随着生物信息领域的高速发展，研究者越来越迫切地希望依靠计算机强大的算力来帮助新药物的发现。作为生物功能的体现者，蛋白质与其他分子的相互作用关系是近年来的研究热点。得到蛋白质与配体结合的口袋是药物发现的第一步，也是理解蛋白质功能的关键步骤，所以提高口袋预测的准确率尤为关键。而随着大规模蛋白质的出现以及生物实验流程的日趋复杂，需要有高效智能的分子可视化交互系统为研究者提供直观的蛋白质三维结构和辅助交互功能。本文对这两个任务展开研究和分析。 本文提出了层级图神经网络来解决蛋白质口袋的预测问题。该方法可以分为三个模块，首先为了充分利用蛋白质多层次的生化和几何信息，分别提取原子层和表面层的特征并输入各自的图神经网络，实现对蛋白质整体特征的学习。第二个模块为了能够学习到口袋局部区域的空间和生物信息，定义存在于口袋空间和蛋白质表面的轴点，并构建不同层级的节点与轴点的图网络，轴点通过该网络汇聚来自邻近原子和表面节点的特征。第三个模块为了解决蛋白质大小不一致的问题，在轴点网络学习中引入连通性注意力机制，并最终将输出结果聚类得到蛋白质口袋中心的位置。通过引入更关注正样本的损失函数来解决正负样本不均匀的问题。本文在两个主流的测试数据集上进行了充分的定量和可视化实验，结果表明该方法对蛋白质口袋预测的准确度优于其他已有方法，验证了该方法的有效性。 本文设计并实现了集可视化、交互设计、智能分析为一体的跨平台生物分子可视化交互系统。首先对蛋白质分子进行基础建模，构建了球棒、卡通、溶剂排除表面等模型以更好地体现蛋白质不同层级的三维结构和特征。在此基础上对可视化的效果和性能进行优化，能够支持可视化上百万原子的蛋白质。在可视化效果方面，引入屏幕空间环境光遮蔽技术增加了分子可视化的真实感。在性能方面，针对大规模的蛋白质分子，设计了基于原子的动态层次细节模型（LOD）和基于残基结构的层次聚类算法，并应用可见性剔除技术，有效地加速了分子可视化的过程。在分子交互模块中，本文提出了交互式编辑和设计分子并自动补全氢原子的方法。同时本文实现了分子结构筛选功能，帮助用户基于自定义需求选择分子特定的区域并可视化。最后将提出的口袋预测算法模型集成到可视化系统中。实验表明本文实现的分子可视化交互系统具有良好的可视化效果和交互体验。

With the rapid development of the bioinformatics, researchers are increasingly eager to rely on the powerful computing power of computers to help discover new drugs. As the embodiment of biological functions, the interaction between proteins and other molecules has been a research hotspot in recent years. Obtaining the protein-ligand pocket is the first step in drug discovery and the key to understanding protein function, so improving the accuracy of pocket prediction is particularly critical. With the emergence of largescale proteins and the increasing complexity of biological experimental processes, there is a need for efficient and intelligent interactive molecular visualization system to provide biological researchers with intuitive three-dimensional structures of proteins and auxiliary interactions. This thesis studies and analyzes these two tasks. This thesis proposes a hierarchical graph neural network to solve protein-ligand pocket prediction problem. This method can be divided into three modules. First, in order to make full use of the multi-level biochemical information and geometric features of the protein, the features of the atoms and surfaces will be extracted and input into respective graph neural networks to learn the overall characteristics of the protein. In order to learn the spatial and biological information of the local area of the pocket, the second module defines pivots that exist in the pocket space and the protein surface. The pivots learn and fuse characteristics from neighboring atoms and surfaces through graph networks constructed by pivots and nodes at different levels. In order to solve the problem of protein size inconsistency, the third module introduces the connectivity attention mechanism into pivot network learning, and finally clusters the output results to obtain the position of the protein pocket center. The problem of uneven positive and negative samples is solved by introducing a loss function that pays more attention to positive samples. This thesis conducts sufficient quantitative and visual comparative experiments on two mainstream test datasets. The results show that the accuracy of this method in predicting protein pockets is better than other existing methods, verifying the effectiveness of this method. This thesis designs and implements a cross-platform interactive molecular visualization system that integrates visualization, interaction, design, and intelligent analysis. First, basic modeling of protein molecules is carried out. Ball-and-stick models, cartoon models, solvent exclusion surface models, etc. are constructed to better reflect the three dimensional structure and characteristics of proteins at different levels. Then the effectand performance of the visualization are optimized to support the rendering of proteins with millions of atoms. In terms of visualization effects, The screen space ambient occlusion technology is used to increase the realism of molecular visualization. In terms of performance, for large-scale protein molecules, this thesis designs an atom-based dynamic level of detail (LOD) algorithm and a residue-based hierarchical clustering algorithm, and applies visibility culling technology to effectively accelerate molecular visualization. In the molecular interaction module, an interactive method for editing and designing molecules and automatically completing hydrogen is proposed. At the same time,this thesis implements a molecular structure query function to help users select and visualize specific regions of molecules based on custom needs. Finally, the proposed pocket prediction algorithm model is integrated into the visualization system.Experiments show that the interactive molecular visualization system implemented in this thesis has engaging visualization effects and interactive experience.