A new AI model has become so good at predicting how molecules evolve over time that, in the future, it could speed up the costly and time-consuming process of testing new drugs. In the long term, this technology could facilitate the development of medicines and new treatments, as promising drug candidates are able to be identified more quickly and with greater accuracy. The findings are being presented in a new Swedish study published in Science Advances.
Developing a new drug often takes over ten years from the initial idea to the finished medicine before it reaches patients. A large proportion of both the cost and the time involved is concentrated in the early stages, as a multitude of tests must be carried out to identify the most promising candidates. Often, several studies are required in which thousands of molecules are screened – but only a fraction of them go on to the next stage.
Traditionally, the movements of molecules have been simulated using what is known as molecular dynamics, in which researchers calculate the forces between all the atoms step by step and move them a tiny bit at a time. For the calculations to be stable, each step must be extremely short, approximately one femtosecond (10⁻¹⁵ seconds). Since the processes that are of interest for drug development take place over much longer timescales, billions of steps are required, which makes the simulations computationally very demanding.
Major changes brought about by AI
The use of AI now enables researchers to detect molecular changes without having to perform numerical calculations. Machine learning can speed up each step of the calculation, and generative models can be used to directly generate plausible molecular structures without simulating their motion.
A group of researchers from Chalmers University of Technology and the University of Gothenburg, Sweden, has now taken another step forward by developing a new AI model that could, in the long term, make drug development testing even more efficient. The new model is more than 10,000 times faster than conventional simulations.
'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. As far as we know, this is the first time this has been done in a way that works for many different molecules,' says Simon Olsson, research leader and Associate Professor in the Department of Computer Science and Engineering at Chalmers University of Technology and the University of Gothenburg.
Thousands of molecules have been tested
The study examined over 12,500 organic molecules, such as those containing carbon, nitrogen, hydrogen and oxygen atoms. Over a thousand short peptides were also studied: molecules consisting of short chains of amino acids that make up proteins. The AI model learned how the molecules typically behave and was therefore able to fast-forward through the simulations. The results are still consistent with the laws of physics.
'We train the model using simulated examples of how the atoms in a molecule move over time. Based on these sequences, the model learns the underlying rules governing the movement of the molecules and can then predict how new molecules will behave,' says Simon Olsson.
The researchers compared the model's results and conclusions with previous studies of molecular evolution.
'We validated the results using extensive post-processing simulations to corroborate them using standard numerical algorithms. And they are consistent with one another,' says Simon Olsson.
Changes can be predicted
Although the AI model is not based on real images, the researchers describe the results as a way to jump between scenes in 'molecular movies,' instead of watching every frame in sequence.
The AI model forms the basis for the computational predictions that the researchers then make in the laboratory.
'There, we measure very specific things: the properties of the molecules, how "happy" they are to be in a particular solution, or whether, for example, they want to pass through a membrane into a cell---but this still lies in the future,' says Simon Olsson.
One major strength is that the model can be applied to molecules it has never encountered during its training, as it has learnt general rules governing molecular motion rather than memorising individual systems.
'There is a certain pattern that the model helps us to identify. The AI model is based on a number of examples, in which it only observes what happens over a period of tens of nanoseconds. Nevertheless, it can predict the properties and changes in molecules that occur over a period a thousand times longer. So, with the help of artificial intelligence, we can work out what is likely to happen in the 'molecular future'. It can predict how molecules change even though it has never seen the process unfold,' says Simon Olsson.
Of interest for the pharmaceutical industry
'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,' says 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.
Researchers are seeing considerable interest from industry in simulations that more accurately reflect reality and enable new drugs to be developed more quickly. As the new AI model can speed up molecular simulations, where large numbers of potential molecules need to be tested, the research team hopes it will be an important step towards more efficient drug development.
'In the long term, AI models like ours could help to identify promising drug candidates more quickly and improve accuracy in the early stages. The research study shows what is currently possible. This will hopefully pave the way for the development of more general techniques, which may ultimately facilitate the development of new drugs and new treatments, and, in a broader sense, also improve our understanding of diseases,' says Juan Viguera Diez.
More about the AI model
The TITO (Transferable Implicit Transfer Operators) AI model is a deep generative modelling framework that learns the statistical rules governing molecular motion directly from simulation data. It makes it possible to predict how atomic configurations (the way atoms are arranged and relate to one another spatially within a molecule) evolve over time scales much more rapidly than conventional numerical simulations.
The method has currently been tested on small molecular systems in simplified solvent models and at a specific temperature. It is now being developed further for more complex and realistic systems.
More about the research:
The article Transferable generative models bridge femtosecond to nanosecond time-step molecular dynamics has been published in Science Advances. The authors are Juan Viguera Diez, Mathias Schreiner and Simon Olsson, all of whom work at Chalmers University of Technology and the University of Gothenburg.
