Wavelet Assignment Graph Kernel for Drug Virtual Screening

We propose a kernel function called Wavelet Assignment graph kernel for graph classification which has applications in drug discovery. This is an extension of wavelet alignment graph kernel. In this method we use graphs to model chemical compounds. For feature extraction we have applied wavelet analysis to graph structured chemical structure, for each atom we collect features about the atom and its local environment with different scales. For finding the similarity between two graphs, nodes of one graph are aligned with nodes of the other graph. such that total overall similarity is maximized with respect to all possible alignment. For alignment between two graphs we have used Wavelet Assignment graph kernel. We have evaluated the efficiency of our kernel function using Predictive Toxicology Challenge data set. Our results indicate that the new kernel function is more efficient than the existing wavelet alignment graph kernel function.