Fig. 5: Transformer training and inference flows are shown.
From: Exhaustive local chemical space exploration using a transformer model
The training is depicted to left (a), while the inference to the right (b). a The transformer model receives as input pairs of molecules consisting of a source molecule and a target molecule (represented as SMILES strings). The SMILES associated to source and targets are CC(Cc1ccc(Cl)cc1)NCC(O)C(N)=O and CC(Cc1ccc(O)cc1)NCC(O)C(N)=O, respectively. The initial character of the target SMILES provided as input is the starting token ˆ. The model is trained to transform a source molecule to a target molecule by producing its next tokens (in parallel). The last character of the produced output is the ending token $. At training molecular pairs of source and target molecules are used to train the transformer model. b At inference a source molecule is transformed into several target molecules. In the figure, those are sampled target1 and sampled target2 corresponding to SMILES CC(Cc1ccc(Cl)cc1)NCC(O)C(N)=S and CC(Cc1ccc(Cl)cc1)NCC(F)C(N)=O, respectively.
