Latest News

 

Teaching Neural Networks Quantum Chemistry

Published on September 3, 2019 by Aaron Dubrow



A model of the EthR transcriptional regulatory repressor protein in complex with compound GSK1107112A. Simulated with ANI-2x (CHNOSFCl), containing ~35K atoms with explicit water. [Courtesy: Olexandr Isayev et al.]

By now, we've all experienced the uncanny ability of artificial intelligence (AI) to mine real world data to identify faces in photos or predict our buying patterns.

But the next wave of AI will combine computer simulations with machine learning (a subset of AI) to create faster, more accurate predictive models for physics, chemistry and biology, in order to forecast weather and climate, design better materials, and tailor medical treatments to specific patients.

Olexandr Isayev, an assistant professor of chemistry at the University of North Carolina at Chapel Hill, is focused on solving fundamental chemical problems with machine learning, molecular modeling, and quantum mechanics.

"For the past five years, I've looked at how machine learning can help us solve otherwise unsolvable challenges in chemistry," Isayev said.

His work has many potential applications, but among the most impactful is in drug discovery, finding new molecules that can interact with specific proteins to treat and cure diseases.

Olexandr Isayev, assistant professor of chemistry at the University of North Carolina at Chapel Hill and an early user of Frontera.

Chemists have used computers for decades to find and test new drug candidates. Probably every modern drug on the market began in silico, its characteristics predicted on a large computer cluster.

But to truly determine how a molecule will respond to cells in real world conditions — treating diseases but also potentially causing side-effects — often requires an understanding of the quantum mechanical behavior of many interacting atoms.

It was a problem that even today's most powerful computers seemed nowhere close to solving. However, new approaches using AI and supercomputers have proved incredibly promising.

Students and postdocs in Isayev lab trained a neural network that can accurately describe the potential energy of molecules based on their three-dimensional structure. In a recent paper published in Nature Communications, his team and team of Adrian Roitberg from the University of Florida showed that by combining several tricks from machine learning, a system can learn coupled cluster theory — a "gold standard" quantum mechanical method used for describing many-body systems — and transfer this knowledge to a neural network.

"We're using machine learning to accelerate quantum mechanics," Olexandr explained. "We train a neural network to approximate the solution of Schrodinger equation, in our case solving density functional theory (DFT) equations for organic molecules first."

The approach Isayev used is called transfer learning. It combines a large number of less-intensive DFT calculations that provide a rough approximation of the system behavior, with a subset of coupled cluster calculations that refine the details of the model.

"Instead of using 100 million CPU hours, you only use one percent of that amount and rely on cheaper methods," Isayev explained. "We were able to achieve a nine order-of-magnitude speed up for certain applications using neural networks. Once the neural network is trained, you can run pretty accurate calculations, essentially on your laptop in a fraction of a second."

This is important in a real world setting, where researchers want to quickly test millions of new molecules to find a few promising candidates to study further in the lab. But it requires a well-trained model, which demands the use of a supercomputer that can compute for a long duration using many processors at the same time.

Isayev shares both the neural network and the code that created it on GitHub and works with the Molecular Sciences Software Institute, a National Science Foundation-funded virtual organization, to make the tool widely available.

"The tool is being used by industry and by academic scientists and it's been quite successful," he said. "A couple pharmaceutical companies have deployed it in their pipelines to generate three dimensional structures of small drug-like molecules, and essentially replaced the previous generation of standard force fields. It has the same speed, much better accuracy, and it's free."

As an early user of the Frontera supercomputer — the fastest academic supercomputer in the world — Isayev has begun the process of scaling up his research to achieve even greater results for his method.

"Frontera is a really useful machine for us. It's large-scale and contains true Intel Xeons processors, so it can solve quantum mechanics problems really well," he said. "On Frontera, we ran three million calculations in 24 hours, which probably was the record for our group. We need to do a lot of calculations to extract all the data and parameterize other elements from the periodic table in our neural network."

"At the interface of science and AI, what's happening is almost like a Sputnik moment... Our nation will benefit greatly from support of projects like Frontera and investments in science and AI."
Olexandr Isayev, The University of North Carolina at Chapel Hill

On Frontera, Isayev plans to move beyond simple force fields to more complex interactions where chemical reactions and environmental effects determine the ultimate results.

"The pilot application is to simulate complex reactions in condensed phase, for example, molecules in water, or reactions on surfaces, or combustion," he said.

The impact of such a tool could be enormous, not only for pharmaceutical companies but for researchers hoping to create new, stronger materials or better medicines.

"At the interface of science and AI, what's happening is almost like a Sputnik moment. There is a race of nations who might lead, and I think it's very important for the U.S. to be competitive, not only because of national importance, but also because we can really transform science for the good of society," Isayev said. "Our nation will benefit greatly from support of projects like Frontera and investments in science and AI."


This research is supported by the NSF Division Of Chemistry within the NSF Directorate for Mathematical and Physical Sciences: Award #: D3SC: CDS&E: Collaborative Research: Development and application of accurate, transferable and extensible deep neural network potentials for molecules and reactions


Contact

Faith Singer-Villalobos

Communications Manager
faith@tacc.utexas.edu | 512-232-5771

Aaron Dubrow

Science And Technology Writer
aarondubrow@tacc.utexas.edu

Jorge Salazar

Technical Writer/Editor
jorge@tacc.utexas.edu | 512-475-9411