Network-principled deep generative models for designing drug combinations as graph sets.

Area of Science:

• Computational biology
• Drug discovery
• Artificial intelligence

Background:

• Combination therapy improves efficacy and reduces side effects, crucial for overcoming drug resistance.
• Current computational methods for small-molecule drug combinations are limited by vast chemical spaces and unclear design principles.
• Generative models hold potential for accelerating the discovery of resistance-overcoming drug combinations.

Purpose of the Study:

• To develop the first deep generative model for computational drug combination design.
• To accelerate the discovery of drug combinations that overcome resistance.
• To generate disease-specific drug combinations with improved efficacy and reduced toxicity.

Main Methods:

• Developed hierarchical variational graph auto-encoders to embed gene-gene, gene-disease, and disease-disease networks.
• Introduced novel attentional pooling for learning disease representations from gene representations.
• Recast drug combination design as graph-set generation using a deep learning model with chemical validity, generative adversarial, and network principle-based rewards.

Main Results:

• Hierarchical variational graph auto-encoders learned more informative and generalizable disease representations than state-of-the-art methods.
• Deep generative models successfully generated drug combinations adhering to network principles across diseases.
• Case studies demonstrated that network-principled combinations exhibited low toxicity and covered disease modules similarly to FDA-approved drugs.

Conclusions:

• The developed deep generative model enables efficient, network-principled generation of disease-specific drug combinations.
• This approach can explore vast chemical spaces to identify novel drug combinations for overcoming resistance.
• The model offers potential for novel systems pharmacology strategies and improved therapeutic outcomes.