【作者】 李春艳；

【导师】 姚俊峰；

【作者基本信息】 厦门大学 ， 计算机科学与技术， 2022， 博士

【摘要】 传统的药物研发是一个耗时而昂贵、复杂而低成功率的过程,通过各种测量技术所产生的生化数据正在以指数级速度增长,这为通过机器学习方法挖掘数据之间的关联提供了新的机遇。分子属性预测是药物发现领域的一项基本任务,因为它可以帮助确定药物的功能。而分子表征研究是药物筛选和新药发现的重要基础,对分子属性预测、生化活性预测以及抗生素筛选等起决定作用。寻找具有良好药理学、毒理学和药物动力学特性的化学物质是药物发现领域的巨大挑战,药物的三维(Three-Dimension,3D)特征很大程度决定了药物性质以及药物靶点结合特性,但3D拓扑结构受分子构象和相对位置取向的影响很大,导致现有的药物3D表征方法存在花费很高以及准确率不高的问题。同时,目前分子表征方法大多基于一维(One-Dimension,1D)或二维(Two-Dimension,2D)分子特征,忽略了分子3D结构,影响了与药物分子相关下游任务的准确性。由于现有分子表征方法的局限性,需要分子表征学习在充分考虑药物分子空间拓扑结构的情况下实现高效设计,更有效地提取特征。基于上述背景,本文提出多视角学习方法,从药物不同视角下的“原始”3D结构中提取特征,具体来说,本文的主要研究内容包括以下几个方面:·提出了基于时空门注意力机制的分子拓扑特征提取算法,针对长程依赖和语义相似度问题,把注意力机制应用到3D Grid分子结构,提出时空门注意力机制的分子拓扑特征提取算法。时空门注意力模块由空间注意力、通道注意力和门机制组成,空间注意力模块抽取3D分子特征并获得空间注意力分数,通道注意力模块获得不同通道(原子类型)的注意力分数,门机制整合两部分注意力以获得全局分子的3D Grid注意力。·提出了基于3D旋转不变性的图卷积神经网络自适应算法,针对无法满足3D旋转不变性问题,提出了旋转不变的分子特征映射算法。保证在3D空间所构建的深度神经网络满足一定的“不变性”以确保模型的泛化能力。然后提出基于3D旋转不变性的自适应图卷积神经网络3DMol-Net。3DMol-Net是可泛化的3D药物分子表征学习方法,其自适应性体现在3个方面,即:输入任意分子结构自动提取3D特征、自动构建残差图拉普拉斯神经网络、以端到端的方式自动应用在任何场景和任何任务中。·提出了基于多重简化分子线性输入规范(Simplified Molecular Input Line En-try System,SMILES)增强的分子性质预测算法。针对小数据量数据集的过拟合问题,受自然语言处理技术启发,分子SMILES字符串一直被用作深度神经网络模型中的一种字符方法,然而,深度学习模型受到SMILES字符串的非唯一性的阻碍,为了有效地沿所有信息路径学习分子特征、捕获分子不同结构特征,对每个分子编码多重SMILES,作为分子性质预测任务的自动数据扩充,从而缓解了分子性质预测数据集中由于数据量小而导致的过拟合问题。·提出了基于3D分子几何的分子生成算法,针对3D表征的药物设计问题,提出基于几何和深度限制变分编码器的分子生成算法GEOM-CVAE。GEOM-CVAE基于3D结构的分子可视化表征和蛋白质靶标的3D Mesh图表征生成具有特殊化学属性的药物分子。不同于以往基于1D或2D的分子生成方法,3D几何信息对成功的分子生成和设计至关重要。GEOM-CVAE以两阶段方式生成分子:通过转换分子的3D坐标为特别的图像以学习隐空间的表征和基于几何的图卷积提取蛋白质靶标特征作为模型的限制条件以生成特殊属性的分子。综上所述,本文基于深度神经网络,在满足3D旋转不变性基础上充分考虑分子空间结构特征,从3D领域以自适应方式和多视角学习方法探索药物分子表征,预测分子化合物属性、生化活性以及毒性等性质,并在药物分子生成模型方面提供新的思考和见解。通过广泛的实验证明了所提出的方法对药物属性预测任务的有效性和相对于基线方法的优越性,基于3D的分子表征研究具有一定的现实意义和应用前景。更多还原

