Prostate cancer (PCa) is the second most common cancer in men worldwide. The current primary screening tool, the prostate-specific antigen (PSA) test, has well-documented shortcomings: it has a high false-positive rate, cannot reliably distinguish aggressive cancers from slow-growing ones, and its levels are elevated by benign conditions like prostatic hyperplasia and prostatitis. Reducing the PSA threshold to catch more cancers only worsens overdiagnosis and the harms of unnecessary treatment.
There is growing recognition that prostate cancer develops within a biological context that includes metabolic comorbidities. Diabetes mellitus (T2DM) and obesity are both highly prevalent conditions in men who develop prostate cancer, and epidemiological evidence suggests complex relationships between these conditions. For instance, insulin resistance stimulates insulin growth factor (IGF-1) pathways that promote cancer cell proliferation, and excess adipose tissue creates a pro-inflammatory environment that may facilitate tumor development. Understanding the shared genetic underpinnings of PCa, diabetes, and obesity could point toward new, more specific biomarkers.
This study took a bioinformatics-driven approach: mining publicly available genetic variant databases, published genomic sequencing datasets, and gene expression resources to identify genes shared among prostate cancer, diabetes, and obesity. The hypothesis was that genes whose variants appear in multiple disease contexts may represent central molecular players -- and that machine learning applied to clinically annotated genetic data could build a probabilistic risk prediction model to support clinical decision-making.
The study drew on multiple publicly available genomic databases. ClinVar (from NCBI) was used to retrieve catalogues of clinically verified genetic variants for prostate cancer, diabetes, and obesity, each variant classified by pathogenicity (unknown significance, likely benign, benign, likely pathogenic, or pathogenic). The GWAS Central database and the PRACTICAL consortium (an international prostate cancer GWAS consortium) provided genome-wide association study data for overlapping analysis. Published RNA-seq datasets for prostate cancer and Type 2 Diabetes Mellitus, as well as whole exome sequencing data, were also included in the cross-comparison.
A key methodological step was the conversion of clinical variant data into semi-binary numerical scores: variants were encoded as -1 (unknown significance), -0.5 (likely benign), 0 (benign), 0.5 (likely pathogenic), or 1 (pathogenic). This encoding transformed heterogeneous clinical annotations into numerical features suitable for machine learning regression, with higher scores corresponding to higher pathogenic risk. These encoded datasets were then analyzed using five regression algorithms implemented in the Weka machine learning platform: Linear Regression, Multilayer Perceptron (MLP), Random Forest, Random Tree, and REPTree.
Overlapping genes across disease datasets were identified using Venn plot analysis. Gene interaction networks for common candidates were visualized using GeneMANIA, which extends a seed gene list to incorporate predicted interaction partners based on shared biological pathways, co-expression, and physical protein-protein interactions. cBioPortal was used to examine mutation frequency and type in prostate adenocarcinoma, and GEPIA 2 (Gene Expression Profiling Interactive Analysis) was used to compare expression levels between tumor and normal prostate tissue using TCGA and GTEx datasets.
Venn plot analysis of ClinVar data identified 27 genes common between prostate cancer and diabetes, 8 common between prostate cancer and obesity, and 69 shared between diabetes and obesity. Four genes were common to all three conditions: BLM, TMEM67, RFX6, and NUDC. Among these, BLM (Bloom Syndrome Protein) -- a DNA helicase involved in DNA repair, cell cycle control, and genomic stability -- emerged as the most significant shared candidate across all three diseases in multiple analyses.
Comparing RNA-seq data from prostate cancer and Type 2 Diabetes Mellitus datasets identified PPM1B and SFTPC as shared genes. GeneMANIA network analysis of PCa-diabetes common genes identified TMPO as a common interacting partner, while FOXP1 -- a transcription factor involved in B-cell development, lung epithelial differentiation, and inhibition of apoptosis -- appeared as a common interaction partner of PPM1B, SFTPC, and TMEM67. BLM was found to interact physically with FOXP1 and show co-expression with TMPO, positioning it as a central network hub connecting the PCa-diabetes-obesity axis.
GWAS cross-comparison confirmed two variants -- rs721048 in EHBP1 and rs138213197 in HOXB13 -- as consistently appearing in both ClinVar PCa data and GWAS Central prostate cancer datasets. BRCA1, BRCA2, and FANCM variants were also found overlapping between PCa ClinVar and GWAS datasets, consistent with their known roles in DNA damage repair pathways that are mutated in hereditary prostate cancer. These variants provide genetic anchor points connecting the bioinformatic findings to clinically established cancer risk genes.
