融合分子向量图卷积等方法的高通量药物筛选及其在新冠病毒RdRp上的运用

高效准确的计算机辅助药物高通量逐级筛选管道是加速全新药物开发的关键。由于类药可合成小分子的数量庞大，蛋白配体复合物难以被精确获取，导致速度过低，精度不足。目前已有的蛋白-小分子预测模型往往缺乏训练负极，难以有效地表示关键信息，且数据量不足，过于依赖复合物结构，缺乏不同方法有效联合。为解决高速筛选瓶颈，本项目拟将蛋白质的口袋信息和小分子转化成分子向量信息，并引入图卷积神经网络保存关键空间信息，运用密集深度全连接网络进行学习，产生的模型具有精度高，速度快等优点，有望快速准确地将大数据库中的活性小分子富集到小数据集。此外，拟开发多个基于结构的深度学习模型，与分子动力学模拟等技术一起用于精细筛选。最终，拟联合多种方法，用于针对新冠病毒RdRp靶点的千万级别超大规模逐级药物虚拟筛选和设计。本研究将为药物筛选和设计提供关键理论和方法支持，有望应用于新冠肺炎等疾病的药物开发。

Developing efficient and accurate computational-aided extremely large-scale virtual screening platforms would help facilitate novel drug development. Because the diversity of feasible synthesizable drug-like compounds is extremely large, traditional computational methods are limited in fulfilling current requirements. How to fast and accurately enrich the activity compounds into the smaller dataset of top score prediction is the major challenge. We convert the protein active cavity and ligand into a molecular vector representation, using the densely fully-connected neural network and Graph Convolutional Network(GCN) as model architectures. The model allows us to avoid the difficult challenge of obtaining accurate protein-ligand structure complexes. This model is fast and easy to use. It is suitable for large-scale target searching for certain ligands, and is suitable for large-scale virtual screening for lead drugs over drug-like molecule databases with millions of small molecules. On the other hand, we develop several deep learning based protein-ligand prediction models for later stage middle and small-scale virtual screening. The pocket molecular dynamics novel method was also developed and incorporated into the platform for later stage elaborate selection of binding compounds. In this research, we will systematically evaluate the model accuracy on many available experiment data. We also will check the enrichment ability for active compounds in a large compound database, hence find out the suitable type of proteins for its application. We will use this pipeline in step by step virtual screening of compounds for RdRp，which is an important therapeutic target for most RNA virus. Our research will provide key methods and new tools for drug virtual screening and design, and are promising in promoting drug development against SARS-CoV-2.