Introduction

Cancer poses a significant threat to human health and imposes a substantial burden on the global public health. It is the leading cause of death worldwide [1, 2]. Prostate cancer (PCA) is one of the most common malignant tumors in men, and its incidence has increased significantly in recent years [3, 4]. Despite progress in its treatment, such as surgery, radiotherapy, and chemotherapy, the disease remains a challenge, particularly in cases of castrate-resistant prostate cancer [5, 6]. Immunotherapy, which has achieved excellent therapeutic effects in various tumors, is a revolutionary breakthrough in tumor treatment [7, 8]. However, immune checkpoint blockade therapy (ICB) has limited efficacy in unselected patients with PCA, and only a small subgroup of patients may be sensitive to ICB [9]. Therefore, determining patient subtypes that can benefit from immunotherapy is an urgent problem to be solved. Several large-scale clinical trials are currently ongoing, such as the phase III clinical trials of pembrolizumab (KEYNOTE-991) and nivolumab plus docetaxel (Checkmate 7DX), to explore the benefits of immunotherapy with/without other conventional treatments for PCA patients [10]. In recent years, numerous biomarkers that can affect ICB efficiency have been discovered, such as PD-L1, microsatellite instability, tumor mutational burden (TMB), and TCR polymorphism [11]. However, they are far from being ideal biomarkers for PCA [10].

The molecular subty** of tumors can be used to guide the precision diagnosis and treatment of tumor patients. Cellular and molecular characteristics of tumors have shown potential in precision therapy of PCA, such as cancer-associated fibroblasts (CAFs) and their signature genes [12, 13]. Serum prostate-specific antigen (PSA) is the most important marker for PCA; however, it has been criticized for its poor specificity [14]. Existing molecular subty** methods for PCA were designed for specific clinical applications. For example, the PAM50 method is used to guide androgen deprivation therapy [18,19]. There is still lacking subty** methods for PCA, which can effectively and systematically characterize patients from various clinical points of view, e.g., diagnosis, prognosis, recurrence risk, metastasis risk, progression risk, and efficacies of different treatments.

Stemness refers to the self-renewal and differentiation potential of cells. In almost all human malignant tumors, there is a rare subset of cancer cells with stem-like properties, called “cancer stem/stem-like cells” (CSCs) [20]. CSCs are considered the origin cells of tumors and play an important role in the recurrence, metastasis, and treatment resistance of many tumors, including PCA [21,22,23]. They also affect the effectiveness of immunotherapy [24]. Therefore, stemness based subty** holds potential in personalized management of PCA patients.

In this study, we utilized single-cell RNA-seq (scRNA-seq), bulk RNA-seq, methylation array, and whole exon sequencing datasets to systematically assess stemness differences among PCA patients. By integrating scRNA-seq and bulk RNA-seq, we developed a stemness-based subty** model consisting of 18 stemness related gene-sets that separated PCA samples into three subtypes with distinctive clinical and molecular characteristics, functional annotations, prognoses, and treatment responses, especially immunotherapy. Furthermore, a subtype predictor including 9 stemness related genes was constructed which showed great performance in tumor diagnosis, predicting ICB and androgen deprivation therapy (ADT) response, metastasis, recurrence, progression, and prognosis.

Results

Stemness scores are closely correlated with clinical and molecular features

Single-cell and bulk RNA-seq analyses show a positive correlation between stemness and PCA malignancy

The workflow of this study is depicted in Additional file 1: Fig. S1 and Methods S1, and all the datasets used in this study are presented in Additional file 1: Table S1. Prostate epithelial cells were extracted from five PCA scRNA-seq datasets [25,26,27,28,29] (Additional file 1: Fig. S2) and the cytoTRACE algorithm [30] was used to calculate their stemness levels. Consistently, the results showed significantly higher cytoTRACE scores in the malignant epithelium than in the para-cancerous or benign prostate epithelium (Fig. 1a, Additional file 1: Figs. S3a–c). Moreover, among malignant epithelial cells, high-grade PCA (Gleason score [GS] > 7) showed significantly higher stemness scores than low-grade PCA (GS ≤ 7) (Fig. 1b, Additional file 1: Fig. S3d).

