Abstract

Machine learning approaches have had a strong impact on many areas of natural sciences, including the study of proteins. My talk focuses on the use of machine learning approaches in the context of studying a specific property of proteins, their tendency to form biomolecular condensates. Biomolecular condensates are central to a very wide variety of cellular processes ranging from gene expression to protein translation. Recent years have given new insight into the factors that affect protein phase behaviour but many of the specifics of the factors that gover process remain not understood. In my presentation, I will show how we have been able to use both biophysics-based insight as well as hypothesis-free language models for describing the tendency of proteins to form condensates. The results illustrate that pre-trained language models have the capability to capture the specifics of important cellular processes at a high accuracy and comparably to physics-based models.