AI-Designed Family of Dual COX-2/mPGES-1 Inhibitors to Treat Inflammation with Greater Precision and Effectiveness

Cyclooxygenase-2 (COX-2) selective inhibitors represent a major class of anti-inflammatory drugs, but their clinical utility remains constrained by cardiovascular risks and incomplete efficacy. Dual inhibition of COX-2 and microsomal prostaglandin E synthase-1 (mPGES-1) offers a promising strategy for enhanced safety and effectiveness. However, traditional drug discovery timelines for multi-target inhibitors are prohibitively long. Here, we demonstrate an AI-driven computational pipeline that compressed the design-to-candidate stage from years to one month, generating a novel family of dual COX-2/mPGES-1 inhibitors. We employed a hybrid reinforcement learning-generative adversarial network (RL-GAN) trained on 308 curated COX-2 inhibitors to explore beyond known chemical space, producing ~50,000 de novo molecules. Multi-stage filtering through ADMET prediction models, billion-scale virtual screening using deep learning-enhanced docking protocols, and structure-based design against four COX-2 crystal structures yielded 23 high-confidence candidates. These molecules exhibit predicted COX-2 binding affinities comparable to or exceeding rofecoxib (-9 to -10.5 kcal/mol), with >100-fold predicted selectivity over COX-1 and concurrent mPGES-1 engagement (-8.3 to -10.3 kcal/mol). Computational patent analysis across 44.3 million records suggests structural novelty (Tanimoto <0.8), with 92.9% of candidates dissimilar to known inhibitors. Rigorous validation of computational models, including Boltz-2 predicted COX-1 structure (backbone RMSD 0.541 Å vs. 6Y3C crystal structure), supports prediction reliability. This work demonstrates that AI-guided de novo design can rapidly identify innovative multi-target therapeutics, offering a paradigm shift in early-stage drug discovery efficiency while highlighting the essential role of experimental validation in clinical translation.