KIP

KIP predicts kinase–small-molecule interactions across 204 kinases using an auxiliary multi-task graph isomorphism network with uncertainty weighting (AMGU) to support lead discovery, drug repurposing, and off-target assessment.

Key Features:

• Multi-Task Graph Isomorphism Network (MT-GIN): The AMGU model is based on an MT-GIN backbone and predicts inhibitory activities across 204 kinases.
• Auxiliary Learning and Uncertainty Weighting: Auxiliary learning and uncertainty weighting are integrated into AMGU to leverage additional signals and to account for prediction confidence.
• Comparative Performance: AMGU outperforms descriptor-based models and state-of-the-art graph neural networks on internal test sets and two external test sets, indicating improved generalizability.
• Interpretability with Edges Masking: A model-agnostic edges masking method was used to interpret GNN predictions and to align insights with known structure–activity relationships, including inhibitors targeting EGFR.

Scientific Applications:

• Lead Discovery: Prediction of kinase–small-molecule inhibitory profiles to prioritize candidate lead compounds across the kinome.
• Drug Repurposing: Identification of polypharmacological profiles of existing compounds to reveal potential new therapeutic uses.
• Side Effect Elucidation: Assessment of potential off-target kinase interactions to inform safety and side-effect analysis.

Methodology:

Computational methods explicitly include an auxiliary multi-task graph isomorphism network with uncertainty weighting (AMGU) employing an MT-GIN backbone, auxiliary learning, uncertainty weighting, a model-agnostic edges masking interpretability approach, and evaluation on internal and two external test sets with comparisons to descriptor-based models and other GNNs.