A graph-based pattern recognition for chemical molecule matching

Pattern recognition techniques that are used for classification of data representations are important components of bioinformatics and chemical informatics. Prediction of the activity of chemical molecules is a significant process that can help saving time and cost devoted to conduct the actual experiments in the laboratory. We present a new method that uses graph-based pattern recognition to compute the similarity between chemical molecules. Our method is used for prediction of the activity of chemical molecules, that is, the prediction of carcinogenicity of molecules. In our method, molecules are depicted as edge-weighted graphs, where each atom corresponds to a vertex and the bonds between the atoms are depicted as edges. The framework performs graph embedding by representing vertices as points in a geometric space. The similarity measure (distance) between the embedded points is computed using the Earth Mover's Distance (EMD) method, which is based on a distribution-based transportation algorithm. Our method shows promising results on the Predictive Toxicology Challenge (PTC) dataset compared to the existing kernels.