The paper introduces Reaction-GFlowNet (RGFN), an extension of the GFlowNet framework that generates small molecules by combining basic chemical fragments through a chain of reactions. RGFN operates directly in the space of chemical reactions, ensuring that the generated molecules are synthesizable while maintaining high quality. The authors propose a curated set of low-cost chemical building blocks and high-yield reactions, which together create a search space orders of magnitude larger than existing screening libraries. They demonstrate that RGFN can produce a diverse set of molecules with low synthesis costs, outperforming existing methods in terms of synthesizability and optimization quality. The effectiveness of RGFN is evaluated across various tasks, including docking score approximation, GPU-accelerated docking, and biological activity estimation. The paper also discusses the limitations and future directions for improving the approach, such as increasing the diversity and complexity of the generated molecules and enhancing the accuracy of scoring methods.The paper introduces Reaction-GFlowNet (RGFN), an extension of the GFlowNet framework that generates small molecules by combining basic chemical fragments through a chain of reactions. RGFN operates directly in the space of chemical reactions, ensuring that the generated molecules are synthesizable while maintaining high quality. The authors propose a curated set of low-cost chemical building blocks and high-yield reactions, which together create a search space orders of magnitude larger than existing screening libraries. They demonstrate that RGFN can produce a diverse set of molecules with low synthesis costs, outperforming existing methods in terms of synthesizability and optimization quality. The effectiveness of RGFN is evaluated across various tasks, including docking score approximation, GPU-accelerated docking, and biological activity estimation. The paper also discusses the limitations and future directions for improving the approach, such as increasing the diversity and complexity of the generated molecules and enhancing the accuracy of scoring methods.