Transfer learning-enabled classification of drugs for Alzheimers Disease

Abstract

Alzheimers disease (AD) is a neurodegenerative disease with a worldwide prevalence of 760.5 per 100,000 inhabitants. Environment factors and genetic predispositions can influence an individuals risk of developing AD. Recent research provided insights into the genetic mechanisms of AD, but a comprehensive database of drug-like molecules that can exacerbate or reduce AD risk remains unavailable. Here, we use machine learning to create a similarity map that reveals new putative drugs structurally similar to existing compounds known to be targeting pathways relevant in AD. We trained an autoencoder on a large drug database of over 14,000 drugs, with features derived from several modalities including molecular fingerprints. We then computed similarity scores based on these reduced dimensions. We show that our model is able to identify new compounds structurally similar to existing drugs linked to AD. We conclude that our model holds the potential to elucidate new compounds based on structural similarity and can be used to identify new drugs that affect critical pathways in AD.