Embedding of FDA Approved Drugs in Chemical Space Using Cascade Autoencoder with Metric Learning

Abstract

Deep learning methods have been successfully used to predict characteristics of small molecules such as physicochemical properties and biological properties. Prediction is typically done by embedding compounds into a low-dimensional chemical space. The goal of our study is to create embedding space that can be used to distinguish approved and withdrawn drugs using compound information only. U.S. Food and Drug Administration (FDA) approved chemical drugs are validated substances in terms of therapeutic effect, toxicity, and side effects. Some of approved drugs are withdrawn due to various reasons, including toxic and disease-causing effects. Our study aims to propose a framework that embed FDA approved chemical drugs on chemical space by integrating representation of chemical structure from various encoding methods. Because withdrawn drugs were approved drugs, distinguishing them using compound information is quite challenging. Our proposed framework consists of three stacked deep autoencoder modules and effectively integrates the information of the chemical compounds by cascade modeling that continuously use latent representation learned from previous modules. Results showed that FDA approved chemical compounds have discriminative regions in the embedding space and complex representation information to understand the embedding of FDA drugs were incorporated well. Such results showed that our framework can be used as an embedding method for determining whether or not drug candidates will be approved.