In Journal of chemical information and modeling
Deep learning approaches have become popular in recent years in the ﬁeld of de novo molecular design. While a variety of diﬀerent methods are available, it is still a challenge to assess and compare their performance. A particularly promising approach for automated drug design is to use recurrent neural network (RNN) as SMILES generators and train them with the learning procedure called 'transfer learning'. This involves ﬁrst training the initial model on a large generic data set of molecules, to learn the general syntax of SMILES, followed by ﬁne-tuning on a smaller set of molecules, coming from e.g. a lead optimization program. In order to create a well-performing transfer learning application which can be automated, it is important to understand how the size of the second data set aﬀects the training process. In addition, extensive post-ﬁltering using similarity metrics of the molecules generated after transfer learning should be avoided, as it can introduce new biases towards the selection of drug candidates. Here we present results from the application of a GRU-RNN to transfer learning on data sets of varying sizes and complexity. Analysis of the results has allowed us to provide some general guidelines for transfer learning. In particular, we show that data set sizes containing at least 190 molecules are needed for eﬀective GRU-RNN based molecular generation using transfer learning. The methods presented here should be applicable generally to other deep learning methodologies.
Amabilino Silvia, Pogany Peter, Pickett Stephen D, Green Darren