ReactionAtlas maps 47,000 chemical reactions using ML to explore origins of life

ReactionAtlas is a new machine learning framework that automatically maps chemical reaction networks “ab origine” — starting from just a handful of seed molecules without any hand-crafted rules. Developed by Stefan Gugler, Max Eissler, and colleagues, the system uses a generative model to propose reactions from kinetically sampled candidate compounds, then filters them through a DFT-trained machine-learned force field (MLFF) to validate transition states. Products from valid reactions re-enter the search as new seeds, allowing the network to grow organically. Starting with eight pre-biotic molecules (CH2O, H2O, OH-, H3O+, CO2, H2CO3, HCO3-, H), ReactionAtlas uncovered approximately 47,000 elementary reactions connecting about 12,000 distinct chemical compounds.

The MLFF’s transition state predictions match expensive PBE0 density functional theory references within 0.5 Å RMSD in 85% of cases, and can be easily refined to full PBE0 accuracy. This enables mapping of small carbohydrate chemistry up to C4H8O4 with unprecedented scale and detail, including charge and stereo information. The framework already provides novel insights into well-studied pathways like the formose cycle, which is central to the chemical origins of life, and reveals alternative reaction pathways previously unknown. ReactionAtlas promises to accelerate research in catalysis, combustion, and prebiotic chemistry by automating a task that was previously impractical with traditional methods.