ReactionAtlas Maps Chemical Networks with Machine Learning

A new machine learning framework called ReactionAtlas has been developed to automatically map complex chemical reaction networks from scratch. Traditionally, constructing these networks, which involve identifying minima, transition states, and elementary reactions, has been computationally prohibitive due to the vast number of transition states requiring characterization by slow methods like Density Functional Theory (DFT). ReactionAtlas overcomes this by using a generative model to propose reactions from initial seed molecules, eliminating the need for pre-defined rules. The framework employs a DFT-trained machine-learned force field (MLFF) to efficiently filter these proposed reactions, identifying valid transition states. The products of these valid reactions then become new seeds, allowing for an iterative and expansive exploration of chemical space. Starting with just eight pre-biotic molecules, ReactionAtlas successfully discovered approximately 47,000 reactions among 12,000 compounds. This achievement provides unprecedented scale and accuracy for mapping small carbohydrate chemistry, including detailed charge and stereo information. The MLFF-identified transition states closely match high-fidelity references, and the framework offers novel insights into well-studied pathways like the formose cycle, which is central to the chemical origins of life.