De novo drug design using self attention mechanism
Vedang Mandhana, Rutuja Taware · 2020
The search space involved in drug discovery is huge. Hence, measures are being taken to perform this search more efficiently with the help of machine learning. Generative models have tremendously increased efficiency of de novo design of drug molecules. This paper proposes to use self-attention mechanism to generate novel molecules through language modeling. Language modeling enables the generated molecules to have properties similar to the molecules present in the training set. Self-attention enables the model to identify the relevant context between the various elements generated in the molecular sequence. This takes care of the dependencies involved in molecular structures. In this generative approach, the molecules are represented using formal molecule notation known as SMILES and the Transformer-XL architecture is used for training and sampling of novel molecules. The Transformer-XL architecture is successful in modeling molecular sequences of variable length.