Fig. 8: Substructure-level sequence-to-sequence learning.
From: Single-step retrosynthesis prediction by leveraging commonly preserved substructures
Both the Transformer encoder and decoder have L identical blocks. The virtual number labeled atoms and substructures are highlighted in green. During training, the product side (input) is converted to substructures and fragments from the product, the reactants side (output) is converted to fragments from reactants only. “∣” in the converted input is a special character marking the beginning of SMILES strings for fragments present in the product. The model is trained on the converted input and output. During inference, we only predict the fragments of reactants; finally predicted reactants are obtained by merging the substructures with the predicted fragments. a Substructure-level sequence-to-sequence learning model. b Our training and inference workflow.
