Molecule Classification Using Visualization and Convolutional Neural Network

István Lakatos, András Hajdu, Balázs Harangi

Abstract: In this paper, we propose a procedure that provides solid performance regarding molecule classification. Our solution can predict with high accuracy the toxicity and activity of different unknown molecules based on their compounds and structural information. As for the methodological contribution, our approach takes the commonly used SMILES strings and generates the three dimensional model of the investigated molecule. After that, we project this model to the two-dimensional plane from different points of view and a pre-trained convolutional neural network classifies all of these generated 2D images. The final class label is derived as an ensemble of these classification outputs. For the ensemble of class labels and the applied visualization method, we have reached 90.66% classification accuracy with ROC-AUC 0.9629.