Using machine learning to discover DNA metabolism biomarkers that direct prostate cancer treatment

Abstract

DNA metabolism genes play pivotal roles in the regulation of cellular processes that contribute to cancer progression, immune modulation, and therapeutic response in prostate cancer (PC). Understanding the mechanisms by which these genes influence the tumor microenvironment and immune evasion is crucial for identifying prognostic biomarkers and developing targeted therapies. We performed an integrative analysis using transcriptomic data from the TCGA cohort and external validation datasets. Differentially expressed genes (DEGs) were identified using the edgeR algorithm with an FDR < 0.01 and a minimum fold change of 1.5. Gene enrichment analysis was conducted through GO and KEGG pathways to explore the biological significance of DNA metabolism genes in PC. In addition, clustering analyses, machine learning models, and single-cell RNA sequencing (scRNA-seq) were employed to investigate the immune characteristics, prognostic value, and therapeutic relevance of these genes. A total of 536 DEGs were identified across six subtypes of prostate cancer, with key DNA metabolism genes such as POLD2, RAD9A, REV3L, MSH6, and WRNIP1 highlighted as critical players. Gene enrichment analyses revealed that these DEGs were significantly associated with pathways involved in DNA repair, cellular aging, and telomere maintenance. Clustering analysis identified two distinct subgroups (C1 and C2) based on DNA metabolism gene expression, with C1 exhibiting a more aggressive phenotype, higher immune infiltration, and poorer prognosis. Machine learning models, particularly the CoxBoost algorithm, identified 21 key genes contributing to an effective prognostic model. Furthermore, scRNA-seq analysis confirmed the upregulation of DNA metabolism genes in PC cells compared to normal cells. Our findings highlight the importance of DNA metabolism genes in the progression and immune dynamics of PC. These genes not only serve as potential biomarkers for prognosis but also offer promising targets for personalized therapies. The integration of multi-omics data and advanced computational models provides new insights into the molecular underpinnings of PC and holds potential for improving treatment strategies.

Introduction

Prostate cancer (PC) is one of the most frequently diagnosed malignancies in men globally, with its incidence continuing to rise^1,2. Current treatment options include surgery, chemotherapy, radiotherapy, endocrine therapy, and PSA-based prognostic testing. However, the considerable heterogeneity of PC, along with limited early detection tools, results in more than 30% of patients experiencing biochemical recurrence (BCR) after treatment³. This underscores the need for innovative biomarkers and advanced tools to improve prognosis prediction and guide precision medicine strategies.

DNA metabolism proteins are increasingly recognized as critical in cancer biology, particularly in the context of tumorigenesis, progression, and therapeutic resistance⁴. Disruptions in DNA replication, repair, and recombination pathways are linked to the aggressive nature of prostate cancer and its tendency to recur after treatment⁴. These proteins are emerging not only as potential therapeutic targets but also as prognostic biomarkers that could significantly enhance patient stratification and treatment response prediction.

Weighted Gene Co-Expression Network Analysis (WGCNA) has proven to be a powerful method for identifying gene modules correlated with clinical traits, helping researchers discover hub genes central to cancer progression⁵. By analyzing gene co-expression patterns, WGCNA provides insights into the molecular drivers of PC, particularly when integrated with additional biological processes. One such process is epithelial-mesenchymal transition (EMT), a key mechanism in cancer metastasis, which is characterized by epithelial cells acquiring mesenchymal traits, promoting invasion, migration, and resistance to treatment^6,7.

To address the complexity and heterogeneity of PC, our study utilized integrated machine learning approaches combined with WGCNA and molecular insights into DNA metabolism proteins and EMT-related pathways. Integrated machine learning refers to the use of multiple machine learning algorithms in tandem with large-scale genomic datasets to improve predictive accuracy and uncover complex relationships between genes and clinical outcomes^8,9. Machine learning has already demonstrated its ability to analyze vast datasets, identify patterns, and build robust predictive models, but its integration with co-expression analysis and key molecular pathways like DNA metabolism and EMT offers a more comprehensive framework for understanding PC progression¹⁰.

To effectively capture the multifactorial nature of prostate cancer, we integrated multiple computational strategies that synergize to enhance discovery and prediction. WGCNA allowed us to uncover gene co-expression modules closely related to key clinical phenotypes, particularly those involved in DNA metabolism and EMT—two processes known to contribute significantly to tumor progression, recurrence, and therapeutic resistance^7,11–14. By focusing on hub genes within these modules, we aimed to prioritize biologically meaningful targets that could inform both prognosis and therapeutic vulnerability. DNA metabolism proteins, for example, are often dysregulated in aggressive tumors, while EMT markers are closely tied to metastatic potential and immune evasion, making both pathways ideal candidates for integrated molecular profiling.

