Exploring protein biophysics with deep learning

Deep learning approaches are becoming increasingly useful for studying protein biophysics. For example, AlphaFold is famously one of the best currently available tools for predicting protein structure from sequence. Beyond structure prediction, deep learning approaches can help to predict mutational effects, protein function, or ligand binding. In all these applications, from a physics perspective, the primary challenge is to understand what the biophysical meaning is of the predictions produced by machine learning, how the machine learning algorithms make their predictions, and how we can curate and select the right training data for obtaining good results. Here, I will discuss several projects in this field that we are currently pursuing in my lab. First, I will describe how machine learning methods can be used to identify sites that are primed for mutation. Second, I will discuss the differences between models trained purely on sequence data versus on structure data. Finally, I will demonstrate how protein embedding models can be used to search sequence data bases for proteins with specific biophysical characteristics.