ActivityDiff: A diffusion model with Positive and Negative Activity Guidance for De Novo Drug Design

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• De novo drug design aims to optimize target activity while minimizing off-target interactions for safety and selectivity.
• Current generative models often neglect negative activity data, hindering the suppression of undesired effects.
• Off-target interactions are a major cause of drug attrition and adverse events.

Purpose of the Study:

• To introduce ActivityDiff, a novel framework for activity-controlled molecular generation.
• To integrate both positive and negative guidance for improved drug design.
• To enhance the safety and efficacy of generated drug candidates.

Main Methods:

• Developed a classifier-guided diffusion framework named ActivityDiff.
• Incorporated separately trained drug-target classifiers for positive and negative guidance.
• Utilized explicit negative guidance to suppress harmful off-target interactions during generation.

Main Results:

• ActivityDiff successfully generated molecules for single- and dual-target design tasks.
• The framework supported fragment-constrained dual-target generation and selective generation for specificity.
• Demonstrated significant reduction in off-target effects, improving overall drug safety.
• Showcased the effectiveness of negative guidance in jointly optimizing efficacy and safety.

Conclusions:

• Classifier-guided diffusion with explicit negative guidance is an effective strategy for molecular design.
• ActivityDiff enables simultaneous optimization of desired activity and minimization of off-target effects.
• This approach represents a significant advancement in developing safer and more selective drug candidates.