Fig. 1
figure 1

Correlation of stemness levels with clinical, pathological, and molecular features in patients with prostate cancer (PCA). a t-distributed stochastic neighbor embedding (t-SNE) plot of malignant and benign epithelial cells from GSE193337 dataset (medium), along with their corresponding stemness scores (cytoTRACE, left), and the comparison of these scores between two groups (right). b t-SNE plot of high and low grade PCA cells from GSE141445 dataset (medium), along with their corresponding cytoTRACE scores (left), and the comparison of these scores between two groups (right). c Comparison of stemness scores (mRNAsi, mDNAsi) between benign and malignant prostate samples, as well as between high (Gleason score [GS] > 7) and low (GS < 7) grade PCA samples from TCGA-PRAD. d Kaplan–Meier (K–M) analysis demonstrated a correlation between the mRNAsi scores and the prognosis of PCA patients from TCGA-PRAD. OS overall survival, PFI progression-free interval, DFI disease-free interval, DSS disease-specific survival. Dashed line: median survival time. Color range: 95% confidence interval (CI)

Full size image

In parallel, there are stemness indices calculated by the one-class logistic regression (OCLR) algorithm based on specific stemness probes, including mRNAsi obtained from bulk RNA-seq data and mDNAsi, EREG_mDNAsi, and DMPsi obtained from DNA methylation data [31]. We computed the stemness indices of 553 patients from The Cancer Genome Atlas-Prostate Adenocarcinoma (TCGA-PRAD). These results further confirmed the findings obtained from the scRNA-seq (Fig. 1c, Additional file 1: Fig. S3e). Likewise, these results were consistently replicated in six other PCA bulk RNA-seq datasets [25,26,41, 42] (Additional file 1: Fig.S5c, Additional file 2: Data S1).

Identification of three stemness subtypes

Subsequently, based on the ssGSEA scores of these 18 signatures, we used unsupervised hierarchical clustering to classify the 553 samples from TCGA-PRAD into three subtypes (Fig. 2b, c, Additional file 1: Fig. S6a). We defined the group with the highest scores, which contained 56 samples (10.1%), as the “High Stemness” (HS) subtype, while the group with the lowest scores, consisting of 261 samples (47.2%), was named the “Low Stemness” (LS) subtype. The medium score group, comprising 236 samples (42.7%), was referred to as the “Medium Stemness” (MS) subtype, indicating that they were in a transitional state with the potential to transform into either a high or low stemness status. To validate our classification results, we compared the four aforementioned stemness indices across the three subtypes. Compared to LS, mRNAsi, mDNAsi, EREG_mDNAsi, DMPsi scores, and tumor purity were all significantly higher in HS and intermediate in MS (Fig. 2d, Additional file 1: Fig. S6b).

Significant differences in prognosis among patients with three stemness subtypes of PCA

K–M analysis revealed that HS patients had the worst PFI, whereas LS patients had the longest PFI (Fig. 2e, p = 0.00015), indicating that HS patients may experience disease progression earlier after treatment. However, there were no significant differences in OS, DFI, or DSS among the three subtypes (Additional file 1: Fig. S6c). Univariate COX analysis showed that LS and HS were protective factors and risk factors for PFI, respectively (Fig. 2f). Moreover, the multivariate COX analysis further demonstrated that LS was an independent protective factor for PFI (Fig. 2g).

Fig. 2
figure 2

Identification of three PCA stemness subtypes based on stemness signatures. a CircosPlot shows 288 stemness marker genes obtained from scRNA-seq data that are significantly positively correlated with cytoTRACE and significantly upregulated in both malignant and high-grade PCA cells (left), and 220 stemness marker genes derived from bulk RNA-seq data that are significantly positively correlated with mRNAsi and significantly upregulated in PCA samples (right). b Unsupervised hierarchical clustering based on the activity scores of the 18 stemness signatures classified PCA patients from TCGA-PRAD into three subtypes: low stemness (LS), medium stemness (MS), and high stemness (HS) subtypes. c 3D projection of the principal components obtained through PCA analysis. d Levels and trends of stemness score (mRNAsi) within three stemness subtypes. e Three stemness subtypes of TCGA-PRAD exhibits distinct PFI outcomes. f, g Univariate (f) and multivariate (g) Cox regression analysis of the three stemness subtypes, and clinical and molecular characteristics. *p < 0.05, **p < 0.01, ***P < 0.001, ****p < 0.0001

Full size image

Finally, we further validated the stability and universality of the stemness classification method by using two additional algorithms, non-negative matrix factorization (NMF) [43] and CensusClusterPlus [44], along with seven independent PCA bulk RNA-seq datasets [32, 33, 37,38,39,40, 45] (Additional file 1: Fig. S6d–u, Remark C).

