Multi-Objective Structure-Based Drug Design Using Causal Discovery.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Structure-based drug design (SBDD) is crucial for identifying drug candidates.
• Deep generative models are increasingly vital in SBDD.
• Current SBDD methods often fail to address multi-objective optimization due to property interdependencies.

Purpose of the Study:

• To develop a multi-objective SBDD algorithm capable of optimizing multiple drug properties concurrently.
• To address the limitations of existing methods that neglect property relationships or focus on single objectives.
• To generate drug candidates with improved overall potential by considering complex property interactions.

Main Methods:

• A novel multi-objective SBDD algorithm based on diffusion models.
• Parallel training of multiple expert networks for property prediction and gradient transmission.
• Construction of a causal graph using causal discovery algorithms to model property relationships.
• Decomposition of joint property distributions guided by the causal graph during molecule generation.

Main Results:

• The proposed algorithm effectively optimizes multiple objectives, including binding affinity and other crucial drug properties.
• It successfully generates molecules with higher drug potential compared to baseline models.
• Demonstrated ability to handle conflicting property relationships during the generative process.

Conclusions:

• The developed diffusion model-based SBDD algorithm offers a significant advancement in multi-objective drug design.
• Considering causal relationships among properties is key to generating superior drug candidates.
• This approach enhances the feasibility of designing molecules that meet diverse and potentially conflicting criteria.