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.

Evaluation of AlphaFold 3’s Protein–Protein Complexes for Predicting Binding Free Energy Changes upon Mutation

Benchmarking AlphaFold3's protein-protein complex accuracy and machine learning prediction reliability for binding free energy changes upon mutation

AlphaFold3, a secret sauce for predicting mutational effects on protein-protein interactions

Comprehensive Evaluation of AlphaFold-Multimer, AlphaFold3 and ColabFold, and Scoring Functions in Predicting Protein-Peptide Complex Structures

AlphaFold2 can predict single-mutation effects

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

Accurate structure prediction of biomolecular interactions with AlphaFold 3

Can AlphaFold2 predict protein-peptide complex structures accurately?

AF2-Mutation: Adversarial Sequence Mutations against AlphaFold2 on Protein Tertiary Structure Prediction

Benchmarking AlphaFold for protein complex modeling reveals accuracy determinants

AF2Complex predicts direct physical interactions in multimeric proteins with deep learning

Improved protein complex prediction with AlphaFold-multimer by denoising the MSA profile

Benchmarking HelixFold3-Predicted Holo Structures for Relative Free Energy Perturbation Calculations

AlphaFold predictions of fold-switched conformations are driven by structure memorization

Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes

How accurately can one predict drug binding modes using AlphaFold models?

Using AlphaFold to predict the impact of single mutations on protein stability and function

AlphaFold illuminates half of the dark human proteins

Overview of AlphaFold2 and breakthroughs in overcoming its limitations

AlphaFold2 knows some protein folding principles

Predicting direct physical interactions in multimeric proteins with deep learning