The RNA expression data and clinical data of four independent PAAD cohorts(TCGA-PAAD, n = 182 [178 tumors, 4 normal]; GSE102238, n = 100 [50 tumors, 50 normal]; GSE57495, n=63[63 tumors]; ICGC-PACA-CA, n = 262 [262 tumors]) were downloaded from the following public databases: The Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/), the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) and the International Cancer Genomics Consortium (ICGC) (https://dcc.icgc.org/). Cell lines gene expression matrices for primary PAAD were obtained from the Cancer Cell Line Encyclopedia(CCLE) dataset (https://portals.broadinstitute.org/ccle/about). TCGA and GSE102238 cohorts are used to build the signature, and the ICGC and GSE57495 cohorts are used to externally verify the signature. To merge TCGA with the GSE102238 cohort, we proceeded as follows. First, we converted the fragments per kilobase of transcript per million value of the RNA expression data of the TCGA cohort to transcripts per kilobase million value. Then, the intersection genes of TCGA and GSE102238 cohorts were extracted, and the corresponding expression data were matched. Next, we removed normal samples of the TCGA and GSE102238 cohorts, converted the TCGA TPM data to log₂(TPM+1) data, and normalized the data from the GSE102238 cohort. Finally, we merged the processed TCGA with the GSE102238 cohort into a cohort, and used the ComBat function of the “SVA” package to remove batch effects on the merged data(n=227) (22). To reduce the probability of non-cancer death, we excluded patients with PAAD with a survival time of<30 days.

52 PRGs ( Supplementary Table 1 ) were from previous researches (23–25). The TCGA and GSE102238 cohorts were combined into one cohort (n=227), and the expression data of 23 PRGs were obtained after matching with 52 PRGs. k-means clustering algorithm was used to obtain the different molecular subtypes associated with PRGs expression. Every molecular subtype was named pyroptosis-related gene cluster (PRGcluster). Single sample gene set enrichment analysis algorithm (ssGSEA) and Normalized enrichment score (NES) were used to quantify the infiltrating abundance of 23 immune cells in different PRGclusters (26). Principal component analysis (PCA) can determine whether the three PRGclusters can be separated. Gene set variation analysis (GSVA) enrichment analysis was used to discover the underlying biological functions between PRGclusters. |log₂FC|>0.585 and false discovery rate (FDR)<0.05 were considered as the criteria for screening differentially expressed genes (DEGs) among the three PRGclusters (27).

The Gene Ontology(GO) and Kyoto Encyclopedia of Genes and Genomes(KEGG) gene enrichment analysis was used to explore the relevant cytological functions and pathways of DEGs. q-value<0.05 was the cutoff criterion for determining whether a gene is significantly enriched.

Univariate Cox regression analysis helped screen prognosis-related differentially expressed genes (PRDEGs). The TCGA+GSE102238 cohort composed of transcriptome data of PRDEGs and the corresponding survival information were randomly divided into the training cohort and testing cohort. The least absolute shrinkage and selection operator (LASSO) Cox regression analysis was performed on PRDEGs in the training cohort to construct a refined prognostic signature (28). Here is the formula for the risk score:

“Coef(Xi)”, “exp(Xi)”, and “n” represent the gene coefficient, expression level, and number of genes, respectively. The median risk score of the training cohort divided patients in the training cohort and validation cohorts into the low- and high-risk groups. Kaplan-Meier survival curves and time-dependent ROC were used to evaluate the prognostic predictive performance of pyroptosis molecular subtypes-related prognostic signature (PMSRPS). To analyze PMSRPS’ independent prognostic performance, univariate and multivariate Cox regression analyses were performed.

The Cell-type Identification By Estimating Relative Subsets Of RNA Transcripts (CIBERSORT) algorithm was used to quantify immune cell infiltration. Based on the Estimation of STromal and Immune cells in Malignant Tumours using the Expression data (ESTIMATE) algorithm (29), we calculated the stromal score, immune score, ESTIMATE score (sum of Stromal and immune scores) for each sample to quantify tumors Microenvironment. Potential immune checkpoint molecules refer to published papers (30, 31) (n=40, Supplementary Table 2 ). Immunophenoscore (IPS) was downloaded from The Cancer Immune Atlas (TCIA) (https://www.tcia.at/home) to compare responses to immune checkpoint inhibitors(ICIs) across different risk groups. TMB, the number of gene mutations, and the type of gene mutation were derived from the somatic mutation data(n=182) of the TCGA database. The gene copy number (n=185) was downloaded from the specific website(https://xena.ucsc.edu) to observe the copy number variation of PRGs. The RNA stemness scores (RNAss) were downloaded from the Pan-Cancer Atlas Hub (https://pancanatlas.xenahubs.net) (32).

