Chemical and Textual Embeddings for Drug Repurposing

Abstract Drug approval is a long and expensive process, that can take 10-15 years and more than 2 billion dollars. Therefore alternative techniques, such as drug repositioning, to identify new uses for approved drugs, has been gaining increasing attention. We examine the employment of different drug embeddings to predict successful drug repositioning. We study the employment of drug molecular structure and show that using larger chemical construct, such as large functional chemical groups, is much more effective than small sub-structures. We then study embeddings that are based on textual medical publications and compare them with the chemical-structure-based embeddings. We eventually present a novel embedding technique to combine the merit of the textual and chemical-based approaches. We provide empirical results on a repositioning benchmark set. Additionally, we present an application of such embedding as part of an ongoing repositioning research conducted with a major health care supplier, and identify a novel drug and indication. The pair has been verified on a corpus of 1.5 million patient EHR data.