Generative Lead Optimization of de novo Molecules: Case Study in Discovery of Potent, Selective Aurora Kinase Inhibitors with Favorable Secondary Pharmacology

AI Applications in Drug Discovery and Development 2019 | Boston, MA

Date: September 10-12, 2019

Speaker: Andrew Weber, MS

Authors: Andrew Weber [1, 2], Jason Deng [1, 2], Kevin McLoughlin [1,3], Thomas D. Sweitzer [1, 4], Juliet McComas [1, 4], Margaret Tse [1, 2], Derek Jones [1, 3], Jonathan Allen [1, 3], Stacie Calad-Thomson [1, 2], Jim Brase [1, 3], and Tom Rush [1, 4]

1 ATOM, San Francisco, CA

2 GlaxoSmithKline, San Francisco, CA

3 Lawrence Livermore National Laboratory, Livermore, CA

4 GlaxoSmithKline Collegeville, PA

Title: Generative Lead Optimization of de novo Molecules: Case Study in Discovery of Potent, Selective Aurora Kinase Inhibitors with Favorable Secondary Pharmacology

Abstract: De novo design of therapeutic agents is currently a slow, expensive process generally relying on a large high throughput screen and several follow up cycles of iterative design to enhance the potency, eliminate safety liabilities, and enable favorable pharmacokinetics. Computer aided drug discovery (CADD) has contributed significantly to addressing these challenges though structure and ligand-based design techniques, however current application often focuses on optimizing a single molecular property at a time, leading to long design cycle times. Recent advances in generative models for chemistry, including variational autoencoders (VAEs) and generative adversarial networks (GANs) have enabled a continuous representation of chemical space, allowing for application of numerical optimization techniques to this multi-factorial problem. As a proof of concept exercise, optimization of generative network across a diverse range of pharmacological properties was performed. Specifically, the framework was challenged to design a potent molecular inhibitor of Aurora Kinase B with selectivity against Aurora Kinase A while simultaneously maintaining high solubility, low microsomal clearance, and limiting potency the hERG ion channel and BSEP transporter (known off-targets linked to cardiac and liver toxicity). The framework was able to generate >200 molecules with predicted properties meeting the acceptability criteria. Initial validation against a hold-out set of Aurora Kinase potency data further demonstrates the promise of this technique.