Building on this foundation, we applied a high-throughput machine learning framework to systematically evaluate and optimize prognostic signatures. By employing an ensemble of 101 machine learning models—including CoxBoost, random forest, and support vector machines—we were able to compare model performance and ensure robustness across diverse datasets. This ensemble-based strategy allowed us to identify gene combinations with the highest predictive accuracy while minimizing overfitting. Importantly, integrating these predictive features with single-cell RNA sequencing data enabled validation at cellular resolution, revealing functional heterogeneity across tumor subpopulations and immune landscapes. This multi-tiered approach not only enhances model reliability but also improves our ability to interpret the biological underpinnings of PC progression and therapy response.

In this study, we leveraged single-cell RNA sequencing (scRNA-seq) to identify key marker genes related to DNA metabolism and EMT. We applied WGCNA to construct gene co-expression networks and used integrated machine learning to develop and validate a prognostic signature. By utilizing a diverse set of 101 machine learning models, we rigorously optimized the predictive accuracy of our model, assessing its clinical applicability across multiple patient cohorts. This integration of machine learning techniques enabled us to refine the model for better prediction of treatment responses and patient outcomes. Our approach combines WGCNA, scRNA-seq, and integrated machine learning to provide a more nuanced understanding of the molecular mechanisms driving prostate cancer aggressiveness. This strategy aims to develop a clinically applicable prognostic tool that not only improves patient stratification but also informs personalized therapeutic decisions, ultimately enhancing treatment efficacy and patient care.

Material and methods

Data sources and preprocessing

The gene expression dataset (GSE55945¹⁵, GSE104749¹⁶, GSE46602¹⁷, and GSE32571¹⁸) was obtained from the Gene Expression Omnibus (GEO) website (https://www.ncbi.nlm.nih.gov/geo/) for the present investigation. A total of 36 prostate cancer specimens and 36 normal prostate specimens were included in the gene expression collection. An overview of the datasets is shown in Table 1. The research obtained from cBioPortal confirmed the gene’s expression in certain organs. To validate our findings, we also analyzed four additional datasets: GSE21036, GSE34933, GSE48430, and GSE212215, which were included to ensure the robustness and reproducibility of the identified differentially expressed DNA metabolism genes across independent cohorts. Sample inclusion was based on the availability of both tumor and adjacent normal tissue samples, adequate sample size, and consistent platform annotation. In cases where datasets contained more samples, we selected a matched subset with complete metadata and expression values. All analyses were conducted using RStudio (version 2022.12.0 or specify the actual version used) with R version 4.2.2. The following R packages were used with their respective versions: edgeR (v3.38.4), limma (v3.52.4), Seurat (v4.3.0), clusterProfiler (v4.6.0), survival (v3.5-5), CoxBoost (v1.4), SuperPC (v1.16), and ggplot2 (v3.4.0). Additional analyses were conducted using the STRING database (v11.5) and Cytoscape (v3.9.1) for PPI network construction and visualization.

Table 1.

Summary of public datasets used in this study.

Dataset ID	Data type	Platform	Sample type	Number of samples	Application in study
GSE55945	Microarray	GPL570	Prostate tissue samples	13 tumor, 8 normal	DEG analysis and integration
GSE104749	Microarray	GPL10558	Prostate tissue samples	30 tumor, 8 normal	DEG validation and clustering analysis
GSE46602	Microarray	GPL570	Prostate tissue samples	36 tumor, 14 normal	DEG validation
GSE32571	Microarray	GPL570	Prostate tissue samples	20 tumor, 12 normal	Prognostic model validation
GSE168826	scRNA-seq	10 × Genomics (filtered matrix)	Prostate cancer tissue	6 tumor samples	Single-cell analysis of DNA metabolism gene expression

Identification of DNA metabolism genes

To guarantee high-quality cells, we used the Seurat package in R to identify genes involved in DNA metabolism. This package also helped with object construction. Cells with less than 55 detectable genes, cells with more than 5% mitochondrial genes, or cells with less than three detectable genes were filtered out. After normalizing the gene profiles, the 1600 genes with the highest levels of variability shown by JackStraw analysis were subjected to principal component analysis (PC). After that, R's FindClusters function was used to cluster the data using a resolution value of 0.5. The t-distributed stochastic neighbor embedding (t-SNE) approach was used for the visualization. The FindAllMarkers (version 2.3.1) function was used in combination with the Wilcoxon–Mann–Whitney test to identify marker genes for each cluster. These genes were defined as having an adjusted P value < 0.01 and |log FC|> 1.3. The test evaluated the differences in gene expression between each cluster and all other clusters. The cell types were also annotated and visualized using the SingleR program^19,20.

PPI network construction and GO enrichment analysis

To explore the interactions among DNA metabolism-related genes, we constructed a protein–protein interaction (PPI) network using the STRING database (v11.5; https://string-db.org/), which provides known and predicted interactions based on experimental data, text mining, co-expression, and other criteria. Genes with a confidence interaction score ≥ 0.4 (medium confidence) were included in the network. The resulting network was visualized using Cytoscape (v3.9.1), and the top hub genes were identified using the CytoHubba plugin based on the degree algorithm. Gene Ontology (GO) enrichment analysis of the PPI network was then conducted using the clusterProfiler R package, focusing on biological process (BP), cellular component (CC), and molecular function (MF) categories.

