This thesis describes novel deep generative neural networks and their applications for the de novo design of molecules with optimized properties. It discusses the background and motivation for using deep generative models for de novo drug design, covering the advantages and limitations of existing techniques such as virtual screening of molecular libraries, genetic algorithms, and combinatorial enumeration of molecules from a set of building blocks. Next, it describes novel deep generative neural network architectures for producing molecules in three commonly used molecular representations – SMILES strings, 2D molecular graphs, and 3D molecular graphs, with details of the design, implementation, and computational experiments. It continues with a slight detour portraying how a generative model can provide an empirical estimate for the number of bioactive compounds in the chemical space and describes the experiment performed to obtain such an estimate. Next, it introduces Reinforcement Learning based strategy to optimize the values of a property of interest for generated molecules. It also proposes several heuristics for more efficient exploration of the chemical space. Finally, it describes how the proposed models and optimization algorithms were used to virtually design and then experimentally confirm novel hits for multiple kinase proteins.

136 pages

Thesis Committee:
Olexandr Isayev (Chair)
Christopher Langmead
David Koes (University of Pittsburgh)
Alexander Tropsha (University of North Carolina at Chapel Hill)

Russell Schwartz, Head, Computational Biology Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu