In selective autophagy, cargo recruitment is mediated by the binding of LC3 interacting regions (LIRs) to LC3 proteins covalently attached to the phagophore membrane. The LIRs are located either in the cargo itself or in cargo receptor proteins. Post-translational modifications, especially phosphorylation, can modulate the LC3-LIR interaction, and, thereby, regulate cargo recruitment. Identifying LIRs and generating structural models of the bound states – the LC3-LIR complex – is a pivotal step towards understanding the regulation of the cargo recruitment process. Recently, machine learning models, such as AlphaFold (AF) 2 and 3, have emerged as powerful tools to predict the structure and properties of protein complexes.
We explore a fragmentation-and-phosphorylation strategy to further improve the prediction of LC3-LIR complex structures in AF. We show that these predictions can be modulated by fragment length and the introduction of phosphomimetic mutations. We apply these findings in the context of a computational protein fragment screen that yields structural models for interactions that AF fails to predict for full-length sequences. Leveraging recent advances in machine learning, we improve the scalability of our approach to further increase its utility as a screening tool.