Toxicity Prediction Using Pre-trained Autoencoder

Abstract

Toxicology in the 21st Century (Tox21) is a collaborative initiative whose purpose is to investigate and develop efficient testing approaches to predict the impact chemical compounds have on Humans. In this paper we investigate how a pre-trained auto-encoder can be used to build classifiers capable of predicting the toxicity property of chemical compounds. Using a Deep Learning approach, we performed experiments to deter-mine if chemical compound fingerprints can be used to predict active and inactive compounds based on simplified molecular-input line-entry system (SMILES) in twelve selected assays. We conducted these experiments using data from ChEMBL and Tox21 to investigate how the latent layer produced by an auto-encoder can be used to train a classifier. All experimental results are compared against the winning teams of the Tox21 challenge, where positives and limitations of the proposed approaches are discussed.