Kinact

Kinact predicts kinase-activating missense mutations using a machine learning framework that integrates structural and sequence features to identify mutations relevant to cancer biology.

Key Features:

• Machine learning model: Uses a machine learning framework to predict kinase-activating missense mutations.
• Structural and sequence integration: Integrates both structural and sequence data to inform predictions.
• Environment residue: Analyzes surrounding residues that influence the effects of mutations.
• Stability change predictions: Estimates mutation-induced changes in protein stability.
• Atomic interactions: Evaluates atomic-level interactions to assess functional impact.
• Graph-based signatures: Represents structural and sequence information as graph-based signatures capturing protein architecture.
• Residue conservation: Incorporates residue conservation to assess evolutionary importance of mutations.
• Performance: Reports precision of 87% (cross-validation) and 94% (blind tests) with AUCs of 0.89 and 0.92 respectively, and outperforms established methods (P < 0.01).
• Homology model tolerance: Maintains high accuracy with homology models built from templates with as low as 33% sequence identity.

Scientific Applications:

• Cancer mutation prioritization: Prioritizes kinase-activating missense mutations in cancer datasets for research and clinical interpretation.
• Variant functional interpretation: Interprets the likely functional impact of missense variants in protein kinases to guide experimental follow-up.
• Structural interpretation: Enables evaluation of mutation effects using experimental structures or homology models down to 33% sequence identity.
• Support for therapeutic research: Informs studies of cancer biology and the development of therapeutic strategies targeting kinases.

Methodology:

Integrates structural and sequence data into machine learning models using features explicitly including environment residue, stability change predictions, atomic interactions, graph-based signatures, and residue conservation, and can operate on experimental structures or homology models (down to ~33% sequence identity).