Three stemness subtypes have distinct clinical features, mutational events, and functional annotations

Significant differences in certain clinicopathological features

Next, we compared the demographic and clinicopathological characteristics of the three stemness subtypes in patients with PCA. We found that LS had the highest proportion of para-cancerous samples (Fig. 3a, b; LS:MS:HS = 18%:1%:2%, p = 6.6e−11). The proportion of high-grade PCAs (GS > 7), pT3 and pT4 PCAs, and lymph node metastasis (N1) patients gradually increased in the LS, MS, and HS subtypes (Fig. 3a and b). There were no significant differences in body weights among the three subtypes (Fig. 3c). Compared with LS and MS, HS patients were older at diagnosis (Fig. 3c), suggesting that PCA patients diagnosed at an older age may have a relatively higher stemness status. Moreover, Gleason score, PSA, and androgen receptor were all significantly higher in HS patients (Fig. 3c), indicating that HS patients with the highest malignancy may benefit more from ADT treatment.

Significant differences in mutation events

Genome changes are research hotspots; therefore, we performed somatic mutation analysis. TMB, fragment genomic alteration (FGA), and amplification were all significantly higher in the HS subtype (Fig. 3d), suggesting that the HS subtype may be more likely to respond to immunotherapy [46]. The deletion was significantly higher in the MS subtype than in the LS subtype (Fig. 3d). Exon imbalance scores did not differ between stemness subtypes (Fig. 3d). Oncoplot revealed that the HS subtype had the highest rate of genomic alterations (Fig. 3e–g, LS:MS:HS = 66.82%:79.83%:92.73%). Although the top mutated genes of the three subtypes are similar, there are great differences in their mutation rates (Fig. 3e–g). We also observed that TP53 and ABCA13 mutations co-occurred, as did TTN and HERC2 mutations in the HS subgroup (Additional file 1: Fig. S7a). Silencing of TP53, a tumor suppressor gene, and inactivation of ABCA13, a cholesterol transporter protein, may be possible mechanisms contributing to the highly malignant phenotype of HS [47, 48]. In the LS subtype, we found a co-mutation of TTN and FAT3 (Additional file 1: Fig. S7b), while SPOP and TP53 mutations were mutually exclusive in the MS subtype (Additional file 1: Fig. S7c). The proportion of more SCNAs cluster gradually increased in the stemness subtypes (Fig. 3h, LS:MS:HS = 16%:38%:72%, p = 2.7e−14), while the high methylation cluster decreased successively (Fig. 3h, p = 2.6e−8). We also investigated the status of mutations and copy number alterations (CNAs) in common biomarkers of PCA. We found that the proportion of TP53 mutations and CNAs (heterozygously deleted [hetloss] + homozygously deleted [homdel]) increased progressively in the three subtypes (Fig. 3h, p = 0.06 and 0.0098, respectively). Compared with LS, the proportion of PTEN mutations was significantly higher in MS and HS (p = 0.03), and the proportion of CNAs increased in these three subtypes (Fig. 3h, p = 0.0022). Similarly, both the mutations and CNAs of CDKN1B and CNAs of RAD51C, FAM175A, CHD1, RB1, FANCC, and SPOPL were significantly higher in HS than in LS (Additional file 1: Fig. S7d). However, we found no significant differences in the mutations and CNAs of BRCA1 and BRCA2 between the three subtypes (Additional file 1: Fig. S7d). Understanding the mutation landscape of the aforementioned stemness subtypes is beneficial for uncovering potential mechanisms underlying tumor development, and provides a basis for discovering potential therapeutic targets and biomarkers.