Consensus clustering analysis

To categorize patients in the TCGA cohort into discrete groups based on DNA metabolism gene expression patterns, we used a combination of agglomerative pam clustering, a 1-Pearson correlation distance metric, and 80% sample resampling for 1000 repeats. The consistency matrix, relative change in the area under the cumulative distribution function (CDF) curve, and cumulative distribution function (CDF) were used to determine the ideal number of clusters. The PC and Kaplan–Meier methods were used to evaluate the variations and biochemical recurrence-free survival (bRFS) rates among the various clusters. We also used the chi-square test to look for correlations between the clusters and age, Gleason score, PSA levels, pathological N (pN) stage, and pathological T (pT) stage, among other clinicopathological variables. Additionally, in order to inquire about any genomic variations, the existence of copy number variation (CNV) was examined among the clusters.

Gene set variation analysis (GSVA)

Using the single-sample gene set enrichment analysis (ssGSEA, version 0.3.11) technique, the TCGA cohort’s gene expression profile for DNA metabolism genes was calculated with the use of the GSVA package in R (version 4.1). The gene sets used for this estimate were sourced from the Kyoto Encyclopedia of Genes and Genomes and the Gene Ontology. After that, we calculated the enrichment score for every gene set’s pathway. We generated heatmaps by determining the differential enrichment scores of pathways between the clusters and then selecting the top 8 routes for each cluster that had a statistically significant (adjusted P value < 0.01) score.

Identification of DNA metabolism genes and co-expression network construction

To identify DNA metabolism-related genes with potential clinical relevance, we first curated a list of genes associated with DNA metabolic processes from Gene Ontology (GO:0006259). We then applied Weighted Gene Co-expression Network Analysis (WGCNA) (version 1.73) using the “WGCNA” R package to construct a gene co-expression network based on the TCGA prostate cancer expression dataset. Genes with the top 25% highest variance were selected as input. An appropriate soft-thresholding power (β) was determined using the scale-free topology criterion. A topological overlap matrix (TOM) was constructed to measure network connectivity, and hierarchical clustering was applied to identify modules of co-expressed genes. Each module was summarized by its eigengene, and module-trait relationships were assessed by correlating eigengenes with clinical traits (e.g., Gleason score, recurrence status). Modules significantly correlated with clinical parameters were further analyzed. DNA metabolism genes within these modules were considered hub candidates for downstream prognostic modeling and functional assessment^{7,12,13,21–25}.

Tumor immune microenvironment analysis

Prior to comparison, the two groups’ tumor purity, stromal score, immunological score, and ESTIMATE score were estimated. Secondly, using data from an earlier work, we computed the scores of immune infiltrations for 28 different kinds or routes of immune cells using the ssGSEA algorithm and the Mann–Whitney test. In addition to ssGSEA, seven other algorithms were used to guarantee the reliability and consistency of the results: TIMER²⁰, CIBERSORT, CIBERSORT-ABS, QUANTISEQ, MCPCOUNTER, XCell, and EPIC. Following the methodology used in an earlier research, the cancer-immunity cycle phases’ scores were examined using Tracking Tumor Immunophenotype (https://biocc.hrbmu.edu.cn/TIP/). Third, 91 immune modulator marker genes were compared across the two clusters. These genes included immunostimulators (n = 30), immunoinhibitors (n = 21), MHC (n = 20), receptors (n = 15), and chemokines (n = 25), all of which were derived from an earlier work^21,23,24,26.

Machine learning-based signature construction and validation

We used a holistic strategy that included 10 separate machine learning algorithms in 101 unique permutations. The goal was to create a predictive signature that was remarkably stable and accurate. The study employed ten original machine learning algorithms: CoxBoost, Enet, survival-SVM, Lasso, plsRcox, Ridge, RSF, stepwise Cox, SuperPC, and GBM. Notably, feature selection skills were shown by a few of these algorithms, including CoxBoost, Lasso, RSF, and stepwise Cox. To conclude, the following is the machine learning method that was used in this study: (1) First, we used univariate Cox regression analysis to identify ECMGs in the TCGA cohort that could have a prognostic impact. (2) Leave-One-Out Cross-Validation Framework: Afterwards, we ran 101 models on the potential prognostic ECMGs using a leave-one-out cross-validation framework inside the TCGA cohort. We set out to create a prediction signature that is both reliable and resilient. (3) Thorough Testing in Separate Validation Cohorts: We conducted extensive testing in four separate validation cohorts to assess how well the built signatures worked. (4) Choosing the Best Model: We computed Harrell’s concordance index (C-index) for all cohorts for each model. The best model was determined to be the one with the greatest mean C-index. The best model used the median risk scores from the TCGA cohort and four independent validation cohorts to stratify patients into low-risk and high-risk categories, respectively. Afterwards, the distribution of the ideal model in the training cohort and validation cohort were examined using t-SNE and principal component analysis, respectively. We used receiver operating characteristic (ROC) curves and Kaplan–Meier curves to assess the best model’s prognostic utility and prediction accuracy.

