For those tired of having to book a supercomputer every time they want to design a new drug molecule, researchers at the University of North Carolina at Chapel Hill and the University of Florida have created a shortcut that can make the process up to a million times faster.
When creating a treatment for a disease, scientists look for biological targets that control specific processes in the body. These targets can be thought of as “locks,” and researchers need to design extremely complex molecular keys to fit these locks. To design a molecular key, scientist need to be able to calculate the energy levels of the molecule.
“In the atomic world, energy is everything,” said Olexandr Isayev, Ph.D., an assistant professor at the UNC Eshelman School of Pharmacy, and one of the paper’s authors. He is a member of the Laboratory for Molecular Modeling at the pharmacy school. “Energy levels determine the properties of any molecular systems — binding, physical properties, 3D structure, interactions, behavior over time, conformations — everything.”
The traditional way of solving the energy of molecular system is to use Schrodinger’s equation, a beast of a linear partial differential equation that scales cubically. This means that when you double the number of atoms in a molecule, you increase the computing time needed to solve the structure of the molecule by eight. Schrodinger’s equation deals with the quantum realm where the rules of uncertainty mean it can only give you the probability that the molecule will be in a certain state at a certain time, further complicating the process.
To speed up and simplify energy calculations, the UNC and Florida
researchers — all self-proclaimed Star Wars fans — created the Accurate NeurAl networK engINe for Molecular Energies or ANAKIN-ME. They call it ANI-1 for short. ANI-1 is a form of artificial intelligence called a deep neural network that is loosely modeled on the human brain. The scientists trained ANI on density functional theory calculations, which are used to investigate the structural, magnetic and electronic properties of molecules and materials and is a way to get an approximate solution to Schrodinger’s equation.
ANI-1 works through pattern recognition. Its neural net has been trained density functional theory solutions for 17 million small drug-like molecules. It uses that data to extrapolate the solution to a new molecular problem instead of doing the math. It’s like photo recognition software that can identify a photo of a dog. It doesn’t know what a dog is, but it has seen enough photos of dogs to recognize a new one based on common characteristics. This method sacrifices a slight bit of accurate for a huge increase in speed, Isayev said, but the loss is negligible.
ANI-1 can solve the structure of the breast-cancer drug Taxol with its 112 atoms in a fraction of a second on a laptop computer, Isayev said. Doing it the traditional way would take a few hours on a supercomputer.
The team’s work was published in Chemical Sciences. The researchers plan to make their code and database freely available.