Fig. 3
figure 3

Comparison of clinicopathological and molecular features among three PCA stemness subtypes. a Sankey diagram showing sample flow for stemness subtype, sample type, and GS. b Comparison of sample type, grade, pT and pN among the three stemness subtypes. c Comparison of patient weight, age at diagnosis, GS, PSA, and AR among three stemness subtypes. d Comparison of TMB, fraction genome altered, amplifications, deletions, and exon imbalance scores among three stemness subtypes. e–g Oncoplots showing the top 10 mutated genes in LS (e), MS (f) and HS (g). h Stacked histograms showing comparisons of somatic copy number alterations (SCNAs), DNA methylation clustering, TP53 mutations and CNAs, and PTEN mutations and CNAs among the three stemness subtypes. *p < 0.05, **p < 0.01, ***P < 0.001, ****p < 0.0001

Full size image

HS enriches oncogenic signaling pathways

Subsequently, we performed gene set variation analysis [49] (GSVA), GSEA [50], and ingenuity pathway analysis (IPA) to investigate the functional annotations, signaling pathways, and underlying mechanisms associated with the PCA stemness subtypes. As shown in Additional file 1: Fig. S8, the LS subtype mainly enriched nonspecific pathways. In contrast, HS exhibited enrichment in cell cycle-related signaling pathways [51], but IPA showed that the activity of these signaling pathways was suppressed. MS-enriched signaling pathways were similar to those of HS, but with slightly lower expression levels. See Additional file 1: Remark D for more details.

Three stemness subtypes have different treatment sensitivity and TIME patterns

Three stemness subtypes retain sensitivity to specific drugs

Given that the three stemness subtypes have unique functional pathways, we used the oncoPredict package [52] to predict the sensitivity of the three stemness subtypes to drugs and quantified sensitivity using half-maximal inhibitory concentrations (IC50). For the selection of conventional drugs, HS and MS subtypes were more sensitive to ADT (bicalutamide, abiraterone) (Fig. 4a, Additional file 1: Fig. S9a); for chemotherapy drugs, HS was more sensitive to taxanes (paclitaxel, docetaxel) (Fig. 4b, Additional file 1: Fig. S9b), etoposide (Fig. 4c), and gemcitabine (Fig. 4d), but resistant to mitoxantrone (Fig. 4e) and platinum drugs (Fig. 4f). In terms of recommended drug selection, HS demonstrated greater sensitivity to epidermal growth factor receptor (EGFR) inhibitors such as sunitinib (Fig. 4g), sorafenib, and imatinib (Additional file 1: Fig. S9c, d), while showing lower sensitivity to cabozantinib, afatinib, and erlotinib (Additional file 1: Fig. S9e–g). Additionally, HS showed increased sensitivity to poly (ADP-ribose) polymerase (PARP) inhibitors, including olaparib and talazoparib (Fig. 4h, Additional file 1: Fig. S9h), as well as to BI-2536 (Fig. S9j), histone deacetylase (HDAC) inhibitors (tacedinaline) (Additional file 1: Fig. S9k), and TGF-β receptor inhibitors (SB-431542) (Additional file 1: Fig. S9l). Conversely, MS exhibited greater sensitivity to anti-EGFR (cetuximab) (Additional file 1: Fig. S9m) and LS was more responsive to cyclin-dependent kinase (CDK) inhibitors (AZD5438) (Additional file 1: Fig. S9n).

Fig. 4
figure 4

Comparison of drug sensitivities and TIME patterns among three PCA stem subtypes. a–h Comparisons of sensitivities of three stemness subtypes to clinically preferred and recommended drugs. i Differences in TME scores among the three stemness subtypes. j Hypergeometric tests reveal an association between stemness subtypes and TIME subtypes. k Boxplots showing comparisons of immunocyte abundance among the three stemness subtypes. l Correlation heatmap showing the correlation between stemness indices and expression levels of immune checkpoint molecules. m Stacked histogram showing differences in responsiveness of the three PCA stemness subtypes to immune-checkpoint blockade (ICB) therapy. Evaluated by TIDE algorithm. n Submap analysis reflects the sensitivity of the three PCA stemness subtypes to an-PD-1, anti-PD-L1 and anti-CTLA-4 treatments. *p < 0.05, **p < 0.01, ***P < 0.001, ****p < 0.0001

