Current drug development processes are slow, costly, and high-risk, posing challenges for advancing safe and effective therapeutics. Here, we present a strategy using high-throughput proteomics to guide the development of androgen receptor (AR)-targeting heterobifunctional degraders (HBDs), a promising but yet-to-be-approved modality for treatment-resistant prostate cancer. We generated a chemically diverse HBD library and leveraged the absence of AR expression in human liver cells to assess off-target effects. In the absence of AR, HBDs induced proteomic responses which reflected degrader chemistries and revealed the off-target toxicity mechanism—respiratory chain complex I inhibition. A proteome-trained machine learning model identified candidates that effectively degraded AR while avoiding off-target interactions. These optimized candidates exhibited reduced hepatotoxicity, enhanced prostate cancer selectivity, and delayed tumor formation in treatment-resistant xenograft models. Our findings pave the way for safer AR degraders and provide a blueprint for proteome-guided drug development.