Development of Enhanced Sampling Methods for Biomolecular Simulations applying Machine Learning methods.

对于小蛋白质和简单的酶，只需几微秒的分子动力学（MD）模拟就可以实现罕见事件的生物分子模拟。然而，当处理更大、更复杂的生物分子时，在蛋白质的自由能景观中取样是很困难的。其中一些事件的过渡态（TS）和激活势垒可能太高，无法通过自由MD攀登，因此近年来开发了许多不同的方法来克服这些障碍并对它们之间的TS区域进行采样。到目前为止，目标是开发新的方法，通过处理结果数据并更新方法或直接将其应用于方法，从而在机器学习（ML）的帮助下改进采样方面。例如，复制交换（Rex）方法正被用于Mg电荷获得大约30%的交换概率，我们将进一步研究它的行为和配置从第一个窗口运行到结束并返回所需的时间，因此我们可以改进计算时间。这是用Crispr-Cas9蛋白系统完成的，上面附着一条DNA链。在77个窗口中，通过指数衰减，MG电荷从有意义的电荷（2.0）到低电荷（0.4）不等。有了这个，我们希望我们将能够使用这种方法在活性位点找到正确的金属配位，并开发一个管道，使其易于应用于任何其他金属酶。除了Rex外，我们还试图提出最佳反应坐标来描述具有良好抑制剂的CDK2系统的结合/解结合事件的FEL。我们正在通过测试一些神经网络（NN）算法（如卷积神经网络或多层感知器神经网络）来测试基于监督学习的ML方法。然后使用最优反应坐标来调整采样方法，以找到到达事件TS的最佳路径。

Biomolecular simulations of rare events are accessible for small proteins and simple enzymes with few microseconds of Molecular Dynamics (MD) simulations. Nevertheless, It is way harder to sample the interesting events in the free energy landscape of a protein when dealing with bigger and way more complex biomolecules. The Transition State (TS) and activation barrier of some of those events may be too high to climb by means of free MD, thus many different methods to overcome the barriers and sample the TS regions between them had been developed in the recent years. So far, the goal is to develop new methods that could improve on the sampling aspect with the aid of Machine Learning (ML) by treating the resulting data and updating the method or either applying it directly in the method. For instance, a Replica Exchange (Rex) method is being adapted for Mg charges to get an approximate of 30% probability of exchange, we will further study its behaviour and the amount of time required for a configuration to run from the first window to the end and get back, so we can improve on calculation time. This is done with for Crispr-Cas9 protein system with a DNA strand attached to it. The MG charge ranges from a meaningful charge (2.0) to a low one (0.4) by an exponential decay in 77 windows. With this we hope we will be able to use this method to find the correct metal coordination for this protein in the active site, and develop a pipeline to apply it easily to any other metal enzymes. Besides the Rex, we are also trying to come up with the optimal reaction coordinates to describe the FEL of a binding/unbinding event for the CDK2 system with a good inhibitor. We are testing a ML approach based on supervised learning by testing some Neural Network (NN) algorithms such as convolutional NN or Multi-layer Perceptron NN. The optimal reaction coordinates will be used then to adapt the sampling method to find the best path to the TS of the event.