Modeling Industrial ADMET Data with Multitask Networks
Vertex Pharmaceuticals is a global pharmaceutical company engaged in the discovery of novel medicines to treat unmet medical needs. Our drug discovery process leverages cutting edge technologies, including advanced medicinal chemistry, CRISPR gene-editing and computationally based predictive models. One aspect of our lead optimization and discovery cycle involves the prediction of ADME properties and off-target effects of our putative drug candidates. Our production prediction system, based upon logistic regression, is refined on a weekly basis by incorporating new experimental data from our in-house assays. Motivated by external reports of remarkable gains in performance by deep-learning approaches in various domains, we engaged in evaluating multitask deep-neural networks in an industrial pharmaceutical setting. Using our in-house ADME datasets, we compared deep multitask neural networks to our standard baseline models. We analyzed multitask learning effects with both random cross-validation and a more relevant temporal validation scheme. Although the multitask neural networks provided modest gains over our baseline models, the gains were not sufficient to justify incorporating these models into our production system. This talk will review our study and discuss issues relevant to applying deep learning methodology to small molecule drug discovery in an industrial pharmaceutical setting.
Brian Goldman is a Research Fellow II at Vertex Pharmaceuticals in Boston Massachusetts where he leads the methods group in the Modeling & Informatics department. His group develops algorithms and software for molecular modeling, cheminformatics, and image analysis. Dr.Goldman is particularly interested in applying machine learning methods to impact drug discovery programs. Before joining Vertex in 1998, Dr. Goldman earned his Ph.D. in Chemistry from the University of California at Santa Cruz where he studied differential geometry as applied to molecular surfaces and molecular similarity calculations. Prior to joining Vertex, Dr. Goldman worked at CombiChem, a computational drug discovery firm based in California.