Deep learning assisted identification of ATP5J and ALDH1A2 combination in RNA-sequencing data as a novel specific potential diagnostic biomarker in prostate cancer

Abstract

Abstract Prostate cancer is ranked as the second most prevalent disease among men globally. The timely diagnosis of this cancer is crucial in reducing morbidity rates. Unfortunately, due to the limitations of current diagnostic methods, which often lack specificity and accuracy, prostate cancer is frequently diagnosed at advanced stages, leading to less effective treatment strategies. Therefore, our primary objective was to identify valuable diagnostic biomarkers through the application of bioinformatics and artificial intelligence. To achieve this goal, we utilized three prostate cancer expression datasets, aiming to pinpoint differentially expressed genes (DEGs) associated with prostate cancer. Subsequently, we harnessed deep learning, a subset of artificial intelligence, to unveil the most significant genes from the pool of 3875 common DEGs implicated in prostate cancer's pathogenesis. The deep learning model's performance was evaluated using six key metrics: Mean Squared Error (MSE) with a value of 0.03, R-squared (R²) at 0.83, Area Under the Curve (AUC) of 0.97, Accuracy at 87.7%, Root Mean Squared Error (RMSE) of 0.18, and Precision-Recall AUC (PR-AUC) at 0.93, demonstrating the model's exceptional performance. Furthermore, gene enrichment analysis shed light on ten candidate genes with pivotal roles in prostate cancer development. Additionally, Protein-Protein Interaction (PPI) network analysis revealed ATP5J, GJA1, AMACR, and B3GAT1 as hub genes, with AMACR and B3GAT1 exhibiting an intriguing interaction. Further validation through Receiver Operating Characteristic (ROC) analysis of the ten key genes identified by deep learning unveiled ATP5J, ALDH1A2, and AMACR as promising diagnostic biomarkers for prostate cancer. Notably, the combined use of ATP5J and ALDH1A2 demonstrated remarkable accuracy, with an accuracy rate of 0.75, sensitivity of 0.73, and specificity of 0.71, comparable to common prostate cancer diagnostic biomarkers such as PSA, PCA3, and PHI. The validation of these biomarkers in prostate cancer was carried out using the PCaDB database, lending support to the potential clinical utility of these markers. In conclusion, our findings underscore the importance of further research and validation to establish the clinical applicability of ATP5J and ALDH1A2 as promising diagnostic biomarkers for prostate cancer, offering a potential breakthrough in early detection and more effective management of this prevalent disease.