Challenges in predicting PROTAC-mediated Protein-Protein Interfaces with AlphaFold

Abstract

In the last few years, predicting the structure of PROTAC-mediated complexes emerged as a fundamental step toward the design of therapeutic modalities for cancer and other conditions. The development of AlphaFold2 (AF2) caused a paradigm shift in the structural biology community. From then onwards, further developments enabled the prediction of multimeric protein structures while improving calculation efficiency, leading to the widespread usage of AF2 and to the recent release of AF3. However, AF2 does not consider ligands, suggesting that ligand-mediated protein-protein interfaces (PPIs) are challenging to predict. One of the main claims of AF3 is that this limitation has been addressed, but the currently released webserver provides only a few ligands and no PROTACs are available. In this article, we benchmark the performance of AF2 on a test set of 28 PROTAC-mediated dimers, as well as a set of 326 protein hetero-dimers orthogonal to AF2 training set, with a special attention to the interface size and presence of ligand at the interface. We then evaluated whether the newly released AF3 model is able to outperform AF2 on the prediction of PROTAC-mediated complexes, despite not being able to include PROTAC ligands in the prediction. In this letter, we aimed at identifying possible reasons why AF-based methods fail to predict PROTAC-mediated interfaces. Our results show that AF2-multimer predictions are sensitive to the size of the interface to predict, with the majority of models being incorrect for the smallest interfaces. While it performs reasonably well for ligand-mediated interfaces in the absence of the ligand, AF2 is unable to predict PROTAC-mediated interfaces reliably. We also found that AF3 does not significantly improve upon the accuracy of AF2, as it is still unable to correctly predict PROTAC-mediated interfaces for the large majority of cases.