EMBER

EMBER predicts kinase-substrate phosphorylation events using deep learning to assign kinases to phosphorylation motifs and to map cellular signaling relationships.

Key Features:

• Multi-Label Classification Model: Implements a multi-label classifier that jointly predicts phosphorylation across 134 kinase families.
• Integration of Kinase Phylogeny Information: Incorporates kinase phylogenetic relationships via a kinase phylogeny-weighted loss function to modulate predictions based on evolutionary relatedness.
• Motif Dissimilarity and Vector Representations: Uses a Siamese network to generate vector embeddings of motif sequences that capture motif dissimilarities.
• Comparison with Existing Embeddings: Evaluates the Siamese-derived motif embeddings against previously proposed peptide embeddings to assess representational performance.
• Input Data Utilization: Combines Siamese-derived motif embeddings with one-hot encoded motif sequences as model inputs.

Scientific Applications:

• Phosphorylation Network Mapping: Predicts kinase assignments for phosphorylation motifs to build preliminary maps of kinase-substrate networks.
• Experimental Prioritization: Supports prioritization of candidate kinases for experimental validation in studies of signal transduction.
• Cellular Process Investigation: Aids investigation of phosphorylation-driven processes such as cell cycle regulation, apoptosis, and differentiation.

Methodology:

Train a deep learning multi-label classifier across 134 kinase families using Siamese network-derived motif embeddings together with one-hot motif encodings and optimize with a kinase phylogeny-weighted loss that integrates phylogenetic relationships.