Application of various graph kernels for finding molecular similarity in ligand based drug discovery

There have been various measures and techniques which have currently been used in order to measure the similarity between graphs [1]. In this paper, molecules have been represented as graphs with atoms of the molecule being represented as nodes and bonds between atoms being represented as edges [2]. We will first see how various available similarity coefficients perform in order to measure the similarity between molecules (graphs) [3]. Then we will move to various popular and widely used Graphlet kernels, kernels based on subtree patterns and then we will see how each algorithm performs against molecular data which is represented as a graph structure. Graph kernel-based virtual screening has very been popular but due to very high runtime complexity of algorithms (even best algorithm is of the order cubic O(n3)), graph kernel based virtual screening has not yet been fully implemented in a full-fledged manner. Even if we apply the best graphlet kernel algorithm to molecular data in order to obtain similarity and to distinguish from active molecules from inactive ones, we just cannot replace the traditional clinical testing procedure. In this paper, we will be comparing all the popular and widely accepted graph kernels. Finally, we will be studying how multi kernel learning (MKL) helps to solve the same problem more efficiently. We will also be seeing how the amount of training data and a number of relevant features contribute towards the computation time.