Full size image

Three stemness subtypes have distinct TIME patterns and immunotherapy responsiveness

As demonstrated above, the stemness scores exhibited an inverse correlation with immune infiltration in PCA (Additional file 1: Fig. S4). To further investigate the relationship between stemness subtypes and their response to immunotherapy, we examined the TIME patterns. Our results indicated that the LS subtype had the highest Stromal, Immune, and ESTIMATE scores compared to the MS and HS subtypes (Fig. 4i). This indicated that the LS subtype had the most abundant immune infiltration. Moreover, we observed a significant association between the LS subtype and the High.immu subtype, whereas the MS subtype was significantly associated with the Low.immu subtype (Fig. 4j, Additional file 1: Fig. S9o).

We utilized the CIBERSORT algorithm [53] to assess immunocyte infiltration levels in the 553 PCA samples. Our findings revealed that CD4T, M1 and M2 macrophages had the highest abundance, while mast and plasma cells showed the lowest infiltration in HS (Fig. 4k). Moreover, Spearman correlation analysis showed a dramatically negative correlation between the stemness indices and the majority of immune checkpoint molecules (Fig. 4l), and most molecules were least expressed in HS (Additional file 1: Fig. S9p). These differences in TIME patterns and immune checkpoint molecule expression could potentially affect the efficacy of immunotherapy. Therefore, we applied the TIDE algorithm [54] to estimate the response of the patients to immunotherapy. The results showed that the proportion of responders in the HS was higher than that in the LS and MS subtypes (Fig. 4m, Additional file 1: Fig. S9q, p = 0.08). Additionally, the submap analysis [55] revealed that HS was the most sensitive to anti-PD-L1 (Fig. 4n).

Construction and validation of stemness subtype predictor

Construction of stemness subtype predictor

To facilitate the clinical application of our findings, we developed a stemness subtype predictor with high sensitivity and specificity, using the process outlined in Additional file 1: Fig. S10a (see Additional file 1: Methods S1 for details). We performed weighted gene co-expression network analysis (WGCNA) [56] (Fig. 5a, Additional file 1: Fig. S10b–g), protein–protein interaction (PPI) analysis (Fig. 5b), and Venn diagram plotting in sequence, and ultimately obtained 9 most critical stemness marker genes (Fig. 5c, Additional file 1: Table S2). We observed that these genes were significantly overexpressed in tumors (compared to normal, Additional file 1: Fig. S10h) and high-grade PCAs (compared to low-grade, Additional file 1: Fig. S10i).

We conducted a literature review on the research status of genes in PCA. CDK1 [57], KIF4A [58], TPX2 [59], BUB1 [60], and TOP2A [61] have been reported to promote PCA progression. However, there is a lack of evidence regarding the roles of SKA3, DLGAP5, NCAPG, and HMMR in PCAs. We collected 60 PCA samples and performed immunohistochemistry (IHC) to verify the expression of these four proteins. Our findings showed that their expression gradually increased in benign prostate tissues, and in low-grade and high-grade PCAs (Fig. 5d).

Fig. 5
figure 5

Construction and validation of stemness subtype predictor. a Correlation analysis between module eigengenes and stemness subtypes of TCGA-PRAD (left). The highest correlation between GS and MM in the pink module. Dots within the pink rectangle were defined as HS hub genes (right). b Protein–protein interaction (PPI) network of the 40 genes of Core.Sig, and these proteins were divided into three clusters based on the MCL inflation parameter. c Venn diagram identified the nine most critical stemness subtype marker genes that were intersected by 4 datasets and 76 machine learning algorithms (MLs). d Immunohistochemistry (IHC) staining shows the protein levels of four critical stemness subtype marker genes (SKA3, DLGAP5, NCAPG, HMMR) in benign, low-grade, and high grade PCA samples. Representative images are shown. e Histogram shows the performances of the 9-gene predictor in distinguishing benign from malignant tumors and predicting androgen deprivation therapy (ADT) response, metastasis, biochemical recurrence and progression via 100 MLs. The top 10 MLs with the best performance are exhibited. f K–M analysis shows the effect of 9-gene-based stemness-related risk score (SRS) on PFI of PCA patients from TCGA-PRAD. g Multivariate COX analysis showed that SRS was the most important independent risk factor for PCA patients

Full size image

Predictor can effectively distinguish stemness subtypes

Consistently, 100 MLs and unsupervised hierarchical clustering based on the 9-gene stemness subtype predictor showed that our predictor is highly effective in accurately distinguishing the stemness subtypes and can be used as an alternative clinically practical marker panel instead of the 18 stemness signatures (Additional file 1: Fig. S11a–d, Remark E).

