Deep Evolutionary Generative Molecular Modeling for RNA Aptamer Drug Design

Abstract

Deep Aptamer Evolutionary Model (DAPTEV Model). Typical drug development processes are costly, time consuming and often manual with regard to research. Aptamers are short, single-stranded oligonucleotides (RNA/DNA) that bind to, and inhibit, target proteins and other types of molecules similar to antibodies. Compared with small-molecule drugs, these aptamers can bind to their targets with high affinity (binding strength) and specificity (designed to uniquely interact with the target only). The typical development process for aptamers utilizes a manual process known as Systematic Evolution of Ligands by Exponential Enrichment (SELEX), which is costly, slow, and often produces mild results. The focus of this research is to create a deep learning approach for the generating and evolving of aptamer sequences to support aptamer-based drug development. These sequences must be unique, contain at least some level of structural complexity, and have a high level of affinity and specificity for the intended target. Moreover, after training, the deep learning system, known as a Variational Autoencoder, must possess the ability to be queried for new sequences without the need for further training. Currently, this research is applied to the SARS-CoV-2 (Covid-19) spike protein’s receptor-binding domain (RBD). However, careful consideration has been placed in the intentional design of a general solution for future viral applications. Each individual run took five and a half days to complete. Over the course of two months, three runs were performed for three different models. After some sequence, score, and statistical comparisons, it was observed that the deep learning model was able to produce structurally complex aptamers with strong binding affinities and specificities to the target Covid-19 RBD. Furthermore, due to the nature of VAEs, this model is indeed able to be queried for new aptamers of similar quality based on previous training. Results suggest that VAE-based deep learning methods are capable of optimizing aptamer-target binding affinities and specificities (multi-objective learning), and are a strong tool to aid in aptamer-based drug development.