AlphaFold-Multimer struggles in predicting PROTAC-mediated protein-protein interfaces

AlphaFold2 (AF2) made its debut in the CASP14 competition, generating structures which could rival experimentally determined ones and causing 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. However, previous work noted that AF2 does not consider ligands and thus suggesting that ligand-mediated protein-protein interfaces (PPIs) are challenging to predict. In this letter, we explore this hypothesis by evaluating AF-Multimers’ accuracy on four datasets, composed of: (i) 31 large PPIs, (ii) 31 small PPIs, (iii) 31 PPIs mediated by ligands and (iv) 28 PROTAC-mediated PPIs. Our results show that AF-Multimer is able to accurately predict the structure of the majority of the protein-protein complexes within the first three datasets (DockQ: 0.7-0.8) but fails to do so for the PROTAC-mediated set (DockQ < 0.2). One explanation is that AF-Multimers’ underlying energy function was trained on naturally occurring complexes and PROTACs mediate interactions between proteins which do not naturally interact with each other. As these “artificial” interfaces fall outside AFs’ applicability domain, their prediction is challenging for AF-Multimer.