A robust prognostic signature was developed using 10 machine learning algorithms across 101 combinations: CoxBoost, Lasso, Ridge, RSF, Enet, GBM, plsRcox, stepwise Cox, survival-SVM, and SuperPC. First, univariate Cox regression identified prognostic DNA metabolism genes. These were input into the model training using leave-one-out cross-validation in the TCGA cohort. The model with the highest average C-index across internal and four external validation cohorts was selected. Patients were stratified into high- and low-risk groups using the median risk score. Model performance was evaluated using Kaplan–Meier curves, ROC analysis, PC, and t-SNE. Associations between the risk score and clinicopathological features were analyzed using chi-square and Cox regression. The signature’s prognostic value was also tested in bladder and renal cancer cohorts. For benchmarking, published PC gene signatures were re-evaluated on our cohorts using their original formulas, and C-index was used for comparison.

Evaluation of the clinical significance of DNA metabolism genes

Using a chi-square test, we looked for differences between DNA METABOLISM GENES risk scores and clinicopathological features. Onwards to the subgroups, we ran a stratified survival analysis. Multivariate and univariate Cox regression analysis were run in order to discover independent prognostic markers. We also used TCGA-derived mRNA expression and survival data for renal clear cell carcinoma and bladder cancer to explore the potential of DNA METABOLISM GENES in other urothelial malignancies. Kaplan–Meier curves were used for further analysis of these datasets.

Comparison of published signatures in PC

In order to compare the efficacy of DNA metabolism genes with known signatures, we performed a thorough literature search on PubMed for model publications that predicted PC outcomes up to June 1, 2023. Various methods, including Lasso and RSF, were used to fit these gathered signals, which included a wide range of biological relevance. After that, we used the genes or RNA and the coefficients given in the publications to determine the risk scores for each of the five cohorts. Next, the C-index was used to compare the performance in predicting PC BCR.

Immunotherapy response and drug sensitivity

This study began by comparing the two groups’ mutation profiles using TCGA somatic mutation data processed using the VarScan platform. Afterwards, the IMvigor 210 cohort was used to assess the variations in immunotherapy response and survival results between the low-risk and high-risk individuals identified by DNA metabolism genes. This cohort was treated with anti-PD-L1. The TIDE method, which stands for tumor immune dysfunction and exclusion, was also used to forecast how well immune checkpoint drugs will work on DNA metabolism genes. Last but not least, we utilized the IC50 values retrieved from the Genomics of Drug Sensitivity in Cancer database (https://www.cancerrxgene.org/) to compare the two groups’ responses to ten anticancer medications.

Data collection and preprocessing of single cells

The Gene Expression Omnibus (GEO) database was queried for transcriptome data and clinical data of PC patients who had neoadjuvant chemotherapy (NAC). You may find the data set at https://www.ncbi.nlm.nih.gov/geo/ under the IDs GSE94577, GSE82225, and GSE176031. With a grand total of 306, the discovery cohort had the largest number of BC samples. Fifteen BC samples from GSE94577 and twenty BC samples from GSE82225 made up the separate validation cohorts. Excluded from the study were specimens for which full survival data was not available. A total of fourteen BC samples’ worth of scRNA-seq data came from a research by Qian et al. The information came from lambrechtslab-Laboratory of Translational Genetics’ website (vib.be). A version 4.0.0 of the Seurat (version 1.73)⁵⁶ R program was used to preprocess the scRNA-seq data. We retained cell samples with an expression rate of mitochondrial genes below 5% and an expression rate of over 200 genes. The scRNA-seq dataset was normalized using the “NormalizedData” tool, and 2000 genes with considerable variability were identified using the “FindVariableFeatures” technique. To lessen the effects of the batch, we used the R package Harmony. In order to organize and depict cells using uniform manifold approximation and projection (UMAP), principal component analysis (PC) was performed after data standardization. We then utilized the “DotPlot” program to graphically show the expression level of marker genes inside a particular cluster. Marker genes decide which known cell lineages these clusters belong to. The “FindClusters” function and the K-nearest neighbour (KNN) approach are used to identify cell clusters with a 0.2 resolution.

Single-cell validation of DNA metabolism gene expression

To validate the expression of DNA metabolism-related genes at the single-cell level, we used scRNA-seq data from 14 prostate cancer samples (GSE168826). Seurat (v4.0.0) was used for quality control, normalization, and clustering. Cells expressing fewer than 200 genes or with > 5% mitochondrial content were removed. PCA and UMAP were applied for dimensionality reduction, and clustering was performed at a resolution of 0.2. Marker genes were identified using the Wilcoxon test and annotated with the SingleR package.

Results

DEGs screening

Using the edgeR software, we were able to accurately identify the DEGs across the different subgroups of prostate cancer. The minimum needed fold change was 1.5, and the adjusted P-value (FDR) cutoff was set at less than 0.01. Analyzing data from four kinds of prostate cancer across six potential combinations led to the identification of 536 genes with differential expression (Fig. 1).