【Abstract】 Traditional drug development is a time-consuming,expensive,complicated process with low-success rate.The biochemical data generated by various measurement technologies is increasing at an exponential rate,which provides new opportunities for mining the associations among data using machine learning methods.Molecular properties prediction is a basic task in the field of drug discovery,because it can help to determine the function of drugs.While molecular representation research is an important basis for drug screening and new drug discovery,and plays a decisive role in the prediction of molecular properties and biochemical activity,as well as antibiotic screening.Finding chemical substances with good pharmacological,toxicological and pharmacokinetic properties is a huge challenge in the field of drug discovery.The three-dimensional(3D)characteristics of molecules largely determine molecular properties and binding characteristics of target.However,the 3D topological structure is greatly affected by the molecular conformation and relative position orientation,leading to high cost and low accuracy in the existing 3D molecular representation methods.At the same time,the current molecular representation methods are mostly based on one-dimensional(1D)or two-dimensional(2D)molecular properties,ignoring molecular 3D structure,which affects the accuracy of downstream tasks related to molecules.Due to the limitations of existing molecular representation methods,the overall modeling of molecular representation learning is required to achieve effective design and extract features more effectively while fully considering the spatial topological structure of molecules.Based on the above background,this dissertation proposes a multi-perspective learning method to extract features from the "original" 3D structure of drugs from different perspectives.Specifically,the main research contents of this dissertation include the following aspects:·This dissertation proposes a molecular topology feature extraction algorithm based on the spatial-temporal gated attention module.Aiming at the problem of long-range dependency and semantic similarity,the attention mechanism is applied to the 3D Grid molecular structure.Therefore,the molecular topology feature extraction algorithm is proposed using the spatial-temporal gated attention mechanism.The spatial-temporal gated attention module consists of spatial attention,channel attention and gate mechanism.The spatial attention module extracts 3D molecular features and obtains the spatial attention score,the channel attention module obtains the attention scores of different channels(atom types),and the gate mechanism integrates above two parts of attention to obtain 3D grid attention of the global-level molecule.·This dissertation proposes an adaptive algorithm of graph convolutional neural network based on 3D rotation invariance.Aiming at the problem that 3D rotation invariance cannot be satisfied,a rotation invariant molecular feature mapping algorithm is proposed.It can ensure that the deep neural network constructed in 3D space meets a certain "invariance" for the generalization ability of the model.Then an adaptive graph convolutional neural network 3DMol-Net is proposed based on 3D rotation invariance,which is a general 3D drug molecular representation method,and its self-adaptability is reflected in three aspects,namely:inputting any molecular structure,automatically extracting 3D features,automatically constructing residual graph Laplacian neural network,and can be automatically applied to any task in any scene in an end-to-end manner.·This dissertation proposes a molecular property prediction algorithm based on the enhancement of multiple Simplified Molecular Input Line Entry System(SMILES).Aiming at the over fitting problem of small datasets,inspired by natural language processing technology,molecular SMILES string has been used as a kind of character methods in deep neural network model.However,deep learning model is hindered by the nonuniqueness of SMILES string.In order to effectively learn molecular features along all information paths and capture different structural features of molecules,multiple SMILES are encoded for each molecule as an automatic data expansion in the task of molecular property prediction,which alleviates the over fitting problem caused by the small amount of data in the dataset of molecular property prediction.·This dissertation proposes a molecular generation algorithm based on 3D molecular geometry.Aiming at the drug design problem using 3D representation,the molecular generation algorithm GEOM-CVAE is proposed based on geometric and constrained variational autoencoder.GEOM-CVAE considers 3D structure-based molecular visualization representation and protein 3D mesh-based graph representation to generate special molecules.3D geometric information is essential for successful molecular generation and design,that is different from the previous 1D or 2D-based molecular generation methods.GEOM-CVAE generates molecules in a two-stage manner:by transforming the 3D coordinates of molecules into special images to learn hidden space characterization,by using geometric-based graph convolution to extract protein features as the constraints of the model for generating molecules with special properties.In summary,this dissertation focuses on the characteristics of molecular spatial structure on the basis of 3D rotation invariance and deep neural network,explores molecular representation from 3D field in an adaptive and multi-perspective learning method for predicting molecular properties,biochemical activity and toxicity.Meanwhile this dissertation provides new thinking and insights on the molecular generation model.The extensive experiments have proved the effectiveness of proposed methods on the task of molecular properties prediction and its superiority compared to the baseline methods.The 3D-based molecular representation research has certain practical significance and application prospects.更多还原