Automating Deep Learning-Based Generation and Evaluation of De Novo Chemical Reaction with ChemRxnSAGE

The generation and evaluation of chemical reactions remain challenging with limited comprehensive studies addressing these issues. We introduce the Chemical Reaction (Rxn) Systematic Assessment of Generation and Evaluation (ChemRxnSAGE) framework, an adaptable end-to-end approach for evaluating the quality, validity, and diversity of machine-generated chemical reactions. Combining automated validity filters with quality metrics and expert insights, ChemRxnSAGE systematically eliminates invalid reactions. We tested its robustness using generative models, including Recurrent Neural Networks and Variational Autoencoders, followed by validation using a chemical "Turing test" with domain experts. Additionally, we assess reaction feasibility through thermodynamic analysis and compare the generated reactions against existing literature to ensure relevance and novelty. By combination of computational tools with expert-driven metrics, ChemRxnSAGE offers a comprehensive and extendable solution that advances the state of chemical reaction generation and evaluation.