Proteins are nanoscale molecular machines whose function is determined not only by their structure, but by how they move and change shape over time. Understanding these dynamics is essential for explaining biological processes and for designing new therapeutics. Molecular dynamics simulations can model protein motion at atomic resolution, generating vast, high-dimensional datasets that capture these conformational changes. Analysing and exploiting this data at scale is a major computational challenge.
In this talk, I will introduce how molecular simulations are used to study protein dynamics, and then present a machine-learning approach based on convolutional neural networks that learns a compact, continuous representation of large simulation datasets. Implemented in our software molearn, the model reduces millions of atomic degrees of freedom to a low-dimensional latent space that preserves the key dynamical features. I will show how this representation can be used to generate new molecular conformations and explore transition pathways between states.