Integrative Structure Modeling of Protein Assemblies in the Age of Deep Learning

Integrative structure modeling is often used to characterize structures and dynamics of large macromolecular assemblies by relying on multiple types of input information. The individual proteins or domains are represented by atomic resolution structures or low-resolution sphere models and data from a variety of sources, such as cross-linking mass spectrometry, cryo-Electron Microscopy, Small Angle x-ray scattering is used to assemble the subunits. Recent progress in protein folding enabled by deep learning by AlphaFold2 and RosettaFold provided an improved structural coverage for domains, and even protein-protein interactions used in Integrative Structure Modeling. However, these methods depend on multiple sequence alignment (MSA), that is not available for immune response complexes, such as antibody-antigen interactions. Moreover, there is a limitation on the size of the modeled proteins or complexes. Our research is focused on addressing these problems. We have developed deep learning models for accurate end-to-end modeling of immune response complexes without MSA, including antibody-antigen and peptide-MHC interactions. In addition, we have developed methods for integrative modeling of large complex assemblies using cross-linking mass spectrometry data, subunit models and pairwise subunit-subunit interaction models predicted by AlphaFold2.