Machine learning algorithm creates successful chemical reactions from failed experiments

Nature : Machine learning algorithms have been used to help create complex molecules based on input of known successful reactions. Now Alex Norquist, Sorelle Friedler, and Joshua Schrier of Haverford College in Pennsylvania and colleagues have included unsuccessful experiments in the data provided to a computer algorithm that predicts whether a specific set of reagents will react to form a crystalline material. The researchers provided the algorithm with data from nearly 4000 experiments, including both published successes and failed reactions, which they had to transcribe from unpublished notes. The team then asked the computer to identify the principles that distinguished the successful reactions from the failures. When subsequently provided with previously untried combinations of reactants, the algorithm suggested nearly 500 reactions, 89% of which resulted in a crystalline product. In comparison, the educated guesses of the researchers were successful only 78% of the time.