Fig. 1.

Identification of DEGs associated with BIVM expression in prostate cancer. (A) Valcano plot of GSE21036, GSE34933, GSE48430, and GSE212215. (B) Venn diagram of the microarray data.

Identification of DNA metabolism genes by microarray

In downstream analyses of the identified DEGs, we found that key DNA metabolism-related genes—POLD2, REV3L, RAD9A, MSH3, WRNIP1, and MSH6—were significantly associated with immune and stromal components of the prostate cancer microenvironment (see Fig. 1).

Establishment of DNA metabolism genes consensus clusters and proteins interaction

Using the TCGA dataset’s expression patterns of DNA metabolism genes, PC was able to differentiate between normal and PC samples, indicating that epithelial cells in PC and normal play different regulatory functions. We next used clustering analysis to look at how the TCGA cohort’s patients’ expression levels of genes related to DNA metabolism varied. Two ideal clusters (k = 2, C1 = 179, C2 = 167) were selected by analyzing the area under the CDF curve, the declining trend of the CDF delta, and the average consistency among patient clusters (Additional file 1: Fig. 1). Following this, further principal component analysis (PC) showed a clear clustering pattern (Fig. 2), suggesting a notable difference in the distribution of genes involved in DNA metabolism between the two groups. Using edgeR, we identified 536 differentially expressed genes (DEGs) across six prostate cancer subgroups, with a log2 fold change > 1.5 and FDR < 0.01 (Fig. 2).

Fig. 2.

Correlation between cell–cell gene expression and protein interaction. (A) Correlation between cell–cell gene expression in prostate cancer and (B) collect gene involved in EMT gene in PC.

Gene enrichment of DEGs

By submitting 156 frequently used DEGs for enrichment analysis by GO and KEGG, the roles of DEGs were examined. Macromolecular biosynthesis, radiation response, double-strand break repair by homologous recombination, and macromolecule metabolic process were the five most abundantly occurring Gene Ontology (GO) terms in the BP of DEGs, as shown in Fig. 2. In the BP of DEG’s PC, the five most abundant GO keywords were: DNA helicase activity, nucleic acid binding, damaged DNA binding, and control of ATPase activity. In the cellular component (CC) of DEG’s PC, the five most abundant GO keywords were genomic area, telomeric region, membrane-enclosed lumen, chromosomal region, and nuclear chromosome (Fig. 3). Microarray analysis of normal and tumor tissues revealed DNA metabolism-related DEGs implicated in the EMT pathway. Key genes included POLD2, REV3L, RAD9A, MSH3, WRNIP1, and MSH6 (Fig. 3).

Fig. 3.

Biological processes, and molecular functions in the genes. (A) analysis of biological processes of PPI, (B) analysis of the molecular function of PPI.

KEGG pathway enrichment

In order to find possible signaling pathways linked to the 10 hub genes, we re-examined the KEGG pathway using the DAVID program (P < 0.05). Figure 4 shows that the genes are highly related to six signaling pathways: cellular aging, immortality, DNA repair, the DNA-PK pathway in nonhomologous end joining, and telomere maintenance. To examine the relative levels of expression for these four genes, we consulted the TCGA database. When compared to healthy individuals, the findings showed that expression levels were significantly higher in patients with BC-adjacent and PC. We used the pc-GenExMiner program to mine co-expression data sets in order to get a comprehensive knowledge of the underlying mechanisms of these four genes in prostate cancer. Figure 3 demonstrates that all seven genes are up-or down-regulated in prostate cancer tissues, which supports the existence of a signaling network (Fig. 3).

Fig. 4.

Signaling pathway analysis. (A) Compare the findings of ssGSEA with those of seven more algorithms: TIMER, CIBERSORT, CIBERSORT-ABS, QUANTISEQ, MCPCOUNTER, XCell, and EPIC. (B) The correlation signaling route is shown in the heatmap. (C) four signaling pathways implicated in PC..

Construction of a prognostic signature based on machine learning

Based on the expression patterns of 134 genes involved in DNA metabolism, we constructed a predictive signature. Our first univariate Cox regression analysis revealed 22 genes associated with DNA metabolism and bRFS. We then fitted 90 prediction models using eight different ML algorithms: CoxBoost, Enet, plsRcox, Ridge, RSF, stepwise Cox, SuperPC, and survival-SVM. To evaluate the robustness of these models and identify the most effective prognostic indicator with the highest mean C-index, we used a tenfold cross-validation strategy. The evaluation comprised the TCGA training cohort in addition to four external validation cohorts, namely PSH3, WRNIP1, RAD9A, POLD2, and REV3L, as shown in Fig. 5A. From 134 DNA metabolism genes, 22 were linked to bRFS. A robust prognostic model was developed using CoxBoost and SuperPC, resulting in a 21-gene signature with the highest mean C-index (0.706) across five cohorts (Figs. 5, 6, 7).

Fig. 5.

