Combining Enhanced Sampling simulations and Deep Learning for the study of Intrinsically Disordered Proteins

The biological functions of proteins intimately depend on their conformational dynamics. This aspect is particularly evident for intrinsically disordered proteins (IDP) that lack a fixed three-dimensional structure and for which structural ensembles often offer more useful representations than individual conformations.

However, obtaining these ensembles of conformations is experimentally challenging, so computational simulations are often used. Enhanced Sampling simulations offer an advantage because they allow for a more comprehensive sampling of possible conformations. However, they generate incredible amounts of data, which are often difficult to analyze, and it is unclear what are the most representative metrics and parameters.

Thus, extracting useful information regarding the most relevant states and conformational transitions requires dimensionality reduction techniques that project high-dimensional data (protein conformations) into low-dimensional representations. These low-dimensional maps can be more easily interpreted and form a basis for clustering the simulation data into conformational states.

In this colloquium, I will explain how we employed extensive enhanced sampling Temperature Replica-Exchange atomistic simulations (TREMD) and deep learning dimensionality reduction to study the conformational ensembles of the human chaperone Heat Shock Protein B8 and its neuropathological mutant K141E, for which no experimental 3D structures are available.