The paper introduces POLYGON, a deep generative chemistry approach for designing polypharmacology drugs, which are compounds that target multiple proteins. POLYGON uses a variational autoencoder (VAE) to embed chemical space and a reinforcement learning system to iteratively sample and reward molecular structures based on their predicted ability to inhibit two protein targets, drug-likeness, and ease of synthesis. The model achieves 82.5% accuracy in recognizing polypharmacology interactions among 100,000 compounds. The authors generate de novo compounds targeting ten pairs of co-dependent proteins and validate 32 compounds targeting MEK1 and mTOR, showing >50% reduction in kinase activity and cell viability at 1-10 μM. These results demonstrate the potential of generative modeling in polypharmacology.The paper introduces POLYGON, a deep generative chemistry approach for designing polypharmacology drugs, which are compounds that target multiple proteins. POLYGON uses a variational autoencoder (VAE) to embed chemical space and a reinforcement learning system to iteratively sample and reward molecular structures based on their predicted ability to inhibit two protein targets, drug-likeness, and ease of synthesis. The model achieves 82.5% accuracy in recognizing polypharmacology interactions among 100,000 compounds. The authors generate de novo compounds targeting ten pairs of co-dependent proteins and validate 32 compounds targeting MEK1 and mTOR, showing >50% reduction in kinase activity and cell viability at 1-10 μM. These results demonstrate the potential of generative modeling in polypharmacology.