ReactionAtlas: Ab origine exploration of chemical reaction networks with machine learning

Short summary

ReactionAtlas uses machine learning to discover chemical reaction networks from seed molecules without hand-crafted rules, discovering ~47,000 reactions among 12,000 compounds with 85% accuracy. The system combines a generative ML model with a DFT-trained force field to identify valid transition states. Results illuminate carbohydrate chemistry and the formose cycle, with implications for prebiotic chemistry.

• •ML-based system discovers 47,000 reactions from 8 seed molecules without manual rules
• •Achieves 85% accuracy matching DFT references, scales to carbohydrate chemistry
• •Reveals alternative formose cycle pathways relevant to chemical origins of life

Generated with AI, which can make mistakes.