Automating Components of Molecule Design with Machine Learning with Trent Hauck
Details
Title
Automating Components of Molecule Design with Machine Learning
Abstract
Automatic design of molecules that confer certain properties has the potential to make huge impacts in medical and industrial applications -- for example, a molecule that binds to certain receptors in the brain, binding affinity being a property, could be be used as the next pain killer. We’re likely still a ways away from purely automatic design, however, borrowing ideas from image and natural language processing, and, of course, ones specific to chemistry, inroads are being made.
During this talk we'll cover the application of sequence generation models to chemical design, which are often used for natural language generation. We’ll also look at graph convolution networks, which are related to convolutional layers applied commonly in image analysis. Generally, the goal for these steps is to build a statistical model for how molecules are created -- a generative model. Once this generative model has been created, the next task is to design molecules in silico that probably have those properties of interest. To this end we’ll review the optimization steps for searching through the "chemical space", and open source tools which are bridging the gap between machine learning and automated molecule design.
Speaker Bio
Trent Hauck is currently a Senior Data Scientist at Zymergen applying statistics and machine learning to areas such as high throughput screening and systems biology. He's worked the last 8 years in different domains including biology, natural language processing, chemistry, and recommender systems. Previously, he authored two books about using Python for machine learning applications, Scikit-Learn Cookbook and Instant Data Intensive Apps with pandas How-to. A Seattle transplant, he hails from Kansas.