Half maximal inhibitory concentration (IC50) was using for predicting the sensitivity of chemotherapy drugs in the high- and low-risk groups. |log₂FC|>0.585 and FDR<0.05 were the screening criteria for DEGs between high- and low-risk groups. After uploading the up-regulated genome (log₂FC>0) and down-regulated genome (log₂FC<0) to the L1000FWD website (https://maayanlab.cloud/L1000FWD), a table including drug name, similarity score, and q-value among others will be obtained. Furthermore, 2D and 3D images of small molecule compounds were obtained from the PubChem website (https://pubchem.ncbi.nlm.nih.gov/).

The human PAAD cell line CFPAC-1 was derived from Procell (CL-0059, Wuhan, China) and was cultured in a special medium (Procell, CM-0059) in a 37°C, 5% CO2 incubator. CFPAC-1 cells have been identified by short tandem repeat (STR) sequences. The SFTA2-specific siRNA and negative control used in this study were from GenePharma (Shanghai, China). The sequences of the siRNAs are shown in Supplementary Table 3 . CFPAC-1 cells were transfected using Lipofectamine 3000 reagent (Invitrogen, Waltham, USA) according to standard guidelines.

The tumor tissue and paired adjacent tissue of PAAD diagnosed after operation were collected and stored at -80°C. Ethical approval was obtained from the Medical Research Ethics Committee of the Second Affiliated Hospital of Nanchang University. Consistent with a previous study (33), total RNA was extracted from tissues and transfected cells, respectively, and reverse transcription and real-time PCR were performed. The following primers used in the experiments are also shown in Supplementary Table 3 .

This study used R software version 4.1.1 to analyze data. Log-rank test was used for survival analysis. Spearman correlation was used to analyze the correlation of risk scores with immune cell infiltration abundance, TMB, and ssRNA. Wilcoxon rank-sum test and chi-square test were used to judge the significant difference(P<0.05) between two or more groups.

Figure 1 shows the overall process of the study. Because of 52 PRGs obtained in this study were derived from ovarian cancer, esophageal adenocarcinoma and glioma studies (23–25), whether they have an effect on PAAD is uncertain. Therefore, we decided to conduct preliminary verification of these PRGs. Somatic mutation and gene expression data have been reported to help infer cancer progression (34). Beside gene mutation, about 95% of patients with PAAD have other genetic changes such as gene amplification and deletion (35). Therefore, we focused on analyzing the genetic variation and expression of PRGs in PAAD. The waterfall plot ( Figure 2A ) showed the frequency and type of somatic mutation of 52 PRGs in TCGA-PAAD. We found that among 158 samples, there are 90 mutations with a frequency of 56.96%, and TP53 demonstrated the highest mutation frequency, and most samples had the nonsense mutation(a type of point mutation in which a change in a base mutates a codon representing an amino acid into a stop codon, resulting in premature termination of peptide synthesis). The copy number variation (CNV) is an important basis and source of genetic variation (36). Figure 2B revealed that CNV changes occurred in all 52 PRGs. Of these, 27 PRGs(51.92%) were dominated by copy loss, 18 PRGs(34.62%) were dominated by copy gain, and the remaining PRGs had the same copy loss and copy gain frequency. The CNV of these genes on the chromosome is shown in Figure 2C . Next, we compared the expression of 52 PRGs in PAAD and normal pancreatic tissues by combining TCGA and the Genotype-Tissue Expression project (GTEx)(https://xenabrowser.net/datapages/) database. The results showed that 36 PRGs were significantly overexpressed and only 11 underexpressed in PAAD ( Figure 2D ). We also explored the prognostic value of PRGs matched with survival information and found that the expression of the majority of PRGs (16/23) correlated with the prognosis of patients with PAAD ( Supplementary Figure 1 ). In total, the genetic variation, expression level and prognostic changes presented by 52 PRGs in PAAD indicated that they may play a role in the occurrence and development of PAAD, and it is convincing to use them for further analysis.

