Fröhlking Thorben, Rizzi Valerio, Aureli Simone, Gervasio Francesco Luigi
School of Pharmaceutical Sciences, University of Geneva, Rue Michel Servet 1, 1206 Genève, Switzerland.
Institute of Pharmaceutical Sciences of Western Switzerland (ISPSO), University of Geneva, 1206 Genève, Switzerland.
J Chem Phys. 2024 Sep 21;161(11). doi: 10.1063/5.0226721.
Path-like collective variables (CVs) can be very effective for accurately modeling complex biomolecular processes in molecular dynamics simulations. Recently, we have introduced DeepLNE (deep-locally non-linear-embedding), a machine learning-based path-like CV that provides a progression variable s along the path as a non-linear combination of several descriptors. We have demonstrated the effectiveness of DeepLNE by showing that for simple models such as the Müller-Brown potential and alanine dipeptide, the progression along the path variable closely approximates the ideal reaction coordinate. However, DeepLNE is computationally expensive for realistic systems needing many descriptors and limited in its ability to handle multi-state reactions. Here, we present DeepLNE++, which uses a knowledge distillation approach to significantly accelerate the evaluation of DeepLNE, making it feasible to compute free energy landscapes for large and complex biomolecular systems. In addition, DeepLNE++ encodes system-specific knowledge within a supervised multitasking framework, enhancing its versatility and effectiveness.
路径类集体变量（CVs）在分子动力学模拟中对复杂生物分子过程进行精确建模时可能非常有效。最近，我们引入了DeepLNE（深度局部非线性嵌入），这是一种基于机器学习的路径类CV，它将沿着路径的进展变量s作为几个描述符的非线性组合来提供。我们通过表明对于诸如穆勒-布朗势和丙氨酸二肽等简单模型，沿着路径变量的进展紧密近似理想反应坐标，证明了DeepLNE的有效性。然而，对于需要许多描述符的实际系统，DeepLNE计算成本高昂，并且在处理多态反应的能力方面有限。在此，我们提出了DeepLNE++，它使用知识蒸馏方法来显著加速DeepLNE的评估，从而使得为大型复杂生物分子系统计算自由能景观变得可行。此外，DeepLNE++在监督多任务框架内编码特定于系统的知识，增强了其通用性和有效性。