Stemness subtype predictor has excellent performance in predicting malignancy, recurrence, progression, metastasis, and treatment response

Subsequently, we investigated the performance of our predictor in distinguishing malignancy and predicting tumor recurrence, progression, metastasis, and ADT efficacy. In TCGA-PRAD, GSE21034 [55], and separate clustering analysis of each cancer type (sample size > 50) (Additional file 1: Fig. S13a–g, Remark G).

Subsequently, univariate logistic regression analysis was performed based on the RNA-seq data of pre-ICB treatment samples to evaluate the effect of the three stemness subtypes on immunotherapy outcomes. Our findings revealed LS as a risk factor and HS as a promoting factor for response (Fig. 6d). Similar observations were also made in separate analyses of bladder cancer (BLCA), breast cancer (BRCA), non-small cell lung cancer (NSCLC), melanoma (SKCM), and head and neck squamous cell carcinoma (HNSCC) (Additional file 1: Fig. S13h). Nonetheless, the stemness subtypes did not significantly affect immunotherapy in clear cell renal cell carcinoma (ccRCC) (Additional file 1: Fig. S13h). We further compared the stemness subtypes with other ICB predictive biomarkers using the IMvigor210 dataset. The results showed that the stemness subtype was an effective predictor of ICB responsiveness (Additional file 1: Fig. S13i–l, Remark G).

Next, we attempted to further explore the effect of immunotherapy on stemness. We collected bulk RNA-seq data of paired samples from the same patients before and on/post ICB treatment [80, 82, 86, 93, 96, 98,99,100, 102]. Using hierarchical clustering to distinguish these samples into three subtypes (Additional file 1: Fig. S14a), we found that the proportion of HS decreased significantly after ICB therapy, while the proportion of LS increased significantly (Additional file 1: Fig. S14b). Further distinguishing between responders and non-responders, we found that most of the HS patients who responded to ICB treatment shifted to LS after ICB treatment (Fig. 6e, Additional file 1: Fig. S14c), while most of the HS patients who were resistant remained unchanged after treatment (Fig. 6f, Additional file 1: Fig. S14d). These observations were consistent across analyses of HNSCC, esophageal adenocarcinoma (EAC), and SKCM (Additional file 1: Fig. S14e–g). We also validated our findings using a single-cell dataset of HNSCC, which demonstrated a significant decrease in cytoTRACE scores following ICB treatment (Fig. 6g).

Fig. 6
figure 6

Interactions between stemness subtypes and ICB treatment in pan-tumors. a Unsupervised hierarchical clustering based on the stemness activity scores of the 18 stemness signatures clustered the baseline pan-tumor patients treated with ICB into three subtypes. b Hypergeometric test collaborates an association between stemness subtypes of ICB pan-tumors and responsiveness of ICB therapy. c Stacked histogram showing differences in responsiveness of the three pan-tumor stemness subtypes to ICB. d Univariate logistic regression shows the effect of the three stemness subtypes on ICB efficacy. e, f Sankey diagram showing sample flow for pre-treatment and on/post-treatment of ICB. Separate presentation for responders (e) and non-responders (f). g t-SNE plot of pre-treatment and post-treatment tumor cells of head and neck squamous cell carcinoma (HNSCC, medium), along with their corresponding cytoTRACE scores (left), and the comparison of these scores between two groups (right)

Full size image

Discussion

We applied single-cell and bulk RNA/DNA-seq technologies to evaluate the stemness status of patients based on stemness signatures. This allowed us to subtype the patients into three stemness subtypes. Specifically, HS patients are sensitive to androgen deprivation therapy, taxanes, and immunotherapy and have the highest stemness, malignancy, TMB levels, worst prognosis, and immunosuppression. LS patients are sensitive to platinum-based chemotherapy but resistant to immunotherapy and have the lowest stemness, malignancy, and TMB levels, best prognosis, and the highest immune infiltration. MS patients represent an intermediate status of stemness, malignancy, and TMB levels with a moderate prognosis. Our 9-gene stemness subtype predictor demonstrated high sensitivity and specificity, and could be easily used to identify stemness subtypes of patients through real-time quantitative PCR or IHC. This simple procedure provides a powerful tool for clinical tumor ty** and treatment selection.

