The article introduces POLYGON, a deep generative chemistry model that enables the de novo design of multi-target compounds. POLYGON uses generative reinforcement learning to generate molecular structures that can inhibit multiple proteins, with rewards based on predicted inhibitory activity, drug-likeness, and synthesizability. The model was trained on a large dataset of small molecules and validated using binding data for over 100,000 compounds, achieving 82.5% accuracy in recognizing polypharmacology interactions. POLYGON was used to generate compounds targeting ten pairs of synthetically lethal cancer proteins, with docking analysis showing that top compounds bind their targets with low free energies and similar 3D orientations to canonical inhibitors. 32 compounds targeting MEK1 and mTOR were synthesized, with most showing >50% reduction in protein activity and cell viability at 1–10 µM. These results support the potential of generative modeling for polypharmacology. The study highlights the challenges of designing polypharmacology compounds and the potential of machine learning in this area. POLYGON addresses the initial phases of polypharmacology design, and further optimization can proceed through classical techniques such as structure-activity relationships. The study also discusses opportunities to improve the generative capacity of the POLYGON algorithm for dual protein inhibitors.The article introduces POLYGON, a deep generative chemistry model that enables the de novo design of multi-target compounds. POLYGON uses generative reinforcement learning to generate molecular structures that can inhibit multiple proteins, with rewards based on predicted inhibitory activity, drug-likeness, and synthesizability. The model was trained on a large dataset of small molecules and validated using binding data for over 100,000 compounds, achieving 82.5% accuracy in recognizing polypharmacology interactions. POLYGON was used to generate compounds targeting ten pairs of synthetically lethal cancer proteins, with docking analysis showing that top compounds bind their targets with low free energies and similar 3D orientations to canonical inhibitors. 32 compounds targeting MEK1 and mTOR were synthesized, with most showing >50% reduction in protein activity and cell viability at 1–10 µM. These results support the potential of generative modeling for polypharmacology. The study highlights the challenges of designing polypharmacology compounds and the potential of machine learning in this area. POLYGON addresses the initial phases of polypharmacology design, and further optimization can proceed through classical techniques such as structure-activity relationships. The study also discusses opportunities to improve the generative capacity of the POLYGON algorithm for dual protein inhibitors.