A multitask GNN-based interpretable model for discovery of selective JAK inhibitors

Area of Science:

• Biochemistry
• Computational Chemistry
• Pharmacology

Background:

• The Janus kinase (JAK) family is crucial in cytokine-mediated inflammatory and autoimmune responses through JAK/STAT signaling.
• JAK inhibitors represent a promising therapeutic strategy for diseases like COVID-19.
• Designing selective JAK inhibitors is challenging due to high homology among JAK isoforms.

Purpose of the Study:

• To develop an interpretable GNN multitask regression model for simultaneously predicting pIC50 values of compounds across all JAK subtypes.
• To identify key atoms and substructures for designing selective JAK inhibitors.

Main Methods:

• Construction of an interpretable GNN multitask regression model.
• Simultaneous prediction of pIC50 values for all JAK subtypes.
• Calculation and visualization of atom weights.
• Rank sum tests and local mean comparisons to identify key molecular features.

Main Results:

• The GNN model achieved high performance with R2 values of 0.96 (training), 0.79 (validation), and 0.78 (test sets).
• Key atoms and substructures influencing JAK inhibition selectivity were identified.
• Case studies confirmed the model's feasibility and ability to learn protein-small molecule interactions.

Conclusions:

• The developed interpretable GNN model offers a novel approach for discovering and designing selective JAK inhibitors.
• The model's ability to identify key molecular features can guide the fine-tuning of JAK inhibitors for improved selectivity.
• This approach provides a valuable tool for researchers in drug discovery for inflammatory and autoimmune diseases.