Visual Convolutional Neural Graph Fingerprints in Theano

Adam Riesselman, Harvard University

Photo of Adam Riesselman

Machine-learning techniques can be used to predict properties of unknown molecules using known training examples. Computational characterization of molecules has traditionally been done using molecular fingerprints, in which bits of a vector code for specific fragments of the molecule. Here we report visual convolutional neural graph fingerprints: Representing the molecule as a graph, this algorithm provides state-of-the-art predictive power for molecular properties and an atom-by-atom visualization of the molecule for additional inference. Written in Theano, this code is portable between CPU and GPU architectures to allow quicker fitting of larger models.

Abstract Author(s): A. Riesselman, D. Duvenaud, D. Marks