Machine Learning Based Androgen Receptor Regulatory Gene-related Random Forest Survival Model for Precise Treatment Decision in Prostate Cancer

Background: It has been demonstrated that aberrant androgen receptor (AR) signaling contributes to the pathogenesis of prostate cancer (PCa). To date, the most efficacious strategy for the treatment of PCa remains to target the AR signaling axis. However, numerous PCa patients still face the issue of overtreatment or undertreatment. The establishment of a precise risk prediction model is urgently needed to distinguish patients with high-risk and select appropriate treatment modalities.

Methods: In this study, a consensus AR regulatory gene-related signature (ARS) was developed by integrating a total of 101 algorithm combinations of 10 machine learning algorithms. We evaluated the value of ARS in predicting patient prognosis and the therapeutic effects of the various treatments. Additionally, we conducted a screening of therapeutic targets and agents for high-risk patients, followed by the verification in vitro and in vivo.

Results: ARS was an independent risk factor for biochemical recurrence and distant metastasis in PCa patients. The enhanced and consistent prognostic predictive capability of ARS across various platforms was confirmed when compared with 44 previously published signatures. More importantly, PCa patients in the ARS group benefit more from PARP inhibitors and immunotherapy, while chemotherapy, radiotherapy, and AR-targeted therapy are more effective for ARS patients. The results of in silico screening suggest that AURKB could potentially serve as a promising therapeutic target for ARS patients.

Conclusions: Collectively, this prediction model based on AR regulatory genes holds great clinical translational potential to solve the dilemma of treatment choice and identify potential novel therapeutic targets in PCa.