Abstract P-47: Recognition of Single Particle’s Diffraction Images Generated in X-ray Free Electron Laser Experiments

Abstract

Background: X-ray free electron lasers (XFEL) proved to be powerful instruments for solving the problem of structural analysis of biological objects and studying its functional structural dynamics. The combination of huge intensity and short flashing time leads to unique results. On the one hand the molecule tears up under Coulomb’s forces caused by the flash. On the other hand the molecule leaves a diffraction image before destruction. XFELs are widely used in traditional X-ray diffraction analysis for the purpose of obtaining the 3D structure of crystals. Methods: The paper proposes an algorithm for obtaining molecule’s 3D structure. Suppose we have a finite set of different molecular structures. The goal of the paper is to obtain the 3D structure by its diffraction image. Each corresponding molecule can be approximated by a number of simple solids like cylinders, spheres, tori. The diffraction images of these solids including a set of them are defined by an explicit formula. Thus, the problem of 3D reconstruction comes to the problem of multi class image classification which is suggested to be solved by means of convolutional neural networks (CNN). CNN is a widely known example of supervised learning i.e. it requires a training dataset with labels. The training dataset is suggested to be generated be means of direct simulation. Results: The proposed method allows to classify the corresponding molecules into up to 50 disjoint classes. It shows the accuracy about 80% on the testing dataset. Conclusion: Many interesting biological objects (including alpha helices, viruses, DNA molecules, etc.) can be approximated by means of a combination of solids. This approach allows us to solve the problem of obtaining the 3D space structure of the corresponding objects.

Keywords

• structural biology
• x-ray lasers
• image recognition
• neural networks