Conformational Analysis of Macrocyclic Compounds Using a Machine-Learned Interatomic Potential

Macrocyclic compounds play a vital role in many chemical and biological systems, yet their conformational analysis remains a significant challenge. In this work, we investigate the conformational landscape of macrocyclic compounds using a machine-learned interatomic potential (MLIP) based on a Nequip-like graph neural network. This MLIP is trained on the energy differences between ωB97XD3 and GFN1-xTB. The model not only reproduces the DFT relative conformer energies of the macrocycles with high fidelity but also yields optimized structures that are practically identical to those obtained via density functional theory. Furthermore, when integrated into a metadynamics-based conformational sampling framework (CREST), we recover structures that very closely match the structure obtained after gas-phase optimization with DFT starting from the crystal structure. These results underscore the potential of machine learning to overcome longstanding challenges in the conformational analysis of complex macrocyclic systems.

大环化合物在诸多化学与生物体系中发挥着关键作用，但其构象分析仍是一项极具挑战性的课题。本研究基于类Nequip图神经网络构建机器学习原子间势（machine-learned interatomic potential，MLIP），用以探究大环化合物的构象势能面。该MLIP以ωB97XD3与GFN1-xTB之间的能量差作为训练数据进行训练。此模型不仅能够高精度复现大环化合物的密度泛函理论（DFT）相对构象能，还可得到与密度泛函理论优化结果几乎完全一致的优化结构。此外，当将该模型集成至基于元动力学的构象采样框架（CREST）中时，从晶体结构出发经气相DFT优化后得到的结构，与该框架所恢复的结构高度吻合。上述结果凸显了机器学习在攻克复杂大环体系构象分析这一长期存在的难题方面的巨大潜力。