EMBER: multi-label prediction of kinase-substrate phosphorylation events through deep learning

Area of Science:

• Cellular signaling and molecular biology.
• Bioinformatics and computational biology.

Background:

• Protein phosphorylation is central to cellular signal transduction, regulating key processes like cell cycle and apoptosis.
• Despite over 10^5 known phosphorylation events, the specific kinases responsible are identified for less than 5%.
• Accurate prediction of kinase-substrate relationships is vital for experimental design and mapping cellular signaling networks.

Purpose of the Study:

• To develop a deep learning method for predicting kinase-motif phosphorylation events.
• To address the limitations of single-label classification by reframing the problem as multi-label classification.
• To create a unified model for predicting phosphorylation events across 134 kinase families.

Main Methods:

• Introduced EMBER (Embedding-based multi-label prediction of phosphorylation events), a deep learning approach.
• Utilized Siamese neural networks to generate embeddings for peptide motif sequences.
• Integrated kinase phylogenetic information and motif dissimilarity into a multi-label classification model with a phylogeny-weighted loss function.

Main Results:

• EMBER successfully predicts kinase-motif phosphorylation events by treating the task as a multi-label problem.
• The novel peptide embedding method shows promise compared to existing approaches.
• The integrated model leverages kinase phylogeny to improve prediction accuracy.

Conclusions:

• The EMBER method offers a significant advancement in predicting kinase-substrate phosphorylation events.
• This approach enhances the mapping of the kinome signaling network.
• The developed method and associated code are publicly available for research use.