Prevalent intratumoral and intertumoral heterogeneity has been revealed an important cause of poor prognosis of PAAD. Therefore, in recent years, increasing evidence supports the subtyping of pancreatic tumors based on their molecular characteristics to improve the accuracy of clinical decision-making on treatment (21). Among them, transcriptome subtyping has been used by more and more people because of its unbiased classification that is robust and reproducible. Therefore, we conducted consensus clustering and typing of 23 PRGs that matched transcriptome data and found that the intergroup correlations were the lowest and the intragroup correlations were the highest when clustering variable k = 3, indicating that the joint cohort (n = 227, merged by TCGA and GSE102238) could be well clustered into three subtypes namely PRGclusters A, B, and C according to the expression of the 23 PRGs ( Figure 3A ). The results of PCA analysis also confirmed the rationality of this classification ( Figure 3B ). The results of the Kaplan-Meier survival analysis showed that the OS of PRGcluster C was significantly better than that of PRGclusters A and B ( Figure 3C ). The results of ssGSEA analysis showed that PRGcluster B had the highest infiltration abundance in most immune cells, which not only included anti-tumor immune cells, but also pro-tumor immunosuppressor cells and immune cells with function of anti-tumor and pro-tumor, followed by PRGcluster A and PRGcluster C (P<0.05, Figure 3D ). Only the infiltrate abundance of Type.17.T.helper. cell and CD56dim.natural. killer(NK) did not differ significantly among the three PRGclusters. Next, we tried to find the correlation between the prognosis of different PRGclusters and immune cell infiltration and found that PRGcluster C had the least infiltration in some immunosuppressive cells that promoted tumors: MDSC, immature B cell, mast cell, plasmacytoid. dendritic cell, regulatory T cell(Tregs) and Type.2.T.helper.cell, which seem to help understand why PRGcluster C has the best prognostic effect. However, PRGcluster B has a lower prognosis than PRGcluster C despite having the highest invasion abundance among most antitumor immune cells. We speculate that this is the result of the checks and balances and interactions between different functional immune cells in PRGclsuter. Together these results suggest a complex association between immune cell infiltration and prognosis of different pyroptosis molecular subtypes of PAAD. The heatmap presents the clinicopathological features of the three PRGclusters ( Figure 3E ). In addition, we performed GSVA analysis on three PRGclusters and found that PRGcluster A was highly expressed in apoptosis, linoleic acid metabolism, and P53 signaling pathway; PRGcluster B was highly expressed in infection, immunity, pyroptosis, and various tumors; and PRGcluster C was up-regulated in olfactory transduction and glycine serine and threonine metabolism ( Figure 3F ). These results indicated that different types of pyroptosis had significantly different effects on the prognosis, immune activity, and biological function of patients with PAAD. To highlight these differences, a differential analysis of the expression levels of three PRGclusters was performed, and a total of 828 DEGs were obtained ( Figure 3G ).

To better understand the potential function of 828 DEGs extracted, we performed GO and KEGG analyses. The GO enrichment results showed that DEGs were significantly enriched in leukocyte mediated immunity and lymphocyte mediated immunity (biological process), immunoglobulin complex, and external side of the plasma membrane (cellular component), as well as antigen binding and immunoglobulin receptor binding (molecular function) ( Figure 4A ). The KEGG enrichment results revealed that DEGs prominently enriched in staphylococcus aureus infection and hematopoietic cell lineage among others ( Figure 4B ). Therefore, DEGs may have biological functions that are relevant to cellular immunity and pyroptosis.

Next, we did a univariate Cox regression analysis for the DEGs combined with survival information and got 276 PRDEGs( Supplementary Table 4 ). Prior to grouping and modelling PRDEGs, it was realized that PMSRPS were established based on pyroptosis molecular subtypes(i.e. PRGclusters), failure to confirm the stability of subtypes inevitably affects the credibility of the signature, prompting us to perform another consensus clustering and typing of PRDEGs that match transcriptional data (37). The optimal cluster number supported the existence of three distinct and robust geneClusters in patients with PAAD ( Figure 3H ). Among these three geneClusters, the significant difference in OS was strikingly consistent with the result of PRGclusters ( Figure 3I ). Also, the infiltration abundance of 23 immune cells was highly in accordance with the differences among the PRGclusters ( Figure 3J ). More interestingly, we found that the majority of the geneClusters’ patients with PAAD came from the corresponding PRGclusters. In the survival analysis, the proportions of PRGclusters A, B and C in geneClusters A, B and C were 76.69%, 67.12% and 91.67%, respectively. In the immunoinfiltration analysis, the proportions of PRGclusters A, B and C in geneClusters A, B and C was 77.54%, 67.53% and 91.67%, respectively. This may provide a reasonable explanation for the high similarity in the survival and immune cell infiltration’ results between two clusters. All in all, these results suggest that pyroptosis molecular subtypes we identified are robust and reliable.

