Chemotherapy and hormonal therapy play a vital role in the treatment of castration-resistant prostate cancer(CRPC). The effectiveness of current medications is determined by measuring the blood prostate specific antigen (PSA) concentration after the fourth cycle of treatment (after 3 months) during the medication process. If this drug is ineffective for the patient, the patient would have to bear the financial burden of expensive anti-cancer drugs and the treatment process with ineffective drugs. In previous studies, miRNAs, especially in exosome, have been demonstrated to be helpful in the diagnosis and prognosis of prostate cancer patients. In this study, we aim to find out the critical miRNA to be a drug efficacy predictor before treatment, and explore the potential to incorporate machine learning to generate a prediction model.

Blood specimens were prospectively collected from 40 CRPC patients before and 6 weeks after either Docetaxel, Enzalutamide or Abiraterone treatment. Exosome isolation was performed and miRNA was isolated from both plasma and exosome samples. Reverse transcription of miRNA was then performed followed by Quantitative Real-Time PCR (qPCR). Quantitative results of miRNA were then matched with patient treatment and drug efficacy for further analysis. Machine learning was then introduced and a model was derived from out training dataset.

7 miRNAs, labelled miRNA-A through miRNA-G, were selected after literature review with an additional 3 used as reference and controls. 40 patients were included with 24 of whom responding to treatment. Higher expression change of miRNA-F was noted in exosome samples of patients responding to Docetaxel, Enzalutamide or Abiraterone treatment, while similar trends of miRNA-B, C and D were also noted in patients effectively treated with Enzalutamide or Abiraterone. Using the post-treatment expression level of these 4 miRNAs, a Linear Discriminant model could be used to predict drug effectiveness with an accuracy of 90.32% in the training set and 85.71% in the testing set. Ensemble Subspace Discriminant Linear SVM algorithm, a model to predict the drug effectiveness was obtained with an accuracy of 87.1% in the training set, and 85.71% in the testing set.

7 CRPC-associated miRNAs in plasma and exosome were quantified by qPCR and further analyzed by machine learning in this study. Among these 7 miRNAs, 4 could be used to determine therapeutic effectiveness of CRPC treatment, with the Linear Discriminant model yielding an accuracy of 90.32% in the training set and 85.71% in the testing set. More specimens’ verification could further improve the accuracy of the model.