Virtual data augmentation method for reaction prediction in small dataset scenario

Abstract

To improve the performance of data-driven reaction prediction models, a new data augmentation method for augmenting data volumes is presented that aims to add fake data in training dataset. This method is only pay attention to small-scale reactions, and manually generates fake data which is chemical and credible molecules by replacing functional groups in reaction sites. And we call this method as virtual data augmentation. Additionally, the transformer model is introduced to explore the effectiveness of virtual data augmentation method in the task of reaction prediction based on small data sets. We apply our method to five classic coupling reactions, the results show that the overall performance of the transformer-baseline model and transformer-transfer learning model combined with virtual data augmentation method is obviously improved, compared to raw datasets. Especially for Suzuki reaction, combining transfer learning strategy and virtual data augmentation method, reaches top-1 accuracy of 97.8%. To sum up, virtual data augmentation can be used as a measure to face up the problem of insufficient data and significantly improves the performance of reaction prediction.