276 PRDEGs carrying transcriptomic data were subsequently matched with survival information to get a TCGA+GSE102238 cohort(n=218). The cohort was equally divided into a training cohort (n = 109) and a testing cohort (n = 109), and a significant difference was not found in proportion of clinical data and database sources between the two cohorts ( Table 1 ). Next, we performed a LASSO Cox regression analysis on these PRDEGs in the training cohort ( Figure 5A ) and obtained a PMSRPS with the best fitting effect. The signature was composed of SFTA2, NCAM1, and SPRR1B. According to the risk coefficients of these three genes, we got a formula: Risk score=expression (SFTA2)×0.118+expression(NCAM1)×(-0.256)+expression(SPRR1B)×0.109. The median risk score of the training cohort, 1.053, divided the cohort of patients with PAAD into the high- and low-risk groups. The Venn diagram showed the relationship between clusters and risk scores, as well as survival status ( Figure 5B ). By comparison, we observed that both PRGcluster A and geneCluster A patients had significantly higher risk scores than other clusters (P<0.001, Figures 5C, D ). In addition, the expression of 16 PRGs was also significantly different between the high- and low-risk groups (P<0.05, Figure 5E ).

The Kaplan-Meier survival curve showed that in the training cohort, the OS of patients with PAAD in the high-risk group was worse than that in the lower-risk group (P=0.001, Figure 6A ). Also, the survival rate was obviously decreased in patients with PAAD with increasing risk scores ( Figure 6A ). Time-dependent ROC curves showed that the area under the curves (AUCs) for predicting 1-, 3-, and 5-year survival were 0.732, 0.684, and 0.606, respectively, in the training cohort ( Figure 6A ). As expected, we observed highly consistent results in the testing and TCGA+GSE102238 cohorts (all P<0.05, Figures 6B, C ). This indicated that PMSRPS had a good predictive performance for the prognosis of patients with PAAD.

Subsequently, we verified the independence, clinical correlation and applicability of PMSRPS in the TCGA+GSE102238 cohorts. The results showed that risk scores could independently predict poor outcomes in patients with PAAD (P< 0.001, Figure 6D ). That is, for all patients with PAAD, the higher the risk score, the worse the prognosis. There was also a significant relationship between risk scores and T stage, and we found that the risk scores of T3-4 were significantly higher than those of T1-2(P=0.029, Figure 6E ), which also seemed to help explain the propensity of high risk scores for poor prognosis. In addition, results of applicability analysis showed that the survival rate was lower in the high-risk group than in the low-risk group when patients with age<=65 years (P= 0.001, Figure 6F ) or >65 years(P= 0.082, Figure 6G ), the gender of female or male (all P< 0.01, Figures 6H, I ), and the tumor T stage of T1-2 or T3-4 (all P< 0.05, Figures 6J, K ). This showed that PMSRPS could accurately predict and distinguish the prognosis of high- and low-risk groups at different ages, different genders, and different T stages. Taken together, the above results suggested that PMSRPS was an independent predictor of poor prognosis in patients with PAAD, and has general applicability.

To confirm the robustness of PMSRPS, we calculated the risk scores of patients with PAAD in the ICGC and GSE57495 cohorts using the same formula, and divided patients into the high- and low-risk groups based on median risk score(1.053) ( Figures 7A, F ). Likewise, we observed that the number of survivors decreased with increasing risk scores ( Figures 7A, F ), and patients in the high-risk group had significantly lower OS than those in the low-risk group (all P< 0.01, Figures 7B, G ). Time-dependent ROC curves showed that AUCs for PMSRPS to predict 1, 3 and 5-year survival rates for patients in ICGC cohort were 0.602, 0.610 and 0.738, respectively, while AUCs for predicting 1, 3 and 5-year survival rates for patients in GSE57495 cohort were all >0.7 ( Figures 7C, H ). Cox regression analysis revealed that risk score were independent prognostic factors in patients with PAAD (all P< 0.01, Figures 7D, E, I , J ). Also, we verified the clinical relevance of PMSRPS in the ICGC cohort and found that the expression of the risk gene SFTA2 was higher in patients with PAAD aged 65 years and younger than in patients older than 65 years ( Figure 7K ). Moreover, the risk gene SPRR1B was highly expressed in the disease state of tumor recurrence/progression ( Figure 7L ).

