Purchase of a GPU cluster for deep learning applications in protein-protein interaction and supercomplex prediction and biochemical literature annotation.

项目总结 尽管在过去的二十年里见证了对细胞成分的大规模分析，例如外显子组，但它们的 对药物发现和精准医学的影响一直不大。例如，6/7的候选药物没有通过 在FDA最近的临床试验中，安全性和3/4的疗效失败。这些尚未解决但相关的问题，安全性和有效性， 反映了对疾病之间的三角相互关系、分子功能、 和药物治疗。当代药物发现的一个关键概念限制是经常隐含地假设 以单一药物为单一蛋白靶点的疾病模型。事实上，大多数疾病都是由多种原因引起的 有故障的分子。无论是疾病治疗还是精准医学诊断，往往都有 无法识别疾病相关的作用模式(MOA)蛋白。为了开始解决这些问题，请参阅 目前的Mira建议，我们开发了一种很有前途的蛋白质结构和基于网络的人工智能(AI) 方法，MEDICASCY，预测疾病相关MOA蛋白、药物适应症、副作用和疗效； 然而，还需要做更多的工作。在这里，我们建议在我们成功的基础上，开发一个完整的 基于人工智能的方法，MEDICASCY-X，解决以下问题：确定药物MOA的第一步 而靶外相互作用就是识别它的蛋白质靶点。这需要所有人类蛋白质的结构和 他们的情结。虽然我们预测了97%的人类蛋白质中至少一个结构域的合适模型，但使用 深度学习，我们将预测缺失结构域的结构，结构域域取向和蛋白质- 蛋白质复合体。我们将扩展小分子虚拟配体筛选(VLS)来预测结合亲和力 基于相互作用的环-蛋白质亚袋几何结构和化学是保守的 蛋白质家族，通常是具有特权的化学结构，可能是低自由能复合体。隐蔽蛋白 最近被确认为重要药物目标的口袋将被预测并包括在我们的VLS方法中。 基于抗体的免疫疗法是有效的，但具有与小分子相似的安全性和有效性问题； 因此，它们的安全性和有效性将通过MEDICASCY-X进行预测。而MEDICASCY的工作对象是 人类“，MEDICASCY-X将考虑个体遗传和表观遗传特征，使其成为真正的精确度 医药工具。我们将预测哪些MOA蛋白应该成为目标，以及如果蛋白质的MOA是由于丢失或 功能的获得。同样的框架将预测药物与药物之间的协同作用。确定优先顺序的另一种方法 MOA蛋白是由疾病共病引起的：在多种疾病中出现的蛋白可能是重要的。如果疾病 共病是可以预测的，我们将构建疾病的系统发育树(S)，这将有助于 更深入地了解疾病之间的相互关系。作为算法有效性的原则证明 正在开发中，将开发针对各种顽固性疾病的新的临床前治疗方法。因此，这一点 项目可以提高药物发现和精准医学的成功率，同时减少时间和成本。

PROJECT SUMMARY Although the past two decades witnessed the large-scale analyses of cellular components, e.g. exomes, their impact on drug discovery and precision medicine has been modest. For example, 6/7 drug candidates failed safety and 3/4 failed efficacy in recent FDA clinical trials. These unsolved, but related issues, safety and efficacy, reflect significant gaps in understanding of the triangular interrelationship between diseases, molecular function, and drug treatments. A key conceptual limitation of contemporary drug discovery is the often implicitly assumed single drug for a single protein target disease model. In reality, most diseases are caused by multiple malfunctioning molecules. Whether it be disease treatment or precision medicine diagnostics, there is often an inability to identify disease-associated mode of action (MOA) proteins. To begin to address these issues, in the current MIRA proposal, we developed a promising protein structure and network-based Artificial Intelligence (AI) approach, MEDICASCY, to predict disease-associated MOA proteins, drug indications, side effects and efficacy; however, much more needs to be done. Here, we propose to build on our successes and develop an integrated AI-based approach, MEDICASCY-X, that addresses the following: The first step in determining a drug’s MOA and off-target interactions is to identity its protein targets. This requires the structures of all human proteins and their complexes. While we predicted suitable models for at least one domain in 97% of human proteins, using deep learning, we will predict the structures of the missing domains, domain-domain orientations and protein- protein complexes. We will extend small molecule virtual ligand screening (VLS) to predict binding affinities based on the insight that interacting ring-protein subpocket geometries and chemistry are conserved across protein families, are often privileged chemical structures and are likely low free energy complexes. Cryptic protein pockets, recently recognized as important drug targets, will be predicted and included in our VLS approach. Antibody-based immunotherapies are powerful but have similar safety and efficacy issues as small-molecules; thus, their safety and efficacy will be predicted by MEDICASCY-X. While MEDICASCY works on an “averaged human”, MEDICASCY-X will consider individual genetic and epigenetic profiles to make it a true precision medicine tool. We will predict which MOA proteins should be targeted and if a protein’s MOA is due to a loss or gain of function. The same framework will predict synergistic drug-drug interactions. Another way to prioritize MOA proteins is by disease comorbidity: proteins occurring in multiple diseases are likely important. If disease comorbidity can be predicted, we will construct the “Phylogenetic” Tree(s) of Diseases that would facilitate a deeper understanding of disease interrelationships. As proof of principle of the effectiveness of the algorithms being developed, novel preclinical treatments for a variety of intractable diseases will be developed. Thus, this project could enhance the success rates of drug discovery and precision medicine while reducing time and cost.