Abstract
MOTIVATION: The structural characterization of protein-protein interactions is a key step in understanding the functions of living cells. Here, I show that AlphaFold3 often fails to predict protein complexes that are either weak or dependent on the presence of a cofactor that is not included in a prediction. RESULTS: To address this problem, I developed gapTrick, an AlphaFold2-based approach that uses multimeric templates to improve prediction reliability. I demonstrate that gapTrick improves predictions of weak and incomplete complexes based on low-accuracy templates, such as individual protein models that have been rigid-body fitted into cryo-EM reconstructions. I also show that gapTrick identifies residue-residue interactions with high precision. These interaction predictions are a very strong indicator of model correctness. The approach can aid in the interpretation of challenging experimental structures and the computational identification of protein-protein interactions. AVAILABILITY AND IMPLEMENTATION: The gapTrick source code is available at https://github.com/gchojnowski/gapTrick and requires only a standard AlphaFold2 installation to run. The repository also provides a Colab notebook that can be used to run gapTrick without installing it on the user's computer.