DiPol-GAN: Generating Molecular Graphs Adversarially with Relational Differentiable Pooling
Pablo Rivas, Michael J. Guarino, Alexander Shah · 2019
Advances in deep generative modeling applied to irregular structures, such as graphs, have led to exciting advances specifically in the generation of graph structured data. This has been of particular importance to drug discovery as it directly applies to the problem of finding new molecular compounds. Many previous approaches to molecule generation have represented molecules using the SMILE (Simplified Molecular Input Line Entry System) strings format [Weininger 1988] rather than representing molecules directly as a graphs. Graph Neural Networks (GNNs) have shown state-of-the-art performance on many graph related tasks such as graph classification [Xinyi 2018] and link prediction [Balavzevic 2019]. In this work we introduce DiPol-GAN, a generative adversarial network (GAN) approach to implicitly learning to generate molecular graphs. Using Differentiable Pooling, DiPol-GAN learns hierarchical representations of molecular graphs leading to more robust discriminator performance [Ying 2018]. This work also proposes an extension of DIFFPOOL allowing it to handle graphs with multiple relation types such as different bond types that occur between atoms. To enhance the utility of this method we also constrain the learned latent representation with a reinforcement learning objective to shift the generation towards a targeted chemical property. Furthermore, we benchmark against other comparable models with similar claims. Preliminary results indicate that the proposed approach is competitive and for specific properties better than the benchmark.