End-Point Affinity Estimation of Galectin Ligands by Classical and Semiempirical Quantum Mechanical Potentials

The use of quantum mechanical potentials in protein–ligand affinity prediction is becoming increasingly feasible with growing computational power. To move forward, validation of such potentials on real-world challenges is necessary. To this end, we have collated an extensive set of over a thousand galectin inhibitors with known affinities and docked them into galectin-3. The docked poses were then used to systematically evaluate several modern force fields and semiempirical quantum mechanical (SQM) methods up to the tight-binding level under consistent computational workflow. Implicit solvation models available with the tested methods were used to simulate solvation effects. Overall, the best methods in this study achieved a Pearson correlation of 0.7–0.8 between the computed and experimental affinities. There were differences between the tested methods in their ability to rank ligands across the entire ligand set as well as within subsets of structurally similar ligands. A major discrepancy was observed for a subset of ligands that bind to the protein via a halogen bond, which was clearly challenging for all the tested methods. The inclusion of an entropic term calculated by the rigid-rotor-harmonic-oscillator approximation at SQM level slightly worsened correlation with experiment but brought the calculated affinities closer to experimental values. We also found that the success of the prediction strongly depended on the solvation model. Furthermore, we provide an in-depth analysis of the individual energy terms and their effect on the overall prediction accuracy.随着计算能力的持续提升，量子力学势能在蛋白质-配体亲和力预测中的应用正变得愈发可行。为推动该领域进一步发展，针对此类势能开展真实场景挑战下的验证工作实属必要。为此，我们整理了一套涵盖千余种已知亲和力半乳糖凝集素（galectin）抑制剂的大型数据集，并将其对接至半乳糖凝集素3（galectin-3）靶点。随后，我们基于统一的计算流程，利用所得对接构象对多款现代分子力场以及紧束缚层级的半经验量子力学（semiempirical quantum mechanical, SQM）方法开展系统性评估。测试方法所配套的隐式溶剂化模型被用于模拟溶剂化效应。本研究中表现最优的方法，其计算所得亲和力与实验亲和力间的皮尔逊相关系数可达0.7至0.8。不同测试方法在对全部配体集以及结构相似配体子集进行排序的能力上存在差异。其中，针对通过卤键与蛋白结合的配位子集，我们观测到了显著偏差，此类配体对所有测试方法而言均颇具挑战性。在半经验量子力学层级引入基于刚性转子-谐振子近似计算得到的熵项，虽小幅降低了与实验结果的相关性，却使计算所得亲和力更贴近实验实测值。我们还发现，预测效果极大程度上依赖于所选用的溶剂化模型。此外，本研究针对各能量项及其对整体预测精度的影响展开了深入分析。