Using GANs with adaptive training data to search for new molecules

摘要

New molecules produced for different replacement strategies. For control (blue), the training data is fixed. For random (red), molecules from the generator randomly replace molecules in the training data. For drug (green), molecules from the generator only replace training samples if they have a higher drug-likeness score. a As training progresses, control stops producing a substantial number of new molecules, but random and drug replacement strategies continue production. Plot shows average over three training runs for each selection type. b Although drug produces less overall new molecules than random, it generates more top performers. Plot shows average over three runs for each selection type with error bars showing one standard deviation