Lancaster University graduate Jonathan De Vita has a BSc (Hons) in Computer Science, having studied the field of artificial intelligence and coding. This article will look at generative artificial intelligence (GenAI), a field that is paving the way for huge breakthroughs in the molecular world.
As the capabilities of GenAI grow, the discipline poses huge potential to help biologists and chemists explore static molecules such as DNA and proteins. Models like AlphaFold have been developed to predict molecular structures to advance drug discovery, while the MIT-assisted RFdiffusion technology can actually help develop new proteins.
One persistent challenge for molecular scientists lies in the fact that molecules are constantly moving, an important characteristic to model when developing new proteins and drugs. Molecular dynamics uses physics to simulate these motions on a computer. However, the technique is incredibly expensive, demanding billions of time steps on supercomputers.
A new system integrating GenAI has been created to emulate the dynamics of molecules, transforming blurry images into videos. A team from MIT Department of Mathematics and Computer Science & Artificial Intelligence Laboratory (CSAIL) have developed a GenAI model that learns from prior data. The system, known as ‘MDGen’, takes a frame of a 3D molecule, simulating that will happen next by connecting separate stills to create a video, even filling in missing frames. By hitting the ‘play button’ on molecules, researchers could achieve huge advancements, enabling them to study and design new molecules and assess how drug prototypes would interact with the molecular structures they are designed to impact.
Molecular dynamics plays a pivotal role in drug discovery by providing insights into the movement and behaviour of molecules over time. Unlike image data, molecular dynamics simulations capture the dynamic behaviour of proteins, revealing conformational changes, interacting forces and binding pathways. In terms of assessing how drugs interact with their targets, these insights are crucial, particularly in complex biological environments where motion and flexibility significantly influence binding affinity and specificity.
Bowen Jing is the co-lead author on the MIT MDGen project. He points to the team’s discoveries as early proof of concept, highlighting their findings as the potential beginning of an exciting new research direction. Whereas early GenAI models produced rudimentary videos of their subjects – for example, a person blinking – today models like Sora or Veo can be used in all kinds of interesting ways. Bowen Jing and his team hope to instil a similar vision in the molecular world, building on current architecture and available data to develop a separate machine-learning method capable of speeding up the data collection process.
MDGen presents an exciting inroad in modelling molecular changes, enabling chemists to delve deeper into the behaviour of medicine prototypes designed to treat a range of diseases, including everything from cancer to tuberculosis, with potentially game-changing results for the medical world.