Drugs Discovery by Shape Similarity Using Deep Learning

Abstract

Searching for one or several molecules in a database using their shapes has great interest from a biochemical point of view, but requires a huge computational effort due to the complexity of the algorithms and the sizes of the databases in the pharmaceutical industry. This work uses Deep Learning by training neural networks with hundreds of images of each molecule, rendered by projections (using GPUs) on planes whose normal vectors are equally distributed in the 3D space (using Fibonacci spirals). The results obtained, both in accuracy and time, exceeded expectations, opening a hopeful path of research.