Labels as a feature: Network homophily for systematically annotating human GPCR drug-target interactions

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Human G protein-coupled receptors are crucial targets for FDA-approved drugs, but comprehensive drug-target interaction testing is limited by cost and technical hurdles.
• Unexplored off-target effects of drugs pose significant risks to patient safety.
• Traditional drug discovery models often focus on exploring new chemical spaces rather than optimizing predictions within known spaces.

Purpose of the Study:

• To develop a novel machine learning model, Chemical Space Neural Networks (CSNN), for accurate prediction of drug-target interactions.
• To investigate the role of network homophily and labels as features in enhancing in-distribution prediction accuracy.
• To validate the CSNN model in a high-throughput experimental system for discovering novel drug-target interactions.

Main Methods:

• Developed a neighborhood-to-prediction model termed Chemical Space Neural Networks (CSNN).
• Utilized network homophily and training-free graph neural networks with labels as features.
• Integrated labeled data during inference to enhance prediction accuracy.
• Validated the model using a high-throughput yeast biosensing system with 3773 drug-target interactions, 539 compounds, and 7 human G protein-coupled receptors.

Main Results:

• CSNN's prediction accuracy strongly correlates with network homophily.
• Using labels as features significantly enhances a machine learning model's in-distribution prediction capacity.
• The model successfully identified novel drug-target interactions for FDA-approved drugs within the experimental system.

Conclusions:

• Chemical Space Neural Networks offer a reliable approach to enhance in-distribution prediction accuracy for drug-target interactions.
• Leveraging network homophily and labeled data is crucial for building robust predictive models in drug discovery.
• This work provides a foundation for guiding experimental verification and expanding the understanding of drug-target interactions, ultimately improving drug safety and discovery pipelines.