Enhancing prostate cancer diagnosis: a machine learning-based biomarker approach.

Abstract

BACKGROUND: The lack of reliable screening biomarkers for prostate cancer (PC) diagnosis makes tissue biopsy the gold-standard strategy. However, its frequent inconclusive results often lead to repeated procedures, increasing patient burden and healthcare costs. In this context, machine learning (ML) presents a novel approach for identifying gene signatures associated with tumor presence, offering a promising avenue for improved PC detection. Therefore, the present study explored the diagnostic potential of an ML-based gene signature and its applicability in tissue and plasma samples. METHODS: This study evaluated the clinical applicability of an ML-based algorithm developed using TCGA data (n = 608) and tested in an independent dataset (n = 349). The eleven candidate genes that contributed most to the predictive model were initially profiled by targeted RNA-Seq in prostate tissue validation cohort (n = 141) and further validated in replication tissue (n = 75) and plasma (n = 50) cohorts by qPCR and dPCR, respectively. RESULTS: Gene expression analysis in prostate tissue led to the identification of a six-gene signature (DLX1, TDRD1, AMACR, HPN, HOXC6, and OR51E2) with high diagnostic performance (AUC = 95.9%). Expression patterns supported the gene signature’s potential to identify false-negative cases and correctly classify inconclusive biopsy results. In plasma, AMACR demonstrated added diagnostic value as a non-invasive biomarker when integrated with clinical parameters (AUC = 93.21%). CONCLUSIONS: These findings demonstrate that our ML-based gene signature can accurately distinguish PC from non-tumoral tissue and resolve ambiguous biopsy results. Its integration alongside histopathology has the potential to reduce diagnostic uncertainty, improving PC early detection and guiding clinical decision-making.