The five machine learning regression algorithms applied to the semi-binary encoded ClinVar datasets showed broadly similar performance. For the prostate cancer dataset, linear regression, random forest, random tree, and REPTree all achieved nearly identical RMSE (root mean square error) values of approximately 0.54, while multilayer perceptron performed slightly worse (RMSE 0.543). For diabetes, linear regression had the lowest RMSE (0.5975), while for obesity, tree-based algorithms outperformed the others. RMSE measures the average deviation of model predictions from actual values -- lower indicates better accuracy.
The similar RMSE values across most algorithms for the prostate cancer dataset suggest that the predictive signal in the semi-binary encoded variant data is consistently captured by multiple algorithms, and that no single algorithm provides a large advantage. The study acknowledges that these results are preliminary -- derived from a pilot-scale dataset -- and that including more independent variables (additional clinical and genetic parameters) would likely improve predictive performance substantially.
cBioPortal mutation frequency analysis of the candidate genes in prostate adenocarcinoma showed that FOXP1 had the highest alteration frequency at 8% (including amplification, deep deletions, and missense mutations), followed by TMEM67 at 6%, SFTPC at 5%, and PPP2CA at 1.8%. BLM showed 0.3% mutations including missense variants and truncations classified as putative drivers -- a lower frequency but with biologically meaningful mutation types that may drive cancer rather than being passenger events. TMPO showed only 0.1% truncation mutations.
Gene expression analysis in GEPIA 2 compared expression levels of the candidate genes between 492 prostate adenocarcinoma tumor samples and 152 normal prostate tissue samples from the TCGA PRAD dataset. PPP2CA (a phosphatase that regulates multiple cancer-relevant targets including MYC and FOXO3) and FOXP1 showed the highest expression levels relative to normal tissue, with Z-scores above 4.9. PPARG and TMEM67 showed lower expression, while TMPO, SFTPC, and PPM1B showed the least change from normal tissue.
Kaplan-Meier survival analysis identified two genes with statistically significant associations with patient outcomes. FOXP1 was associated with disease-free survival (log-rank p = 0.005), and BLM showed the most significant association with disease-free survival (log-rank p = 0.00065). These results suggest that high BLM expression may be associated with worse or better disease-free outcomes in prostate cancer patients -- a finding consistent with BLM's dual role in maintaining genome stability (protective) and potentially enabling cancer cell survival under replication stress.
The survival significance of both FOXP1 and BLM, combined with their central positions in the interaction network connecting PCa to diabetes and obesity comorbidities, makes them the most compelling candidates from this analysis for follow-up functional validation. FOXP1 is of particular interest given its anti-apoptotic function in B cells and its known overexpression in several cancer types, while BLM's role in DNA damage repair pathways overlaps with established prostate cancer risk genes like BRCA1 and BRCA2.
This pilot study demonstrates a framework for identifying candidate prostate cancer biomarkers by integrating machine learning with multi-source genomic data analysis spanning ClinVar variants, GWAS datasets, RNA-seq expression data, and whole exome sequencing. The analysis consistently identified BLM as the only gene common to prostate cancer, diabetes mellitus, and obesity, interacting with FOXP1 and TMPO in gene networks, and showing significant association with disease-free survival in prostate cancer patients.
The machine learning models, while preliminary in their current form, establish a proof-of-concept that semi-binary encoded clinical variant data can be used in regression algorithms to predict cancer risk probability. The authors emphasize that broader clinical deployment would require more extensive variable inclusion, prospective validation, and larger patient cohorts. They position this as a first step toward precision scale modeling for prostate cancer prognosis.
The broader landscape of prostate cancer biomarker discovery is evolving rapidly. FDA-approved tests now include PSA, PHI (Prostate Health Index), and PCA3, while tests like the Decipher, Oncotype DX, and Prolaris gene panels offer tissue-based risk stratification. Emerging biomarkers from liquid biopsy approaches (circulating tumor cells, microRNAs, exosomes) remain in earlier validation stages. The genes identified in this study -- particularly BLM and FOXP1 -- represent computational candidates that could contribute to multi-gene panels or serve as targets for further functional and clinical validation studies.