Can structural information help with phosphorylation prediction?

Yannick Kiefl, Gamza Gamouh, Michael Heinzinger, David Hoksza, Marian Novotny
Affiliations:
 Department of Cell Biology, Faculty of Science, Charles University, Czech Republic,
 Department of Software Engineering, Faculty of Mathematics and Physics, Charles University, Czech Republic,
 Department for Bioinformatics and Computational Biology, Faculty of Informatics, Technical University of Munich, Germany.

Phosphorylation is one of the most frequent and important mechanisms the proteins have to react to changing conditions or environment. Deregulation of phosphorylation often leads to pathological conditions. Detecting phosphorylation and describing sites influenced by phosphorylation is however experimentally challenging due to the transient nature of this modification. Various sequence-based approaches have been developed to predict phosphorylation sites, but in this work we tested whether we could use the growing amount and availability of 3D structural information to predict phosphorylation sites using structural context and sequence features. We built a small set of 360 protein structures with at least one known phosphorylation site in each to train a graph neural network that will use structural and sequence features to predict phosphorylation sites. The structural information is used to give the sequence features additional context. Our initial data shows that structural features can be beneficial for prediction of phosphorylation sites, but that it has to be fine-tuned to provide competitive results. The final predictor achieves a precision of 0.47 on a class imbalance of 1/34 while retaining a low false negative rate of 0.1.