Simulation of Multi-Protein systems

研究的主要领域，由赠款R01 GM118078支持，题为“模拟多蛋白质 系统”，正在研究细胞组装中蛋白质之间的相互作用。阐明这种 相互作用将进一步机械理解细胞过程，并将使有效的 治疗发展。要求NIGMS行政补充，以获得高绩效 GPU集群用于开发物理感知深度学习模型。通常，原子模拟是 使用基于物理或机器学习的方法进行，结果由两个 方法彼此正交。项目的目标不是简单的方法组合， 开发特殊的深度神经网络架构，编码的几何和物理的 底层系统这种网络的推理将执行逼真的物理模拟， 在可用数据上训练参数，以最佳地再现实验观察结果。培训 这样的模型需要大量的GPU资源，因此证明了对集群的需要。新款车型 将被应用于了解健康和患病细胞的分子机制， 基因组学规模。此外，Kozakov实验室的研究结果也可供使用。 通过大分子建模服务器MIPPro，一个拥有超过25000用户的公共资源， 每个月都要进行上万次的计算。在服务器上部署这些方法还 需要大量的GPU资源。因此，购买的GPU集群将服务于双重目标，首先提供 培训时间来开发和应用新的模型，然后添加到XNUMX Pro后端。

The main area of research, supported by the grant R01 GM118078, entitled “Simulation of Multi-Protein Systems”, is investigating the interactions between proteins in cellular assemblies. Elucidating such interactions will further mechanistic understanding of cellular processes, and will enable effective therapy development. NIGMS administrative supplement is requested to acquire a high performance GPU cluster to develop physics aware deep learning models . Typically atomic simulations are performed using either physics or machine learning based approaches, with results obtained by the two methods being orthogonal to each other. Instead of a trivial combination of methods, the project goal is to develop special deep neural network architectures, which encode geometry and physics of the underlying system. The inference of such network will be performing realistic physical simulations, with parameters trained on the available data to optimally reproduce experimental observations. Training of such models require significant GPU resources thus justifying the need for the cluster. The new models will be applied to understanding the molecular mechanisms in healthy and diseased cells on the genomics scale. In addition, the results of the research in the Kozakov lab are made available through the macromolecular modeling server ClusPro, a public resource with more than 25000 users running tens of thousands calculation each month. Deploying these methods on the server is also requiring significant GPU resources. Thus the purchased GPU cluster will serve dual goals, first provide training time to develop and apply the new models and then adding to the ClusPro backend.