DrugEx

DrugEx generates small-molecule candidates for multi-objective drug design in polypharmacology to optimize target selectivity and minimize off-target interactions.

Key Features:

• Multi-Objective Optimization: Implements multi-objective optimization to generate drug-like molecules tailored for interactions with multiple biological targets while minimizing off-target interactions.
• Reinforcement Learning Framework: Uses a reinforcement learning framework with pre-trained Recurrent Neural Network (RNN) agents that interact with an environment composed of machine learning predictors to iteratively improve molecule generation.
• Evolutionary Algorithm Integration: Integrates evolutionary concepts, specifically crossover and mutation operations, executed by the same deep learning model acting as the agent to enhance diversity and quality of generated molecules.
• Pareto-Based Ranking System: Produces batches of SMILES and evaluates them using environment-assigned multi-objective scores to construct Pareto ranks via non-dominated sorting and Tanimoto-based crowding distance calculations using chemical fingerprints.
• GPU Acceleration: Leverages GPU acceleration to expedite the computationally intensive Pareto optimization and training convergence.
• Reward-Based Training: Determines final rewards for molecules based on their Pareto ranking to guide the agent toward desired generation outcomes.

Scientific Applications:

• Polypharmacology multi-target design: Designs compounds with diverse selectivity profiles to modulate multiple pathways simultaneously in polypharmacological contexts.
• Adenosine receptor profiling: Generates candidate compounds optimized for activity profiles against adenosine receptors A1AR and A2AAR.
• hERG liability reduction: Enables generation of compounds with reduced predicted interaction with the potassium ion channel hERG to lower toxicity risk.

Methodology:

Pre-trained RNN agents operate within a reinforcement learning framework interacting with machine learning predictors to generate batches of SMILES; the environment assigns multi-objective scores used to construct Pareto ranks via non-dominated sorting and Tanimoto-based crowding distance on chemical fingerprints; crossover and mutation operations are performed by the deep learning agent; final rewards are derived from Pareto rankings, and GPU acceleration is applied to Pareto optimization.