Prognostic and Predictive Value of Machine Learning-Based Biomarker and Pathomics Signatures in Patients With Prostate Cancer

Abstract

Recurrence and the potential development of castration resistance after radical prostatectomy (RP) are significant challenges in the management of prostate cancer (PCa). Despite the development of advanced prognostic models, few have been clinically applied. Five machine learning algorithms (LASSO, RSF, SVM-RFE, Boruta, and XGBoost) were used to identify biomarkers for PCa using transcriptome data from multicenters (TCGA, MSKCC, DKFZ, and GSE70770) for constructing and validating the metastasis-associated prognostic risk score (MAPRS), which revealed the molecular biological heterogeneity and was confirmed with in-house histopathological samples. The pathomics score (PSpc), derived from a machine learning framework (XGBoost, RSF, GBM, plsRCox, CoxBoost, Enet, Ridge, LASSO, SVM, and superPC) using hematoxylin and eosin (H&E)-stained digital pathology, quantified tumor morphological heterogeneity. The MAPRS correlated with poorer recurrence-free survival (RFS) and was associated with the tumor microenvironment and pathogenic variants. A higher MAPRS may indicate sensitivity to treatments such as PARP inhibitors, docetaxel, and oxaliplatin. Pathology-based evaluations of MAPRS, PSpc, and their combination effectively predicted RFS in patients who underwent RP. MAPRS also predicted progression-free survival in patients receiving androgen deprivation therapy when combined with clinical indicators, whereas PSpc demonstrated limited efficacy. The digital pathology-based signatures showed superior predictive efficacy compared to other tools.

Trial Registration: Chinese Clinical Trial Registry number: ChiCTR2400085748 (June 18, 2024).