Abstract
AlphaFold2 has achieved relatively high structure prediction accuracy on proteins. However, it is reported that directly feeding coordinates into deep learning models cannot achieve ideal results on downstream tasks. Therefore, how to process the predicted results into an effective form that deep learning networks can understand to improve the performance of downstream tasks is worth exploring. In this study, taking single-sequence PPI site prediction as an example, we verified the effects of three processing strategies of coordinates, namely spatial Altering, SVD20, and the rASA feature calculation. The experiment results showed that spatial filtering and the rASA feature were two effective and suitable ways to encode structural information for deep learning models. Besides, we also performed a case study of a mutated protein. The results proved that spatial filtering might potentially introduce structural changes into HHblits profiles and deep learning networks when protein mutations occur. This work provides new insight into the downstream tasks, such as predicting the binding sites of proteins or predicting the effects of mutations.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have