New Methods and Tools for Computational Drug Discovery

项目摘要 我的目标是为药物发现开发有效和高效的计算方法，应用这些方法， 寻找新的和有效的药物来治疗疾病的方法，并将这些方法部署在易于使用的开放 源工具。我的研究小组率先开发和集成了深度神经网络， 用于基于结构的药物设计的用户友好的分子对接软件，以预测药物的位姿和效力。 小分子与它们的分子靶点结合。我们将在基础工作的基础上， 学习同时解决评分和抽样问题，这将克服可扩展性 现有方法的固有局限性。 我们建议为基于结构的药物设计开发第一个深度生成模型。不像tra- 经过筛选，生成建模并不局限于预定义的化学空间。在生成模式下- eling，一个深度神经网络学习分子结构和性质的潜在分布， 被描绘成一个潜在的空间。可以从这个习得的潜在空间中提取新的结构，以具有 理想的性能。理想情况下，生成模型将产生新颖的、接近最佳的分子结构 几乎是同时发生的我们假设使用现有的3D蛋白质和蛋白质训练生成模型， 配体结构将使我们能够创建通用模型，可以有效地应用于新的，结构， 由于蛋白质-配体相互作用的丰富性和普遍性，我们将进一步 开发这些方法，以支持生成优化的潜在客户候选人，其中生成 随着药物发现过程的进展，该过程被更新以包括来自实验测定的结果。 我们将继续应用我们的方法来识别分子相互作用的小分子调节剂， 与正常生理和疾病相关的功能。例如，使用我们目前的工具，我们确定了艾德 第一个profilin-actin相互作用的抑制剂，与癌症相关的抗血管生成靶点， 糖尿病视网膜病变，我们计划进一步改善这些化合物，目的是确定Candi- 临床试验的日期。我们将应用我们的方法来解决其他未开发的分子靶点， 例如NFATc 2，其与癌症和自身免疫性疾病有关。这些潜在的应用 我们的方法将提供公正和现实的评估，进一步了解他们的发展。 最后，我们所有的代码和经过训练的深度神经网络模型都将部署为新工具， 用于生成建模或增强我们广泛使用的计算药物开源工具 discovery：（1）PHARMIT，用于基于结构的药物发现的交互式网络应用;（2）GNINA， 用于分子对接的C/C++深度学习框架;以及（3）新发布的LIBMOLGRID， 用于加速分子网格化的Python库，与流行的深度学习工具包集成。 这些工具和方法将使药物发现过程更容易获得和更有效。

Project Summary My goal is to develop effective and efficient computational methods for drug discovery, apply these methods to find new and efficacious drugs to treat diseases, and deploy these methods in easy-to-use open source tools. My research group pioneered the development and integration of deep neural networks in user-friendly molecular docking software for structure-based drug design to predict poses and potency of small molecules binding to their molecular targets. We will build on our foundational work by using deep learning to simultaneously solving the scoring and sampling problems, which will overcome scalability limitations inherent in current approaches. We propose to develop the first deep generative models for structure-based drug design. Unlike tra- ditional screening, generative modeling is not limited to a predefined chemical space. In generative mod- eling, a deep neural network learns an underlying distribution of molecular structures and properties represented as a latent space. New structures can be extracted from this learned latent space to have desirable properties. Ideally, a generative model will produce novel, near-optimal molecular structures almost instantaneously. We hypothesize that training generative models using existing 3D protein and ligand structures will allow us to create general models that can be productively applied to new, struc- turally enabled targets due to the richness and universality of protein-ligand interactions. We will further develop these methods to support the generation of optimized lead candidates, where the generative process is updated to include results from experimental assays as the drug discovery process progresses. We will continually apply our methods to identify small molecule modulators of molecular interac- tions relevant to normal physiology and disease. For example, using our current tools, we identified the first inhibitors of the profilin-actin interaction, an anti-angiogenesis target with relevance to cancer and diabetic retinopathy, and we plan to further improve these compounds with the goal of identifying candi- dates for clinical testing. We will apply our methods to address other under-explored molecular targets, such as NFATc2, which is implicated in cancer and autoimmune diseases. These prospective applications of our methods will provide unbiased and realistic evaluations that further inform their development. Finally, all of our code and trained deep neural network models will be deployed either as new tools for generative modeling or as enhancements to our widely used open source tools for computational drug discovery: (1) PHARMIT, an interactive web application for structure-based drug discovery; (2) GNINA, a C/C++ deep learning framework for molecular docking; and (3) the newly released LIBMOLGRID, a Python library for accelerated molecular gridding that integrates with popular deep learning toolkits. These tools and methods will make the drug discovery process more accessible and efficient.