For the past decade, generative models have become an established research direction with many computer vision and natural language processing applications. As the chemical data is getting bigger, these models find applications in computational chemistry as well. In this thesis proposal, we will discuss various deep generative neural networks for molecular design. The talk will be focused on the different representations of molecules such as SMILES strings, 2D molecular graphs, and 3D conformations. For each representation, we will introduce a suitable generative model. We will also discuss how these models can be combined with reinforcement learning techniques to optimize the properties of generated molecules and showcase the results on the examples of designing novel kinase inhibitors.
Olexandr Isayev (Chair)
David Koes (University of Pittsburgh)
Alexander Tropsha (University of North Carolina - Chapel Hill)
Zoom Participation. See announcement.