De novo drug design with deep generative models : an empirical study

Mehdi cherti, Balazs Kegl, Akin kazakci

Abstract: We present an empirical study about the usage of RNN generative models for stochastic optimization in the context of de novo drug design. We study different kinds of architectures and we find models that can generate molecules with higher values than ones seen in the training set. Our results suggest that we can improve traditional stochastic optimizers, that rely on random perturbations or random sampling by using generative models trained on unlabeled data, to perform knowledge-driven optimization.

Conflicts: lri.fr

Keywords: Unsupervised Learning, Optimization