Background: Protein folding remains one of computational biology's grand challenges, with the search space growing exponentially with sequence length. 5Monte Carlo methods have dominated stochastic optimization since 1946, but struggle with the rugged energy landscapes characteristic of protein folding problems. 6Methods: We introduce the Forgetting Engine (FE), a novel optimization algorithm that treats elimination as an active computational primitive rather than a limitation. 7FE strategically eliminates the worst-performing 30% of candidate solutions each generation while paradoxically retaining a small subset of these "contradictory" states that show structural promise despite high energy. 8We validated FE against standard Monte Carlo (MC) methods using the 3D hydrophobic-polar (HP) lattice model with a 20-residue test sequence (HPHPHPHΗΡΗΗΗΡHPPPHPH) across 4, 000 independent trials. 9Results: FE achieved a 25. 4% success rate compared to MC's 5. 5%, representing a 362% improvement (4. 62-fold increase). 10This superiority was statistically overwhelming (p<10^-12, Cohen's d = 1. 056, 99% confidence intervals non-overlapping). 11FE also achieved 19. 5% better mean energy (-8. 408 vs -7. 037) and 12. 8% lower standard deviation. 12Mechanistic analysis revealed that paradox retention was active in 100% of FE trials (mean 15 retentions per trial) and correlated with improved energy optimization (r=-0. 087, p=0. 0001). 13Conclusions: The Forgetting Engine represents a paradigm shift in optimization by making "forgetting" an active tool rather than a limitation. 14The 362% improvement over Monte Carlo, combined with pharmaceutical-grade statistical validation (p<10^-12), establishes FE as a transformative approach for protein folding and potentially other NP-hard optimization problems. 15Our pre-registered protocol and fixed random seeds ensure full reproducibility.
Derek Angell (Thu,) studied this question.