Furthermore, to highlight the advantages of PMSRPS, we compared the performance of PMSRPS with three PAAD prognostic signatures recently published. The first is a pyroptosis-related risk signature constructed by Bai et al. (38). The second is a DNA-methylation-driven genes based prognostic signature (GPRC5A, SOWAHC, S100A14, ARNTL2) created by Xiao et al(39). The third is the m6A-related RNA signature(AP005233.2, AC092171.3, AC010175.1, CASC8, TP53TG1, SNAI3.AS1, FLRT1, AC022098.1, DCST1.AS1) constructed by Wu et al (40). The results showed that the AUCs of PMSRPS (0.676, 0.631, 0.600) for predicting the 1-, 3-, and 5-year overall survival of patients with PAAD were higher than those of Bai-Sig (0.609, 0.594, 0.542), Xiao-Sig (0.623, 0.614, 0.557) and Wu-Sig (0.667, 0.630, 0.551) ( Figures 7M–O ). This reveals that the predictive performance of PMSRPS is better than some existing prognostic signatures.

The TIME of dynamic evolution during tumor progression is an important histological feature of PAAD, which makes us eager to know whether there are TIME differences between high - and low-risk groups with different survival outcomes. We first analyzed the infiltration of immune cells. Results of the this study suggested that the abundance of Tregs, NK cells activated, Macrophages M0, and Mast cells activated infiltration was distinctly higher in the high-risk group than in the low-risk group (all P< 0.01, Figure 8A ), and with increasing risk scores, their infiltration abundance increased prominently (all P< 0.01, Figure 8B ). However, the abundance of T cells CD8, T cells CD4 memory resting, B cells naïve, Monocytes, and Mast cells resting infiltration in the high-risk group was significantly lower than that in the low-risk group (all P< 0.05, Figure 8A ), and they were less abundant with greater risk scores (all P< 0.01, Figure 8B ). In addition to infiltrating immune cells, infiltrating stromal cells are also vital components of tumor TIME, which together perturb the tumor signal and play an important role in cancer biology (29). Therefore, we used the ESTIMATE method to assess the distribution of these components between the high- and low-risk groups, and found that the stromal score, immune score, and estimated score were significantly lower in the high-risk group than in the low-risk group (all P<0.05, Figure 8C ), which meant PMSRPS could accurately predict and distinguish TIME in the different risk groups, and that the high-risk group had higher tumor purity and more typical immunosuppressive “cold” tumor microenvironment. Furthermore, researchers found the expression of immune checkpoint molecules on a significant percentage of tumor cells, and they have been shown to promote epithelial-mesenchymal transformation, resistance to apoptosis and antitumor drugs, and propensity to spread and metastasize (41). In this study, we observed that about half (19/40) of the immune checkpoint molecules had significant expression differences between the two groups. Among them, a total of 12 genes, including PDCD1, BTLA, and CD28, were significantly highly expressed in the low-risk group, while 7 genes, including TNFRSF14, LGALS9, and CD276, were significantly highly expressed in the high-risk group (all P< 0.05, Figure 8D ). To overcome the immunosuppressive power of checkpoint molecules, a new immunotherapy called ICIs works by blocking these immunosuppressive molecules and reactivating effector T cells to specifically kill tumor cells. Currently, the main ICIs used in clinical practice are CTLA4 inhibitors, PD-1 inhibitors and PD-L1 inhibitors. We evaluated the response of patients with PAAD in different risk groups to ICIs by IPS scores and found that the IPS (PD-1/PD-L1/PD-L2(-) and CTLA-4(-)) scores of patients in the high-risk group were significantly higher in the lower-risk group(P< 0.01, Figure 8E ), rather than IPS-PD1/PDL1/PDL2 blocker score, IPS-CTLA4 blocker score, and IPS-CTLA4 and PD1/PDL1/PDL2 blocker score ( Supplementary Figure 2 ). This suggested that patients with PAAD in the high-risk group might be more suitable for novel ICIs therapy rather than PD1/CTLA4 immunotherapy. Besides PMSRPS, we were surprised to find that the model gene had good predictive performance for the existing immunotherapy cohort via a web server for Comprehensive Analysis on Multi-Omics of Immunotherapy in Pan-cancerimmune checkpoint inhibitors (CAMOIP) (http://www.camoip.net/). The results showed that the highly expressed prognostic risk gene SPRR1B had significantly worse OS (HR: 5.5, P=0.001, Figure 8F ) in the melanoma immunotherapy cohort of Hugo. et al. (42). And the highly expressed prognostic protective gene NCAM1 had significantly better OS (HR: 0.35, P=0.022, Figure 8G ) in the melanoma immunotherapy cohort of Auslander et al. (43). Taken together, the above results manifested that PMSRPS might be a reliable tool for predicting immune activity and immunotherapy response in patients with PAAD.

