Section: New Results

Deep Learning for Symmetry detection

Participants : Guillaume Pagès, Sergei Grudinin.

We are working on a fully-structural method for detecting symmetries in molecular structures. This will allow us to detect tandem repeats, or even symmetry in density maps. We created a method based on neural network and deep learning, inspired by the advances in computer vision in the past decade. According to our first tests on simulated examples, our method is able to detect the order of a cyclic symmetry (which can be 1 for asymmetric structure) with a 92% accuracy, and guesses the direction of the axis of symmetry with an average error of $3^{\circ }$. We are still working on improving it and doing tests on more realistic examples.