Drug-target interaction (DTI) databases comprise millions of manually curated data points, yet there are missed opportunities for repurposing established interaction networks to infer DTIs. To address this gap, we first collected DTIs on 128 unique G protein-coupled receptors across 187K molecules to establish an all-vs-all chemical space network. We next developed a chemical space neural network (CSNN), which operates on the graph structure of chemical space rather than on the graphs of compounds, to infer drug bioactivity classes with up to 98% accuracy. We combined this virtual library screen with a cost-efficient experimental platform to validate our predictions and discovered 14 novel DTIs in the process. Altogether, our platform integrates virtual library screening and experimental validation for fast and efficient coverage of missing DTIs.