Evaluations of the Perturbation Resistance of the\nDeep-Learning-Based Ligand Conformation Optimization Algorithm

In\nrecent years, the deep learning (DL) technique has rapidly developed\nand shown great success in scoring the protein–ligand binding\naffinities. The protein–ligand conformation optimization based\non DL-derived scoring functions holds broad application prospects,\nfor instance, drug design and enzyme engineering. In this study, we\nevaluated the robustness of a DL-based ligand conformation optimization\nprotocol (DeepRMSD+Vina) for optimizing structures with input perturbations\nby examining the predicted ligand binding poses and scoring. Our results\nclearly indicated that compared to traditional optimization algorithms\n(such as Prime MM-GBSA and Vina optimization), DeepRMSD+Vina exhibits\nhigher performance when treating diverse protein–ligand cases.\nThe DeepRMSD+Vina is robust and can always generate the correct binding\nstructures even if perturbations (up to 3 Å) are introduced to\nthe input structure. The success rate is 62% for perturbation with\na RMSD within 2–3 Å. However, the success rate dramatically\ndrops to 11% for large perturbations, with RMSD extending to 3–4\nÅ. Furthermore, compared to the widely used optimization protocol\nof AutoDock Vina, the DL-generated conformation shows a balanced performance\nfor all of the systems under examination. Overall, the DL-based DeepRMSD+Vina\nis unarguably more reliable than the traditional methods, which is\nattributed to the physically inspired design of the neural networks\nin DeepRMSD+Vina where the distance-transformed features describing\nthe atomic interactions between the protein and the ligand have been\nexplicitly considered and modeled. The outstanding robustness of the\nDL-based ligand conformational optimization algorithm further validates\nits superiority in the field of conformational optimization.