The next step in our global fight against coronavirus diseases is to develop new therapeutic strategies with broad antiviral activity in preparation for future strains or severe variants. Presumably, host-based proteins may offer an alternative targeting approach with a high barrier to drug resistance. Here, we identify novel COVID-19 host targets and drug repurposing candidates, using two complementary machine learning technologies. Non-obvious, host-based targets were first identified using a Graph Convolutional Network (GCN) trained to model clinically relevant network proximity distances in a multiscale interactome. The multiscale interactome combines relationships between genes, proteins, drugs, biological pathways and disease, including viral-host protein-protein interactions networks. Additional drug repurposing candidates were then retrieved from PolypharmDB for several GCN-identified host targets. PolypharmDB is a database of 10,244 clinically-tested drugs cross-screened with 8535 human proteins, based on a deep learning Drug Target Interaction (DTI) prediction model. Twenty-six FDA-approved drugs identified by this screen were selected for cellular infectivity assays, of which four had demonstrable bioactivities. One notable hit demonstrated potent antiviral activity against five genetically different human coronaviruses, while its predicted target was also confirmed by siRNA gene silencing. Together, we present a promising drug repurposing candidate and a new therapeutic target for future drug design programs.