Identification of Potential Biomarkers in Prostate Cancer Microarray Gene Expression Leveraging Explainable Machine Learning Classifiers.

Background and Objective: Prostate cancer remains one of the most prevalent and potentially lethal malignancies among men worldwide, and timely and accurate diagnosis, along with the stratification of patients by disease severity, is critical for personalized treatment and improved outcomes for this cancer. One of the tools used for diagnosis is bioinformatics. However, traditional biomarker discovery methods often lack transparency and interpretability, which means that clinicians find it difficult to trust biomarkers for their application in a clinical setting. Methods: This paper introduces a novel approach that leverages Explainable Machine Learning (XML) techniques to identify and prioritize biomarkers associated with different levels of severity of prostate cancer. The proposed XML approach presented in this study incorporates some traditional machine learning (ML) algorithms with transparent models to facilitate understanding of the importance of the characteristics for bioinformatics analysis, allowing for more informed clinical decisions. The proposed method contains the implementation of several ML classifiers, such as Naive Bayes (NB), Random Forest (RF), Decision Tree (DT), Support Vector Machine (SVM), Logistic Regression (LR), and Bagging (Bg); followed by SHAPly values for the XML pipeline. In this study, for pre-processing of missing values, imputation was applied; SMOTE (Synthetic Minority Oversampling Technique) and the Tomek link method were applied to handle the class imbalance problem. The k-fold stratified validation of machine learning (ML) models and SHAP values (SHapley Additive explanations) were used for explainability. Results: This study utilized a novel tissue microarray data set that has 102 patient data comprising prostate cancer and healthy patients. The proposed model satisfactorily identifies genes as biomarkers, with highest accuracy obtained being 81.01% using RF. The top 10 potential biomarkers identified in this study are DEGS1, HPN, ERG, CFD, TMPRSS2, PDLIM5, XBP1, AJAP1, NPM1 and C7. Conclusions: As XML continues to unravel the complexities within prostate cancer datasets, the identification of severity-specific biomarkers is poised at the forefront of precision oncology. This integration paves the way for targeted interventions, improving patient outcomes, and heralding a new era of individualized care in the fight against prostate cancer.