Prediction of Drug Metabolites Using a Deep Learning Language Model

Abstract

The understanding of metabolism is essential in drug development, but conducting drug metabolism experiments is resource-intensive. To support this, in silico experiments using machine learning have been explored, with several tools available, but these rely on rule-based assessments and are restricted in their scalability. To build a better model for metabolite prediction in drug discovery, a deep neural network model called the Focused Transformer has been explored.

For the model, metabolite data was gathered and curated. Several strategies were explored to improve the model’s performance, including a novel pretraining strategy involving pairs of structurally analogous molecules termed matched molecular pairs. The best derived model managed to find one true metabolite and had a validity of 4.5% when evaluated on an internal test set. While the model shows reasonable prediction for metabolite prediction, there is potential to achieve higher performance in future work and we conclude by suggesting several potential strategies that can be explored further, such as handling of data during training.