Drug discovery by virtual screening based on novel molecular shape similarity computational algorithms

M.Phil.In drug discovery industry, molecular shape has been widely acknowledged as a key factor for biological activity and thus is regarded as an important pattern of biology information. The molecular shape similarity measurement, which is a hot topic in this field, has a wide range of applications with the ability to filter out a group of novel chemical compounds from the library with their shape structure. Molecular shape similarity research helps scientists gain a better knowledge of the biological interaction between ligands and receptors, which will help understand and explain the multiple functions and unexpected side effects of existing drugs.The molecular shape similarity can be categorized into two kinds of approaches: Structural alignment, also called molecular superposition and shape recognition. Structural alignment approaches use atom structural information and attempt to find the best alignment between two molecules and it is generally computationally intensive. However, shape recognition converts the molecular structural information of ligands into a description vector, and the similarity between two molecules can be calculated as a mathematical expression of the description vectors, which will be more efficient compared with structural alignment methods.The structural alignment method is straightforward, the difficulty lies in reducing the search space. We had initial success in structural alignment method by combining genetic algorithm( GA) and ultrafast shape recognition(USR), which is a method capable to convert a ligand into a fixed-sized vector and make the virtual screening possible for millions or even billions of ligands with extremely fast speed. We utilized rotatable bonds information to generate substructures and solved the conformer problem in molecular shape comparison. Experimental result indicates that the algorithm has achieved a better performance comparing to existing approaches.Apart from computationally intensive methods, we succeeded in quantizing the similarity of molecule shape by utilizing the topology and neighborhood information. The algorithm with topology information converts three-dimensional structure into a weighted vector containing vertex and edge connectivity information, then calculates the similarity by greedy method. The algorithm takes the advantage of moment-based algorithms for preserving ligands into vectors for further duplicated usage and large-scale virtual screening. The experimental result shows a significant performance improvement on the benchmark dataset.The third novel algorithm, which only uses the neighborhood information, extracts the atom type of the vertex and edge distance information and hence adds more varieties into consideration. By comparing with other methods, the superiority of the algorithm in accuracy and efficiency is demonstrated.在藥物發現領域，分子形狀(Molecule shape) 被認為是決定分子生物活性以及生物多動的一個重要因素。分子形狀相似性是其中一個很熱門的研究方向，對這個方向的研究，能幫助我們解釋部分藥物的藥理學原因，以及發現新的能和其有相似功效的藥物並從巨大的數據庫中篩選出來。分子形狀相似性算法主要分為兩種，結構匹配算法(Structuralalignment) 以及形狀識別算法(Shape recognition)。結構匹配算法，主要是利用已知的原子位置，對整個結構進行精確的幾何學比較。形狀識別算法將已知的形狀信息轉化成數字向量，該數字向量可以用來量化兩個分子形狀之間的相似性。結構匹配算法的原理都是相似的，區別在於如何減小搜索空間以提高效率。我們提出了基於形狀識別算法Ultrafast ShapeRecognition(USR) 和遺傳算法(Genetic algorithm) 的一種混合算法USRT，利用旋轉鍵將分子分為多個子結構，然後利用USR 對子結構的相似性進行評估，最後再利用遺傳算法獲得整個結構的相似度值。實驗結果顯示，USRT 能夠解決其他形狀識別算法無法解決的構象異構體(Conformer) 問題。基於分子之間的拓撲(Topology) 信息，我們提出了算法TopSim。TopSim 將分子的三維結構(3D Structure) 轉化成一個含有該分子頂點以及邊的信息的向量，再利用貪心算法對不同分子進行比較。實驗顯示TopSim 繼承了形狀識別算法的高效的優點，並且在基準數據集上取得了非常好的表現。我們還提出了基於最近鄰(Nearest neighbor) 信息的算法NNLSC，該算法中利用了結構中最近鄰頂點和對應邊的信息，對已有的形狀識別算法進行改進，在基準數據集中獲得了極佳的表現。我們認為NNLSC 在準確度和效率中找到了平衡點，顯示出其非常適合作為虛擬篩選(Virtual screening) 算法的潛質。Wang, Yichen.Thesis M.Phil. Chinese University of Hong Kong 2018.Includes bibliographical references (leaves 91-96).Abstracts also in Chinese.Title from PDF title page (viewed on …)