DeepKinZero

DeepKinZero applies zero-shot learning to predict which kinases phosphorylate protein sites, enabling kinase assignment for phosphosites when no prior phosphosite information exists.

Key Features:

• Zero-shot learning: Enables prediction for kinases with no documented phosphosites by transferring information from characterized kinases.
• Bidirectional recurrent neural network (RNN): Uses a bidirectional RNN to model sequence patterns around phosphosites.
• Kinase-specific positional amino acid preferences: Learns positional amino acid preferences specific to individual kinases to inform predictions.
• Proteome-level phosphosite compatibility: Operates on phosphosites identified at the proteome level.
• Improved predictive performance: Produces higher accuracy than baseline models and existing methods as reported.
• Addresses phosphoproteome data sparsity: Targets the issue that over 95% of reported human phosphosites lack known kinase associations.

Scientific Applications:

• Phosphoproteome mapping: Assigns kinases to phosphosites to aid construction of the phosphoproteome atlas.
• Characterization of understudied kinases: Expands knowledge of kinases with limited or no prior substrate information.
• Investigation of cellular processes and disease mechanisms: Supports research into kinase roles in signaling and diseases including cancer.

Methodology:

DeepKinZero employs a zero-shot learning framework using a bidirectional recurrent neural network to learn kinase-specific positional amino acid preferences and transfer knowledge from well-characterized kinases to predict kinases for phosphosites with no prior annotations.