Researchers at Chalmers University of Technology and the University of Gothenburg have developed an AI model that can predict molecular motion more than 10,000 times faster than conventional molecular dynamics simulations, according to a Swedish study published in Science Advances.
The model, called TITO, short for Transferable Implicit Transfer Operators, is a deep generative modelling framework designed to learn statistical rules governing how molecules move over time. The researchers say the system can predict changes in molecular configurations across longer timescales than those directly observed during training.
Molecular dynamics simulations are widely used in early-stage drug development to study how molecules behave. Traditional methods calculate the forces between atoms step by step, using extremely short time intervals of about one femtosecond to maintain stability. Because many processes relevant to drug discovery occur over much longer timescales, these simulations can require billions of steps and substantial computing resources.
The researchers tested the AI model on more than 12,500 organic molecules, including molecules containing carbon, nitrogen, hydrogen and oxygen atoms, as well as more than 1,000 short peptides. The model was trained on simulated examples of atomic motion and used those sequences to learn general patterns in molecular behaviour.
“What sets our AI model apart is that it learns the underlying dynamics over longer time scales. It not only provides insights into the shapes that molecules take on, but also into how quickly and through which pathways these molecular transitions occur,” said Simon Olsson, associate professor in the Department of Computer Science and Engineering at Chalmers University of Technology and the University of Gothenburg.
The research team compared the model’s results with previous studies of molecular evolution and used post-processing simulations based on standard numerical algorithms to corroborate the findings. Olsson said the results were consistent with those calculations.
The model is intended to help researchers predict how molecules may change over time without simulating every intermediate step. The researchers compare this to moving between selected scenes in a molecular “movie” rather than watching every frame in sequence.
Juan Viguera Diez, an industrial doctoral student at AstraZeneca in the Department of Computer Science and Engineering at Chalmers and the University of Gothenburg and lead author of the article, said the work demonstrates that AI can be used to learn aspects of molecular physics in a transferable way.
“In order to be able to predict the physical phenomena exhibited by molecules, we need to understand the underlying physics of how the system behaves. I believe we are among the first to demonstrate this in a general sense and show that it is possible,” Viguera Diez said.
The researchers said the model may eventually support drug development by helping identify promising drug candidates more quickly in early-stage testing. At present, the method has been tested on small molecular systems in simplified solvent models and at a specific temperature. The team is continuing work to extend the approach to more complex and realistic systems.
