学术文献库

De novo design of molecules has recently enjoyed the power of generative deep neural networks. Current approaches aim to generate molecules either resembling the properties of the molecules of the training set or molecules that are optimized with respect to specific physicochemical properties. None of the methods generates molecules specific to a target protein. In the approach presented here, we introduce a method which is conditioned on the protein target sequence to generate de novo molecules that are relevant to the target. We use an implementation adapted from Google's "Show and Tell" image caption generation method, to generate SMILES strings of molecules from protein sequence embeddings generated by a deep bi-directional language model ELMo. ELMo is used to generate contextualized embedding vectors of the protein sequence. Using reinforcement learning, the trained model is further optimized through augmented episodic likelihood to increase the diversity of the generated compounds compared to the training set. We used the model to generate compounds for two major drug target families, i.e. for GPCRs and Tyrosine Kinase targets. The model generated compounds which are structurally different form the training set, while also being more similar to compounds known to bind to the two families of drug targets compared to a random set of molecules. The compounds further display reasonable synthesizability and drug-likeness scores.