Although ICIs is considered a breakthrough in cancer treatment, the benefit to patients is limited, so it makes sense to understand the underlying indicators that influence response to ICIs treatment. Cancer stemness was reported to be significantly positively associated with ICIs resistance in cancer (44). TMB predicts clinical response to ICIs in a variety of tumors and is associated with improved survival after receiving ICIs (45). Genomic stability was the third proven predictor of ICIs response (46). Therefore, we explored these indicators in the high- and low-risk groups with different immunotherapy responses. As shown in Figures 9A, B , the TMB and RNAss of the high-risk group were visibly higher than those of the low-risk group (all P< 0.05), and TMB and RNAss had distinctly increased with the increase of the risk score (all P ≤ 0.01, Figures 9C, D ). In addition, the genomic mutation rate in the high-risk group was significantly higher than that in the low-risk group (90.48% [76/84] vs 59.26% [32/54], P<0.001, Figures 9E, F ). Among them, gene mutation rates of the top two genes, KRAS and TP53, increased more significantly from the low- to the high-risk groups (KRAS: 71% [60/84] vs. 30% [16/54], P<0.001; TP53: 67% [56/84] vs 37% [20/54], P=0.001). These results suggest that high TMB, cancer stemness, and genomic instability are characteristic of high-risk patients compared with low-risk patients, which can help evaluate and predict immunotherapy effects in patients with different risk groups from multiple perspectives.

Immunotherapy has important therapeutic value for some malignancies, but in most phase I and II trials, It has failed to show good clinical efficacy in patients with PAAD when used alone, unless combined with chemotherapy (47). This suggests that it is necessary to simultaneously screen for more effective chemotherapy drugs. The results of chemotherapeutic drug sensitivity analysis in this study revealed that the high-risk group was more sensitive to commonly used chemotherapy drugs for PAAD (Paclitaxel, Doxorubicin, Docetaxel among others) than the low-risk group (all P< 0.05, Figure 10 ). Moreover, PMSRPS can predict chemotherapy responses in different risk populations to other cancers (Bicalutamide and Lenalidomide among others). (all P< 0.001, Figure 10 ).

In addition, to expand and search for potential small-molecule drugs for PAAD, we extracted DEGs between the high- and low-risk groups. Next, the up-regulated and down-regulated DEGs were uploaded separately to the L1000FWD website and matched with small molecule therapies. This study presented the 10 most important potential small molecule drugs, followed by similarity scores, q-values, and mechanism of action( Table 2 ). Among them, the top three drugs with negative similarity scores were tamoxifen, ZM-241385, and BRD-A24021119, which were expected to negatively regulate the expression of DEGs. 2D and 3D images of these three drugs could be seen in Supplementary Figure 3 . These potential small-molecule drugs might reverse the high expression of high-risk group genes and guide the development of PAAD-targeted drugs. However, the exact efficacy of these drugs needs to be further confirmed in future prospective studies.

Finally, to further verify the reliability of the PMSRPS, we performed expression and in vitro functional experiments on the SFTA2 gene with the highest correlation coefficient. As shown in Figure 11A , we found that SFTA2 was highly expressed in PAAD tissues compared with the corresponding paracancerous tissues (P<0.01). Further, we selected PAAD CFPAC-1 cell lines with high expression through CCLE database for SFTA2 gene interference ( Figure 11B ). Figure 11C shows that si-SFTA2#1 and si-SFTA2#2 had more obvious inhibitory effects, which was used for further functional experiments. Through CCK-8 experiments, we found that interfering with the expression of SFTA2 inhibited the cell viability of PAAD cells ( Figure 11D ). By EdU staining, we found that compared with the control group, the proliferation rate of the SFTA2-inhibited group was significantly reduced ( Figure 11E ). Through transwell migration experiments, we found that interfering with SFTA2 expression would inhibit the migration of PAAD cells ( Figure 11F ).