Development and verification of DNA metabolism genes with machine learning. (A) A total of eight machine learning methods applied to five different cohorts yielded a C-index of 89. From 51 prognostic ECMGs, the CoxBoost approach was used to calculate coefficients for 21 model genes. (B) POLD2, (C) REV3L, (D) RAD9A, (E) MSH3, and (F) WRNIP1 immunophenotypic evaluation.

Fig. 6.

A regional association map is shown for the gene area encompassing the POLD2, REV3L, RAD9A, MSH3, WRNIP1, and MSH6. (A) There are a considerable number of SNPs in this area that have varying degrees of linkage disequilibrium (LD), and they are all linked to an increased risk of prostate cancer. The studied gene is given below the x-axis, which reflects the chromosomal location of the SNPs. After accounting for histologic variables and fourteen expression main components, the y-axis displays the − log10(P value) derived from a regression analysis of normalized gene expression levels on the quantity of minor alleles of each SNP genotype. (B) A dotted red vertical line represents the PrCa-risk single nucleotide polymorphism (SNP), and a diamond represents the expression quantitative trait locus (eQTL) result. A regional association map is shown for the following genes: MSH3, WRNIP1, POLD2, REV3L, RAD9A, and MSH6.

Fig. 7.

Comparison between DNA metabolism genes. (A) The 40 samples were divided into four separate groups using the consensus clustering matrix. Multiple categories were used to categorize the standard samples. (B) UMAP maps showing the different types of cells and the genetic markers linked to them. The cell cycle and clustering of cells are shown in the heatmap in (C). In (D) POLD2 and (E) REV3L, the gene expression of the marker genes in scRNA-seq is shown. Panel of canonical marker genes (EPCAM for epithelial cells, CD3D for T cells, CD14 for myeloid cells with a dot plot visualization.

Immunophenotypic analysis based on DNA metabolism genes clusters

Significant differences were seen in immune infiltration and the different phases of the cancer-immunity cycle between C1 and C2. With the exception of tumor purity, C1 showed much greater stromal score, immunological score, and ESTIMATE score compared to C2. C1 also showed significantly more immune-related cells and pathways infiltration and greater immunological activity throughout 16 of the 23 stages of the cancer-immunity cycle compared to C2 (Fig. 4B–D). In order to verify that our analytical approach was not biased while creating these two clusters, we used seven other algorithms to check that the ssGSEA findings were stable and reliable: CIBERSORT-ABS, CIBERSORT, EPIC, MCPCOUNTER, QUANTISEQ, TIMER, and XCell. Compared to C2, C1 showed much increased expression of the majority of immune modulators (Fig. 6). These results show that elevated immunological patterns are associated with ECMG malignancy and poor prognosis in PC. This led us to conclude that C1 cancers were immunologically hot and C2 tumors were immunologically chilly.

According to the survival study, C1’s prognosis was worse than C2’s. We looked at the CNV frequency in the C1 and C2 patient groups since CNVs are important indicators of the evolution of malignant tumors. Notably, CNV was more common in C1 patients compared to C2, suggesting that C1 individuals had a more aggressive behavior and worse prognosis in PC. Additionally, C1 patients had a higher Gleason score and advanced pN stage, which further supports C1 as a subgroup with a high malignant potential. Using the GSVA technique, we dug more and found that C1 was strongly enriched in a number of signaling pathways, most of which were involved in cancer and development. The number of immunomodulatory signaling pathways and functions was much larger in C1 compared to C2. Thus, it seems that immunological features may significantly impact the cancerous nature and bad outcome of ECMGs in PC, according to our findings.

The ideal model was determined to be the one that combined the CoxBoost and SuperPC algorithms since it had the greatest mean C-index (0.706). As shown in Fig. 5, a very dependable prognostic model called DNA METABOLISM GENES was developed after the CoxBoost method, which was used for machine learning, discovered 21 genes of paramount importance. These genes were further optimized using the SuperPC algorithm, which improved the model’s performance.

Evaluation of the clinical features of DNA metabolism genes

In order to investigate how clinicopathological characteristics relate to the prognostic significance of genes involved in DNA metabolism, we evaluated risk scores across several stratification criteria. Except for the GSE70769 cohort, which did not show a correlation between high-risk scores and a higher Gleason score (P < 0.01), the TCGA cohort consistently showed this correlation. Statistically significant changes were seen across all groups for the Gleason score, the sole clinical characteristic. Stratified survival analysis showed that in all cohorts, patients with higher Gleason scores also had poorer bRFS. We can now better understand what went wrong with the high-risk group of DNA metabolism genes, thanks to these findings (Fig. 5).

Gene-based quantitative trait locus (eqtl)

Based on the level of similarity (LD) between the PrCa-risk SNP and the strongest eQTL signal, we classified the six target gene regions into three groups. Figure 6 shows that the classification used both primary and secondary analysis. Using the Pearson correlation coefficient, we assessed the linkage disequilibrium (LD) between the peak eQTL signal, which is the single nucleotide polymorphism most strongly connected with gene expression level, and the PrCa-risk SNP, which is the SNP associated with prostate cancer risk. Statistical analysis was performed to determine the degree of association between each single nucleotide polymorphism (SNP) in the target gene region and the primary expression quantitative trait locus (eQTL) signal SNP. This allowed us to detect the presence of multiple regulatory SNPs on the target genes (Fig. 6).

