Deep learning has drawn a lot of attentions in different areas including drug discovery. In pharmaceutical industry, one of the impactful applications of deep learning is to predict the molecular activities, including physical chemical properties, potency, pharmacokinetics, and toxicity, using molecules’ chemical structure. This is known as quantitative structure activity relationship (QSAR) modeling. Generally, a machine learning algorithm is used to model such relationship. Deep learning is expected to outperform other machine learning algorithms, especially with big data sets. This presentation summarized some of our recent works focusing on answering following questions: 1. What advantages deep learning can bring to us? 2. What is the best way for molecular structure representation? 3. How to conduct hyperparameter tuning for deep learning QSAR models? 4. How to improve the generalization ability for deep learning QSAR models? In order to answer those questions, we have conducted systematic evaluations on 24 industry ADME data sets with over 50000 deep learning QSAR models. We have also explored several ways of building deep learning based QSAR models with large industry ADME datasets, including adding different types of noises into the data and model, using multi-task learning and transfer learning. Based on the data and analysis, we have proposed a best practice of building deep learning QSAR models.

Industry perspective: Deep learning for QSAR models

2019 Spring National Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.