KSFinder

KSFinder predicts kinase–substrate links by embedding a phosphoproteome knowledge graph and classifying embeddings to identify phosphorylation relationships across over 85% of known human kinases, including 68 Illuminating the Druggable Genome "dark" kinases.

Key Features:

• Knowledge Graph Embedding: Uses a knowledge graph embedding algorithm to represent phosphoproteome nodes as low-dimensional vectors capturing semantic relationships among proteins.
• Multilayer Perceptron Classifier: Processes embedded vectors with a multilayer perceptron (MLP) classifier trained to distinguish kinase–substrate links.
• Strategic Negative Generation: Generates negative training samples by combining experimentally validated non-interacting protein pairs, proteins in different subcellular compartments, and random sampling to reduce bias.
• Coverage of Human Kinases: Provides predictive coverage for over 85% of known human kinases.
• Generalization Capability: Assesses performance across four distinct datasets and demonstrates superior generalization compared to other kinase–substrate prediction models.
• Focus on "Dark" Kinases: Targets substrates of 68 understudied kinases identified by the Illuminating the Druggable Genome program.
• Literature Evidence Integration: Integrates RLIMS-P text-mining with manual curation to search for literature evidence supporting predictions.
• Case Study Novel Interactions: Identified 17 novel kinase–substrate interactions with probability scores ≥0.7 in a reported case study.

Scientific Applications:

• Functional Enrichment Analysis: Enables functional enrichment analysis of predicted substrates for dark kinases such as HIPK3 and CAMKK1, revealing biological processes including extracellular matrix regulation, epigenetic gene expression, lipid storage regulation, and glucose homeostasis.
• Novel Therapeutic Targets: Generates candidate kinase–substrate links for understudied kinases to inform discovery of potential therapeutic targets in diseases associated with aberrant phosphorylation.

Methodology:

Constructs and embeds a phosphoproteome knowledge graph; trains an MLP classifier on embeddings using strategically generated negative samples (experimentally validated non-interacting pairs, subcellular compartment differences, and random sampling); evaluates generalization across four distinct datasets; and validates predictions via RLIMS-P text-mining, manual curation, and functional enrichment analysis.