Put AI to work
June 26-27, 2017: Training
June 27-29, 2017: Tutorials & Conference
New York, NY

AI-assisted computational chemistry: Predicting chemical properties with minimal expert knowledge

Garrett Goh (Pacific Northwest National Lab)
1:45pm2:25pm Thursday, June 29, 2017
Verticals and applications
Location: Gramercy East/West Level: Intermediate
Secondary topics:  Health care

Prerequisite Knowledge

  • General familiarity with machine learning concepts
  • A basic understanding of convolutional neural networks
  • No specific chemistry knowledge required

What you'll learn

  • Learn how deep learning can be used as an AI tool to make predictions and new discoveries in chemistry
  • Understand a real-world example of how to build and apply AI to a technical field that has limited historical AI applications

Description

The use of machine learning in chemistry is not new, although its influence in the field has waxed and waned over the decades. Historically using expert “chemical intuition” and engineered features (molecular descriptors), computational chemistry has achieved modest success in creating models for predicting a broad range of chemical properties. Nevertheless, glass ceilings, particularly in the more challenging biological and materials applications, still remain.

Garrett Goh demonstrates how to use deep learning to construct computational chemistry models that compare favorably to existing state-of-the-art models developed by expert practitioners—with virtually no expert knowledge—proving the potential of AI assistance to accelerate the scientific discovery process from a typical span of years to a matter of months. Garrett offers an overview of a “hands-free” end-to-end approach for developing models for chemical predictions by integrating the latest algorithmic and architecture developments in traditional deep learning coupled with automated Bayesian hyperparameter tuning. Starting with a modest dataset of a few thousand chemicals, Garrett and his team have constructed deep convolutional neural network models for predicting a wide range of physical, biological, and materials properties.

Garrett then explores AI-assisted computational chemistry models that have been in development for less than a year and explains how they are achieving, and in some cases exceeding, the performance of state-of-the-art models developed by expert practitioners that build on decades of chemistry research and domain knowledge. These findings suggest the impact of AI assistance in accelerating the scientific discovery process and the generalizability of AI-assisted scientific research, which will find its future not just in chemistry but in affiliated applications, such as biotechnology, pharmaceuticals, and consumer goods, and perhaps other technical fields as well.

Photo of Garrett Goh

Garrett Goh

Pacific Northwest National Lab

Garrett Goh is a scientist in the Advanced Computing, Mathematics, and Data division at the Pacific Northwest National Lab (PNNL), where he holds the Pauling fellowship, which supports his research combining deep learning and artificial intelligence with traditional computational chemistry applications. His current interest is in AI-assisted computational chemistry—the application of deep learning to predict chemical properties and the discovery of new chemical insights using minimal expert knowledge. Previously, Garrett held a Howard Hughes Medical Institute (HHMI) fellowship, which supported his PhD in computational chemistry at the University of Michigan.