PCSCs have been shown to play a significant role in the occurrence, development, treatment resistance, and recurrence of PCA [21]. In our study, the HS subgroup exhibited a significantly worse prognosis (Fig. 2e), significantly higher Gleason score, and expressed the highest level of serum PSA (Fig. 3c), indicating the highest malignancy in HS patients. It is essential to study the signaling pathways and functional annotations associated with PCSCs to understand the molecular mechanisms of carcinogenesis and to identify potential drug targets. Several signaling pathways are closely related to the regulation of PCSCs, such as the PI3K/AKT/mTOR pathway and the Wnt/β-catenin pathway [103, 104]. Studies have shown that PCSCs have a relatively slow proliferation rate compared to ordinary PCA cells and exist in a quiescent status [105, 106]. GSVA and GSEA revealed that cell cycle-related signaling pathways, including cell cycle checkpoints, were enriched in HS (Additional file 1: Fig. S8a, b). However, IPA showed that the cell growth, proliferation, and development of HS were significantly inhibited (Additional file 1: Fig. S8e). These results suggested that while the signaling pathways and molecules related to the cell cycle are upregulated in HS, their proliferative activity is suppressed or static. This phenomenon may be attributed to the self-renewal and multidirectional differentiation ability of PCSCs, such as negative feedback regulation after excessive proliferation (enrichment of cell cycle checkpoints) and activation disorders of cell cycle-related proteins [106,107,108,109].

The PCA treatment continues to be a major research focus. Radical prostatectomy and radiation therapy, with or without ADT, are the standard treatments for localized PCA. ADT remains the primary treatment for advanced disease [10, 110]. Although many treatments initially eradicate cancer cells, cancer often recurs owing to the presence of resistant PCSCs [111]. Therefore, targeting CSCs is critical for preventing tumor recurrence. The significance of molecular subty** in tumors lies in its ability to help doctors better understand the biological characteristics, molecular mechanisms, development trends, and treatment responses of tumors, and to provide personalized treatment plans for patients. Existing molecular subty** methods for PCA include the PAM50 method [18], which is used to guide ADT, and the Decipher method [19], which is used to guide radiotherapy and surgical treatment. By contrast, our stemness subty** method can be used to guide various treatments. We found that different stemness subtypes exhibited different sensitivities to drugs. HS was more sensitive to ADT (bicalutamide), PARP inhibitors (olaparib), EGFR inhibitors (sunitinib and sorafenib), immunotherapy, and chemotherapeutic drugs such as docetaxel, etoposide, and gemcitabine. In contrast, LS was more sensitive to platinum drugs, erlotinib, and CDK inhibitors (AZD5438) (Fig. 4, Additional file 1: Fig. S9). Previously, the effectiveness of immunotherapy for PCA has been limited. However, with the development of immunology and cutting-edge molecular diagnostic tools, immunotherapy is expected to become a viable treatment option for PCA [112]. Unlike tumors, such as melanoma, bladder cancer, and NSCLC, which are highly responsive to immunotherapy and characterized by infiltrating lymphocyte proliferation, PCA is considered a “cold” tumor with an immunosuppressive TME [9, 112]. Although the effectiveness of ICB monotherapy for PCA is limited, combined strategies with other standard treatments (ADT, chemotherapy, PARP inhibitors, radium-223, and tyrosine kinase inhibitors) have shown some positive effects [9]. Overall, these findings underscore the importance of molecular subty** for guiding cancer treatment. By tailoring therapies to the specific molecular characteristics of tumors, doctors can improve treatment outcomes and help patients achieve the best possible outcomes.

