Ligand-based drug design methods are thought to require large experimental datasets to become useful for virtual screening. In this work, we propose a computational strategy to design novel inhibitors of coronavirus main protease, M
pro
. The pipeline integrates publicly available screening and binding affinity data in a two-stage machine-learning model using the recent MACAW embeddings. Once trained, the model can be deployed to rapidly screen large libraries of molecules
in silico
. Several hundred thousand compounds were virtually screened and 10 of them were selected for experimental testing. From these 10 compounds, 8 showed a clear inhibitory effect on recombinant M
pro
, with half-maximal inhibitory concentration values (IC
50
) in the range 0.18–18.82 μM. Cellular assays were also conducted to evaluate cytotoxic, haemolytic, and antiviral properties. A promising lead compound against coronavirus M
pro
was identified with dose-dependent inhibition of virus infectivity and minimal toxicity on human MRC-5 cells.