Integrative machine learning models predict prostate cancer diagnosis and biochemical recurrence risk: Advancing precision oncology

Abstract

Prostate cancer (PCa) ranks among the most prevalent cancers in men worldwide. Biochemical recurrence (BCR) presents a major clinical challenge in PCa management, with significant prognostic heterogeneity observed among patients post-recurrence. This study aimed to develop machine learning models for predicting both the diagnosis and prognosis of PCa patients. Using WGCNA, we initially identified 16 BCR-related target genes. Cluster analysis revealed these genes were significantly associated with PCa prognosis, drug sensitivity, and immune infiltration. We constructed a robust diagnostic model integrating multiple machine learning algorithms, demonstrating strong predictive capability for PCa. Furthermore, a BCR-related prognostic model built using the LASSO algorithm also yielded satisfactory performance. Among the differentially expressed BCR-associated prognostic genes, COMP emerged as a critical regulatory factor. Both in vitro and in vivo experiments confirmed COMP's role in influencing PCa progression. Additionally, COMP demonstrates significant potential as a dual biomarker for both the diagnosis and recurrence prediction of PCa.

Fig. 1. Identification of 16 BCR-associated genes based on WGCNA.

A Plot scale independence. B Plot average connectivity. C Draw sample clusters. D Dividing the GSE116918 sample into 11 modules. E Module-phenotype correlation heat map. F Scatter plot of correlation between gene significance and module membership. G Heatmap of BCR-related gene expression in TCGA-PRAD. H Prognostic forest plot of BCR-related genes in TCGA-PRAD. I Heatmap of correlation of biochemical relapse-related genes in TCGA-PRAD. J Differential expression of BCR-related genes in BCR and non-BCR samples in GSE116918.

Fig. 2. BCR-related genes play an important role in PRAD.

A KEGG and GO analysis of BCR-related genes in PRAD. B GSCA database-based analysis of BCR-associated gene function in PRAD. C Analysis of the correlation between BCR related genes and the frequency of copy number variants. D–F Differential expression of BCR-related genes in different pathological stages of TCGA-PRAD. ***p < 0.001.

Fig. 3. BCR-related genes are significantly associated with prognostic and pathologic parameters in PRAD patients.

A, B Consensus map of NMF clustering. C Survival differences between clusters. D Differences in the expression of BCR related genes between different clusters. E–H Differences in the distribution of different subgroups in different pathological stages of PRAD. ****p < 0.0001.

Fig. 4. BCR-related genes are strongly associated with immune infiltration in PRAD.

A, B Analysis of BCR-related genes correlating with PRAD immune infiltration. C Heat map of different immune cell infiltration levels. D Analysis of biochemical relapse-related genes and chemotherapy sensitivity in PRAD patients. E Gene enrichment analysis of two clusters. ^ns på 0.05, *p < 0.05; **p < 0.01; ***p < 0.001; ****p < 0.0001.

Fig. 5. The LASSO + LDA algorithm combination is considered to be the best combination for constructing diagnostic models.

A An examination of how biologic recurrence-related genes predict diagnosis in PCa patients. B AUC values for diagnostic models created using various algorithm combinations. C The count of genes included in diagnostic models that were formed with different algorithm combinations.

Fig. 6. Constructing BCR-related prognostic models.

A Screening of prognosis-related genes by univariate Cox analysis. B Genes were screened through lasso/stepwise regression, and a model was constructed using multivariate Cox analysis. C ROC curve analysis of the model’s predictive ability for BCR-related prognosis in the entire dataset. D The Kaplan-Meier survival curves for the high-risk and low-risk groups in the entire dataset. E Risk factor plot of the test dataset. F Risk factor plot of the train dataset. G ROC curve analysis of the model’s predictive ability for BCR-related prognosis in patients within the test dataset. H The Kaplan-Meier survival curves for the high-risk and low-risk groups in the test dataset. I ROC curve analysis of the model’s predictive ability for BCR-related prognosis in patients within the train dataset. J The Kaplan-Meier survival curves for the high-risk and low-risk groups in the train dataset.

Fig. 7. COMP is identified as a key regulatory gene for BCR.

A, B Xgboost algorithm identifies key regulatory genes for BCR. C Friend analysis of dominant genes in BCR-regulated genes. D, E Correlation analysis of COMP with the level of immune cell infiltration. F Molecular docking of COMP with drugs. G, H Functional analysis of COMP in PRAD. I Correlation Analysis Between COMP and Immunotherapy. *p < 0.05; **p < 0.01; ***p < 0.001.

Fig. 8. COMP is highly expressed in PRAD.

A–C The expression of COMP in normal prostate tissue. D–F Expression of COMP in non-recurrence PRAD tissues. G–I Expression of COMP in recurrence PRAD tissues. J Differences in COMP expression between normal prostate tissue and PRAD. K The predictive value of COMP in the diagnosis of PRAD patients. L The expression difference of COMP in recurrence and non-recurrence PRAD samples. M The predictive value of COMP for recurrence in PRAD patients. *p < 0.05; ***p < 0.001.

Fig. 9. Knockdown of COMP can inhibit the proliferation and metastasis of PCa.

A Verification of COMP knockdown efficiency in PCa cells. B, C Colony formation assay used to evaluate the effect of COMP on PCa cell proliferation. D, E Transwell assays to assess the effect of COMP on the metastatic ability of PCa cells. F Tumor modeling in male BALB/c nude mice. G, H Subcutaneous tumor volume change curve. I Final subcutaneous tumor weight. J Final subcutaneous tumor size. K, L IHC staining images of COMP and Ki-67 in subcutaneous tumors. M, N Measurement of lung metastasis luciferase activity via in vivo imaging system. **p < 0.01; ***p < 0.001.