risk model, we performed Lasso regression analysis on 22 PARGs and selected CDK6, GNB3, MET, PPP2R3A, PPP2R1B and TSC1 to construct a prediction model ( Figures 3A, B ). Patients were classified into training and test groups, and analysis of the clinicopathological characteristics demonstrated no significant differences between them ( Table 1 ). Figure 3 presents the risk scores, survival status and expression of the six PARGs in the training group ( Figures 3C, E, G ) and the test group ( Figures 3D, F, H ).

Survival analysis demonstrated significantly lower OS in the HRS group ( Figures 4A, B ). In addition, the area under the ROC curve (AUC) data for 1-year, 3-year and 5-year OS were 0.760, 0.834, and 0.976 (training group) and 0.735, 0.697, and 0.893 (test group), respectively ( Figures 4C, D ). Risk scores were significant predictors of poorer survival relative to age, sex and stage, suggesting that the model based on PARG risk scores may be more valid than pathological agents for predicting patient survival ( Figures 4E–H ). We also analyzed OS in the HRS and LRS groups classified into age, sex and tumor stage subgroups and found that OS was shorter in the HRS group regardless of age, sex (male) and tumor stage (T1-2, T3-4, M0, N0, N1, Stage I-II, G1-2 and G3-4) ( Figures 5A–H , S1A-F ). In conclusion, PARG-based risk models can be applied to older and younger patients, to women and men, and to patients with early and advanced disease.

Comparison of risk scores between different subgroups demonstrated significant differences between M0 and M1, N0 and N1, C1 and C2, and high and low immune scores. Patients with C1, no metastases or high immune scores had lower risk scores and a better prognosis ( Figures 6A–I ). In addition, CDK6, MET and PPP2R3A were highly expressed in the HRS group, whereas GNB3, PPP2R3B and TSC1 exhibited the opposite trend ( Figure 6J ). These discoveries suggest that PARGs are correlated with immunological status and distant metastasis in pancreatic cancer.

To assess the association of risk scores and pancreatic cancer immunity, we calculated the abundance of 22 immune-related cell types and demonstrated that the predominant tumor-infiltrating immune cells in the HRS group were Tregs, activated NK cells and M0 and M2 macrophages ( Figure 7A ). In contrast, naive B cells, plasma cells, activated CD8 T cells and M1 macrophages were more abundant in the LRS group. Therefore, these results suggest a potential association between HRS and immunosuppression.

We further tested the transcript expression of immunosuppressive molecules between the HRS and LRS groups and demonstrated that APC_co_inhibition, CCR and parinflamation were markedly different between them, with APC_co_inhibition being the most pronounced ( Figure 7B ). We then further analyzed the immune molecules included in this group. NECTIN3, CD274 and PDCD1LG2 were hyperexpressed in the HRS group, whereas LAGLS9 was low ( Figure 7C ), indicating that inhibitors targeting NECTIN2, CD272 and PDCD1LG2 were more effective in the HRS group. Moreover, we analyzed the correlation between APC_co_inhibition component molecules and six PARGs and determined that CD274 and NECTIN3 were positively correlated with the HRS molecules CDK6, MET and PPP2R3A, whereas LGALS9 was negatively correlated with GNB3 and TSC1 ( Figure 7D ). To verify the accuracy of the results, we collected specimens from 20 clinical pancreatic cancer patients for qPCR validation, and the results were generally consistent with the above findings ( Figure 7E ), indicating that the data were reliable. Finally, we compared the transcript expression of immunosuppressive biomarkers, including immune checkpoint markers and secreted immunosuppressive factors, between the HRS and LRS groups. The results of the correlation analysis demonstrated that CCL2, CCL4, CD226, CD48 CD80, CD86, HAVCR2, IDO1, IL10, LAG3, PDCD1 (PD-1), TGFB3 and TIGIT were positively correlated with HRS, with the exception of CTLA4, which was negatively correlated ( Figure 7F ). Taken together, the above results suggest that six PARGs are instrumental in regulating the immune status of tumors.

Radiotherapy and chemotherapy are the first-line treatments in the clinical management of pancreatic cancer (22, 23). To further investigate whether risk features correlate with treatment response, we compared OS in patients treated with radiotherapy, gemcitabine or 5-FU and untreated patients in the HRS and LRS groups. Specifically, the relationship between radiation therapy ( Figures 8A–C ) or guitarcitabine chemotherapy ( Figures 8D–F ) and risk score exhibited a similar pattern, with both treatments improving the prognosis of patients with HRS, but there was no significant difference between the treated and nontreated groups for patients with LRS. For 5-FU, there was no significant improvement in OS with this therapy in either the HRS or LRS group ( Figures 8G–I ). The above results demonstrate that the HRS may be a valid marker for pancreatic cancer patients undergoing radiotherapy and gemcitabine treatment. However, these results are inconclusive for 5-FU treatment.

To gain more insight into the association between risk score and drug response, we analyzed sensitivity to therapeutic agents using the pRRophetic package. We found that the HRS group may be resistant to AMPK activators (phenylephrine), Chk inhibitors (AZD7762) and Src inhibitors (bostatinib) that regulate cell growth and differentiation but may possibly be sensitive to PI3K/AKT/mTOR signaling pathway inhibitors (e.g., PIK-93, GSK2126458, CAL-101 and rapamycin) and ATP synthase inhibitors (thapsigargin) ( Figure 8J ). Together, these results suggest that the PARG-based risk model can be used to develop novel treatments for pancreatic cancer.

PPP2R3A belongs to the PP2A regulatory subunit B” family, and its role in pancreatic cancer tumorigenesis is rarely reported (24). Our results demonstrate that PPP2R3A is highly expressed in the HRS group and in C1 ( Figures 3G, H , 6J ), suggesting that this protein may be a tumor-promoting factor. To verify this hypothesis, we used Kaplan−Meier analysis to explore the correlation between PPP2R3A expression and OS; patients with high PPP2R3A expression had a shorter OS ( Figure 9A ).

To further investigate the role of PPP2R3A, we selected two pancreatic cancer cell lines with high metastatic activity, AsPC-1 and Capan-1, and established stable PPP2R3A knockdown cell lines by transfection with PPP2R3A shRNA lentivirus. We then detected the relative proliferation rate and clone formation ability of the cells. After knockdown of PPP2R3A ( Figures 9B, C , top), the proliferation rate of cells decreased ( Figures 9B, C , bottom), the clonal sphere size decreased, and the number of clones decreased ( Figures 9D–F ). Importantly, wound healing assays and transwell results demonstrated that tumor cell migration ( Figures 9G–J ) and invasion ( Figures 9K–M ) were also significantly reduced after PPP2R3A knockdown. Furthermore, knockdown of PPP2R3A in AsPC-1 cells significantly inhibited xenograft tumor growth in mouse models ( Figures 10A–E ). In addition, we analyzed the expression of PPP2R3A in clinical human pancreatic cancer tissue samples ( Figure 10F ) and found that upregulation of PPP2R3A expression was significantly associated with tumor recurrence ( Figure 10G ). Overall, these results suggest that PPP2R3A is a tumor-promoting factor in pancreatic cancer.