Drug optimization requires balancing multiple competing objectives: high on-target binding affinity, minimal off-target binding, favorable drug-likeness (QED), synthetic accessibility (SA), and scaffold similarity. Current approaches suffer from three limitations: (1) manual off-target panel curation, (2) lack of systematic off-target identification, and (3) fragmented multi-objective optimization.
Off-target identification is non-trivial. Considering all human proteins is computationally intractable. Instead, off-targets should be selected based on biological relevance---proteins interacting with the primary target in PPI networks or known safety targets. This requires integration of biological knowledge bases with molecular generation pipelines.
Our approach integrates knowledge graph-guided protein-protein interaction (PPI) analysis with reinforcement learning-based molecular optimization to generate off-target aware drug candidates. The system leverages the STRING database to identify interacting proteins for a given drug target, automatically trains quantitative structure-activity relationship (QSAR) models using ChEMBL bioactivity data, and employs DrugEx's graph transformer architecture with multi-objective reinforcement learning to optimize molecules across on-target affinity, off-target penalty, drug-likeness (QED), synthetic accessibility (SA), and structural similarity constraints.
The technical implementation builds on DrugEx's graph transformer generator and reinforcement learning framework, enhanced with automated QSAR model training using Random Forest classifiers, PPI network querying via the STRING API, and ChEMBL data integration for bioactivity retrieval. The multi-objective optimization employs reward shaping and either weighted sum or Pareto ranking schemes to balance competing objectives, enabling the generation of optimized molecules that maintain structural similarity to input scaffolds while maximizing therapeutic efficacy and minimizing off-target interactions.