Landscape of scRNA-seq analysis

On a molecular level, we also looked at up/down gene expression in PC extensively. Forty samples were analyzed using scRNA-seq, with twenty samples representing PC and twenty samples representing normalcy. Following quality control, normalization, and batch effect removal, we discovered 45,122 cells of varying hues. These cells belonged to several samples. Based on the findings of prior studies, we have successfully classified seven distinct cell types using the UMAP method. Due to the uniform distribution of cells from PC and normal samples, batch effect was eliminated. A number of cell types, including as peritubular and epithelial cells, are located differently in PC and normal samples. By analyzing RNA from individual cells, we were able to show that PC is linked to an upregulation of many genes, including REV3L, MSH6, POLD2, RAD9A, MSH3, and WRNIP1 (Figs. 7, 8).

Fig. 8.

Expression of RAD9A, MSH3, WRNIP1, and MSH6 in scRNA-seq in PC. Gene expression of sc-Seq for the marker genes in scRNA-seq in (A) MSH6, (B) WRNIP1, (C) MSH3, and (D) RAD9A.

Predictive value of DNA metabolism genes in therapy

We want to investigate the possibility of these genes in predicting the response to immunotherapy and medicine in order to make the most practical use of our prior research demonstrating that elevated immunological patterns promote aggressive and poor prognosis in PC DNA metabolism genes. A much higher percentage of samples from the low-risk group (21% vs. 12% in the high-risk group) initially tested positive for SPOP mutations. Also, using the IMvigor cohort, an immunotherapy-focused dataset, we discovered that low-risk urothelial cancer patients had longer survival times when we directly built the genes involved in DNA metabolism (P < 0.05). Treatment with anti-PD-L1 was more likely to be effective in patients who routinely scored lower on the risk assessment. Then, TIDE verified that higher risk assessments were associated with a higher probability of immune evasion. When considered as a whole, these findings strongly suggest that immunotherapy has a better chance of benefiting the low-risk group identified by DNA metabolism genes.

Lastly, we compared the IC50 values of 10 commonly used chemotherapeutic and targeted medications to see if the low-risk and high-risk groups were sensitive to any of them. We were able to explore the possibility of using genes related to DNA metabolism in PC therapy selection with more precision and individualization because of this. The fact that seven commonly used anti-PC medications had such wide ranges of sensitivity was unexpected. When comparing the two groups, it was found that paclitaxel, gemcitabine, methotrexate, and cisplatin were more sensitive in the low-risk group, but RDEA119, AR-42, WZ3105, and BX-912 were more sensitive in the high-risk group. The findings add to the growing body of research suggesting that genes involved in DNA metabolism may aid in the selection of more effective personalized drugs and the development of more tailored treatment regimens. Low-risk groups showed better immunotherapy response (e.g., SPOP mutation enrichment, anti-PD-L1 sensitivity). Drug sensitivity analysis revealed distinct chemotherapeutic responses by risk group—e.g., low-risk tumors responded better to paclitaxel and methotrexate, while high-risk tumors were more sensitive to targeted agents (Fig. 9).

Fig. 9.

Predict drugs for purpose docking and drug delivery.

Discussion

The identification of DEGs associated with DNA metabolism in prostate cancer (PC) provides a significant leap in understanding the molecular mechanisms driving cancer progression. The 536 DEGs identified through edgeR analysis, particularly the DNA metabolism genes POLD2, REV3L, RAD9A, MSH3, WRNIP1, and MSH6, highlight their pivotal role in shaping the tumor microenvironment. These genes were shown to influence key processes like homologous recombination, DNA repair, and chromosomal maintenance, all of which are crucial for cancer cell survival and proliferation²⁷. Furthermore, our findings underscore the importance of DNA metabolism genes in the immune landscape of PC, which could have profound implications for immunotherapy. This study provides a comprehensive bioinformatic exploration of DNA metabolism genes in prostate cancer (PC), revealing their influence on tumor behavior, immune modulation, and therapeutic response. Based on integrated analyses, we propose the following collective hypothesis: Dysregulation of DNA metabolism genes defines molecular subtypes of PC with distinct immunological profiles and clinical behaviors, which can inform prognosis and treatment selection.

