Tumor-antigens and immune landscapes identification for prostate adenocarcinoma mRNA vaccine

A total of 733 overexpressed genes in TCGA-PRAD samples were identified (Fig. S1A, B) and their distribution in chromosomes was shown in Fig. 1A. We then identified 10881 genes that potentially encode tumor-specific antigens through calculating the fraction of altered genome and tumor mutational counts (Fig. 1B, C). Ten genes with the highest altered genome fractions and mutation counts were displayed in Fig. S1C and D. The 733 overexpressed genes were then intersected with the 10881 mutated tumor antigen-encoding genes, and 311 genes were identified afterwards (Fig. S1E). Cox regression revealed that 13 genes were significantly associated with OS (Fig. 1D) and 70 genes were significantly associated with disease free interval (DFI) (Fig. 1E). Further intersection analysis indicated eight genes, including KLHL17, CPT1B, IQGAP3, LIME1, YJEFN3, MSH5, CELSR3 and KIAA1529, were correlated with both OS and DFI (Fig. 1F). The Kaplan–Meier survival curves of OS for those eight genes were shown in Fig. 1G, and the higher expression of them was indicative of worse survival. The correlation between them and the infiltration of major APCs, including B cells, macrophages as well as dendritic cells (DCs), was also analyzed. Figure 1H showed a significantly positive correlation of IQGAP3, CELSR3, and KIAA1529 with APCs, whereas negative for CPT1B (Fig. S1F). Taken together, our evidence identified tumor-specific antigens owning potentiality of mRNA vaccine development for PRAD.

To the best of our knowledge, our study is the first to systematically screen suitable tumor antigens for mRNA vaccine development in PRAD. Despite the non-application of these antigens in mRNA vaccine development until the present, some of these antigens have been functionally explored in previous studies. For instance, CPT1B silencing could reduce cell proliferation and invasion in PRAD cell lines and its expression level might be regulated by androgen receptors [8]. IQGAP3 was upregulated in most cancer types and predicted a poor prognosis. It might also participate in the Paris forrestii antitumor effect [9]. Besides, CELSR3 downregulation significantly suppresses PRAD cell proliferation and migration [10]. MSH5 has been reported as a pleiotropic susceptibility locus for several cancers and was identified as a novel candidate gene warranting additional follow-up as a prospective PRAD risk locus [11]. However, KLHL17, KIAA1529, LIME1, and YJEFN3 were not fully elucidated in PRAD or other cancers. Their function in cancers especially PRAD warrants further exploration.

A total of 13426 immunogenic genes were obtained from the MSigDB c7 datasets, 23 of which were associated with predictive survival outcomes through Lasso regression. The PAM algorithm accordingly identified the optimal number of clusters as three based on the training cohort (Fig. 2A). The accumulative curve and delta area of clustering were displayed in Fig. S2A and B. Principal component analysis showed the distribution of TCGA-PRAD individuals in each cluster (Fig. S2C), and heatmap revealed the differential expression of partial immunogenic genes across the three clusters (Fig. S2D). Importantly, the PRAD immune subtype 1 (PIS1) consistently had better survival outcomes compared to the PIS2 and PIS3 in both training cohort (P < 0.0001) and validation cohort (P = 0.041) (Fig. 2B and S2C).

Clinical features of the PRAD immune subtypes were assessed. The predicted response to immunotherapy of the subtypes indicated that PIS2 and PIS3 were more likely to respond to anti-PD-L1 treatment (Fig. 2D). PIS3 had a higher frequency of biochemical recurrence (Fig. 2E) and pathological N1 stage and higher pathological T stage (Fig. S2E, F). Moreover, patients in PIS3 also had a higher risk of receiving radiation therapy (Fig. S2G). This evidence implies that PIS2 and PIS3 are associated with more aggressive clinical features and more suitable for immunotherapy.

The correlation of existing PRAD biomarkers with PRAD immune subtypes was evaluated. The expression of HOXC6, whose higher expression indicated short survival and higher recurrence rate of PRAD [12], was significantly higher in PIS2 and PIS3 than in PIS1 (Fig. S3A). Nevertheless, PDK4 and STAT3, two classical genes whose low expression was associated with worse survival of PRAD [13, 14], were significantly less expressed in PIS2 and PIS3 (Fig. S3B and C). Notably, our PRAD immune subtypes were also compared with previously published pan-cancer immune subtypes. Figure 2F demonstrated that the C3 subtype had a decreasing distribution, and C1 and C2 showed an increasing tendency across PIS1 to PIS3. Interestingly, C3 was claimed to be a positive marker of prognosis but C1 and C2 were negatively associated with survival [15], which was consistent with our findings.

In terms of mutation status, PIS3 and PIS2 presented more frequent CNV (either gain or loss) across the chromosomes (Figs. 2G and S3D). Similarly, it could be seen from the mutation landscape of the immune subtypes that PIS3 had a more frequent mutation of the top 20 most mutated genes (Fig. S3E), and PIS2 and PIS3 had significantly higher tumor mutation counts (Fig. 2H). In addition, tumor mutation burden (TMB) was also found to be significantly heavier in PIS2 (P = 0.0053) and PIS3 (P < 0.0001) than in PIS1 (Fig. 2I). Our results also demonstrated that PIS2 and PIS3 had higher telomeric allelic imbalance (HRD-NtAI), large scale transition (HRD-LST) and loss of heterozygosity (HRD-LOH) and combined homologous recombination deficiency (HRD scores) (Fig. 2J, S3F-H). Moreover, mRNA stemness index (mRNAsi) was also higher in PIS2 and PIS3 compared to that of PIS1 (Fig. 2K).