The significance of pan-cancer research lies in the application of diagnosis and treatment to more tumors through cross-tumor similarities [113]. We validated the conserved characteristics of the three subtypes in pan-tumors and their responsiveness to immunotherapy using 28,381 pan-tumor samples and 2641 ICB pretreatment samples. This indicates that our findings have implications beyond PCA, and can potentially benefit a wider range of patients. Furthermore, develo** a predictive model using multiple MLs and datasets can enhance the model’s generalization ability and prevent overfitting or underfitting issues, which is an effective way to improve the accuracy and robustness of the predictors. We jointly developed a 9-gene stemness subtype predictor with high sensitivity, specificity, and excellent generalization ability using four datasets and 76 machine learning algorithms. Significantly, this predictor can be further developed into a kit for clinical application. We believe that this predictor has great potential for clinical application, as it offers rapid and reliable molecular diagnosis and prognosis for patients with PCA and guides personalized treatment decisions. Moreover, this predictor can facilitate the enrollment of PCA patients into clinical trials for immunotherapy or other targeted therapies based on their stemness subtype, and it may also be applicable to other cancer types with the same stemness subtypes as PCA.

Although this study yielded valuable insights, it is important to acknowledge some of its limitations. First, the sample size for our real-world validation was only 60, and we did not include any pan-cancer samples. Expanding the sample size and including pan-cancer samples for verification in future studies are essential. Second, there is a lack of RNA-seq data of PCA treated with ICB. Although the pan-tumor ICB treatment cohort validated the relationship between stemness subtypes and immunotherapy responsiveness, further validation is still needed for PCA immunotherapy datasets.

Conclusions

In conclusion, tumor molecular ty** is of great significance for understanding how cancer develops and progresses, as well as for guiding clinical treatment and the development of new anticancer drugs. By categorizing PCA patients into three stemness subtypes, we can systematically characterize patients from various points of view, e.g., stemness, prognosis, clinical pathological features, mutation patterns, malignancy degree, immune infiltration levels, and efficiencies of different treatments, including immunotherapy. This classification method is also applicable to pan-tumor analyses. Furthermore, the 9-gene stemness subtype predictor we developed is expected to be a clinically useful tool for precision oncology. Significantly, our method provides a pipeline for the development of cancer classification that can be applied to various tumors based on different research hotspots.

Methods

Patients and datasets collection

Five PCA scRNA-seq datasets [25,26,1: Fig. S5a. We utilized scRNA-seq data to identify stemness marker genes in malignant and high-grade PCA epithelial cells, and used bulk RNA-seq and DNA methylation data to identify stemness marker genes in malignant PCA samples. Our approach involved analyzing multiple datasets, performing Spearman correlation analysis and DE analysis, and selecting genes that were positively correlated with stemness scores and were upregulated in malignant and high-grade PCA. We then analyzed gene sets on MSigDB [114] with the stemness marker genes obtained and identified significant risk signaling pathways using univariate COX analysis for subsequent stemness classification. See Additional file 1: Methods S1 for further details.

Immunohistochemistry (IHC)

Paraffin sections underwent dewaxing, antigen retrieval, and serum blocking. They were incubated with primary antibodies overnight at 4 °C, secondary antibodies and SABC for 30 min at 37 °C. Sections were stained with DAB and counterstained with hematoxylin. The primary antibodies used were the NCAPG Polyclonal antibody (24563-1-AP, Proteintech, Wuhan, China), HMMR-specific polyclonal antibody (15820-1-AP, Proteintech, Wuhan, China), HURP polyclonal antibody (12038-1-AP, Abcam, Proteintech, Wuhan, China), and SKA3 antibody (SC-390326, Santa, California, USA).

Statistical analysis

Numerical variables were compared using t-tests or ANOVA, categorical variables using χ2, Fisher’s exact or Kruskal–Wallis tests. Non-normally distributed variables were compared using non-parametric tests. Correlations were evaluated using Pearson or Spearman tests. Survival differences were compared using the log-rank test. Confidence intervals (CIs) were reported as 95% and significance was set at P < 0.05. Analyses were performed using R (v4.2.1), Python (v3.10), and Origin 2022 software.

More methods and details can be seen in Additional file 1: Methods S1.