The integration of multiple machine learning models, such as CoxBoost and SuperPC, in our study enabled the construction of a robust prognostic model. This approach not only allowed us to identify 21 key genes from 134 DNA metabolism genes but also significantly improved the predictive accuracy of the model (C-index = 0.706). These findings align with the work of Lee²⁸ et al., who demonstrated that machine learning approaches are effective in developing prognostic models for cancer based on genomic data. The importance of such models cannot be overstated, as they provide a powerful tool for stratifying patients based on their risk, thus enabling personalized therapeutic strategies. Our study also explored the immune characteristics of PC, revealing a stark contrast between two clusters of patients (C1 and C2). Cluster C1, characterized by higher stromal and immune scores, exhibited enhanced immune activity and worse clinical outcomes compared to C2. This is consistent with recent findings that immune-rich tumors, despite their aggressive behavior, may not always respond favorably to conventional therapies. The association of high immune infiltration with poor prognosis has been previously reported by Smyth et al.²⁹, who showed that certain immune cell populations, such as myeloid-derived suppressor cells (MDSCs), can promote tumor progression rather than inhibit it. Our analysis further supports the notion that immune modulation in PC is complex and requires a deeper understanding to optimize therapeutic interventions.

The significant association between high-risk scores and advanced Gleason scores across multiple cohorts suggests that DNA metabolism genes play a crucial role in determining the aggressiveness of PC. Our results align with the work of Maekawa et al.³⁰, who reported that genomic alterations in DNA repair pathways are frequent in high-grade PC and contribute to tumor progression. Furthermore, the enrichment of homologous recombination and DNA repair pathways in C1 suggests that DNA metabolism genes may serve as valuable biomarkers for therapeutic targeting. Emerging evidence supports the use of DNA repair inhibitors, such as PARP inhibitors, in treating patients with defects in these pathways³¹. Our findings suggest that patients in the C1 cluster, with high DNA repair activity, may benefit from such targeted therapies.

Single-cell RNA sequencing (scRNA-seq) revealed significant upregulation of key DNA metabolism genes, including POLD2 and RAD9A, in PC cells. This is in line with the findings of Horning et al.³², who demonstrated that single-cell transcriptomics can uncover novel cancer cell subpopulations and their associated molecular signatures. The heterogeneity observed in our scRNA-seq data highlights the complexity of PC and emphasizes the need for therapies that can address this intra-tumoral diversity. Our study further contributes to the growing body of literature on the utility of scRNA-seq in cancer research, offering insights into the cellular dynamics that drive PC progression³³. Clustering analyses revealed two molecular subtypes (C1 and C2) with distinct immune microenvironments. C1 tumors were immunologically active yet clinically more aggressive, displaying higher immune infiltration and stromal content. This supports a model where immune-enriched tumors may harbor suppressive cell types or signaling pathways that promote tumor progression despite immune presence. Our findings align with the emerging concept that immune composition, rather than quantity alone, determines prognosis and therapy response.

The potential of DNA metabolism genes in predicting immunotherapy response and drug sensitivity is another key finding of this study. We observed that patients with lower risk scores were more likely to benefit from anti-PD-L1 therapy, a result consistent with the work of Klempner et al.³⁴, who found that immune checkpoint inhibitors are particularly effective in patients with a low tumor mutational burden. Additionally, our drug sensitivity analysis revealed that high-risk patients were more sensitive to specific targeted therapies, such as RDEA119 and AR-42, whereas low-risk patients responded better to traditional chemotherapeutic agents like cisplatin and paclitaxel. This is in line with studies showing that molecular profiling can inform drug selection and enhance treatment outcomes in cancer (Fig. 10)³⁵.

Fig. 10.

Overview of the bioinformatics workflow for identifying and characterizing DNA metabolism genes in prostate cancer.

Drug sensitivity analyses showed distinct patterns based on risk classification. Low-risk tumors were more responsive to immunotherapy and standard chemotherapeutics, while high-risk tumors showed increased sensitivity to targeted inhibitors such as AR-42 and RDEA119. These insights support a model of precision therapy based on DNA metabolism gene signatures, where molecular profiling guides treatment selection. In summary, we propose that DNA metabolism gene dysregulation underlies a functional axis in PC that simultaneously drives genomic instability and shapes the tumor-immune interface. These genes offer promising biomarkers for risk stratification and therapeutic targeting. Future validation in prospective clinical cohorts will be essential to translate these findings into precision oncology applications.

Conclusion

This study set out to investigate the role of DNA metabolism genes in PC progression, prognosis, and therapy response. By integrating bulk transcriptomics, machine learning, and scRNA-seq data analysis, we identified key genes—such as POLD2, RAD9A, and MSH6—that are differentially expressed and associated with immune modulation and clinical outcomes. Our findings revealed two distinct molecular subtypes of PC with different immune profiles and drug sensitivities, underscoring the relevance of DNA metabolism genes as biomarkers for risk stratification and therapeutic decision-making. These insights suggest that targeting DNA repair pathways could benefit high-risk patients, particularly through PARP inhibitors and tailored immunotherapy. While promising, these results warrant validation in larger, independent cohorts to confirm their clinical utility and further refine personalized treatment strategies in PC.

Author contributions

A.S.A.: Machine learning, bioinformatics analysis and writing, M.D.: Bioinformatics analysis, and H.Z.: Responsible for conceptualizing the study and editing the paper. All authors have read and agreed to the published version of the manuscript.

Funding

This work was completed without funding.

Data availability

The data that support the findings of this study are available from the corresponding author upon reasonable request.

Competing interests

The authors declare no competing interests.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-025-11457-1.