As shown in Fig. S4A and D, the tendency of stromal score, immune score, and tumor purity was variable across the subtypes in both training and validation cohorts. In the training cohort, PIS3 had richer infiltration of M2 macrophages and memory B cells, but PIS1 had more infiltration of naïve B cells and memory-resting CD4+ T cells (Fig. S4B, C). Consistently, PIS3 still had a higher degree of memory B cell infiltration compared to PIS1 and PIS2 in the validation cohort (Fig. S4E, F). Hence, PIS3 implies having a better performance of presenting tumor antigen during the immune response.

The anticancer immune activity of the three immune subtypes was calculated with the TIP analysis. PIS1 performed better at recruiting CD4+ T cells, Th22 cells, and monocytes, whereas PIS2 and PIS3 were still proved to be better at recruiting B cells (Fig. S5A). These outcomes may explain the better survival of PIS1 and also indicate the suitability of PIS2 and PIS3 for receiving tumor vaccines. As for the immune modulators, a total of 37 ICP genes and 25 ICD genes were analyzed, which revealed that 31 ICP genes and 18 ICD genes were differentially expressed across the immune subtypes in the training cohort from TCGA datasets (Fig. S5B, C). Interestingly, fewer ICPs and ICDs were significantly differentially expressed among three clusters in the GEO cohorts, and the PIS3 cluster showed markedly lower ICP and ICD expression (Fig. S5D, E). These findings indicated that the immunotyping showed a distinct expression pattern of ICPs and ICDs, and these modulators could be utilized as potential markers for treatment with mRNA vaccines.

The gene expression value of each patient across the three PRAD immune subtypes was used to build the immune landscape of PRAD (Fig. 2L), with PIS1, PIS2 and PIS3 were generally distributed in different branches of the tree. Principal component 1 (horizontal axis) was positively correlated with plasma cells and M2 macrophages, but negatively correlated with naïve B cells, resting DCs and M1macrophages. Interestingly, principal component 2 (vertical axis) had a positive correlation with DCs and naïve B cells, but a negative correlation with M0 and M1 macrophages (Fig. 2M). The general distribution of PIS1 can be observed to contrast PIS3. Also, individuals within the same immune subtype of PIS1 and PIS3 showed opposing distribution. Therefore, PIS1 and PIS3 were further divided (Fig. 2N), which turned out that PIS1A had a generally higher enrichment score regarding activated DCs and memory B cells compared to PIS1B (Fig. S6A). Similarly, PIS3A had a higher enrichment score of activated DCs, activated B cells and memory B cells than PIS3B (Fig. S6B). Therefore, tumor antigen may be more effective in PIS1A and PIS3A compared to PIS1B and PIS3B, respectively. Besides, individuals with extreme distribution in the immune landscape (Fig. 2O) were taken into further survival analysis. Group N1 was associated with the worst survival and group N3 had the best survival outcomes (P = 0.0011) (Fig. 2P). The immune subtype-based landscape can potentially designate the precise mRNA vaccine therapeutics for patients with PRAD by identifying immune components of patients with PRAD and predict survival.

WGCNA with a fixed soft threshold of nine was used to construct the immunogenic gene co-expression network of PRAD (Fig. S7A-C). Eventually, 9 co-expression modules were obtained (Fig. S7D, E). Distribution analysis showed that PRAD immune subtypes were differentially distributed in most of the modules (Fig. S7F). The negative prognostic value of the pink (Hazard ratio (HR) 2.30, 95% CI 1.59–3.32), magenta (HR 2.19, 95% CI 1.56–3.08), and purple modules (HR 1.61, 95% CI 1.16–2.24) were presented in Fig. S8A. The biological function of the prognostic modules, including B cell activation, and regulation of adaptive immune response were also displayed in Fig. S8B–D. 62 hub genes in the prognostic modules were identified, and three of them (CDC20, ESPL1, MAPK8IP3) were eventually selected after multivariate Cox regression (Fig. S8E, F). Patients were divided into the high-risk and low-risk groups. KM curve demonstrated that the high-risk group had worse survival (P = 0.0011) and the area under the receiver operating curve (AUC) was 0.852, indicating a good accuracy of the model (Fig. S8G, H). Thus, this risk model based on immunogenic genes co-expression network may work as a novel biomarker for predicting the prognosis.

391 DEGs across PRAD immune subtypes were found and displayed (Fig. S9A, B). The prognostic value of these DEGs was calculated and 93 DEGs were prognostic. Lasso regression reduced the dimension of these DEGs and 21 genes were finally used to construct the risk model (Fig. S9C–F). The risk of each patient was calculated based on the expression value of the 21 genes and their coefficients in Lasso regression (Fig. S9G, H). Patients were categorized into high-risk or low-risk groups, and the high-risk group had worse survival with an AUC of 0.892 (Fig. S10A, B). Consistently, Fig. S10C and D summarized that PIS2 and PIS3 had higher risk scores with more PIS1 were distributed in the low-risk group, which in reverse proved the accuracy of PRAD immune subtype in predicting PRAD prognosis.

Methods and materials used in our study are attached as Supplementary information. All data are freely available from the public databases and the other necessary and reasonable information could be obtained from the corresponding author.