Differentiable Neural Computers for in silico molecular design: Benchmarks of architectures in generative modeling of molecules

Abstract

In the area of in silico drug discovery, deep learning has grown immensely as a field of research. Recurrent neural networks (RNN) is one of the most common approaches used for generative modeling, but recently the differentiable neural computer (DNC) has been shown to give considerable improvement over the RNN for modeling of sequential data. In this thesis, a DNC has been implemented as an extension to REINVENT, an RNN based model that has already been successfully shown to generate molecules with high validity. The model was benchmarked on its capacity to learn the SMILES language on the GDB-13 and ChEMBL datasets. The DNC shows some improvement on all tests conducted at the cost of greatly increased computational time and memory consumption, which puts its practical use into question. This project also gives some insight into the effect of the DNC hyperparameters for the task of generative modeling of molecules.