2025/03/14 by Mehran Jalaie, Jindřich Fanfrlík, Adam Pecina +2 · 2 voices
Engineering · Materials Science · Biochemistry, Genetics and Molecular Biology · #Molecular Junctions and Nanostructures #Machine Learning in Materials Science #Advanced Electron Microscopy Techniques and Applications
paper · pdf · doi:10.26434/chemrxiv-2025-38lf5
Single-structure scoring functions have been considered inferior to expensive ensemble free energy methods in predicting protein-ligand affinities. We are revisiting this dogma with the recently developed semiempirical quantum-mechanical (SQM)-based scoring function, SQM2.20, comparing its performance to the standard scoring functions on one hand and state-of-the-art molecular dynamics (MD)-based free-energy methods on the other hand. The comparison is conducted on a well-established Wang dataset comprising eight protein targets with 200 ligands. The initial low correlation of SQM2.20 scores with the experimental binding affinities of R² = 0.21 was improved to R² = 0.47 by a systematic refinement of the input structures. Consequently, SQM2.20 representing accurate single-structure scoring functions, exhibited an average performance comparable to that of MD-based methods (R² = 0.52) and surpassed the performance of standard scoring functions (R² = 0.26). The per-target analysis highlighted the pivotal role of high-quality input structures on the outcomes of single-structure methods. In the instances where such structures are available, SQM2.20 scoring has been shown to rival or even exceed MD-based methods in predicting protein-ligand binding affinities, while exhibiting significantly reduced computation time.