Applying Deep Learning to Discover Highly Functionalized Nucleic Acid Polymers That Bind to Small Molecules

Abstract

Developing novel binders for small molecule and protein targets has been at the core of a number of recent medical breakthroughs. A popular developmental method focuses on monoclonal antibodies, currently a $98B industry. Antibodies, however, are costly to produce and difficult to manufacture. DNA aptamers – oligonucleotides that bind to a specific target – present a promising alternative, as they can be manufactured at scale for lower cost. Unfortunately, existing experimental methods for identifying functional aptamers typically take months, and can only sample a small fraction (1 in 10^10 possibilities) of the theoretical search space. Deep learning techniques offer a novel solution to the challenge of aptamer discovery. In addition to developing a theoretical model for quantifying aptamer binding affinity, this thesis demonstrates that a conditional variational autoencoder (CVAE) can be used to generate novel high-binding aptamers for daunomycin, a chemotherapeutic agent. After training on eight rounds of experimental data, a CVAE was able to successfully generate entirely original aptamers that performed as well as, or better than, those generated through conventional selections(KD≈10-30nM).This thesis shows the power of using deep learning techniques, coupled with intimate domain knowledge, to accelerate the process of screening aptamers, and thereby advance the field towards concrete therapies and diagnostics.